diff --git "a/openai_whisper-large-v3/TextDecoder.mlmodelc/model.mil" "b/openai_whisper-large-v3/TextDecoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/openai_whisper-large-v3/TextDecoder.mlmodelc/model.mil" @@ -0,0 +1,5345 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "2.2.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "7.1"}})] +{ + func main(tensor cache_length, tensor decoder_key_padding_mask, tensor encoder_output_embeds, tensor input_ids, tensor key_cache, tensor kv_cache_update_mask, tensor value_cache) { + tensor var_80_axis_0 = const()[name = tensor("op_80_axis_0"), val = tensor(0)]; + tensor var_80_batch_dims_0 = const()[name = tensor("op_80_batch_dims_0"), val = tensor(0)]; + tensor embed_tokens_weight_to_fp16 = const()[name = tensor("embed_tokens_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor var_80_cast_fp16 = gather(axis = var_80_axis_0, batch_dims = var_80_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor("op_80_cast_fp16")]; + tensor var_84_axis_0 = const()[name = tensor("op_84_axis_0"), val = tensor(0)]; + tensor var_84_batch_dims_0 = const()[name = tensor("op_84_batch_dims_0"), val = tensor(0)]; + tensor embed_positions_weight_to_fp16 = const()[name = tensor("embed_positions_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132777088)))]; + tensor var_84_cast_fp16 = gather(axis = var_84_axis_0, batch_dims = var_84_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor("op_84_cast_fp16")]; + tensor hidden_states_1_cast_fp16 = add(x = var_80_cast_fp16, y = var_84_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; + tensor var_98_axes_0 = const()[name = tensor("op_98_axes_0"), val = tensor([2])]; + tensor var_98_cast_fp16 = expand_dims(axes = var_98_axes_0, x = hidden_states_1_cast_fp16)[name = tensor("op_98_cast_fp16")]; + tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([3])]; + tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_98_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280])]; + tensor var_103_axis_0 = const()[name = tensor("op_103_axis_0"), val = tensor(1)]; + tensor var_103_cast_fp16_0, tensor var_103_cast_fp16_1, tensor var_103_cast_fp16_2, tensor var_103_cast_fp16_3, tensor var_103_cast_fp16_4, tensor var_103_cast_fp16_5, tensor var_103_cast_fp16_6, tensor var_103_cast_fp16_7, tensor var_103_cast_fp16_8, tensor var_103_cast_fp16_9, tensor var_103_cast_fp16_10, tensor var_103_cast_fp16_11, tensor var_103_cast_fp16_12, tensor var_103_cast_fp16_13, tensor var_103_cast_fp16_14, tensor var_103_cast_fp16_15, tensor var_103_cast_fp16_16, tensor var_103_cast_fp16_17, tensor var_103_cast_fp16_18, tensor var_103_cast_fp16_19, tensor var_103_cast_fp16_20, tensor var_103_cast_fp16_21, tensor var_103_cast_fp16_22, tensor var_103_cast_fp16_23, tensor var_103_cast_fp16_24, tensor var_103_cast_fp16_25, tensor var_103_cast_fp16_26, tensor var_103_cast_fp16_27, tensor var_103_cast_fp16_28, tensor var_103_cast_fp16_29, tensor var_103_cast_fp16_30, tensor var_103_cast_fp16_31 = split(axis = var_103_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor("op_103_cast_fp16")]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280])]; + tensor var_138_axis_0 = const()[name = tensor("op_138_axis_0"), val = tensor(1)]; + tensor var_138_cast_fp16_0, tensor var_138_cast_fp16_1, tensor var_138_cast_fp16_2, tensor var_138_cast_fp16_3, tensor var_138_cast_fp16_4, tensor var_138_cast_fp16_5, tensor var_138_cast_fp16_6, tensor var_138_cast_fp16_7, tensor var_138_cast_fp16_8, tensor var_138_cast_fp16_9, tensor var_138_cast_fp16_10, tensor var_138_cast_fp16_11, tensor var_138_cast_fp16_12, tensor var_138_cast_fp16_13, tensor var_138_cast_fp16_14, tensor var_138_cast_fp16_15, tensor var_138_cast_fp16_16, tensor var_138_cast_fp16_17, tensor var_138_cast_fp16_18, tensor var_138_cast_fp16_19, tensor var_138_cast_fp16_20, tensor var_138_cast_fp16_21, tensor var_138_cast_fp16_22, tensor var_138_cast_fp16_23, tensor var_138_cast_fp16_24, tensor var_138_cast_fp16_25, tensor var_138_cast_fp16_26, tensor var_138_cast_fp16_27, tensor var_138_cast_fp16_28, tensor var_138_cast_fp16_29, tensor var_138_cast_fp16_30, tensor var_138_cast_fp16_31 = split(axis = var_138_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor("op_138_cast_fp16")]; + tensor var_176 = const()[name = tensor("op_176"), val = tensor(3)]; + tensor var_183 = const()[name = tensor("op_183"), val = tensor(1)]; + tensor var_184 = const()[name = tensor("op_184"), val = tensor(true)]; + tensor var_196 = const()[name = tensor("op_196"), val = tensor([1])]; + tensor channels_mean_1_cast_fp16 = reduce_mean(axes = var_196, keep_dims = var_184, x = inputs_1_cast_fp16)[name = tensor("channels_mean_1_cast_fp16")]; + tensor zero_mean_1_cast_fp16 = sub(x = inputs_1_cast_fp16, y = channels_mean_1_cast_fp16)[name = tensor("zero_mean_1_cast_fp16")]; + tensor zero_mean_sq_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = zero_mean_1_cast_fp16)[name = tensor("zero_mean_sq_1_cast_fp16")]; + tensor var_200 = const()[name = tensor("op_200"), val = tensor([1])]; + tensor var_201_cast_fp16 = reduce_mean(axes = var_200, keep_dims = var_184, x = zero_mean_sq_1_cast_fp16)[name = tensor("op_201_cast_fp16")]; + tensor var_202_to_fp16 = const()[name = tensor("op_202_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_203_cast_fp16 = add(x = var_201_cast_fp16, y = var_202_to_fp16)[name = tensor("op_203_cast_fp16")]; + tensor denom_1_epsilon_0_to_fp16 = const()[name = tensor("denom_1_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_1_cast_fp16 = rsqrt(epsilon = denom_1_epsilon_0_to_fp16, x = var_203_cast_fp16)[name = tensor("denom_1_cast_fp16")]; + tensor out_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = denom_1_cast_fp16)[name = tensor("out_1_cast_fp16")]; + tensor obj_1_mean_0_to_fp16 = const()[name = tensor("obj_1_mean_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133924032)))]; + tensor obj_1_variance_0_to_fp16 = const()[name = tensor("obj_1_variance_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133926656)))]; + tensor obj_1_gamma_0_to_fp16 = const()[name = tensor("obj_1_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133929280)))]; + tensor obj_1_beta_0_to_fp16 = const()[name = tensor("obj_1_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133931904)))]; + tensor obj_1_epsilon_0_to_fp16 = const()[name = tensor("obj_1_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor("obj_1_cast_fp16")]; + tensor var_218 = const()[name = tensor("op_218"), val = tensor([1, 1])]; + tensor var_220 = const()[name = tensor("op_220"), val = tensor([1, 1])]; + tensor query_1_pad_type_0 = const()[name = tensor("query_1_pad_type_0"), val = tensor("custom")]; + tensor query_1_pad_0 = const()[name = tensor("query_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133934528)))]; + tensor layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137211392)))]; + tensor query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = var_220, groups = var_183, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = var_218, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("query_1_cast_fp16")]; + tensor var_224 = const()[name = tensor("op_224"), val = tensor([1, 1])]; + tensor var_226 = const()[name = tensor("op_226"), val = tensor([1, 1])]; + tensor current_key_1_pad_type_0 = const()[name = tensor("current_key_1_pad_type_0"), val = tensor("custom")]; + tensor current_key_1_pad_0 = const()[name = tensor("current_key_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137214016)))]; + tensor current_key_1_cast_fp16 = conv(dilations = var_226, groups = var_183, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = var_224, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("current_key_1_cast_fp16")]; + tensor var_231 = const()[name = tensor("op_231"), val = tensor([1, 1])]; + tensor var_233 = const()[name = tensor("op_233"), val = tensor([1, 1])]; + tensor current_value_1_pad_type_0 = const()[name = tensor("current_value_1_pad_type_0"), val = tensor("custom")]; + tensor current_value_1_pad_0 = const()[name = tensor("current_value_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140490880)))]; + tensor layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143767744)))]; + tensor current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = var_233, groups = var_183, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = var_231, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("current_value_1_cast_fp16")]; + tensor var_237_axes_0 = const()[name = tensor("op_237_axes_0"), val = tensor([1])]; + tensor var_237_cast_fp16 = expand_dims(axes = var_237_axes_0, x = kv_cache_update_mask)[name = tensor("op_237_cast_fp16")]; + tensor var_238_axes_0 = const()[name = tensor("op_238_axes_0"), val = tensor([2])]; + tensor var_238_cast_fp16 = expand_dims(axes = var_238_axes_0, x = var_237_cast_fp16)[name = tensor("op_238_cast_fp16")]; + tensor var_240_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_240_cast_fp16")]; + tensor var_177_to_fp16 = const()[name = tensor("op_177_to_fp16"), val = tensor(0x1p+0)]; + tensor var_241_cast_fp16 = sub(x = var_177_to_fp16, y = var_238_cast_fp16)[name = tensor("op_241_cast_fp16")]; + tensor var_242_cast_fp16 = mul(x = var_103_cast_fp16_0, y = var_241_cast_fp16)[name = tensor("op_242_cast_fp16")]; + tensor key_1_cast_fp16 = add(x = var_240_cast_fp16, y = var_242_cast_fp16)[name = tensor("key_1_cast_fp16")]; + tensor var_244_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_244_cast_fp16")]; + tensor var_246_cast_fp16 = mul(x = var_138_cast_fp16_0, y = var_241_cast_fp16)[name = tensor("op_246_cast_fp16")]; + tensor value_1_cast_fp16 = add(x = var_244_cast_fp16, y = var_246_cast_fp16)[name = tensor("value_1_cast_fp16")]; + tensor var_249 = const()[name = tensor("op_249"), val = tensor([1, 20, 64, -1])]; + tensor var_250_cast_fp16 = reshape(shape = var_249, x = query_1_cast_fp16)[name = tensor("op_250_cast_fp16")]; + tensor var_251_to_fp16 = const()[name = tensor("op_251_to_fp16"), val = tensor(0x1p-3)]; + tensor var_252_cast_fp16 = mul(x = var_250_cast_fp16, y = var_251_to_fp16)[name = tensor("op_252_cast_fp16")]; + tensor var_253 = const()[name = tensor("op_253"), val = tensor([1, 20, 64, -1])]; + tensor var_254_cast_fp16 = reshape(shape = var_253, x = key_1_cast_fp16)[name = tensor("op_254_cast_fp16")]; + tensor mh_w_1_transpose_x_0 = const()[name = tensor("mh_w_1_transpose_x_0"), val = tensor(true)]; + tensor mh_w_1_transpose_y_0 = const()[name = tensor("mh_w_1_transpose_y_0"), val = tensor(false)]; + tensor mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_252_cast_fp16, y = var_254_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; + tensor var_258_axes_0 = const()[name = tensor("op_258_axes_0"), val = tensor([1])]; + tensor var_258_cast_fp16 = expand_dims(axes = var_258_axes_0, x = decoder_key_padding_mask)[name = tensor("op_258_cast_fp16")]; + tensor var_259_axes_0 = const()[name = tensor("op_259_axes_0"), val = tensor([2])]; + tensor var_259_cast_fp16 = expand_dims(axes = var_259_axes_0, x = var_258_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; + tensor var_262_cast_fp16 = softmax(axis = var_176, x = mh_w_3_cast_fp16)[name = tensor("op_262_cast_fp16")]; + tensor var_263 = const()[name = tensor("op_263"), val = tensor([1, 20, 64, -1])]; + tensor var_264_cast_fp16 = reshape(shape = var_263, x = value_1_cast_fp16)[name = tensor("op_264_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(true)]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_264_cast_fp16, y = var_262_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_267 = const()[name = tensor("op_267"), val = tensor([1, 1280, 1, -1])]; + tensor input_1_cast_fp16 = reshape(shape = var_267, x = attn_1_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_271 = const()[name = tensor("op_271"), val = tensor([1, 1])]; + tensor var_273 = const()[name = tensor("op_273"), val = tensor([1, 1])]; + tensor obj_7_pad_type_0 = const()[name = tensor("obj_7_pad_type_0"), val = tensor("custom")]; + tensor obj_7_pad_0 = const()[name = tensor("obj_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143770368)))]; + tensor layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147047232)))]; + tensor obj_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = var_273, groups = var_183, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = var_271, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("obj_7_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor var_283 = const()[name = tensor("op_283"), val = tensor([1])]; + tensor channels_mean_3_cast_fp16 = reduce_mean(axes = var_283, keep_dims = var_184, x = inputs_3_cast_fp16)[name = tensor("channels_mean_3_cast_fp16")]; + tensor zero_mean_3_cast_fp16 = sub(x = inputs_3_cast_fp16, y = channels_mean_3_cast_fp16)[name = tensor("zero_mean_3_cast_fp16")]; + tensor zero_mean_sq_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = zero_mean_3_cast_fp16)[name = tensor("zero_mean_sq_3_cast_fp16")]; + tensor var_287 = const()[name = tensor("op_287"), val = tensor([1])]; + tensor var_288_cast_fp16 = reduce_mean(axes = var_287, keep_dims = var_184, x = zero_mean_sq_3_cast_fp16)[name = tensor("op_288_cast_fp16")]; + tensor var_289_to_fp16 = const()[name = tensor("op_289_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_290_cast_fp16 = add(x = var_288_cast_fp16, y = var_289_to_fp16)[name = tensor("op_290_cast_fp16")]; + tensor denom_3_epsilon_0_to_fp16 = const()[name = tensor("denom_3_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_3_cast_fp16 = rsqrt(epsilon = denom_3_epsilon_0_to_fp16, x = var_290_cast_fp16)[name = tensor("denom_3_cast_fp16")]; + tensor out_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = denom_3_cast_fp16)[name = tensor("out_3_cast_fp16")]; + tensor obj_9_gamma_0_to_fp16 = const()[name = tensor("obj_9_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147049856)))]; + tensor obj_9_beta_0_to_fp16 = const()[name = tensor("obj_9_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147052480)))]; + tensor obj_9_epsilon_0_to_fp16 = const()[name = tensor("obj_9_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor("obj_9_cast_fp16")]; + tensor var_305 = const()[name = tensor("op_305"), val = tensor([1, 1])]; + tensor var_307 = const()[name = tensor("op_307"), val = tensor([1, 1])]; + tensor query_3_pad_type_0 = const()[name = tensor("query_3_pad_type_0"), val = tensor("custom")]; + tensor query_3_pad_0 = const()[name = tensor("query_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147055104)))]; + tensor layers_0_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150331968)))]; + tensor query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = var_307, groups = var_183, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = var_305, weight = layers_0_encoder_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("query_3_cast_fp16")]; + tensor var_311 = const()[name = tensor("op_311"), val = tensor([1, 1])]; + tensor var_313 = const()[name = tensor("op_313"), val = tensor([1, 1])]; + tensor key_3_pad_type_0 = const()[name = tensor("key_3_pad_type_0"), val = tensor("custom")]; + tensor key_3_pad_0 = const()[name = tensor("key_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150334592)))]; + tensor key_3_cast_fp16 = conv(dilations = var_313, groups = var_183, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = var_311, weight = layers_0_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_3_cast_fp16")]; + tensor var_318 = const()[name = tensor("op_318"), val = tensor([1, 1])]; + tensor var_320 = const()[name = tensor("op_320"), val = tensor([1, 1])]; + tensor value_3_pad_type_0 = const()[name = tensor("value_3_pad_type_0"), val = tensor("custom")]; + tensor value_3_pad_0 = const()[name = tensor("value_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153611456)))]; + tensor layers_0_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156888320)))]; + tensor value_3_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_bias_to_fp16, dilations = var_320, groups = var_183, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = var_318, weight = layers_0_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_3_cast_fp16")]; + tensor var_324 = const()[name = tensor("op_324"), val = tensor([1, 20, 64, -1])]; + tensor var_325_cast_fp16 = reshape(shape = var_324, x = query_3_cast_fp16)[name = tensor("op_325_cast_fp16")]; + tensor var_326_to_fp16 = const()[name = tensor("op_326_to_fp16"), val = tensor(0x1p-3)]; + tensor var_327_cast_fp16 = mul(x = var_325_cast_fp16, y = var_326_to_fp16)[name = tensor("op_327_cast_fp16")]; + tensor var_328 = const()[name = tensor("op_328"), val = tensor([1, 20, 64, -1])]; + tensor var_329_cast_fp16 = reshape(shape = var_328, x = key_3_cast_fp16)[name = tensor("op_329_cast_fp16")]; + tensor mh_w_5_transpose_x_0 = const()[name = tensor("mh_w_5_transpose_x_0"), val = tensor(true)]; + tensor mh_w_5_transpose_y_0 = const()[name = tensor("mh_w_5_transpose_y_0"), val = tensor(false)]; + tensor mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_327_cast_fp16, y = var_329_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; + tensor obj_13_cast_fp16 = softmax(axis = var_176, x = mh_w_5_cast_fp16)[name = tensor("obj_13_cast_fp16")]; + tensor var_333 = const()[name = tensor("op_333"), val = tensor([1, 20, 64, -1])]; + tensor var_334_cast_fp16 = reshape(shape = var_333, x = value_3_cast_fp16)[name = tensor("op_334_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(true)]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_334_cast_fp16, y = obj_13_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_337 = const()[name = tensor("op_337"), val = tensor([1, 1280, 1, -1])]; + tensor input_3_cast_fp16 = reshape(shape = var_337, x = attn_3_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor var_341 = const()[name = tensor("op_341"), val = tensor([1, 1])]; + tensor var_343 = const()[name = tensor("op_343"), val = tensor([1, 1])]; + tensor obj_11_pad_type_0 = const()[name = tensor("obj_11_pad_type_0"), val = tensor("custom")]; + tensor obj_11_pad_0 = const()[name = tensor("obj_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156890944)))]; + tensor layers_0_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160167808)))]; + tensor obj_11_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = var_343, groups = var_183, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = var_341, weight = layers_0_encoder_attn_o_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("obj_11_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor var_349 = const()[name = tensor("op_349"), val = tensor([1])]; + tensor channels_mean_5_cast_fp16 = reduce_mean(axes = var_349, keep_dims = var_184, x = inputs_5_cast_fp16)[name = tensor("channels_mean_5_cast_fp16")]; + tensor zero_mean_5_cast_fp16 = sub(x = inputs_5_cast_fp16, y = channels_mean_5_cast_fp16)[name = tensor("zero_mean_5_cast_fp16")]; + tensor zero_mean_sq_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = zero_mean_5_cast_fp16)[name = tensor("zero_mean_sq_5_cast_fp16")]; + tensor var_353 = const()[name = tensor("op_353"), val = tensor([1])]; + tensor var_354_cast_fp16 = reduce_mean(axes = var_353, keep_dims = var_184, x = zero_mean_sq_5_cast_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_355_to_fp16 = const()[name = tensor("op_355_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_356_cast_fp16 = add(x = var_354_cast_fp16, y = var_355_to_fp16)[name = tensor("op_356_cast_fp16")]; + tensor denom_5_epsilon_0_to_fp16 = const()[name = tensor("denom_5_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_5_cast_fp16 = rsqrt(epsilon = denom_5_epsilon_0_to_fp16, x = var_356_cast_fp16)[name = tensor("denom_5_cast_fp16")]; + tensor out_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = denom_5_cast_fp16)[name = tensor("out_5_cast_fp16")]; + tensor input_5_gamma_0_to_fp16 = const()[name = tensor("input_5_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160170432)))]; + tensor input_5_beta_0_to_fp16 = const()[name = tensor("input_5_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160173056)))]; + tensor input_5_epsilon_0_to_fp16 = const()[name = tensor("input_5_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = batch_norm(beta = input_5_beta_0_to_fp16, epsilon = input_5_epsilon_0_to_fp16, gamma = input_5_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor var_367 = const()[name = tensor("op_367"), val = tensor([1, 1])]; + tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 1])]; + tensor input_7_pad_type_0 = const()[name = tensor("input_7_pad_type_0"), val = tensor("custom")]; + tensor input_7_pad_0 = const()[name = tensor("input_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_fc1_weight_to_fp16 = const()[name = tensor("layers_0_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160175680)))]; + tensor layers_0_fc1_bias_to_fp16 = const()[name = tensor("layers_0_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173282944)))]; + tensor input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = var_369, groups = var_183, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = var_367, weight = layers_0_fc1_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_mode_0 = const()[name = tensor("input_9_mode_0"), val = tensor("EXACT")]; + tensor input_9_cast_fp16 = gelu(mode = input_9_mode_0, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor var_375 = const()[name = tensor("op_375"), val = tensor([1, 1])]; + tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 1])]; + tensor hidden_states_3_pad_type_0 = const()[name = tensor("hidden_states_3_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_3_pad_0 = const()[name = tensor("hidden_states_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_fc2_weight_to_fp16 = const()[name = tensor("layers_0_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173293248)))]; + tensor layers_0_fc2_bias_to_fp16 = const()[name = tensor("layers_0_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186400512)))]; + tensor hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = var_377, groups = var_183, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = var_375, weight = layers_0_fc2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("hidden_states_3_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor var_390 = const()[name = tensor("op_390"), val = tensor(3)]; + tensor var_397 = const()[name = tensor("op_397"), val = tensor(1)]; + tensor var_398 = const()[name = tensor("op_398"), val = tensor(true)]; + tensor var_410 = const()[name = tensor("op_410"), val = tensor([1])]; + tensor channels_mean_7_cast_fp16 = reduce_mean(axes = var_410, keep_dims = var_398, x = inputs_7_cast_fp16)[name = tensor("channels_mean_7_cast_fp16")]; + tensor zero_mean_7_cast_fp16 = sub(x = inputs_7_cast_fp16, y = channels_mean_7_cast_fp16)[name = tensor("zero_mean_7_cast_fp16")]; + tensor zero_mean_sq_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = zero_mean_7_cast_fp16)[name = tensor("zero_mean_sq_7_cast_fp16")]; + tensor var_414 = const()[name = tensor("op_414"), val = tensor([1])]; + tensor var_415_cast_fp16 = reduce_mean(axes = var_414, keep_dims = var_398, x = zero_mean_sq_7_cast_fp16)[name = tensor("op_415_cast_fp16")]; + tensor var_416_to_fp16 = const()[name = tensor("op_416_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_417_cast_fp16 = add(x = var_415_cast_fp16, y = var_416_to_fp16)[name = tensor("op_417_cast_fp16")]; + tensor denom_7_epsilon_0_to_fp16 = const()[name = tensor("denom_7_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_7_cast_fp16 = rsqrt(epsilon = denom_7_epsilon_0_to_fp16, x = var_417_cast_fp16)[name = tensor("denom_7_cast_fp16")]; + tensor out_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = denom_7_cast_fp16)[name = tensor("out_7_cast_fp16")]; + tensor obj_15_gamma_0_to_fp16 = const()[name = tensor("obj_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186403136)))]; + tensor obj_15_beta_0_to_fp16 = const()[name = tensor("obj_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186405760)))]; + tensor obj_15_epsilon_0_to_fp16 = const()[name = tensor("obj_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor("obj_15_cast_fp16")]; + tensor var_432 = const()[name = tensor("op_432"), val = tensor([1, 1])]; + tensor var_434 = const()[name = tensor("op_434"), val = tensor([1, 1])]; + tensor query_5_pad_type_0 = const()[name = tensor("query_5_pad_type_0"), val = tensor("custom")]; + tensor query_5_pad_0 = const()[name = tensor("query_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186408384)))]; + tensor layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189685248)))]; + tensor query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = var_434, groups = var_397, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = var_432, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("query_5_cast_fp16")]; + tensor var_438 = const()[name = tensor("op_438"), val = tensor([1, 1])]; + tensor var_440 = const()[name = tensor("op_440"), val = tensor([1, 1])]; + tensor current_key_3_pad_type_0 = const()[name = tensor("current_key_3_pad_type_0"), val = tensor("custom")]; + tensor current_key_3_pad_0 = const()[name = tensor("current_key_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189687872)))]; + tensor current_key_3_cast_fp16 = conv(dilations = var_440, groups = var_397, pad = current_key_3_pad_0, pad_type = current_key_3_pad_type_0, strides = var_438, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("current_key_3_cast_fp16")]; + tensor var_445 = const()[name = tensor("op_445"), val = tensor([1, 1])]; + tensor var_447 = const()[name = tensor("op_447"), val = tensor([1, 1])]; + tensor current_value_3_pad_type_0 = const()[name = tensor("current_value_3_pad_type_0"), val = tensor("custom")]; + tensor current_value_3_pad_0 = const()[name = tensor("current_value_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192964736)))]; + tensor layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196241600)))]; + tensor current_value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = var_447, groups = var_397, pad = current_value_3_pad_0, pad_type = current_value_3_pad_type_0, strides = var_445, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("current_value_3_cast_fp16")]; + tensor var_454_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_454_cast_fp16")]; + tensor var_456_cast_fp16 = mul(x = var_103_cast_fp16_1, y = var_241_cast_fp16)[name = tensor("op_456_cast_fp16")]; + tensor key_5_cast_fp16 = add(x = var_454_cast_fp16, y = var_456_cast_fp16)[name = tensor("key_5_cast_fp16")]; + tensor var_458_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_458_cast_fp16")]; + tensor var_460_cast_fp16 = mul(x = var_138_cast_fp16_1, y = var_241_cast_fp16)[name = tensor("op_460_cast_fp16")]; + tensor value_5_cast_fp16 = add(x = var_458_cast_fp16, y = var_460_cast_fp16)[name = tensor("value_5_cast_fp16")]; + tensor var_463 = const()[name = tensor("op_463"), val = tensor([1, 20, 64, -1])]; + tensor var_464_cast_fp16 = reshape(shape = var_463, x = query_5_cast_fp16)[name = tensor("op_464_cast_fp16")]; + tensor var_465_to_fp16 = const()[name = tensor("op_465_to_fp16"), val = tensor(0x1p-3)]; + tensor var_466_cast_fp16 = mul(x = var_464_cast_fp16, y = var_465_to_fp16)[name = tensor("op_466_cast_fp16")]; + tensor var_467 = const()[name = tensor("op_467"), val = tensor([1, 20, 64, -1])]; + tensor var_468_cast_fp16 = reshape(shape = var_467, x = key_5_cast_fp16)[name = tensor("op_468_cast_fp16")]; + tensor mh_w_7_transpose_x_0 = const()[name = tensor("mh_w_7_transpose_x_0"), val = tensor(true)]; + tensor mh_w_7_transpose_y_0 = const()[name = tensor("mh_w_7_transpose_y_0"), val = tensor(false)]; + tensor mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_466_cast_fp16, y = var_468_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; + tensor mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; + tensor var_476_cast_fp16 = softmax(axis = var_390, x = mh_w_9_cast_fp16)[name = tensor("op_476_cast_fp16")]; + tensor var_477 = const()[name = tensor("op_477"), val = tensor([1, 20, 64, -1])]; + tensor var_478_cast_fp16 = reshape(shape = var_477, x = value_5_cast_fp16)[name = tensor("op_478_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(true)]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_478_cast_fp16, y = var_476_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_481 = const()[name = tensor("op_481"), val = tensor([1, 1280, 1, -1])]; + tensor input_11_cast_fp16 = reshape(shape = var_481, x = attn_5_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor var_485 = const()[name = tensor("op_485"), val = tensor([1, 1])]; + tensor var_487 = const()[name = tensor("op_487"), val = tensor([1, 1])]; + tensor obj_21_pad_type_0 = const()[name = tensor("obj_21_pad_type_0"), val = tensor("custom")]; + tensor obj_21_pad_0 = const()[name = tensor("obj_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196244224)))]; + tensor layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199521088)))]; + tensor obj_21_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = var_487, groups = var_397, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = var_485, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("obj_21_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_21_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor var_497 = const()[name = tensor("op_497"), val = tensor([1])]; + tensor channels_mean_9_cast_fp16 = reduce_mean(axes = var_497, keep_dims = var_398, x = inputs_9_cast_fp16)[name = tensor("channels_mean_9_cast_fp16")]; + tensor zero_mean_9_cast_fp16 = sub(x = inputs_9_cast_fp16, y = channels_mean_9_cast_fp16)[name = tensor("zero_mean_9_cast_fp16")]; + tensor zero_mean_sq_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = zero_mean_9_cast_fp16)[name = tensor("zero_mean_sq_9_cast_fp16")]; + tensor var_501 = const()[name = tensor("op_501"), val = tensor([1])]; + tensor var_502_cast_fp16 = reduce_mean(axes = var_501, keep_dims = var_398, x = zero_mean_sq_9_cast_fp16)[name = tensor("op_502_cast_fp16")]; + tensor var_503_to_fp16 = const()[name = tensor("op_503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_504_cast_fp16 = add(x = var_502_cast_fp16, y = var_503_to_fp16)[name = tensor("op_504_cast_fp16")]; + tensor denom_9_epsilon_0_to_fp16 = const()[name = tensor("denom_9_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_9_cast_fp16 = rsqrt(epsilon = denom_9_epsilon_0_to_fp16, x = var_504_cast_fp16)[name = tensor("denom_9_cast_fp16")]; + tensor out_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = denom_9_cast_fp16)[name = tensor("out_9_cast_fp16")]; + tensor obj_23_gamma_0_to_fp16 = const()[name = tensor("obj_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199523712)))]; + tensor obj_23_beta_0_to_fp16 = const()[name = tensor("obj_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199526336)))]; + tensor obj_23_epsilon_0_to_fp16 = const()[name = tensor("obj_23_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor("obj_23_cast_fp16")]; + tensor var_519 = const()[name = tensor("op_519"), val = tensor([1, 1])]; + tensor var_521 = const()[name = tensor("op_521"), val = tensor([1, 1])]; + tensor query_7_pad_type_0 = const()[name = tensor("query_7_pad_type_0"), val = tensor("custom")]; + tensor query_7_pad_0 = const()[name = tensor("query_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199528960)))]; + tensor layers_1_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202805824)))]; + tensor query_7_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = var_521, groups = var_397, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = var_519, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor("query_7_cast_fp16")]; + tensor var_525 = const()[name = tensor("op_525"), val = tensor([1, 1])]; + tensor var_527 = const()[name = tensor("op_527"), val = tensor([1, 1])]; + tensor key_7_pad_type_0 = const()[name = tensor("key_7_pad_type_0"), val = tensor("custom")]; + tensor key_7_pad_0 = const()[name = tensor("key_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202808448)))]; + tensor key_7_cast_fp16 = conv(dilations = var_527, groups = var_397, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = var_525, weight = layers_1_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_7_cast_fp16")]; + tensor var_532 = const()[name = tensor("op_532"), val = tensor([1, 1])]; + tensor var_534 = const()[name = tensor("op_534"), val = tensor([1, 1])]; + tensor value_7_pad_type_0 = const()[name = tensor("value_7_pad_type_0"), val = tensor("custom")]; + tensor value_7_pad_0 = const()[name = tensor("value_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206085312)))]; + tensor layers_1_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209362176)))]; + tensor value_7_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_bias_to_fp16, dilations = var_534, groups = var_397, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = var_532, weight = layers_1_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_7_cast_fp16")]; + tensor var_538 = const()[name = tensor("op_538"), val = tensor([1, 20, 64, -1])]; + tensor var_539_cast_fp16 = reshape(shape = var_538, x = query_7_cast_fp16)[name = tensor("op_539_cast_fp16")]; + tensor var_540_to_fp16 = const()[name = tensor("op_540_to_fp16"), val = tensor(0x1p-3)]; + tensor var_541_cast_fp16 = mul(x = var_539_cast_fp16, y = var_540_to_fp16)[name = tensor("op_541_cast_fp16")]; + tensor var_542 = const()[name = tensor("op_542"), val = tensor([1, 20, 64, -1])]; + tensor var_543_cast_fp16 = reshape(shape = var_542, x = key_7_cast_fp16)[name = tensor("op_543_cast_fp16")]; + tensor mh_w_11_transpose_x_0 = const()[name = tensor("mh_w_11_transpose_x_0"), val = tensor(true)]; + tensor mh_w_11_transpose_y_0 = const()[name = tensor("mh_w_11_transpose_y_0"), val = tensor(false)]; + tensor mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_541_cast_fp16, y = var_543_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; + tensor obj_27_cast_fp16 = softmax(axis = var_390, x = mh_w_11_cast_fp16)[name = tensor("obj_27_cast_fp16")]; + tensor var_547 = const()[name = tensor("op_547"), val = tensor([1, 20, 64, -1])]; + tensor var_548_cast_fp16 = reshape(shape = var_547, x = value_7_cast_fp16)[name = tensor("op_548_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(true)]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_548_cast_fp16, y = obj_27_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_551 = const()[name = tensor("op_551"), val = tensor([1, 1280, 1, -1])]; + tensor input_13_cast_fp16 = reshape(shape = var_551, x = attn_7_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor var_555 = const()[name = tensor("op_555"), val = tensor([1, 1])]; + tensor var_557 = const()[name = tensor("op_557"), val = tensor([1, 1])]; + tensor obj_25_pad_type_0 = const()[name = tensor("obj_25_pad_type_0"), val = tensor("custom")]; + tensor obj_25_pad_0 = const()[name = tensor("obj_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209364800)))]; + tensor layers_1_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212641664)))]; + tensor obj_25_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = var_557, groups = var_397, pad = obj_25_pad_0, pad_type = obj_25_pad_type_0, strides = var_555, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("obj_25_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_25_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor var_563 = const()[name = tensor("op_563"), val = tensor([1])]; + tensor channels_mean_11_cast_fp16 = reduce_mean(axes = var_563, keep_dims = var_398, x = inputs_11_cast_fp16)[name = tensor("channels_mean_11_cast_fp16")]; + tensor zero_mean_11_cast_fp16 = sub(x = inputs_11_cast_fp16, y = channels_mean_11_cast_fp16)[name = tensor("zero_mean_11_cast_fp16")]; + tensor zero_mean_sq_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = zero_mean_11_cast_fp16)[name = tensor("zero_mean_sq_11_cast_fp16")]; + tensor var_567 = const()[name = tensor("op_567"), val = tensor([1])]; + tensor var_568_cast_fp16 = reduce_mean(axes = var_567, keep_dims = var_398, x = zero_mean_sq_11_cast_fp16)[name = tensor("op_568_cast_fp16")]; + tensor var_569_to_fp16 = const()[name = tensor("op_569_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_570_cast_fp16 = add(x = var_568_cast_fp16, y = var_569_to_fp16)[name = tensor("op_570_cast_fp16")]; + tensor denom_11_epsilon_0_to_fp16 = const()[name = tensor("denom_11_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_11_cast_fp16 = rsqrt(epsilon = denom_11_epsilon_0_to_fp16, x = var_570_cast_fp16)[name = tensor("denom_11_cast_fp16")]; + tensor out_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = denom_11_cast_fp16)[name = tensor("out_11_cast_fp16")]; + tensor input_15_gamma_0_to_fp16 = const()[name = tensor("input_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212644288)))]; + tensor input_15_beta_0_to_fp16 = const()[name = tensor("input_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212646912)))]; + tensor input_15_epsilon_0_to_fp16 = const()[name = tensor("input_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = batch_norm(beta = input_15_beta_0_to_fp16, epsilon = input_15_epsilon_0_to_fp16, gamma = input_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor var_581 = const()[name = tensor("op_581"), val = tensor([1, 1])]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 1])]; + tensor input_17_pad_type_0 = const()[name = tensor("input_17_pad_type_0"), val = tensor("custom")]; + tensor input_17_pad_0 = const()[name = tensor("input_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_fc1_weight_to_fp16 = const()[name = tensor("layers_1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212649536)))]; + tensor layers_1_fc1_bias_to_fp16 = const()[name = tensor("layers_1_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(225756800)))]; + tensor input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = var_583, groups = var_397, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = var_581, weight = layers_1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_mode_0 = const()[name = tensor("input_19_mode_0"), val = tensor("EXACT")]; + tensor input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor var_589 = const()[name = tensor("op_589"), val = tensor([1, 1])]; + tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1])]; + tensor hidden_states_5_pad_type_0 = const()[name = tensor("hidden_states_5_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_5_pad_0 = const()[name = tensor("hidden_states_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_fc2_weight_to_fp16 = const()[name = tensor("layers_1_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(225767104)))]; + tensor layers_1_fc2_bias_to_fp16 = const()[name = tensor("layers_1_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238874368)))]; + tensor hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = var_591, groups = var_397, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = var_589, weight = layers_1_fc2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("hidden_states_5_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_604 = const()[name = tensor("op_604"), val = tensor(3)]; + tensor var_611 = const()[name = tensor("op_611"), val = tensor(1)]; + tensor var_612 = const()[name = tensor("op_612"), val = tensor(true)]; + tensor var_624 = const()[name = tensor("op_624"), val = tensor([1])]; + tensor channels_mean_13_cast_fp16 = reduce_mean(axes = var_624, keep_dims = var_612, x = inputs_13_cast_fp16)[name = tensor("channels_mean_13_cast_fp16")]; + tensor zero_mean_13_cast_fp16 = sub(x = inputs_13_cast_fp16, y = channels_mean_13_cast_fp16)[name = tensor("zero_mean_13_cast_fp16")]; + tensor zero_mean_sq_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = zero_mean_13_cast_fp16)[name = tensor("zero_mean_sq_13_cast_fp16")]; + tensor var_628 = const()[name = tensor("op_628"), val = tensor([1])]; + tensor var_629_cast_fp16 = reduce_mean(axes = var_628, keep_dims = var_612, x = zero_mean_sq_13_cast_fp16)[name = tensor("op_629_cast_fp16")]; + tensor var_630_to_fp16 = const()[name = tensor("op_630_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_631_cast_fp16 = add(x = var_629_cast_fp16, y = var_630_to_fp16)[name = tensor("op_631_cast_fp16")]; + tensor denom_13_epsilon_0_to_fp16 = const()[name = tensor("denom_13_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_13_cast_fp16 = rsqrt(epsilon = denom_13_epsilon_0_to_fp16, x = var_631_cast_fp16)[name = tensor("denom_13_cast_fp16")]; + tensor out_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = denom_13_cast_fp16)[name = tensor("out_13_cast_fp16")]; + tensor obj_29_gamma_0_to_fp16 = const()[name = tensor("obj_29_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238876992)))]; + tensor obj_29_beta_0_to_fp16 = const()[name = tensor("obj_29_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238879616)))]; + tensor obj_29_epsilon_0_to_fp16 = const()[name = tensor("obj_29_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor("obj_29_cast_fp16")]; + tensor var_646 = const()[name = tensor("op_646"), val = tensor([1, 1])]; + tensor var_648 = const()[name = tensor("op_648"), val = tensor([1, 1])]; + tensor query_9_pad_type_0 = const()[name = tensor("query_9_pad_type_0"), val = tensor("custom")]; + tensor query_9_pad_0 = const()[name = tensor("query_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238882240)))]; + tensor layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242159104)))]; + tensor query_9_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = var_648, groups = var_611, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = var_646, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("query_9_cast_fp16")]; + tensor var_652 = const()[name = tensor("op_652"), val = tensor([1, 1])]; + tensor var_654 = const()[name = tensor("op_654"), val = tensor([1, 1])]; + tensor current_key_5_pad_type_0 = const()[name = tensor("current_key_5_pad_type_0"), val = tensor("custom")]; + tensor current_key_5_pad_0 = const()[name = tensor("current_key_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242161728)))]; + tensor current_key_5_cast_fp16 = conv(dilations = var_654, groups = var_611, pad = current_key_5_pad_0, pad_type = current_key_5_pad_type_0, strides = var_652, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("current_key_5_cast_fp16")]; + tensor var_659 = const()[name = tensor("op_659"), val = tensor([1, 1])]; + tensor var_661 = const()[name = tensor("op_661"), val = tensor([1, 1])]; + tensor current_value_5_pad_type_0 = const()[name = tensor("current_value_5_pad_type_0"), val = tensor("custom")]; + tensor current_value_5_pad_0 = const()[name = tensor("current_value_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245438592)))]; + tensor layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248715456)))]; + tensor current_value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = var_661, groups = var_611, pad = current_value_5_pad_0, pad_type = current_value_5_pad_type_0, strides = var_659, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("current_value_5_cast_fp16")]; + tensor var_668_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_668_cast_fp16")]; + tensor var_670_cast_fp16 = mul(x = var_103_cast_fp16_2, y = var_241_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor key_9_cast_fp16 = add(x = var_668_cast_fp16, y = var_670_cast_fp16)[name = tensor("key_9_cast_fp16")]; + tensor var_672_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_672_cast_fp16")]; + tensor var_674_cast_fp16 = mul(x = var_138_cast_fp16_2, y = var_241_cast_fp16)[name = tensor("op_674_cast_fp16")]; + tensor value_9_cast_fp16 = add(x = var_672_cast_fp16, y = var_674_cast_fp16)[name = tensor("value_9_cast_fp16")]; + tensor var_677 = const()[name = tensor("op_677"), val = tensor([1, 20, 64, -1])]; + tensor var_678_cast_fp16 = reshape(shape = var_677, x = query_9_cast_fp16)[name = tensor("op_678_cast_fp16")]; + tensor var_679_to_fp16 = const()[name = tensor("op_679_to_fp16"), val = tensor(0x1p-3)]; + tensor var_680_cast_fp16 = mul(x = var_678_cast_fp16, y = var_679_to_fp16)[name = tensor("op_680_cast_fp16")]; + tensor var_681 = const()[name = tensor("op_681"), val = tensor([1, 20, 64, -1])]; + tensor var_682_cast_fp16 = reshape(shape = var_681, x = key_9_cast_fp16)[name = tensor("op_682_cast_fp16")]; + tensor mh_w_13_transpose_x_0 = const()[name = tensor("mh_w_13_transpose_x_0"), val = tensor(true)]; + tensor mh_w_13_transpose_y_0 = const()[name = tensor("mh_w_13_transpose_y_0"), val = tensor(false)]; + tensor mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_680_cast_fp16, y = var_682_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; + tensor mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; + tensor var_690_cast_fp16 = softmax(axis = var_604, x = mh_w_15_cast_fp16)[name = tensor("op_690_cast_fp16")]; + tensor var_691 = const()[name = tensor("op_691"), val = tensor([1, 20, 64, -1])]; + tensor var_692_cast_fp16 = reshape(shape = var_691, x = value_9_cast_fp16)[name = tensor("op_692_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(true)]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_692_cast_fp16, y = var_690_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_695 = const()[name = tensor("op_695"), val = tensor([1, 1280, 1, -1])]; + tensor input_21_cast_fp16 = reshape(shape = var_695, x = attn_9_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_699 = const()[name = tensor("op_699"), val = tensor([1, 1])]; + tensor var_701 = const()[name = tensor("op_701"), val = tensor([1, 1])]; + tensor obj_35_pad_type_0 = const()[name = tensor("obj_35_pad_type_0"), val = tensor("custom")]; + tensor obj_35_pad_0 = const()[name = tensor("obj_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248718080)))]; + tensor layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251994944)))]; + tensor obj_35_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = var_701, groups = var_611, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = var_699, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("obj_35_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_35_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor var_711 = const()[name = tensor("op_711"), val = tensor([1])]; + tensor channels_mean_15_cast_fp16 = reduce_mean(axes = var_711, keep_dims = var_612, x = inputs_15_cast_fp16)[name = tensor("channels_mean_15_cast_fp16")]; + tensor zero_mean_15_cast_fp16 = sub(x = inputs_15_cast_fp16, y = channels_mean_15_cast_fp16)[name = tensor("zero_mean_15_cast_fp16")]; + tensor zero_mean_sq_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = zero_mean_15_cast_fp16)[name = tensor("zero_mean_sq_15_cast_fp16")]; + tensor var_715 = const()[name = tensor("op_715"), val = tensor([1])]; + tensor var_716_cast_fp16 = reduce_mean(axes = var_715, keep_dims = var_612, x = zero_mean_sq_15_cast_fp16)[name = tensor("op_716_cast_fp16")]; + tensor var_717_to_fp16 = const()[name = tensor("op_717_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_718_cast_fp16 = add(x = var_716_cast_fp16, y = var_717_to_fp16)[name = tensor("op_718_cast_fp16")]; + tensor denom_15_epsilon_0_to_fp16 = const()[name = tensor("denom_15_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_15_cast_fp16 = rsqrt(epsilon = denom_15_epsilon_0_to_fp16, x = var_718_cast_fp16)[name = tensor("denom_15_cast_fp16")]; + tensor out_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = denom_15_cast_fp16)[name = tensor("out_15_cast_fp16")]; + tensor obj_37_gamma_0_to_fp16 = const()[name = tensor("obj_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251997568)))]; + tensor obj_37_beta_0_to_fp16 = const()[name = tensor("obj_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252000192)))]; + tensor obj_37_epsilon_0_to_fp16 = const()[name = tensor("obj_37_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor("obj_37_cast_fp16")]; + tensor var_733 = const()[name = tensor("op_733"), val = tensor([1, 1])]; + tensor var_735 = const()[name = tensor("op_735"), val = tensor([1, 1])]; + tensor query_11_pad_type_0 = const()[name = tensor("query_11_pad_type_0"), val = tensor("custom")]; + tensor query_11_pad_0 = const()[name = tensor("query_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252002816)))]; + tensor layers_2_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255279680)))]; + tensor query_11_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_bias_to_fp16, dilations = var_735, groups = var_611, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = var_733, weight = layers_2_encoder_attn_q_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("query_11_cast_fp16")]; + tensor var_739 = const()[name = tensor("op_739"), val = tensor([1, 1])]; + tensor var_741 = const()[name = tensor("op_741"), val = tensor([1, 1])]; + tensor key_11_pad_type_0 = const()[name = tensor("key_11_pad_type_0"), val = tensor("custom")]; + tensor key_11_pad_0 = const()[name = tensor("key_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255282304)))]; + tensor key_11_cast_fp16 = conv(dilations = var_741, groups = var_611, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = var_739, weight = layers_2_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_11_cast_fp16")]; + tensor var_746 = const()[name = tensor("op_746"), val = tensor([1, 1])]; + tensor var_748 = const()[name = tensor("op_748"), val = tensor([1, 1])]; + tensor value_11_pad_type_0 = const()[name = tensor("value_11_pad_type_0"), val = tensor("custom")]; + tensor value_11_pad_0 = const()[name = tensor("value_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258559168)))]; + tensor layers_2_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261836032)))]; + tensor value_11_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_bias_to_fp16, dilations = var_748, groups = var_611, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = var_746, weight = layers_2_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_11_cast_fp16")]; + tensor var_752 = const()[name = tensor("op_752"), val = tensor([1, 20, 64, -1])]; + tensor var_753_cast_fp16 = reshape(shape = var_752, x = query_11_cast_fp16)[name = tensor("op_753_cast_fp16")]; + tensor var_754_to_fp16 = const()[name = tensor("op_754_to_fp16"), val = tensor(0x1p-3)]; + tensor var_755_cast_fp16 = mul(x = var_753_cast_fp16, y = var_754_to_fp16)[name = tensor("op_755_cast_fp16")]; + tensor var_756 = const()[name = tensor("op_756"), val = tensor([1, 20, 64, -1])]; + tensor var_757_cast_fp16 = reshape(shape = var_756, x = key_11_cast_fp16)[name = tensor("op_757_cast_fp16")]; + tensor mh_w_17_transpose_x_0 = const()[name = tensor("mh_w_17_transpose_x_0"), val = tensor(true)]; + tensor mh_w_17_transpose_y_0 = const()[name = tensor("mh_w_17_transpose_y_0"), val = tensor(false)]; + tensor mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_755_cast_fp16, y = var_757_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; + tensor obj_41_cast_fp16 = softmax(axis = var_604, x = mh_w_17_cast_fp16)[name = tensor("obj_41_cast_fp16")]; + tensor var_761 = const()[name = tensor("op_761"), val = tensor([1, 20, 64, -1])]; + tensor var_762_cast_fp16 = reshape(shape = var_761, x = value_11_cast_fp16)[name = tensor("op_762_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(true)]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_762_cast_fp16, y = obj_41_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 1280, 1, -1])]; + tensor input_23_cast_fp16 = reshape(shape = var_765, x = attn_11_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor var_769 = const()[name = tensor("op_769"), val = tensor([1, 1])]; + tensor var_771 = const()[name = tensor("op_771"), val = tensor([1, 1])]; + tensor obj_39_pad_type_0 = const()[name = tensor("obj_39_pad_type_0"), val = tensor("custom")]; + tensor obj_39_pad_0 = const()[name = tensor("obj_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261838656)))]; + tensor layers_2_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265115520)))]; + tensor obj_39_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = var_771, groups = var_611, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = var_769, weight = layers_2_encoder_attn_o_proj_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("obj_39_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_39_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor var_777 = const()[name = tensor("op_777"), val = tensor([1])]; + tensor channels_mean_17_cast_fp16 = reduce_mean(axes = var_777, keep_dims = var_612, x = inputs_17_cast_fp16)[name = tensor("channels_mean_17_cast_fp16")]; + tensor zero_mean_17_cast_fp16 = sub(x = inputs_17_cast_fp16, y = channels_mean_17_cast_fp16)[name = tensor("zero_mean_17_cast_fp16")]; + tensor zero_mean_sq_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = zero_mean_17_cast_fp16)[name = tensor("zero_mean_sq_17_cast_fp16")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor([1])]; + tensor var_782_cast_fp16 = reduce_mean(axes = var_781, keep_dims = var_612, x = zero_mean_sq_17_cast_fp16)[name = tensor("op_782_cast_fp16")]; + tensor var_783_to_fp16 = const()[name = tensor("op_783_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_784_cast_fp16 = add(x = var_782_cast_fp16, y = var_783_to_fp16)[name = tensor("op_784_cast_fp16")]; + tensor denom_17_epsilon_0_to_fp16 = const()[name = tensor("denom_17_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_17_cast_fp16 = rsqrt(epsilon = denom_17_epsilon_0_to_fp16, x = var_784_cast_fp16)[name = tensor("denom_17_cast_fp16")]; + tensor out_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = denom_17_cast_fp16)[name = tensor("out_17_cast_fp16")]; + tensor input_25_gamma_0_to_fp16 = const()[name = tensor("input_25_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265118144)))]; + tensor input_25_beta_0_to_fp16 = const()[name = tensor("input_25_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265120768)))]; + tensor input_25_epsilon_0_to_fp16 = const()[name = tensor("input_25_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_25_cast_fp16 = batch_norm(beta = input_25_beta_0_to_fp16, epsilon = input_25_epsilon_0_to_fp16, gamma = input_25_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor var_795 = const()[name = tensor("op_795"), val = tensor([1, 1])]; + tensor var_797 = const()[name = tensor("op_797"), val = tensor([1, 1])]; + tensor input_27_pad_type_0 = const()[name = tensor("input_27_pad_type_0"), val = tensor("custom")]; + tensor input_27_pad_0 = const()[name = tensor("input_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_fc1_weight_to_fp16 = const()[name = tensor("layers_2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265123392)))]; + tensor layers_2_fc1_bias_to_fp16 = const()[name = tensor("layers_2_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278230656)))]; + tensor input_27_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = var_797, groups = var_611, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = var_795, weight = layers_2_fc1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_mode_0 = const()[name = tensor("input_29_mode_0"), val = tensor("EXACT")]; + tensor input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor var_803 = const()[name = tensor("op_803"), val = tensor([1, 1])]; + tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 1])]; + tensor hidden_states_7_pad_type_0 = const()[name = tensor("hidden_states_7_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_7_pad_0 = const()[name = tensor("hidden_states_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_fc2_weight_to_fp16 = const()[name = tensor("layers_2_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278240960)))]; + tensor layers_2_fc2_bias_to_fp16 = const()[name = tensor("layers_2_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291348224)))]; + tensor hidden_states_7_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = var_805, groups = var_611, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = var_803, weight = layers_2_fc2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("hidden_states_7_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor var_818 = const()[name = tensor("op_818"), val = tensor(3)]; + tensor var_825 = const()[name = tensor("op_825"), val = tensor(1)]; + tensor var_826 = const()[name = tensor("op_826"), val = tensor(true)]; + tensor var_838 = const()[name = tensor("op_838"), val = tensor([1])]; + tensor channels_mean_19_cast_fp16 = reduce_mean(axes = var_838, keep_dims = var_826, x = inputs_19_cast_fp16)[name = tensor("channels_mean_19_cast_fp16")]; + tensor zero_mean_19_cast_fp16 = sub(x = inputs_19_cast_fp16, y = channels_mean_19_cast_fp16)[name = tensor("zero_mean_19_cast_fp16")]; + tensor zero_mean_sq_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = zero_mean_19_cast_fp16)[name = tensor("zero_mean_sq_19_cast_fp16")]; + tensor var_842 = const()[name = tensor("op_842"), val = tensor([1])]; + tensor var_843_cast_fp16 = reduce_mean(axes = var_842, keep_dims = var_826, x = zero_mean_sq_19_cast_fp16)[name = tensor("op_843_cast_fp16")]; + tensor var_844_to_fp16 = const()[name = tensor("op_844_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_845_cast_fp16 = add(x = var_843_cast_fp16, y = var_844_to_fp16)[name = tensor("op_845_cast_fp16")]; + tensor denom_19_epsilon_0_to_fp16 = const()[name = tensor("denom_19_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_19_cast_fp16 = rsqrt(epsilon = denom_19_epsilon_0_to_fp16, x = var_845_cast_fp16)[name = tensor("denom_19_cast_fp16")]; + tensor out_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = denom_19_cast_fp16)[name = tensor("out_19_cast_fp16")]; + tensor obj_43_gamma_0_to_fp16 = const()[name = tensor("obj_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291350848)))]; + tensor obj_43_beta_0_to_fp16 = const()[name = tensor("obj_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291353472)))]; + tensor obj_43_epsilon_0_to_fp16 = const()[name = tensor("obj_43_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor("obj_43_cast_fp16")]; + tensor var_860 = const()[name = tensor("op_860"), val = tensor([1, 1])]; + tensor var_862 = const()[name = tensor("op_862"), val = tensor([1, 1])]; + tensor query_13_pad_type_0 = const()[name = tensor("query_13_pad_type_0"), val = tensor("custom")]; + tensor query_13_pad_0 = const()[name = tensor("query_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291356096)))]; + tensor layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294632960)))]; + tensor query_13_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = var_862, groups = var_825, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = var_860, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("query_13_cast_fp16")]; + tensor var_866 = const()[name = tensor("op_866"), val = tensor([1, 1])]; + tensor var_868 = const()[name = tensor("op_868"), val = tensor([1, 1])]; + tensor current_key_7_pad_type_0 = const()[name = tensor("current_key_7_pad_type_0"), val = tensor("custom")]; + tensor current_key_7_pad_0 = const()[name = tensor("current_key_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294635584)))]; + tensor current_key_7_cast_fp16 = conv(dilations = var_868, groups = var_825, pad = current_key_7_pad_0, pad_type = current_key_7_pad_type_0, strides = var_866, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("current_key_7_cast_fp16")]; + tensor var_873 = const()[name = tensor("op_873"), val = tensor([1, 1])]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([1, 1])]; + tensor current_value_7_pad_type_0 = const()[name = tensor("current_value_7_pad_type_0"), val = tensor("custom")]; + tensor current_value_7_pad_0 = const()[name = tensor("current_value_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297912448)))]; + tensor layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(301189312)))]; + tensor current_value_7_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = var_875, groups = var_825, pad = current_value_7_pad_0, pad_type = current_value_7_pad_type_0, strides = var_873, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("current_value_7_cast_fp16")]; + tensor var_882_cast_fp16 = mul(x = current_key_7_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_882_cast_fp16")]; + tensor var_884_cast_fp16 = mul(x = var_103_cast_fp16_3, y = var_241_cast_fp16)[name = tensor("op_884_cast_fp16")]; + tensor key_13_cast_fp16 = add(x = var_882_cast_fp16, y = var_884_cast_fp16)[name = tensor("key_13_cast_fp16")]; + tensor var_886_cast_fp16 = mul(x = current_value_7_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_886_cast_fp16")]; + tensor var_888_cast_fp16 = mul(x = var_138_cast_fp16_3, y = var_241_cast_fp16)[name = tensor("op_888_cast_fp16")]; + tensor value_13_cast_fp16 = add(x = var_886_cast_fp16, y = var_888_cast_fp16)[name = tensor("value_13_cast_fp16")]; + tensor var_891 = const()[name = tensor("op_891"), val = tensor([1, 20, 64, -1])]; + tensor var_892_cast_fp16 = reshape(shape = var_891, x = query_13_cast_fp16)[name = tensor("op_892_cast_fp16")]; + tensor var_893_to_fp16 = const()[name = tensor("op_893_to_fp16"), val = tensor(0x1p-3)]; + tensor var_894_cast_fp16 = mul(x = var_892_cast_fp16, y = var_893_to_fp16)[name = tensor("op_894_cast_fp16")]; + tensor var_895 = const()[name = tensor("op_895"), val = tensor([1, 20, 64, -1])]; + tensor var_896_cast_fp16 = reshape(shape = var_895, x = key_13_cast_fp16)[name = tensor("op_896_cast_fp16")]; + tensor mh_w_19_transpose_x_0 = const()[name = tensor("mh_w_19_transpose_x_0"), val = tensor(true)]; + tensor mh_w_19_transpose_y_0 = const()[name = tensor("mh_w_19_transpose_y_0"), val = tensor(false)]; + tensor mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_894_cast_fp16, y = var_896_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; + tensor mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; + tensor var_904_cast_fp16 = softmax(axis = var_818, x = mh_w_21_cast_fp16)[name = tensor("op_904_cast_fp16")]; + tensor var_905 = const()[name = tensor("op_905"), val = tensor([1, 20, 64, -1])]; + tensor var_906_cast_fp16 = reshape(shape = var_905, x = value_13_cast_fp16)[name = tensor("op_906_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(true)]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_906_cast_fp16, y = var_904_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_909 = const()[name = tensor("op_909"), val = tensor([1, 1280, 1, -1])]; + tensor input_31_cast_fp16 = reshape(shape = var_909, x = attn_13_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor var_913 = const()[name = tensor("op_913"), val = tensor([1, 1])]; + tensor var_915 = const()[name = tensor("op_915"), val = tensor([1, 1])]; + tensor obj_49_pad_type_0 = const()[name = tensor("obj_49_pad_type_0"), val = tensor("custom")]; + tensor obj_49_pad_0 = const()[name = tensor("obj_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(301191936)))]; + tensor layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(304468800)))]; + tensor obj_49_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = var_915, groups = var_825, pad = obj_49_pad_0, pad_type = obj_49_pad_type_0, strides = var_913, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("obj_49_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_49_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor var_925 = const()[name = tensor("op_925"), val = tensor([1])]; + tensor channels_mean_21_cast_fp16 = reduce_mean(axes = var_925, keep_dims = var_826, x = inputs_21_cast_fp16)[name = tensor("channels_mean_21_cast_fp16")]; + tensor zero_mean_21_cast_fp16 = sub(x = inputs_21_cast_fp16, y = channels_mean_21_cast_fp16)[name = tensor("zero_mean_21_cast_fp16")]; + tensor zero_mean_sq_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = zero_mean_21_cast_fp16)[name = tensor("zero_mean_sq_21_cast_fp16")]; + tensor var_929 = const()[name = tensor("op_929"), val = tensor([1])]; + tensor var_930_cast_fp16 = reduce_mean(axes = var_929, keep_dims = var_826, x = zero_mean_sq_21_cast_fp16)[name = tensor("op_930_cast_fp16")]; + tensor var_931_to_fp16 = const()[name = tensor("op_931_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_932_cast_fp16 = add(x = var_930_cast_fp16, y = var_931_to_fp16)[name = tensor("op_932_cast_fp16")]; + tensor denom_21_epsilon_0_to_fp16 = const()[name = tensor("denom_21_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_21_cast_fp16 = rsqrt(epsilon = denom_21_epsilon_0_to_fp16, x = var_932_cast_fp16)[name = tensor("denom_21_cast_fp16")]; + tensor out_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = denom_21_cast_fp16)[name = tensor("out_21_cast_fp16")]; + tensor obj_51_gamma_0_to_fp16 = const()[name = tensor("obj_51_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(304471424)))]; + tensor obj_51_beta_0_to_fp16 = const()[name = tensor("obj_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(304474048)))]; + tensor obj_51_epsilon_0_to_fp16 = const()[name = tensor("obj_51_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor("obj_51_cast_fp16")]; + tensor var_947 = const()[name = tensor("op_947"), val = tensor([1, 1])]; + tensor var_949 = const()[name = tensor("op_949"), val = tensor([1, 1])]; + tensor query_15_pad_type_0 = const()[name = tensor("query_15_pad_type_0"), val = tensor("custom")]; + tensor query_15_pad_0 = const()[name = tensor("query_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(304476672)))]; + tensor layers_3_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307753536)))]; + tensor query_15_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_bias_to_fp16, dilations = var_949, groups = var_825, pad = query_15_pad_0, pad_type = query_15_pad_type_0, strides = var_947, weight = layers_3_encoder_attn_q_proj_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor("query_15_cast_fp16")]; + tensor var_953 = const()[name = tensor("op_953"), val = tensor([1, 1])]; + tensor var_955 = const()[name = tensor("op_955"), val = tensor([1, 1])]; + tensor key_15_pad_type_0 = const()[name = tensor("key_15_pad_type_0"), val = tensor("custom")]; + tensor key_15_pad_0 = const()[name = tensor("key_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307756160)))]; + tensor key_15_cast_fp16 = conv(dilations = var_955, groups = var_825, pad = key_15_pad_0, pad_type = key_15_pad_type_0, strides = var_953, weight = layers_3_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_15_cast_fp16")]; + tensor var_960 = const()[name = tensor("op_960"), val = tensor([1, 1])]; + tensor var_962 = const()[name = tensor("op_962"), val = tensor([1, 1])]; + tensor value_15_pad_type_0 = const()[name = tensor("value_15_pad_type_0"), val = tensor("custom")]; + tensor value_15_pad_0 = const()[name = tensor("value_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311033024)))]; + tensor layers_3_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314309888)))]; + tensor value_15_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_bias_to_fp16, dilations = var_962, groups = var_825, pad = value_15_pad_0, pad_type = value_15_pad_type_0, strides = var_960, weight = layers_3_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_15_cast_fp16")]; + tensor var_966 = const()[name = tensor("op_966"), val = tensor([1, 20, 64, -1])]; + tensor var_967_cast_fp16 = reshape(shape = var_966, x = query_15_cast_fp16)[name = tensor("op_967_cast_fp16")]; + tensor var_968_to_fp16 = const()[name = tensor("op_968_to_fp16"), val = tensor(0x1p-3)]; + tensor var_969_cast_fp16 = mul(x = var_967_cast_fp16, y = var_968_to_fp16)[name = tensor("op_969_cast_fp16")]; + tensor var_970 = const()[name = tensor("op_970"), val = tensor([1, 20, 64, -1])]; + tensor var_971_cast_fp16 = reshape(shape = var_970, x = key_15_cast_fp16)[name = tensor("op_971_cast_fp16")]; + tensor mh_w_23_transpose_x_0 = const()[name = tensor("mh_w_23_transpose_x_0"), val = tensor(true)]; + tensor mh_w_23_transpose_y_0 = const()[name = tensor("mh_w_23_transpose_y_0"), val = tensor(false)]; + tensor mh_w_23_cast_fp16 = matmul(transpose_x = mh_w_23_transpose_x_0, transpose_y = mh_w_23_transpose_y_0, x = var_969_cast_fp16, y = var_971_cast_fp16)[name = tensor("mh_w_23_cast_fp16")]; + tensor obj_55_cast_fp16 = softmax(axis = var_818, x = mh_w_23_cast_fp16)[name = tensor("obj_55_cast_fp16")]; + tensor var_975 = const()[name = tensor("op_975"), val = tensor([1, 20, 64, -1])]; + tensor var_976_cast_fp16 = reshape(shape = var_975, x = value_15_cast_fp16)[name = tensor("op_976_cast_fp16")]; + tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; + tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(true)]; + tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = var_976_cast_fp16, y = obj_55_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_979 = const()[name = tensor("op_979"), val = tensor([1, 1280, 1, -1])]; + tensor input_33_cast_fp16 = reshape(shape = var_979, x = attn_15_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor var_983 = const()[name = tensor("op_983"), val = tensor([1, 1])]; + tensor var_985 = const()[name = tensor("op_985"), val = tensor([1, 1])]; + tensor obj_53_pad_type_0 = const()[name = tensor("obj_53_pad_type_0"), val = tensor("custom")]; + tensor obj_53_pad_0 = const()[name = tensor("obj_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314312512)))]; + tensor layers_3_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317589376)))]; + tensor obj_53_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = var_985, groups = var_825, pad = obj_53_pad_0, pad_type = obj_53_pad_type_0, strides = var_983, weight = layers_3_encoder_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("obj_53_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_53_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor var_991 = const()[name = tensor("op_991"), val = tensor([1])]; + tensor channels_mean_23_cast_fp16 = reduce_mean(axes = var_991, keep_dims = var_826, x = inputs_23_cast_fp16)[name = tensor("channels_mean_23_cast_fp16")]; + tensor zero_mean_23_cast_fp16 = sub(x = inputs_23_cast_fp16, y = channels_mean_23_cast_fp16)[name = tensor("zero_mean_23_cast_fp16")]; + tensor zero_mean_sq_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = zero_mean_23_cast_fp16)[name = tensor("zero_mean_sq_23_cast_fp16")]; + tensor var_995 = const()[name = tensor("op_995"), val = tensor([1])]; + tensor var_996_cast_fp16 = reduce_mean(axes = var_995, keep_dims = var_826, x = zero_mean_sq_23_cast_fp16)[name = tensor("op_996_cast_fp16")]; + tensor var_997_to_fp16 = const()[name = tensor("op_997_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_998_cast_fp16 = add(x = var_996_cast_fp16, y = var_997_to_fp16)[name = tensor("op_998_cast_fp16")]; + tensor denom_23_epsilon_0_to_fp16 = const()[name = tensor("denom_23_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_23_cast_fp16 = rsqrt(epsilon = denom_23_epsilon_0_to_fp16, x = var_998_cast_fp16)[name = tensor("denom_23_cast_fp16")]; + tensor out_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = denom_23_cast_fp16)[name = tensor("out_23_cast_fp16")]; + tensor input_35_gamma_0_to_fp16 = const()[name = tensor("input_35_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317592000)))]; + tensor input_35_beta_0_to_fp16 = const()[name = tensor("input_35_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317594624)))]; + tensor input_35_epsilon_0_to_fp16 = const()[name = tensor("input_35_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = batch_norm(beta = input_35_beta_0_to_fp16, epsilon = input_35_epsilon_0_to_fp16, gamma = input_35_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor var_1009 = const()[name = tensor("op_1009"), val = tensor([1, 1])]; + tensor var_1011 = const()[name = tensor("op_1011"), val = tensor([1, 1])]; + tensor input_37_pad_type_0 = const()[name = tensor("input_37_pad_type_0"), val = tensor("custom")]; + tensor input_37_pad_0 = const()[name = tensor("input_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_fc1_weight_to_fp16 = const()[name = tensor("layers_3_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317597248)))]; + tensor layers_3_fc1_bias_to_fp16 = const()[name = tensor("layers_3_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(330704512)))]; + tensor input_37_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = var_1011, groups = var_825, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = var_1009, weight = layers_3_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_mode_0 = const()[name = tensor("input_39_mode_0"), val = tensor("EXACT")]; + tensor input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor var_1017 = const()[name = tensor("op_1017"), val = tensor([1, 1])]; + tensor var_1019 = const()[name = tensor("op_1019"), val = tensor([1, 1])]; + tensor hidden_states_9_pad_type_0 = const()[name = tensor("hidden_states_9_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_9_pad_0 = const()[name = tensor("hidden_states_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_fc2_weight_to_fp16 = const()[name = tensor("layers_3_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(330714816)))]; + tensor layers_3_fc2_bias_to_fp16 = const()[name = tensor("layers_3_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343822080)))]; + tensor hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = var_1019, groups = var_825, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = var_1017, weight = layers_3_fc2_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("hidden_states_9_cast_fp16")]; + tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor("inputs_25_cast_fp16")]; + tensor var_1032 = const()[name = tensor("op_1032"), val = tensor(3)]; + tensor var_1039 = const()[name = tensor("op_1039"), val = tensor(1)]; + tensor var_1040 = const()[name = tensor("op_1040"), val = tensor(true)]; + tensor var_1052 = const()[name = tensor("op_1052"), val = tensor([1])]; + tensor channels_mean_25_cast_fp16 = reduce_mean(axes = var_1052, keep_dims = var_1040, x = inputs_25_cast_fp16)[name = tensor("channels_mean_25_cast_fp16")]; + tensor zero_mean_25_cast_fp16 = sub(x = inputs_25_cast_fp16, y = channels_mean_25_cast_fp16)[name = tensor("zero_mean_25_cast_fp16")]; + tensor zero_mean_sq_25_cast_fp16 = mul(x = zero_mean_25_cast_fp16, y = zero_mean_25_cast_fp16)[name = tensor("zero_mean_sq_25_cast_fp16")]; + tensor var_1056 = const()[name = tensor("op_1056"), val = tensor([1])]; + tensor var_1057_cast_fp16 = reduce_mean(axes = var_1056, keep_dims = var_1040, x = zero_mean_sq_25_cast_fp16)[name = tensor("op_1057_cast_fp16")]; + tensor var_1058_to_fp16 = const()[name = tensor("op_1058_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1059_cast_fp16 = add(x = var_1057_cast_fp16, y = var_1058_to_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor denom_25_epsilon_0_to_fp16 = const()[name = tensor("denom_25_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_25_cast_fp16 = rsqrt(epsilon = denom_25_epsilon_0_to_fp16, x = var_1059_cast_fp16)[name = tensor("denom_25_cast_fp16")]; + tensor out_25_cast_fp16 = mul(x = zero_mean_25_cast_fp16, y = denom_25_cast_fp16)[name = tensor("out_25_cast_fp16")]; + tensor obj_57_gamma_0_to_fp16 = const()[name = tensor("obj_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343824704)))]; + tensor obj_57_beta_0_to_fp16 = const()[name = tensor("obj_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343827328)))]; + tensor obj_57_epsilon_0_to_fp16 = const()[name = tensor("obj_57_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_57_cast_fp16 = batch_norm(beta = obj_57_beta_0_to_fp16, epsilon = obj_57_epsilon_0_to_fp16, gamma = obj_57_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor("obj_57_cast_fp16")]; + tensor var_1074 = const()[name = tensor("op_1074"), val = tensor([1, 1])]; + tensor var_1076 = const()[name = tensor("op_1076"), val = tensor([1, 1])]; + tensor query_17_pad_type_0 = const()[name = tensor("query_17_pad_type_0"), val = tensor("custom")]; + tensor query_17_pad_0 = const()[name = tensor("query_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343829952)))]; + tensor layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347106816)))]; + tensor query_17_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_bias_to_fp16, dilations = var_1076, groups = var_1039, pad = query_17_pad_0, pad_type = query_17_pad_type_0, strides = var_1074, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("query_17_cast_fp16")]; + tensor var_1080 = const()[name = tensor("op_1080"), val = tensor([1, 1])]; + tensor var_1082 = const()[name = tensor("op_1082"), val = tensor([1, 1])]; + tensor current_key_9_pad_type_0 = const()[name = tensor("current_key_9_pad_type_0"), val = tensor("custom")]; + tensor current_key_9_pad_0 = const()[name = tensor("current_key_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347109440)))]; + tensor current_key_9_cast_fp16 = conv(dilations = var_1082, groups = var_1039, pad = current_key_9_pad_0, pad_type = current_key_9_pad_type_0, strides = var_1080, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("current_key_9_cast_fp16")]; + tensor var_1087 = const()[name = tensor("op_1087"), val = tensor([1, 1])]; + tensor var_1089 = const()[name = tensor("op_1089"), val = tensor([1, 1])]; + tensor current_value_9_pad_type_0 = const()[name = tensor("current_value_9_pad_type_0"), val = tensor("custom")]; + tensor current_value_9_pad_0 = const()[name = tensor("current_value_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350386304)))]; + tensor layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353663168)))]; + tensor current_value_9_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_bias_to_fp16, dilations = var_1089, groups = var_1039, pad = current_value_9_pad_0, pad_type = current_value_9_pad_type_0, strides = var_1087, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("current_value_9_cast_fp16")]; + tensor var_1096_cast_fp16 = mul(x = current_key_9_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor var_1098_cast_fp16 = mul(x = var_103_cast_fp16_4, y = var_241_cast_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor key_17_cast_fp16 = add(x = var_1096_cast_fp16, y = var_1098_cast_fp16)[name = tensor("key_17_cast_fp16")]; + tensor var_1100_cast_fp16 = mul(x = current_value_9_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1100_cast_fp16")]; + tensor var_1102_cast_fp16 = mul(x = var_138_cast_fp16_4, y = var_241_cast_fp16)[name = tensor("op_1102_cast_fp16")]; + tensor value_17_cast_fp16 = add(x = var_1100_cast_fp16, y = var_1102_cast_fp16)[name = tensor("value_17_cast_fp16")]; + tensor var_1105 = const()[name = tensor("op_1105"), val = tensor([1, 20, 64, -1])]; + tensor var_1106_cast_fp16 = reshape(shape = var_1105, x = query_17_cast_fp16)[name = tensor("op_1106_cast_fp16")]; + tensor var_1107_to_fp16 = const()[name = tensor("op_1107_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1108_cast_fp16 = mul(x = var_1106_cast_fp16, y = var_1107_to_fp16)[name = tensor("op_1108_cast_fp16")]; + tensor var_1109 = const()[name = tensor("op_1109"), val = tensor([1, 20, 64, -1])]; + tensor var_1110_cast_fp16 = reshape(shape = var_1109, x = key_17_cast_fp16)[name = tensor("op_1110_cast_fp16")]; + tensor mh_w_25_transpose_x_0 = const()[name = tensor("mh_w_25_transpose_x_0"), val = tensor(true)]; + tensor mh_w_25_transpose_y_0 = const()[name = tensor("mh_w_25_transpose_y_0"), val = tensor(false)]; + tensor mh_w_25_cast_fp16 = matmul(transpose_x = mh_w_25_transpose_x_0, transpose_y = mh_w_25_transpose_y_0, x = var_1108_cast_fp16, y = var_1110_cast_fp16)[name = tensor("mh_w_25_cast_fp16")]; + tensor mh_w_27_cast_fp16 = add(x = mh_w_25_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_27_cast_fp16")]; + tensor var_1118_cast_fp16 = softmax(axis = var_1032, x = mh_w_27_cast_fp16)[name = tensor("op_1118_cast_fp16")]; + tensor var_1119 = const()[name = tensor("op_1119"), val = tensor([1, 20, 64, -1])]; + tensor var_1120_cast_fp16 = reshape(shape = var_1119, x = value_17_cast_fp16)[name = tensor("op_1120_cast_fp16")]; + tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; + tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(true)]; + tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = var_1120_cast_fp16, y = var_1118_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_1123 = const()[name = tensor("op_1123"), val = tensor([1, 1280, 1, -1])]; + tensor input_41_cast_fp16 = reshape(shape = var_1123, x = attn_17_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor var_1127 = const()[name = tensor("op_1127"), val = tensor([1, 1])]; + tensor var_1129 = const()[name = tensor("op_1129"), val = tensor([1, 1])]; + tensor obj_63_pad_type_0 = const()[name = tensor("obj_63_pad_type_0"), val = tensor("custom")]; + tensor obj_63_pad_0 = const()[name = tensor("obj_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353665792)))]; + tensor layers_4_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356942656)))]; + tensor obj_63_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_bias_to_fp16, dilations = var_1129, groups = var_1039, pad = obj_63_pad_0, pad_type = obj_63_pad_type_0, strides = var_1127, weight = layers_4_self_attn_o_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("obj_63_cast_fp16")]; + tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_63_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; + tensor var_1139 = const()[name = tensor("op_1139"), val = tensor([1])]; + tensor channels_mean_27_cast_fp16 = reduce_mean(axes = var_1139, keep_dims = var_1040, x = inputs_27_cast_fp16)[name = tensor("channels_mean_27_cast_fp16")]; + tensor zero_mean_27_cast_fp16 = sub(x = inputs_27_cast_fp16, y = channels_mean_27_cast_fp16)[name = tensor("zero_mean_27_cast_fp16")]; + tensor zero_mean_sq_27_cast_fp16 = mul(x = zero_mean_27_cast_fp16, y = zero_mean_27_cast_fp16)[name = tensor("zero_mean_sq_27_cast_fp16")]; + tensor var_1143 = const()[name = tensor("op_1143"), val = tensor([1])]; + tensor var_1144_cast_fp16 = reduce_mean(axes = var_1143, keep_dims = var_1040, x = zero_mean_sq_27_cast_fp16)[name = tensor("op_1144_cast_fp16")]; + tensor var_1145_to_fp16 = const()[name = tensor("op_1145_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1146_cast_fp16 = add(x = var_1144_cast_fp16, y = var_1145_to_fp16)[name = tensor("op_1146_cast_fp16")]; + tensor denom_27_epsilon_0_to_fp16 = const()[name = tensor("denom_27_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_27_cast_fp16 = rsqrt(epsilon = denom_27_epsilon_0_to_fp16, x = var_1146_cast_fp16)[name = tensor("denom_27_cast_fp16")]; + tensor out_27_cast_fp16 = mul(x = zero_mean_27_cast_fp16, y = denom_27_cast_fp16)[name = tensor("out_27_cast_fp16")]; + tensor obj_65_gamma_0_to_fp16 = const()[name = tensor("obj_65_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356945280)))]; + tensor obj_65_beta_0_to_fp16 = const()[name = tensor("obj_65_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356947904)))]; + tensor obj_65_epsilon_0_to_fp16 = const()[name = tensor("obj_65_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_65_cast_fp16 = batch_norm(beta = obj_65_beta_0_to_fp16, epsilon = obj_65_epsilon_0_to_fp16, gamma = obj_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor("obj_65_cast_fp16")]; + tensor var_1161 = const()[name = tensor("op_1161"), val = tensor([1, 1])]; + tensor var_1163 = const()[name = tensor("op_1163"), val = tensor([1, 1])]; + tensor query_19_pad_type_0 = const()[name = tensor("query_19_pad_type_0"), val = tensor("custom")]; + tensor query_19_pad_0 = const()[name = tensor("query_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356950528)))]; + tensor layers_4_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360227392)))]; + tensor query_19_cast_fp16 = conv(bias = layers_4_encoder_attn_q_proj_bias_to_fp16, dilations = var_1163, groups = var_1039, pad = query_19_pad_0, pad_type = query_19_pad_type_0, strides = var_1161, weight = layers_4_encoder_attn_q_proj_weight_to_fp16, x = obj_65_cast_fp16)[name = tensor("query_19_cast_fp16")]; + tensor var_1167 = const()[name = tensor("op_1167"), val = tensor([1, 1])]; + tensor var_1169 = const()[name = tensor("op_1169"), val = tensor([1, 1])]; + tensor key_19_pad_type_0 = const()[name = tensor("key_19_pad_type_0"), val = tensor("custom")]; + tensor key_19_pad_0 = const()[name = tensor("key_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360230016)))]; + tensor key_19_cast_fp16 = conv(dilations = var_1169, groups = var_1039, pad = key_19_pad_0, pad_type = key_19_pad_type_0, strides = var_1167, weight = layers_4_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_19_cast_fp16")]; + tensor var_1174 = const()[name = tensor("op_1174"), val = tensor([1, 1])]; + tensor var_1176 = const()[name = tensor("op_1176"), val = tensor([1, 1])]; + tensor value_19_pad_type_0 = const()[name = tensor("value_19_pad_type_0"), val = tensor("custom")]; + tensor value_19_pad_0 = const()[name = tensor("value_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363506880)))]; + tensor layers_4_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366783744)))]; + tensor value_19_cast_fp16 = conv(bias = layers_4_encoder_attn_v_proj_bias_to_fp16, dilations = var_1176, groups = var_1039, pad = value_19_pad_0, pad_type = value_19_pad_type_0, strides = var_1174, weight = layers_4_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_19_cast_fp16")]; + tensor var_1180 = const()[name = tensor("op_1180"), val = tensor([1, 20, 64, -1])]; + tensor var_1181_cast_fp16 = reshape(shape = var_1180, x = query_19_cast_fp16)[name = tensor("op_1181_cast_fp16")]; + tensor var_1182_to_fp16 = const()[name = tensor("op_1182_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1183_cast_fp16 = mul(x = var_1181_cast_fp16, y = var_1182_to_fp16)[name = tensor("op_1183_cast_fp16")]; + tensor var_1184 = const()[name = tensor("op_1184"), val = tensor([1, 20, 64, -1])]; + tensor var_1185_cast_fp16 = reshape(shape = var_1184, x = key_19_cast_fp16)[name = tensor("op_1185_cast_fp16")]; + tensor mh_w_29_transpose_x_0 = const()[name = tensor("mh_w_29_transpose_x_0"), val = tensor(true)]; + tensor mh_w_29_transpose_y_0 = const()[name = tensor("mh_w_29_transpose_y_0"), val = tensor(false)]; + tensor mh_w_29_cast_fp16 = matmul(transpose_x = mh_w_29_transpose_x_0, transpose_y = mh_w_29_transpose_y_0, x = var_1183_cast_fp16, y = var_1185_cast_fp16)[name = tensor("mh_w_29_cast_fp16")]; + tensor obj_69_cast_fp16 = softmax(axis = var_1032, x = mh_w_29_cast_fp16)[name = tensor("obj_69_cast_fp16")]; + tensor var_1189 = const()[name = tensor("op_1189"), val = tensor([1, 20, 64, -1])]; + tensor var_1190_cast_fp16 = reshape(shape = var_1189, x = value_19_cast_fp16)[name = tensor("op_1190_cast_fp16")]; + tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; + tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(true)]; + tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = var_1190_cast_fp16, y = obj_69_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_1193 = const()[name = tensor("op_1193"), val = tensor([1, 1280, 1, -1])]; + tensor input_43_cast_fp16 = reshape(shape = var_1193, x = attn_19_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([1, 1])]; + tensor var_1199 = const()[name = tensor("op_1199"), val = tensor([1, 1])]; + tensor obj_67_pad_type_0 = const()[name = tensor("obj_67_pad_type_0"), val = tensor("custom")]; + tensor obj_67_pad_0 = const()[name = tensor("obj_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366786368)))]; + tensor layers_4_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370063232)))]; + tensor obj_67_cast_fp16 = conv(bias = layers_4_encoder_attn_o_proj_bias_to_fp16, dilations = var_1199, groups = var_1039, pad = obj_67_pad_0, pad_type = obj_67_pad_type_0, strides = var_1197, weight = layers_4_encoder_attn_o_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("obj_67_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = obj_67_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; + tensor var_1205 = const()[name = tensor("op_1205"), val = tensor([1])]; + tensor channels_mean_29_cast_fp16 = reduce_mean(axes = var_1205, keep_dims = var_1040, x = inputs_29_cast_fp16)[name = tensor("channels_mean_29_cast_fp16")]; + tensor zero_mean_29_cast_fp16 = sub(x = inputs_29_cast_fp16, y = channels_mean_29_cast_fp16)[name = tensor("zero_mean_29_cast_fp16")]; + tensor zero_mean_sq_29_cast_fp16 = mul(x = zero_mean_29_cast_fp16, y = zero_mean_29_cast_fp16)[name = tensor("zero_mean_sq_29_cast_fp16")]; + tensor var_1209 = const()[name = tensor("op_1209"), val = tensor([1])]; + tensor var_1210_cast_fp16 = reduce_mean(axes = var_1209, keep_dims = var_1040, x = zero_mean_sq_29_cast_fp16)[name = tensor("op_1210_cast_fp16")]; + tensor var_1211_to_fp16 = const()[name = tensor("op_1211_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1212_cast_fp16 = add(x = var_1210_cast_fp16, y = var_1211_to_fp16)[name = tensor("op_1212_cast_fp16")]; + tensor denom_29_epsilon_0_to_fp16 = const()[name = tensor("denom_29_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_29_cast_fp16 = rsqrt(epsilon = denom_29_epsilon_0_to_fp16, x = var_1212_cast_fp16)[name = tensor("denom_29_cast_fp16")]; + tensor out_29_cast_fp16 = mul(x = zero_mean_29_cast_fp16, y = denom_29_cast_fp16)[name = tensor("out_29_cast_fp16")]; + tensor input_45_gamma_0_to_fp16 = const()[name = tensor("input_45_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370065856)))]; + tensor input_45_beta_0_to_fp16 = const()[name = tensor("input_45_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370068480)))]; + tensor input_45_epsilon_0_to_fp16 = const()[name = tensor("input_45_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = batch_norm(beta = input_45_beta_0_to_fp16, epsilon = input_45_epsilon_0_to_fp16, gamma = input_45_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor var_1223 = const()[name = tensor("op_1223"), val = tensor([1, 1])]; + tensor var_1225 = const()[name = tensor("op_1225"), val = tensor([1, 1])]; + tensor input_47_pad_type_0 = const()[name = tensor("input_47_pad_type_0"), val = tensor("custom")]; + tensor input_47_pad_0 = const()[name = tensor("input_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_fc1_weight_to_fp16 = const()[name = tensor("layers_4_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370071104)))]; + tensor layers_4_fc1_bias_to_fp16 = const()[name = tensor("layers_4_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383178368)))]; + tensor input_47_cast_fp16 = conv(bias = layers_4_fc1_bias_to_fp16, dilations = var_1225, groups = var_1039, pad = input_47_pad_0, pad_type = input_47_pad_type_0, strides = var_1223, weight = layers_4_fc1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor var_1231 = const()[name = tensor("op_1231"), val = tensor([1, 1])]; + tensor var_1233 = const()[name = tensor("op_1233"), val = tensor([1, 1])]; + tensor hidden_states_11_pad_type_0 = const()[name = tensor("hidden_states_11_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_11_pad_0 = const()[name = tensor("hidden_states_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_fc2_weight_to_fp16 = const()[name = tensor("layers_4_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383188672)))]; + tensor layers_4_fc2_bias_to_fp16 = const()[name = tensor("layers_4_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396295936)))]; + tensor hidden_states_11_cast_fp16 = conv(bias = layers_4_fc2_bias_to_fp16, dilations = var_1233, groups = var_1039, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = var_1231, weight = layers_4_fc2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("hidden_states_11_cast_fp16")]; + tensor inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = hidden_states_11_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; + tensor var_1246 = const()[name = tensor("op_1246"), val = tensor(3)]; + tensor var_1253 = const()[name = tensor("op_1253"), val = tensor(1)]; + tensor var_1254 = const()[name = tensor("op_1254"), val = tensor(true)]; + tensor var_1266 = const()[name = tensor("op_1266"), val = tensor([1])]; + tensor channels_mean_31_cast_fp16 = reduce_mean(axes = var_1266, keep_dims = var_1254, x = inputs_31_cast_fp16)[name = tensor("channels_mean_31_cast_fp16")]; + tensor zero_mean_31_cast_fp16 = sub(x = inputs_31_cast_fp16, y = channels_mean_31_cast_fp16)[name = tensor("zero_mean_31_cast_fp16")]; + tensor zero_mean_sq_31_cast_fp16 = mul(x = zero_mean_31_cast_fp16, y = zero_mean_31_cast_fp16)[name = tensor("zero_mean_sq_31_cast_fp16")]; + tensor var_1270 = const()[name = tensor("op_1270"), val = tensor([1])]; + tensor var_1271_cast_fp16 = reduce_mean(axes = var_1270, keep_dims = var_1254, x = zero_mean_sq_31_cast_fp16)[name = tensor("op_1271_cast_fp16")]; + tensor var_1272_to_fp16 = const()[name = tensor("op_1272_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1273_cast_fp16 = add(x = var_1271_cast_fp16, y = var_1272_to_fp16)[name = tensor("op_1273_cast_fp16")]; + tensor denom_31_epsilon_0_to_fp16 = const()[name = tensor("denom_31_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_31_cast_fp16 = rsqrt(epsilon = denom_31_epsilon_0_to_fp16, x = var_1273_cast_fp16)[name = tensor("denom_31_cast_fp16")]; + tensor out_31_cast_fp16 = mul(x = zero_mean_31_cast_fp16, y = denom_31_cast_fp16)[name = tensor("out_31_cast_fp16")]; + tensor obj_71_gamma_0_to_fp16 = const()[name = tensor("obj_71_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396298560)))]; + tensor obj_71_beta_0_to_fp16 = const()[name = tensor("obj_71_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396301184)))]; + tensor obj_71_epsilon_0_to_fp16 = const()[name = tensor("obj_71_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_71_cast_fp16 = batch_norm(beta = obj_71_beta_0_to_fp16, epsilon = obj_71_epsilon_0_to_fp16, gamma = obj_71_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor("obj_71_cast_fp16")]; + tensor var_1288 = const()[name = tensor("op_1288"), val = tensor([1, 1])]; + tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([1, 1])]; + tensor query_21_pad_type_0 = const()[name = tensor("query_21_pad_type_0"), val = tensor("custom")]; + tensor query_21_pad_0 = const()[name = tensor("query_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396303808)))]; + tensor layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(399580672)))]; + tensor query_21_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_bias_to_fp16, dilations = var_1290, groups = var_1253, pad = query_21_pad_0, pad_type = query_21_pad_type_0, strides = var_1288, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("query_21_cast_fp16")]; + tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([1, 1])]; + tensor var_1296 = const()[name = tensor("op_1296"), val = tensor([1, 1])]; + tensor current_key_11_pad_type_0 = const()[name = tensor("current_key_11_pad_type_0"), val = tensor("custom")]; + tensor current_key_11_pad_0 = const()[name = tensor("current_key_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(399583296)))]; + tensor current_key_11_cast_fp16 = conv(dilations = var_1296, groups = var_1253, pad = current_key_11_pad_0, pad_type = current_key_11_pad_type_0, strides = var_1294, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("current_key_11_cast_fp16")]; + tensor var_1301 = const()[name = tensor("op_1301"), val = tensor([1, 1])]; + tensor var_1303 = const()[name = tensor("op_1303"), val = tensor([1, 1])]; + tensor current_value_11_pad_type_0 = const()[name = tensor("current_value_11_pad_type_0"), val = tensor("custom")]; + tensor current_value_11_pad_0 = const()[name = tensor("current_value_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402860160)))]; + tensor layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(406137024)))]; + tensor current_value_11_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_bias_to_fp16, dilations = var_1303, groups = var_1253, pad = current_value_11_pad_0, pad_type = current_value_11_pad_type_0, strides = var_1301, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("current_value_11_cast_fp16")]; + tensor var_1310_cast_fp16 = mul(x = current_key_11_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1310_cast_fp16")]; + tensor var_1312_cast_fp16 = mul(x = var_103_cast_fp16_5, y = var_241_cast_fp16)[name = tensor("op_1312_cast_fp16")]; + tensor key_21_cast_fp16 = add(x = var_1310_cast_fp16, y = var_1312_cast_fp16)[name = tensor("key_21_cast_fp16")]; + tensor var_1314_cast_fp16 = mul(x = current_value_11_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1314_cast_fp16")]; + tensor var_1316_cast_fp16 = mul(x = var_138_cast_fp16_5, y = var_241_cast_fp16)[name = tensor("op_1316_cast_fp16")]; + tensor value_21_cast_fp16 = add(x = var_1314_cast_fp16, y = var_1316_cast_fp16)[name = tensor("value_21_cast_fp16")]; + tensor var_1319 = const()[name = tensor("op_1319"), val = tensor([1, 20, 64, -1])]; + tensor var_1320_cast_fp16 = reshape(shape = var_1319, x = query_21_cast_fp16)[name = tensor("op_1320_cast_fp16")]; + tensor var_1321_to_fp16 = const()[name = tensor("op_1321_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1322_cast_fp16 = mul(x = var_1320_cast_fp16, y = var_1321_to_fp16)[name = tensor("op_1322_cast_fp16")]; + tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([1, 20, 64, -1])]; + tensor var_1324_cast_fp16 = reshape(shape = var_1323, x = key_21_cast_fp16)[name = tensor("op_1324_cast_fp16")]; + tensor mh_w_31_transpose_x_0 = const()[name = tensor("mh_w_31_transpose_x_0"), val = tensor(true)]; + tensor mh_w_31_transpose_y_0 = const()[name = tensor("mh_w_31_transpose_y_0"), val = tensor(false)]; + tensor mh_w_31_cast_fp16 = matmul(transpose_x = mh_w_31_transpose_x_0, transpose_y = mh_w_31_transpose_y_0, x = var_1322_cast_fp16, y = var_1324_cast_fp16)[name = tensor("mh_w_31_cast_fp16")]; + tensor mh_w_33_cast_fp16 = add(x = mh_w_31_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_33_cast_fp16")]; + tensor var_1332_cast_fp16 = softmax(axis = var_1246, x = mh_w_33_cast_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([1, 20, 64, -1])]; + tensor var_1334_cast_fp16 = reshape(shape = var_1333, x = value_21_cast_fp16)[name = tensor("op_1334_cast_fp16")]; + tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; + tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(true)]; + tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = var_1334_cast_fp16, y = var_1332_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([1, 1280, 1, -1])]; + tensor input_51_cast_fp16 = reshape(shape = var_1337, x = attn_21_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor var_1341 = const()[name = tensor("op_1341"), val = tensor([1, 1])]; + tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([1, 1])]; + tensor obj_77_pad_type_0 = const()[name = tensor("obj_77_pad_type_0"), val = tensor("custom")]; + tensor obj_77_pad_0 = const()[name = tensor("obj_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(406139648)))]; + tensor layers_5_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409416512)))]; + tensor obj_77_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_bias_to_fp16, dilations = var_1343, groups = var_1253, pad = obj_77_pad_0, pad_type = obj_77_pad_type_0, strides = var_1341, weight = layers_5_self_attn_o_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("obj_77_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = obj_77_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; + tensor var_1353 = const()[name = tensor("op_1353"), val = tensor([1])]; + tensor channels_mean_33_cast_fp16 = reduce_mean(axes = var_1353, keep_dims = var_1254, x = inputs_33_cast_fp16)[name = tensor("channels_mean_33_cast_fp16")]; + tensor zero_mean_33_cast_fp16 = sub(x = inputs_33_cast_fp16, y = channels_mean_33_cast_fp16)[name = tensor("zero_mean_33_cast_fp16")]; + tensor zero_mean_sq_33_cast_fp16 = mul(x = zero_mean_33_cast_fp16, y = zero_mean_33_cast_fp16)[name = tensor("zero_mean_sq_33_cast_fp16")]; + tensor var_1357 = const()[name = tensor("op_1357"), val = tensor([1])]; + tensor var_1358_cast_fp16 = reduce_mean(axes = var_1357, keep_dims = var_1254, x = zero_mean_sq_33_cast_fp16)[name = tensor("op_1358_cast_fp16")]; + tensor var_1359_to_fp16 = const()[name = tensor("op_1359_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1360_cast_fp16 = add(x = var_1358_cast_fp16, y = var_1359_to_fp16)[name = tensor("op_1360_cast_fp16")]; + tensor denom_33_epsilon_0_to_fp16 = const()[name = tensor("denom_33_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_33_cast_fp16 = rsqrt(epsilon = denom_33_epsilon_0_to_fp16, x = var_1360_cast_fp16)[name = tensor("denom_33_cast_fp16")]; + tensor out_33_cast_fp16 = mul(x = zero_mean_33_cast_fp16, y = denom_33_cast_fp16)[name = tensor("out_33_cast_fp16")]; + tensor obj_79_gamma_0_to_fp16 = const()[name = tensor("obj_79_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409419136)))]; + tensor obj_79_beta_0_to_fp16 = const()[name = tensor("obj_79_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409421760)))]; + tensor obj_79_epsilon_0_to_fp16 = const()[name = tensor("obj_79_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_79_cast_fp16 = batch_norm(beta = obj_79_beta_0_to_fp16, epsilon = obj_79_epsilon_0_to_fp16, gamma = obj_79_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor("obj_79_cast_fp16")]; + tensor var_1375 = const()[name = tensor("op_1375"), val = tensor([1, 1])]; + tensor var_1377 = const()[name = tensor("op_1377"), val = tensor([1, 1])]; + tensor query_23_pad_type_0 = const()[name = tensor("query_23_pad_type_0"), val = tensor("custom")]; + tensor query_23_pad_0 = const()[name = tensor("query_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409424384)))]; + tensor layers_5_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412701248)))]; + tensor query_23_cast_fp16 = conv(bias = layers_5_encoder_attn_q_proj_bias_to_fp16, dilations = var_1377, groups = var_1253, pad = query_23_pad_0, pad_type = query_23_pad_type_0, strides = var_1375, weight = layers_5_encoder_attn_q_proj_weight_to_fp16, x = obj_79_cast_fp16)[name = tensor("query_23_cast_fp16")]; + tensor var_1381 = const()[name = tensor("op_1381"), val = tensor([1, 1])]; + tensor var_1383 = const()[name = tensor("op_1383"), val = tensor([1, 1])]; + tensor key_23_pad_type_0 = const()[name = tensor("key_23_pad_type_0"), val = tensor("custom")]; + tensor key_23_pad_0 = const()[name = tensor("key_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412703872)))]; + tensor key_23_cast_fp16 = conv(dilations = var_1383, groups = var_1253, pad = key_23_pad_0, pad_type = key_23_pad_type_0, strides = var_1381, weight = layers_5_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_23_cast_fp16")]; + tensor var_1388 = const()[name = tensor("op_1388"), val = tensor([1, 1])]; + tensor var_1390 = const()[name = tensor("op_1390"), val = tensor([1, 1])]; + tensor value_23_pad_type_0 = const()[name = tensor("value_23_pad_type_0"), val = tensor("custom")]; + tensor value_23_pad_0 = const()[name = tensor("value_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(415980736)))]; + tensor layers_5_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419257600)))]; + tensor value_23_cast_fp16 = conv(bias = layers_5_encoder_attn_v_proj_bias_to_fp16, dilations = var_1390, groups = var_1253, pad = value_23_pad_0, pad_type = value_23_pad_type_0, strides = var_1388, weight = layers_5_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_23_cast_fp16")]; + tensor var_1394 = const()[name = tensor("op_1394"), val = tensor([1, 20, 64, -1])]; + tensor var_1395_cast_fp16 = reshape(shape = var_1394, x = query_23_cast_fp16)[name = tensor("op_1395_cast_fp16")]; + tensor var_1396_to_fp16 = const()[name = tensor("op_1396_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1397_cast_fp16 = mul(x = var_1395_cast_fp16, y = var_1396_to_fp16)[name = tensor("op_1397_cast_fp16")]; + tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([1, 20, 64, -1])]; + tensor var_1399_cast_fp16 = reshape(shape = var_1398, x = key_23_cast_fp16)[name = tensor("op_1399_cast_fp16")]; + tensor mh_w_35_transpose_x_0 = const()[name = tensor("mh_w_35_transpose_x_0"), val = tensor(true)]; + tensor mh_w_35_transpose_y_0 = const()[name = tensor("mh_w_35_transpose_y_0"), val = tensor(false)]; + tensor mh_w_35_cast_fp16 = matmul(transpose_x = mh_w_35_transpose_x_0, transpose_y = mh_w_35_transpose_y_0, x = var_1397_cast_fp16, y = var_1399_cast_fp16)[name = tensor("mh_w_35_cast_fp16")]; + tensor obj_83_cast_fp16 = softmax(axis = var_1246, x = mh_w_35_cast_fp16)[name = tensor("obj_83_cast_fp16")]; + tensor var_1403 = const()[name = tensor("op_1403"), val = tensor([1, 20, 64, -1])]; + tensor var_1404_cast_fp16 = reshape(shape = var_1403, x = value_23_cast_fp16)[name = tensor("op_1404_cast_fp16")]; + tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; + tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(true)]; + tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = var_1404_cast_fp16, y = obj_83_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_1407 = const()[name = tensor("op_1407"), val = tensor([1, 1280, 1, -1])]; + tensor input_53_cast_fp16 = reshape(shape = var_1407, x = attn_23_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor var_1411 = const()[name = tensor("op_1411"), val = tensor([1, 1])]; + tensor var_1413 = const()[name = tensor("op_1413"), val = tensor([1, 1])]; + tensor obj_81_pad_type_0 = const()[name = tensor("obj_81_pad_type_0"), val = tensor("custom")]; + tensor obj_81_pad_0 = const()[name = tensor("obj_81_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419260224)))]; + tensor layers_5_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422537088)))]; + tensor obj_81_cast_fp16 = conv(bias = layers_5_encoder_attn_o_proj_bias_to_fp16, dilations = var_1413, groups = var_1253, pad = obj_81_pad_0, pad_type = obj_81_pad_type_0, strides = var_1411, weight = layers_5_encoder_attn_o_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("obj_81_cast_fp16")]; + tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_81_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; + tensor var_1419 = const()[name = tensor("op_1419"), val = tensor([1])]; + tensor channels_mean_35_cast_fp16 = reduce_mean(axes = var_1419, keep_dims = var_1254, x = inputs_35_cast_fp16)[name = tensor("channels_mean_35_cast_fp16")]; + tensor zero_mean_35_cast_fp16 = sub(x = inputs_35_cast_fp16, y = channels_mean_35_cast_fp16)[name = tensor("zero_mean_35_cast_fp16")]; + tensor zero_mean_sq_35_cast_fp16 = mul(x = zero_mean_35_cast_fp16, y = zero_mean_35_cast_fp16)[name = tensor("zero_mean_sq_35_cast_fp16")]; + tensor var_1423 = const()[name = tensor("op_1423"), val = tensor([1])]; + tensor var_1424_cast_fp16 = reduce_mean(axes = var_1423, keep_dims = var_1254, x = zero_mean_sq_35_cast_fp16)[name = tensor("op_1424_cast_fp16")]; + tensor var_1425_to_fp16 = const()[name = tensor("op_1425_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1426_cast_fp16 = add(x = var_1424_cast_fp16, y = var_1425_to_fp16)[name = tensor("op_1426_cast_fp16")]; + tensor denom_35_epsilon_0_to_fp16 = const()[name = tensor("denom_35_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_35_cast_fp16 = rsqrt(epsilon = denom_35_epsilon_0_to_fp16, x = var_1426_cast_fp16)[name = tensor("denom_35_cast_fp16")]; + tensor out_35_cast_fp16 = mul(x = zero_mean_35_cast_fp16, y = denom_35_cast_fp16)[name = tensor("out_35_cast_fp16")]; + tensor input_55_gamma_0_to_fp16 = const()[name = tensor("input_55_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422539712)))]; + tensor input_55_beta_0_to_fp16 = const()[name = tensor("input_55_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422542336)))]; + tensor input_55_epsilon_0_to_fp16 = const()[name = tensor("input_55_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_55_cast_fp16 = batch_norm(beta = input_55_beta_0_to_fp16, epsilon = input_55_epsilon_0_to_fp16, gamma = input_55_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor var_1437 = const()[name = tensor("op_1437"), val = tensor([1, 1])]; + tensor var_1439 = const()[name = tensor("op_1439"), val = tensor([1, 1])]; + tensor input_57_pad_type_0 = const()[name = tensor("input_57_pad_type_0"), val = tensor("custom")]; + tensor input_57_pad_0 = const()[name = tensor("input_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_fc1_weight_to_fp16 = const()[name = tensor("layers_5_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422544960)))]; + tensor layers_5_fc1_bias_to_fp16 = const()[name = tensor("layers_5_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435652224)))]; + tensor input_57_cast_fp16 = conv(bias = layers_5_fc1_bias_to_fp16, dilations = var_1439, groups = var_1253, pad = input_57_pad_0, pad_type = input_57_pad_type_0, strides = var_1437, weight = layers_5_fc1_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor var_1445 = const()[name = tensor("op_1445"), val = tensor([1, 1])]; + tensor var_1447 = const()[name = tensor("op_1447"), val = tensor([1, 1])]; + tensor hidden_states_13_pad_type_0 = const()[name = tensor("hidden_states_13_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_13_pad_0 = const()[name = tensor("hidden_states_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_fc2_weight_to_fp16 = const()[name = tensor("layers_5_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435662528)))]; + tensor layers_5_fc2_bias_to_fp16 = const()[name = tensor("layers_5_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448769792)))]; + tensor hidden_states_13_cast_fp16 = conv(bias = layers_5_fc2_bias_to_fp16, dilations = var_1447, groups = var_1253, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = var_1445, weight = layers_5_fc2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("hidden_states_13_cast_fp16")]; + tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor("inputs_37_cast_fp16")]; + tensor var_1460 = const()[name = tensor("op_1460"), val = tensor(3)]; + tensor var_1467 = const()[name = tensor("op_1467"), val = tensor(1)]; + tensor var_1468 = const()[name = tensor("op_1468"), val = tensor(true)]; + tensor var_1480 = const()[name = tensor("op_1480"), val = tensor([1])]; + tensor channels_mean_37_cast_fp16 = reduce_mean(axes = var_1480, keep_dims = var_1468, x = inputs_37_cast_fp16)[name = tensor("channels_mean_37_cast_fp16")]; + tensor zero_mean_37_cast_fp16 = sub(x = inputs_37_cast_fp16, y = channels_mean_37_cast_fp16)[name = tensor("zero_mean_37_cast_fp16")]; + tensor zero_mean_sq_37_cast_fp16 = mul(x = zero_mean_37_cast_fp16, y = zero_mean_37_cast_fp16)[name = tensor("zero_mean_sq_37_cast_fp16")]; + tensor var_1484 = const()[name = tensor("op_1484"), val = tensor([1])]; + tensor var_1485_cast_fp16 = reduce_mean(axes = var_1484, keep_dims = var_1468, x = zero_mean_sq_37_cast_fp16)[name = tensor("op_1485_cast_fp16")]; + tensor var_1486_to_fp16 = const()[name = tensor("op_1486_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1487_cast_fp16 = add(x = var_1485_cast_fp16, y = var_1486_to_fp16)[name = tensor("op_1487_cast_fp16")]; + tensor denom_37_epsilon_0_to_fp16 = const()[name = tensor("denom_37_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_37_cast_fp16 = rsqrt(epsilon = denom_37_epsilon_0_to_fp16, x = var_1487_cast_fp16)[name = tensor("denom_37_cast_fp16")]; + tensor out_37_cast_fp16 = mul(x = zero_mean_37_cast_fp16, y = denom_37_cast_fp16)[name = tensor("out_37_cast_fp16")]; + tensor obj_85_gamma_0_to_fp16 = const()[name = tensor("obj_85_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448772416)))]; + tensor obj_85_beta_0_to_fp16 = const()[name = tensor("obj_85_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448775040)))]; + tensor obj_85_epsilon_0_to_fp16 = const()[name = tensor("obj_85_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_85_cast_fp16 = batch_norm(beta = obj_85_beta_0_to_fp16, epsilon = obj_85_epsilon_0_to_fp16, gamma = obj_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_37_cast_fp16)[name = tensor("obj_85_cast_fp16")]; + tensor var_1502 = const()[name = tensor("op_1502"), val = tensor([1, 1])]; + tensor var_1504 = const()[name = tensor("op_1504"), val = tensor([1, 1])]; + tensor query_25_pad_type_0 = const()[name = tensor("query_25_pad_type_0"), val = tensor("custom")]; + tensor query_25_pad_0 = const()[name = tensor("query_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448777664)))]; + tensor layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452054528)))]; + tensor query_25_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_bias_to_fp16, dilations = var_1504, groups = var_1467, pad = query_25_pad_0, pad_type = query_25_pad_type_0, strides = var_1502, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("query_25_cast_fp16")]; + tensor var_1508 = const()[name = tensor("op_1508"), val = tensor([1, 1])]; + tensor var_1510 = const()[name = tensor("op_1510"), val = tensor([1, 1])]; + tensor current_key_13_pad_type_0 = const()[name = tensor("current_key_13_pad_type_0"), val = tensor("custom")]; + tensor current_key_13_pad_0 = const()[name = tensor("current_key_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452057152)))]; + tensor current_key_13_cast_fp16 = conv(dilations = var_1510, groups = var_1467, pad = current_key_13_pad_0, pad_type = current_key_13_pad_type_0, strides = var_1508, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("current_key_13_cast_fp16")]; + tensor var_1515 = const()[name = tensor("op_1515"), val = tensor([1, 1])]; + tensor var_1517 = const()[name = tensor("op_1517"), val = tensor([1, 1])]; + tensor current_value_13_pad_type_0 = const()[name = tensor("current_value_13_pad_type_0"), val = tensor("custom")]; + tensor current_value_13_pad_0 = const()[name = tensor("current_value_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455334016)))]; + tensor layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(458610880)))]; + tensor current_value_13_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_bias_to_fp16, dilations = var_1517, groups = var_1467, pad = current_value_13_pad_0, pad_type = current_value_13_pad_type_0, strides = var_1515, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("current_value_13_cast_fp16")]; + tensor var_1524_cast_fp16 = mul(x = current_key_13_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1524_cast_fp16")]; + tensor var_1526_cast_fp16 = mul(x = var_103_cast_fp16_6, y = var_241_cast_fp16)[name = tensor("op_1526_cast_fp16")]; + tensor key_25_cast_fp16 = add(x = var_1524_cast_fp16, y = var_1526_cast_fp16)[name = tensor("key_25_cast_fp16")]; + tensor var_1528_cast_fp16 = mul(x = current_value_13_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1528_cast_fp16")]; + tensor var_1530_cast_fp16 = mul(x = var_138_cast_fp16_6, y = var_241_cast_fp16)[name = tensor("op_1530_cast_fp16")]; + tensor value_25_cast_fp16 = add(x = var_1528_cast_fp16, y = var_1530_cast_fp16)[name = tensor("value_25_cast_fp16")]; + tensor var_1533 = const()[name = tensor("op_1533"), val = tensor([1, 20, 64, -1])]; + tensor var_1534_cast_fp16 = reshape(shape = var_1533, x = query_25_cast_fp16)[name = tensor("op_1534_cast_fp16")]; + tensor var_1535_to_fp16 = const()[name = tensor("op_1535_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1536_cast_fp16 = mul(x = var_1534_cast_fp16, y = var_1535_to_fp16)[name = tensor("op_1536_cast_fp16")]; + tensor var_1537 = const()[name = tensor("op_1537"), val = tensor([1, 20, 64, -1])]; + tensor var_1538_cast_fp16 = reshape(shape = var_1537, x = key_25_cast_fp16)[name = tensor("op_1538_cast_fp16")]; + tensor mh_w_37_transpose_x_0 = const()[name = tensor("mh_w_37_transpose_x_0"), val = tensor(true)]; + tensor mh_w_37_transpose_y_0 = const()[name = tensor("mh_w_37_transpose_y_0"), val = tensor(false)]; + tensor mh_w_37_cast_fp16 = matmul(transpose_x = mh_w_37_transpose_x_0, transpose_y = mh_w_37_transpose_y_0, x = var_1536_cast_fp16, y = var_1538_cast_fp16)[name = tensor("mh_w_37_cast_fp16")]; + tensor mh_w_39_cast_fp16 = add(x = mh_w_37_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_39_cast_fp16")]; + tensor var_1546_cast_fp16 = softmax(axis = var_1460, x = mh_w_39_cast_fp16)[name = tensor("op_1546_cast_fp16")]; + tensor var_1547 = const()[name = tensor("op_1547"), val = tensor([1, 20, 64, -1])]; + tensor var_1548_cast_fp16 = reshape(shape = var_1547, x = value_25_cast_fp16)[name = tensor("op_1548_cast_fp16")]; + tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; + tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(true)]; + tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = var_1548_cast_fp16, y = var_1546_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_1551 = const()[name = tensor("op_1551"), val = tensor([1, 1280, 1, -1])]; + tensor input_61_cast_fp16 = reshape(shape = var_1551, x = attn_25_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor var_1555 = const()[name = tensor("op_1555"), val = tensor([1, 1])]; + tensor var_1557 = const()[name = tensor("op_1557"), val = tensor([1, 1])]; + tensor obj_91_pad_type_0 = const()[name = tensor("obj_91_pad_type_0"), val = tensor("custom")]; + tensor obj_91_pad_0 = const()[name = tensor("obj_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(458613504)))]; + tensor layers_6_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461890368)))]; + tensor obj_91_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_bias_to_fp16, dilations = var_1557, groups = var_1467, pad = obj_91_pad_0, pad_type = obj_91_pad_type_0, strides = var_1555, weight = layers_6_self_attn_o_proj_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("obj_91_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_91_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; + tensor var_1567 = const()[name = tensor("op_1567"), val = tensor([1])]; + tensor channels_mean_39_cast_fp16 = reduce_mean(axes = var_1567, keep_dims = var_1468, x = inputs_39_cast_fp16)[name = tensor("channels_mean_39_cast_fp16")]; + tensor zero_mean_39_cast_fp16 = sub(x = inputs_39_cast_fp16, y = channels_mean_39_cast_fp16)[name = tensor("zero_mean_39_cast_fp16")]; + tensor zero_mean_sq_39_cast_fp16 = mul(x = zero_mean_39_cast_fp16, y = zero_mean_39_cast_fp16)[name = tensor("zero_mean_sq_39_cast_fp16")]; + tensor var_1571 = const()[name = tensor("op_1571"), val = tensor([1])]; + tensor var_1572_cast_fp16 = reduce_mean(axes = var_1571, keep_dims = var_1468, x = zero_mean_sq_39_cast_fp16)[name = tensor("op_1572_cast_fp16")]; + tensor var_1573_to_fp16 = const()[name = tensor("op_1573_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1574_cast_fp16 = add(x = var_1572_cast_fp16, y = var_1573_to_fp16)[name = tensor("op_1574_cast_fp16")]; + tensor denom_39_epsilon_0_to_fp16 = const()[name = tensor("denom_39_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_39_cast_fp16 = rsqrt(epsilon = denom_39_epsilon_0_to_fp16, x = var_1574_cast_fp16)[name = tensor("denom_39_cast_fp16")]; + tensor out_39_cast_fp16 = mul(x = zero_mean_39_cast_fp16, y = denom_39_cast_fp16)[name = tensor("out_39_cast_fp16")]; + tensor obj_93_gamma_0_to_fp16 = const()[name = tensor("obj_93_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461892992)))]; + tensor obj_93_beta_0_to_fp16 = const()[name = tensor("obj_93_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461895616)))]; + tensor obj_93_epsilon_0_to_fp16 = const()[name = tensor("obj_93_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_93_cast_fp16 = batch_norm(beta = obj_93_beta_0_to_fp16, epsilon = obj_93_epsilon_0_to_fp16, gamma = obj_93_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_39_cast_fp16)[name = tensor("obj_93_cast_fp16")]; + tensor var_1589 = const()[name = tensor("op_1589"), val = tensor([1, 1])]; + tensor var_1591 = const()[name = tensor("op_1591"), val = tensor([1, 1])]; + tensor query_27_pad_type_0 = const()[name = tensor("query_27_pad_type_0"), val = tensor("custom")]; + tensor query_27_pad_0 = const()[name = tensor("query_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461898240)))]; + tensor layers_6_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465175104)))]; + tensor query_27_cast_fp16 = conv(bias = layers_6_encoder_attn_q_proj_bias_to_fp16, dilations = var_1591, groups = var_1467, pad = query_27_pad_0, pad_type = query_27_pad_type_0, strides = var_1589, weight = layers_6_encoder_attn_q_proj_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor("query_27_cast_fp16")]; + tensor var_1595 = const()[name = tensor("op_1595"), val = tensor([1, 1])]; + tensor var_1597 = const()[name = tensor("op_1597"), val = tensor([1, 1])]; + tensor key_27_pad_type_0 = const()[name = tensor("key_27_pad_type_0"), val = tensor("custom")]; + tensor key_27_pad_0 = const()[name = tensor("key_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465177728)))]; + tensor key_27_cast_fp16 = conv(dilations = var_1597, groups = var_1467, pad = key_27_pad_0, pad_type = key_27_pad_type_0, strides = var_1595, weight = layers_6_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_27_cast_fp16")]; + tensor var_1602 = const()[name = tensor("op_1602"), val = tensor([1, 1])]; + tensor var_1604 = const()[name = tensor("op_1604"), val = tensor([1, 1])]; + tensor value_27_pad_type_0 = const()[name = tensor("value_27_pad_type_0"), val = tensor("custom")]; + tensor value_27_pad_0 = const()[name = tensor("value_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468454592)))]; + tensor layers_6_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471731456)))]; + tensor value_27_cast_fp16 = conv(bias = layers_6_encoder_attn_v_proj_bias_to_fp16, dilations = var_1604, groups = var_1467, pad = value_27_pad_0, pad_type = value_27_pad_type_0, strides = var_1602, weight = layers_6_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_27_cast_fp16")]; + tensor var_1608 = const()[name = tensor("op_1608"), val = tensor([1, 20, 64, -1])]; + tensor var_1609_cast_fp16 = reshape(shape = var_1608, x = query_27_cast_fp16)[name = tensor("op_1609_cast_fp16")]; + tensor var_1610_to_fp16 = const()[name = tensor("op_1610_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1611_cast_fp16 = mul(x = var_1609_cast_fp16, y = var_1610_to_fp16)[name = tensor("op_1611_cast_fp16")]; + tensor var_1612 = const()[name = tensor("op_1612"), val = tensor([1, 20, 64, -1])]; + tensor var_1613_cast_fp16 = reshape(shape = var_1612, x = key_27_cast_fp16)[name = tensor("op_1613_cast_fp16")]; + tensor mh_w_41_transpose_x_0 = const()[name = tensor("mh_w_41_transpose_x_0"), val = tensor(true)]; + tensor mh_w_41_transpose_y_0 = const()[name = tensor("mh_w_41_transpose_y_0"), val = tensor(false)]; + tensor mh_w_41_cast_fp16 = matmul(transpose_x = mh_w_41_transpose_x_0, transpose_y = mh_w_41_transpose_y_0, x = var_1611_cast_fp16, y = var_1613_cast_fp16)[name = tensor("mh_w_41_cast_fp16")]; + tensor obj_97_cast_fp16 = softmax(axis = var_1460, x = mh_w_41_cast_fp16)[name = tensor("obj_97_cast_fp16")]; + tensor var_1617 = const()[name = tensor("op_1617"), val = tensor([1, 20, 64, -1])]; + tensor var_1618_cast_fp16 = reshape(shape = var_1617, x = value_27_cast_fp16)[name = tensor("op_1618_cast_fp16")]; + tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; + tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(true)]; + tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = var_1618_cast_fp16, y = obj_97_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_1621 = const()[name = tensor("op_1621"), val = tensor([1, 1280, 1, -1])]; + tensor input_63_cast_fp16 = reshape(shape = var_1621, x = attn_27_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor var_1625 = const()[name = tensor("op_1625"), val = tensor([1, 1])]; + tensor var_1627 = const()[name = tensor("op_1627"), val = tensor([1, 1])]; + tensor obj_95_pad_type_0 = const()[name = tensor("obj_95_pad_type_0"), val = tensor("custom")]; + tensor obj_95_pad_0 = const()[name = tensor("obj_95_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471734080)))]; + tensor layers_6_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(475010944)))]; + tensor obj_95_cast_fp16 = conv(bias = layers_6_encoder_attn_o_proj_bias_to_fp16, dilations = var_1627, groups = var_1467, pad = obj_95_pad_0, pad_type = obj_95_pad_type_0, strides = var_1625, weight = layers_6_encoder_attn_o_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("obj_95_cast_fp16")]; + tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = obj_95_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; + tensor var_1633 = const()[name = tensor("op_1633"), val = tensor([1])]; + tensor channels_mean_41_cast_fp16 = reduce_mean(axes = var_1633, keep_dims = var_1468, x = inputs_41_cast_fp16)[name = tensor("channels_mean_41_cast_fp16")]; + tensor zero_mean_41_cast_fp16 = sub(x = inputs_41_cast_fp16, y = channels_mean_41_cast_fp16)[name = tensor("zero_mean_41_cast_fp16")]; + tensor zero_mean_sq_41_cast_fp16 = mul(x = zero_mean_41_cast_fp16, y = zero_mean_41_cast_fp16)[name = tensor("zero_mean_sq_41_cast_fp16")]; + tensor var_1637 = const()[name = tensor("op_1637"), val = tensor([1])]; + tensor var_1638_cast_fp16 = reduce_mean(axes = var_1637, keep_dims = var_1468, x = zero_mean_sq_41_cast_fp16)[name = tensor("op_1638_cast_fp16")]; + tensor var_1639_to_fp16 = const()[name = tensor("op_1639_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1640_cast_fp16 = add(x = var_1638_cast_fp16, y = var_1639_to_fp16)[name = tensor("op_1640_cast_fp16")]; + tensor denom_41_epsilon_0_to_fp16 = const()[name = tensor("denom_41_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_41_cast_fp16 = rsqrt(epsilon = denom_41_epsilon_0_to_fp16, x = var_1640_cast_fp16)[name = tensor("denom_41_cast_fp16")]; + tensor out_41_cast_fp16 = mul(x = zero_mean_41_cast_fp16, y = denom_41_cast_fp16)[name = tensor("out_41_cast_fp16")]; + tensor input_65_gamma_0_to_fp16 = const()[name = tensor("input_65_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(475013568)))]; + tensor input_65_beta_0_to_fp16 = const()[name = tensor("input_65_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(475016192)))]; + tensor input_65_epsilon_0_to_fp16 = const()[name = tensor("input_65_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = batch_norm(beta = input_65_beta_0_to_fp16, epsilon = input_65_epsilon_0_to_fp16, gamma = input_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_41_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor var_1651 = const()[name = tensor("op_1651"), val = tensor([1, 1])]; + tensor var_1653 = const()[name = tensor("op_1653"), val = tensor([1, 1])]; + tensor input_67_pad_type_0 = const()[name = tensor("input_67_pad_type_0"), val = tensor("custom")]; + tensor input_67_pad_0 = const()[name = tensor("input_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_fc1_weight_to_fp16 = const()[name = tensor("layers_6_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(475018816)))]; + tensor layers_6_fc1_bias_to_fp16 = const()[name = tensor("layers_6_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488126080)))]; + tensor input_67_cast_fp16 = conv(bias = layers_6_fc1_bias_to_fp16, dilations = var_1653, groups = var_1467, pad = input_67_pad_0, pad_type = input_67_pad_type_0, strides = var_1651, weight = layers_6_fc1_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_mode_0 = const()[name = tensor("input_69_mode_0"), val = tensor("EXACT")]; + tensor input_69_cast_fp16 = gelu(mode = input_69_mode_0, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor var_1659 = const()[name = tensor("op_1659"), val = tensor([1, 1])]; + tensor var_1661 = const()[name = tensor("op_1661"), val = tensor([1, 1])]; + tensor hidden_states_15_pad_type_0 = const()[name = tensor("hidden_states_15_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_15_pad_0 = const()[name = tensor("hidden_states_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_fc2_weight_to_fp16 = const()[name = tensor("layers_6_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488136384)))]; + tensor layers_6_fc2_bias_to_fp16 = const()[name = tensor("layers_6_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(501243648)))]; + tensor hidden_states_15_cast_fp16 = conv(bias = layers_6_fc2_bias_to_fp16, dilations = var_1661, groups = var_1467, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = var_1659, weight = layers_6_fc2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("hidden_states_15_cast_fp16")]; + tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = hidden_states_15_cast_fp16)[name = tensor("inputs_43_cast_fp16")]; + tensor var_1674 = const()[name = tensor("op_1674"), val = tensor(3)]; + tensor var_1681 = const()[name = tensor("op_1681"), val = tensor(1)]; + tensor var_1682 = const()[name = tensor("op_1682"), val = tensor(true)]; + tensor var_1694 = const()[name = tensor("op_1694"), val = tensor([1])]; + tensor channels_mean_43_cast_fp16 = reduce_mean(axes = var_1694, keep_dims = var_1682, x = inputs_43_cast_fp16)[name = tensor("channels_mean_43_cast_fp16")]; + tensor zero_mean_43_cast_fp16 = sub(x = inputs_43_cast_fp16, y = channels_mean_43_cast_fp16)[name = tensor("zero_mean_43_cast_fp16")]; + tensor zero_mean_sq_43_cast_fp16 = mul(x = zero_mean_43_cast_fp16, y = zero_mean_43_cast_fp16)[name = tensor("zero_mean_sq_43_cast_fp16")]; + tensor var_1698 = const()[name = tensor("op_1698"), val = tensor([1])]; + tensor var_1699_cast_fp16 = reduce_mean(axes = var_1698, keep_dims = var_1682, x = zero_mean_sq_43_cast_fp16)[name = tensor("op_1699_cast_fp16")]; + tensor var_1700_to_fp16 = const()[name = tensor("op_1700_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1701_cast_fp16 = add(x = var_1699_cast_fp16, y = var_1700_to_fp16)[name = tensor("op_1701_cast_fp16")]; + tensor denom_43_epsilon_0_to_fp16 = const()[name = tensor("denom_43_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_43_cast_fp16 = rsqrt(epsilon = denom_43_epsilon_0_to_fp16, x = var_1701_cast_fp16)[name = tensor("denom_43_cast_fp16")]; + tensor out_43_cast_fp16 = mul(x = zero_mean_43_cast_fp16, y = denom_43_cast_fp16)[name = tensor("out_43_cast_fp16")]; + tensor obj_99_gamma_0_to_fp16 = const()[name = tensor("obj_99_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(501246272)))]; + tensor obj_99_beta_0_to_fp16 = const()[name = tensor("obj_99_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(501248896)))]; + tensor obj_99_epsilon_0_to_fp16 = const()[name = tensor("obj_99_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_99_cast_fp16 = batch_norm(beta = obj_99_beta_0_to_fp16, epsilon = obj_99_epsilon_0_to_fp16, gamma = obj_99_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_43_cast_fp16)[name = tensor("obj_99_cast_fp16")]; + tensor var_1716 = const()[name = tensor("op_1716"), val = tensor([1, 1])]; + tensor var_1718 = const()[name = tensor("op_1718"), val = tensor([1, 1])]; + tensor query_29_pad_type_0 = const()[name = tensor("query_29_pad_type_0"), val = tensor("custom")]; + tensor query_29_pad_0 = const()[name = tensor("query_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(501251520)))]; + tensor layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504528384)))]; + tensor query_29_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_bias_to_fp16, dilations = var_1718, groups = var_1681, pad = query_29_pad_0, pad_type = query_29_pad_type_0, strides = var_1716, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("query_29_cast_fp16")]; + tensor var_1722 = const()[name = tensor("op_1722"), val = tensor([1, 1])]; + tensor var_1724 = const()[name = tensor("op_1724"), val = tensor([1, 1])]; + tensor current_key_15_pad_type_0 = const()[name = tensor("current_key_15_pad_type_0"), val = tensor("custom")]; + tensor current_key_15_pad_0 = const()[name = tensor("current_key_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504531008)))]; + tensor current_key_15_cast_fp16 = conv(dilations = var_1724, groups = var_1681, pad = current_key_15_pad_0, pad_type = current_key_15_pad_type_0, strides = var_1722, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("current_key_15_cast_fp16")]; + tensor var_1729 = const()[name = tensor("op_1729"), val = tensor([1, 1])]; + tensor var_1731 = const()[name = tensor("op_1731"), val = tensor([1, 1])]; + tensor current_value_15_pad_type_0 = const()[name = tensor("current_value_15_pad_type_0"), val = tensor("custom")]; + tensor current_value_15_pad_0 = const()[name = tensor("current_value_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507807872)))]; + tensor layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511084736)))]; + tensor current_value_15_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_bias_to_fp16, dilations = var_1731, groups = var_1681, pad = current_value_15_pad_0, pad_type = current_value_15_pad_type_0, strides = var_1729, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("current_value_15_cast_fp16")]; + tensor var_1738_cast_fp16 = mul(x = current_key_15_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1738_cast_fp16")]; + tensor var_1740_cast_fp16 = mul(x = var_103_cast_fp16_7, y = var_241_cast_fp16)[name = tensor("op_1740_cast_fp16")]; + tensor key_29_cast_fp16 = add(x = var_1738_cast_fp16, y = var_1740_cast_fp16)[name = tensor("key_29_cast_fp16")]; + tensor var_1742_cast_fp16 = mul(x = current_value_15_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1742_cast_fp16")]; + tensor var_1744_cast_fp16 = mul(x = var_138_cast_fp16_7, y = var_241_cast_fp16)[name = tensor("op_1744_cast_fp16")]; + tensor value_29_cast_fp16 = add(x = var_1742_cast_fp16, y = var_1744_cast_fp16)[name = tensor("value_29_cast_fp16")]; + tensor var_1747 = const()[name = tensor("op_1747"), val = tensor([1, 20, 64, -1])]; + tensor var_1748_cast_fp16 = reshape(shape = var_1747, x = query_29_cast_fp16)[name = tensor("op_1748_cast_fp16")]; + tensor var_1749_to_fp16 = const()[name = tensor("op_1749_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1750_cast_fp16 = mul(x = var_1748_cast_fp16, y = var_1749_to_fp16)[name = tensor("op_1750_cast_fp16")]; + tensor var_1751 = const()[name = tensor("op_1751"), val = tensor([1, 20, 64, -1])]; + tensor var_1752_cast_fp16 = reshape(shape = var_1751, x = key_29_cast_fp16)[name = tensor("op_1752_cast_fp16")]; + tensor mh_w_43_transpose_x_0 = const()[name = tensor("mh_w_43_transpose_x_0"), val = tensor(true)]; + tensor mh_w_43_transpose_y_0 = const()[name = tensor("mh_w_43_transpose_y_0"), val = tensor(false)]; + tensor mh_w_43_cast_fp16 = matmul(transpose_x = mh_w_43_transpose_x_0, transpose_y = mh_w_43_transpose_y_0, x = var_1750_cast_fp16, y = var_1752_cast_fp16)[name = tensor("mh_w_43_cast_fp16")]; + tensor mh_w_45_cast_fp16 = add(x = mh_w_43_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_45_cast_fp16")]; + tensor var_1760_cast_fp16 = softmax(axis = var_1674, x = mh_w_45_cast_fp16)[name = tensor("op_1760_cast_fp16")]; + tensor var_1761 = const()[name = tensor("op_1761"), val = tensor([1, 20, 64, -1])]; + tensor var_1762_cast_fp16 = reshape(shape = var_1761, x = value_29_cast_fp16)[name = tensor("op_1762_cast_fp16")]; + tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; + tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(true)]; + tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = var_1762_cast_fp16, y = var_1760_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_1765 = const()[name = tensor("op_1765"), val = tensor([1, 1280, 1, -1])]; + tensor input_71_cast_fp16 = reshape(shape = var_1765, x = attn_29_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor var_1769 = const()[name = tensor("op_1769"), val = tensor([1, 1])]; + tensor var_1771 = const()[name = tensor("op_1771"), val = tensor([1, 1])]; + tensor obj_105_pad_type_0 = const()[name = tensor("obj_105_pad_type_0"), val = tensor("custom")]; + tensor obj_105_pad_0 = const()[name = tensor("obj_105_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511087360)))]; + tensor layers_7_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514364224)))]; + tensor obj_105_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_bias_to_fp16, dilations = var_1771, groups = var_1681, pad = obj_105_pad_0, pad_type = obj_105_pad_type_0, strides = var_1769, weight = layers_7_self_attn_o_proj_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("obj_105_cast_fp16")]; + tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = obj_105_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; + tensor var_1781 = const()[name = tensor("op_1781"), val = tensor([1])]; + tensor channels_mean_45_cast_fp16 = reduce_mean(axes = var_1781, keep_dims = var_1682, x = inputs_45_cast_fp16)[name = tensor("channels_mean_45_cast_fp16")]; + tensor zero_mean_45_cast_fp16 = sub(x = inputs_45_cast_fp16, y = channels_mean_45_cast_fp16)[name = tensor("zero_mean_45_cast_fp16")]; + tensor zero_mean_sq_45_cast_fp16 = mul(x = zero_mean_45_cast_fp16, y = zero_mean_45_cast_fp16)[name = tensor("zero_mean_sq_45_cast_fp16")]; + tensor var_1785 = const()[name = tensor("op_1785"), val = tensor([1])]; + tensor var_1786_cast_fp16 = reduce_mean(axes = var_1785, keep_dims = var_1682, x = zero_mean_sq_45_cast_fp16)[name = tensor("op_1786_cast_fp16")]; + tensor var_1787_to_fp16 = const()[name = tensor("op_1787_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1788_cast_fp16 = add(x = var_1786_cast_fp16, y = var_1787_to_fp16)[name = tensor("op_1788_cast_fp16")]; + tensor denom_45_epsilon_0_to_fp16 = const()[name = tensor("denom_45_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_45_cast_fp16 = rsqrt(epsilon = denom_45_epsilon_0_to_fp16, x = var_1788_cast_fp16)[name = tensor("denom_45_cast_fp16")]; + tensor out_45_cast_fp16 = mul(x = zero_mean_45_cast_fp16, y = denom_45_cast_fp16)[name = tensor("out_45_cast_fp16")]; + tensor obj_107_gamma_0_to_fp16 = const()[name = tensor("obj_107_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514366848)))]; + tensor obj_107_beta_0_to_fp16 = const()[name = tensor("obj_107_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514369472)))]; + tensor obj_107_epsilon_0_to_fp16 = const()[name = tensor("obj_107_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_107_cast_fp16 = batch_norm(beta = obj_107_beta_0_to_fp16, epsilon = obj_107_epsilon_0_to_fp16, gamma = obj_107_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor("obj_107_cast_fp16")]; + tensor var_1803 = const()[name = tensor("op_1803"), val = tensor([1, 1])]; + tensor var_1805 = const()[name = tensor("op_1805"), val = tensor([1, 1])]; + tensor query_31_pad_type_0 = const()[name = tensor("query_31_pad_type_0"), val = tensor("custom")]; + tensor query_31_pad_0 = const()[name = tensor("query_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514372096)))]; + tensor layers_7_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517648960)))]; + tensor query_31_cast_fp16 = conv(bias = layers_7_encoder_attn_q_proj_bias_to_fp16, dilations = var_1805, groups = var_1681, pad = query_31_pad_0, pad_type = query_31_pad_type_0, strides = var_1803, weight = layers_7_encoder_attn_q_proj_weight_to_fp16, x = obj_107_cast_fp16)[name = tensor("query_31_cast_fp16")]; + tensor var_1809 = const()[name = tensor("op_1809"), val = tensor([1, 1])]; + tensor var_1811 = const()[name = tensor("op_1811"), val = tensor([1, 1])]; + tensor key_31_pad_type_0 = const()[name = tensor("key_31_pad_type_0"), val = tensor("custom")]; + tensor key_31_pad_0 = const()[name = tensor("key_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517651584)))]; + tensor key_31_cast_fp16 = conv(dilations = var_1811, groups = var_1681, pad = key_31_pad_0, pad_type = key_31_pad_type_0, strides = var_1809, weight = layers_7_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_31_cast_fp16")]; + tensor var_1816 = const()[name = tensor("op_1816"), val = tensor([1, 1])]; + tensor var_1818 = const()[name = tensor("op_1818"), val = tensor([1, 1])]; + tensor value_31_pad_type_0 = const()[name = tensor("value_31_pad_type_0"), val = tensor("custom")]; + tensor value_31_pad_0 = const()[name = tensor("value_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520928448)))]; + tensor layers_7_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524205312)))]; + tensor value_31_cast_fp16 = conv(bias = layers_7_encoder_attn_v_proj_bias_to_fp16, dilations = var_1818, groups = var_1681, pad = value_31_pad_0, pad_type = value_31_pad_type_0, strides = var_1816, weight = layers_7_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_31_cast_fp16")]; + tensor var_1822 = const()[name = tensor("op_1822"), val = tensor([1, 20, 64, -1])]; + tensor var_1823_cast_fp16 = reshape(shape = var_1822, x = query_31_cast_fp16)[name = tensor("op_1823_cast_fp16")]; + tensor var_1824_to_fp16 = const()[name = tensor("op_1824_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1825_cast_fp16 = mul(x = var_1823_cast_fp16, y = var_1824_to_fp16)[name = tensor("op_1825_cast_fp16")]; + tensor var_1826 = const()[name = tensor("op_1826"), val = tensor([1, 20, 64, -1])]; + tensor var_1827_cast_fp16 = reshape(shape = var_1826, x = key_31_cast_fp16)[name = tensor("op_1827_cast_fp16")]; + tensor mh_w_47_transpose_x_0 = const()[name = tensor("mh_w_47_transpose_x_0"), val = tensor(true)]; + tensor mh_w_47_transpose_y_0 = const()[name = tensor("mh_w_47_transpose_y_0"), val = tensor(false)]; + tensor mh_w_47_cast_fp16 = matmul(transpose_x = mh_w_47_transpose_x_0, transpose_y = mh_w_47_transpose_y_0, x = var_1825_cast_fp16, y = var_1827_cast_fp16)[name = tensor("mh_w_47_cast_fp16")]; + tensor obj_111_cast_fp16 = softmax(axis = var_1674, x = mh_w_47_cast_fp16)[name = tensor("obj_111_cast_fp16")]; + tensor var_1831 = const()[name = tensor("op_1831"), val = tensor([1, 20, 64, -1])]; + tensor var_1832_cast_fp16 = reshape(shape = var_1831, x = value_31_cast_fp16)[name = tensor("op_1832_cast_fp16")]; + tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; + tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(true)]; + tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = var_1832_cast_fp16, y = obj_111_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_1835 = const()[name = tensor("op_1835"), val = tensor([1, 1280, 1, -1])]; + tensor input_73_cast_fp16 = reshape(shape = var_1835, x = attn_31_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor var_1839 = const()[name = tensor("op_1839"), val = tensor([1, 1])]; + tensor var_1841 = const()[name = tensor("op_1841"), val = tensor([1, 1])]; + tensor obj_109_pad_type_0 = const()[name = tensor("obj_109_pad_type_0"), val = tensor("custom")]; + tensor obj_109_pad_0 = const()[name = tensor("obj_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524207936)))]; + tensor layers_7_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527484800)))]; + tensor obj_109_cast_fp16 = conv(bias = layers_7_encoder_attn_o_proj_bias_to_fp16, dilations = var_1841, groups = var_1681, pad = obj_109_pad_0, pad_type = obj_109_pad_type_0, strides = var_1839, weight = layers_7_encoder_attn_o_proj_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("obj_109_cast_fp16")]; + tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_109_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; + tensor var_1850 = const()[name = tensor("op_1850"), val = tensor([1])]; + tensor channels_mean_47_cast_fp16 = reduce_mean(axes = var_1850, keep_dims = var_1682, x = inputs_47_cast_fp16)[name = tensor("channels_mean_47_cast_fp16")]; + tensor zero_mean_47_cast_fp16 = sub(x = inputs_47_cast_fp16, y = channels_mean_47_cast_fp16)[name = tensor("zero_mean_47_cast_fp16")]; + tensor zero_mean_sq_47_cast_fp16 = mul(x = zero_mean_47_cast_fp16, y = zero_mean_47_cast_fp16)[name = tensor("zero_mean_sq_47_cast_fp16")]; + tensor var_1854 = const()[name = tensor("op_1854"), val = tensor([1])]; + tensor var_1855_cast_fp16 = reduce_mean(axes = var_1854, keep_dims = var_1682, x = zero_mean_sq_47_cast_fp16)[name = tensor("op_1855_cast_fp16")]; + tensor var_1856_to_fp16 = const()[name = tensor("op_1856_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1857_cast_fp16 = add(x = var_1855_cast_fp16, y = var_1856_to_fp16)[name = tensor("op_1857_cast_fp16")]; + tensor denom_47_epsilon_0_to_fp16 = const()[name = tensor("denom_47_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_47_cast_fp16 = rsqrt(epsilon = denom_47_epsilon_0_to_fp16, x = var_1857_cast_fp16)[name = tensor("denom_47_cast_fp16")]; + tensor out_47_cast_fp16 = mul(x = zero_mean_47_cast_fp16, y = denom_47_cast_fp16)[name = tensor("out_47_cast_fp16")]; + tensor input_75_gamma_0_to_fp16 = const()[name = tensor("input_75_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527487424)))]; + tensor input_75_beta_0_to_fp16 = const()[name = tensor("input_75_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527490048)))]; + tensor input_75_epsilon_0_to_fp16 = const()[name = tensor("input_75_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_75_cast_fp16 = batch_norm(beta = input_75_beta_0_to_fp16, epsilon = input_75_epsilon_0_to_fp16, gamma = input_75_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_47_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor var_1868 = const()[name = tensor("op_1868"), val = tensor([1, 1])]; + tensor var_1870 = const()[name = tensor("op_1870"), val = tensor([1, 1])]; + tensor input_77_pad_type_0 = const()[name = tensor("input_77_pad_type_0"), val = tensor("custom")]; + tensor input_77_pad_0 = const()[name = tensor("input_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_fc1_weight_to_fp16 = const()[name = tensor("layers_7_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527492672)))]; + tensor layers_7_fc1_bias_to_fp16 = const()[name = tensor("layers_7_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540599936)))]; + tensor input_77_cast_fp16 = conv(bias = layers_7_fc1_bias_to_fp16, dilations = var_1870, groups = var_1681, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = var_1868, weight = layers_7_fc1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_mode_0 = const()[name = tensor("input_79_mode_0"), val = tensor("EXACT")]; + tensor input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor var_1876 = const()[name = tensor("op_1876"), val = tensor([1, 1])]; + tensor var_1878 = const()[name = tensor("op_1878"), val = tensor([1, 1])]; + tensor hidden_states_17_pad_type_0 = const()[name = tensor("hidden_states_17_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_17_pad_0 = const()[name = tensor("hidden_states_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_fc2_weight_to_fp16 = const()[name = tensor("layers_7_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540610240)))]; + tensor layers_7_fc2_bias_to_fp16 = const()[name = tensor("layers_7_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553717504)))]; + tensor hidden_states_17_cast_fp16 = conv(bias = layers_7_fc2_bias_to_fp16, dilations = var_1878, groups = var_1681, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = var_1876, weight = layers_7_fc2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("hidden_states_17_cast_fp16")]; + tensor inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = hidden_states_17_cast_fp16)[name = tensor("inputs_49_cast_fp16")]; + tensor var_1892 = const()[name = tensor("op_1892"), val = tensor(3)]; + tensor var_1899 = const()[name = tensor("op_1899"), val = tensor(1)]; + tensor var_1900 = const()[name = tensor("op_1900"), val = tensor(true)]; + tensor var_1912 = const()[name = tensor("op_1912"), val = tensor([1])]; + tensor channels_mean_49_cast_fp16 = reduce_mean(axes = var_1912, keep_dims = var_1900, x = inputs_49_cast_fp16)[name = tensor("channels_mean_49_cast_fp16")]; + tensor zero_mean_49_cast_fp16 = sub(x = inputs_49_cast_fp16, y = channels_mean_49_cast_fp16)[name = tensor("zero_mean_49_cast_fp16")]; + tensor zero_mean_sq_49_cast_fp16 = mul(x = zero_mean_49_cast_fp16, y = zero_mean_49_cast_fp16)[name = tensor("zero_mean_sq_49_cast_fp16")]; + tensor var_1916 = const()[name = tensor("op_1916"), val = tensor([1])]; + tensor var_1917_cast_fp16 = reduce_mean(axes = var_1916, keep_dims = var_1900, x = zero_mean_sq_49_cast_fp16)[name = tensor("op_1917_cast_fp16")]; + tensor var_1918_to_fp16 = const()[name = tensor("op_1918_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1919_cast_fp16 = add(x = var_1917_cast_fp16, y = var_1918_to_fp16)[name = tensor("op_1919_cast_fp16")]; + tensor denom_49_epsilon_0_to_fp16 = const()[name = tensor("denom_49_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_49_cast_fp16 = rsqrt(epsilon = denom_49_epsilon_0_to_fp16, x = var_1919_cast_fp16)[name = tensor("denom_49_cast_fp16")]; + tensor out_49_cast_fp16 = mul(x = zero_mean_49_cast_fp16, y = denom_49_cast_fp16)[name = tensor("out_49_cast_fp16")]; + tensor obj_113_gamma_0_to_fp16 = const()[name = tensor("obj_113_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553720128)))]; + tensor obj_113_beta_0_to_fp16 = const()[name = tensor("obj_113_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553722752)))]; + tensor obj_113_epsilon_0_to_fp16 = const()[name = tensor("obj_113_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_113_cast_fp16 = batch_norm(beta = obj_113_beta_0_to_fp16, epsilon = obj_113_epsilon_0_to_fp16, gamma = obj_113_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_49_cast_fp16)[name = tensor("obj_113_cast_fp16")]; + tensor var_1934 = const()[name = tensor("op_1934"), val = tensor([1, 1])]; + tensor var_1936 = const()[name = tensor("op_1936"), val = tensor([1, 1])]; + tensor query_33_pad_type_0 = const()[name = tensor("query_33_pad_type_0"), val = tensor("custom")]; + tensor query_33_pad_0 = const()[name = tensor("query_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553725376)))]; + tensor layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(557002240)))]; + tensor query_33_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_bias_to_fp16, dilations = var_1936, groups = var_1899, pad = query_33_pad_0, pad_type = query_33_pad_type_0, strides = var_1934, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("query_33_cast_fp16")]; + tensor var_1940 = const()[name = tensor("op_1940"), val = tensor([1, 1])]; + tensor var_1942 = const()[name = tensor("op_1942"), val = tensor([1, 1])]; + tensor current_key_17_pad_type_0 = const()[name = tensor("current_key_17_pad_type_0"), val = tensor("custom")]; + tensor current_key_17_pad_0 = const()[name = tensor("current_key_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(557004864)))]; + tensor current_key_17_cast_fp16 = conv(dilations = var_1942, groups = var_1899, pad = current_key_17_pad_0, pad_type = current_key_17_pad_type_0, strides = var_1940, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("current_key_17_cast_fp16")]; + tensor var_1947 = const()[name = tensor("op_1947"), val = tensor([1, 1])]; + tensor var_1949 = const()[name = tensor("op_1949"), val = tensor([1, 1])]; + tensor current_value_17_pad_type_0 = const()[name = tensor("current_value_17_pad_type_0"), val = tensor("custom")]; + tensor current_value_17_pad_0 = const()[name = tensor("current_value_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560281728)))]; + tensor layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563558592)))]; + tensor current_value_17_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_bias_to_fp16, dilations = var_1949, groups = var_1899, pad = current_value_17_pad_0, pad_type = current_value_17_pad_type_0, strides = var_1947, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("current_value_17_cast_fp16")]; + tensor var_1956_cast_fp16 = mul(x = current_key_17_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1956_cast_fp16")]; + tensor var_1958_cast_fp16 = mul(x = var_103_cast_fp16_8, y = var_241_cast_fp16)[name = tensor("op_1958_cast_fp16")]; + tensor key_33_cast_fp16 = add(x = var_1956_cast_fp16, y = var_1958_cast_fp16)[name = tensor("key_33_cast_fp16")]; + tensor var_1960_cast_fp16 = mul(x = current_value_17_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1960_cast_fp16")]; + tensor var_1962_cast_fp16 = mul(x = var_138_cast_fp16_8, y = var_241_cast_fp16)[name = tensor("op_1962_cast_fp16")]; + tensor value_33_cast_fp16 = add(x = var_1960_cast_fp16, y = var_1962_cast_fp16)[name = tensor("value_33_cast_fp16")]; + tensor var_1965 = const()[name = tensor("op_1965"), val = tensor([1, 20, 64, -1])]; + tensor var_1966_cast_fp16 = reshape(shape = var_1965, x = query_33_cast_fp16)[name = tensor("op_1966_cast_fp16")]; + tensor var_1967_to_fp16 = const()[name = tensor("op_1967_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1968_cast_fp16 = mul(x = var_1966_cast_fp16, y = var_1967_to_fp16)[name = tensor("op_1968_cast_fp16")]; + tensor var_1969 = const()[name = tensor("op_1969"), val = tensor([1, 20, 64, -1])]; + tensor var_1970_cast_fp16 = reshape(shape = var_1969, x = key_33_cast_fp16)[name = tensor("op_1970_cast_fp16")]; + tensor mh_w_49_transpose_x_0 = const()[name = tensor("mh_w_49_transpose_x_0"), val = tensor(true)]; + tensor mh_w_49_transpose_y_0 = const()[name = tensor("mh_w_49_transpose_y_0"), val = tensor(false)]; + tensor mh_w_49_cast_fp16 = matmul(transpose_x = mh_w_49_transpose_x_0, transpose_y = mh_w_49_transpose_y_0, x = var_1968_cast_fp16, y = var_1970_cast_fp16)[name = tensor("mh_w_49_cast_fp16")]; + tensor mh_w_51_cast_fp16 = add(x = mh_w_49_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_51_cast_fp16")]; + tensor var_1978_cast_fp16 = softmax(axis = var_1892, x = mh_w_51_cast_fp16)[name = tensor("op_1978_cast_fp16")]; + tensor var_1979 = const()[name = tensor("op_1979"), val = tensor([1, 20, 64, -1])]; + tensor var_1980_cast_fp16 = reshape(shape = var_1979, x = value_33_cast_fp16)[name = tensor("op_1980_cast_fp16")]; + tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; + tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(true)]; + tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = var_1980_cast_fp16, y = var_1978_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_1983 = const()[name = tensor("op_1983"), val = tensor([1, 1280, 1, -1])]; + tensor input_81_cast_fp16 = reshape(shape = var_1983, x = attn_33_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor var_1987 = const()[name = tensor("op_1987"), val = tensor([1, 1])]; + tensor var_1989 = const()[name = tensor("op_1989"), val = tensor([1, 1])]; + tensor obj_119_pad_type_0 = const()[name = tensor("obj_119_pad_type_0"), val = tensor("custom")]; + tensor obj_119_pad_0 = const()[name = tensor("obj_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563561216)))]; + tensor layers_8_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566838080)))]; + tensor obj_119_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_bias_to_fp16, dilations = var_1989, groups = var_1899, pad = obj_119_pad_0, pad_type = obj_119_pad_type_0, strides = var_1987, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("obj_119_cast_fp16")]; + tensor inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_119_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; + tensor var_1999 = const()[name = tensor("op_1999"), val = tensor([1])]; + tensor channels_mean_51_cast_fp16 = reduce_mean(axes = var_1999, keep_dims = var_1900, x = inputs_51_cast_fp16)[name = tensor("channels_mean_51_cast_fp16")]; + tensor zero_mean_51_cast_fp16 = sub(x = inputs_51_cast_fp16, y = channels_mean_51_cast_fp16)[name = tensor("zero_mean_51_cast_fp16")]; + tensor zero_mean_sq_51_cast_fp16 = mul(x = zero_mean_51_cast_fp16, y = zero_mean_51_cast_fp16)[name = tensor("zero_mean_sq_51_cast_fp16")]; + tensor var_2003 = const()[name = tensor("op_2003"), val = tensor([1])]; + tensor var_2004_cast_fp16 = reduce_mean(axes = var_2003, keep_dims = var_1900, x = zero_mean_sq_51_cast_fp16)[name = tensor("op_2004_cast_fp16")]; + tensor var_2005_to_fp16 = const()[name = tensor("op_2005_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2006_cast_fp16 = add(x = var_2004_cast_fp16, y = var_2005_to_fp16)[name = tensor("op_2006_cast_fp16")]; + tensor denom_51_epsilon_0_to_fp16 = const()[name = tensor("denom_51_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_51_cast_fp16 = rsqrt(epsilon = denom_51_epsilon_0_to_fp16, x = var_2006_cast_fp16)[name = tensor("denom_51_cast_fp16")]; + tensor out_51_cast_fp16 = mul(x = zero_mean_51_cast_fp16, y = denom_51_cast_fp16)[name = tensor("out_51_cast_fp16")]; + tensor obj_121_gamma_0_to_fp16 = const()[name = tensor("obj_121_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566840704)))]; + tensor obj_121_beta_0_to_fp16 = const()[name = tensor("obj_121_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566843328)))]; + tensor obj_121_epsilon_0_to_fp16 = const()[name = tensor("obj_121_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_121_cast_fp16 = batch_norm(beta = obj_121_beta_0_to_fp16, epsilon = obj_121_epsilon_0_to_fp16, gamma = obj_121_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_51_cast_fp16)[name = tensor("obj_121_cast_fp16")]; + tensor var_2021 = const()[name = tensor("op_2021"), val = tensor([1, 1])]; + tensor var_2023 = const()[name = tensor("op_2023"), val = tensor([1, 1])]; + tensor query_35_pad_type_0 = const()[name = tensor("query_35_pad_type_0"), val = tensor("custom")]; + tensor query_35_pad_0 = const()[name = tensor("query_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566845952)))]; + tensor layers_8_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570122816)))]; + tensor query_35_cast_fp16 = conv(bias = layers_8_encoder_attn_q_proj_bias_to_fp16, dilations = var_2023, groups = var_1899, pad = query_35_pad_0, pad_type = query_35_pad_type_0, strides = var_2021, weight = layers_8_encoder_attn_q_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("query_35_cast_fp16")]; + tensor var_2027 = const()[name = tensor("op_2027"), val = tensor([1, 1])]; + tensor var_2029 = const()[name = tensor("op_2029"), val = tensor([1, 1])]; + tensor key_35_pad_type_0 = const()[name = tensor("key_35_pad_type_0"), val = tensor("custom")]; + tensor key_35_pad_0 = const()[name = tensor("key_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570125440)))]; + tensor key_35_cast_fp16 = conv(dilations = var_2029, groups = var_1899, pad = key_35_pad_0, pad_type = key_35_pad_type_0, strides = var_2027, weight = layers_8_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_35_cast_fp16")]; + tensor var_2034 = const()[name = tensor("op_2034"), val = tensor([1, 1])]; + tensor var_2036 = const()[name = tensor("op_2036"), val = tensor([1, 1])]; + tensor value_35_pad_type_0 = const()[name = tensor("value_35_pad_type_0"), val = tensor("custom")]; + tensor value_35_pad_0 = const()[name = tensor("value_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(573402304)))]; + tensor layers_8_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(576679168)))]; + tensor value_35_cast_fp16 = conv(bias = layers_8_encoder_attn_v_proj_bias_to_fp16, dilations = var_2036, groups = var_1899, pad = value_35_pad_0, pad_type = value_35_pad_type_0, strides = var_2034, weight = layers_8_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_35_cast_fp16")]; + tensor var_2040 = const()[name = tensor("op_2040"), val = tensor([1, 20, 64, -1])]; + tensor var_2041_cast_fp16 = reshape(shape = var_2040, x = query_35_cast_fp16)[name = tensor("op_2041_cast_fp16")]; + tensor var_2042_to_fp16 = const()[name = tensor("op_2042_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2043_cast_fp16 = mul(x = var_2041_cast_fp16, y = var_2042_to_fp16)[name = tensor("op_2043_cast_fp16")]; + tensor var_2044 = const()[name = tensor("op_2044"), val = tensor([1, 20, 64, -1])]; + tensor var_2045_cast_fp16 = reshape(shape = var_2044, x = key_35_cast_fp16)[name = tensor("op_2045_cast_fp16")]; + tensor mh_w_53_transpose_x_0 = const()[name = tensor("mh_w_53_transpose_x_0"), val = tensor(true)]; + tensor mh_w_53_transpose_y_0 = const()[name = tensor("mh_w_53_transpose_y_0"), val = tensor(false)]; + tensor mh_w_53_cast_fp16 = matmul(transpose_x = mh_w_53_transpose_x_0, transpose_y = mh_w_53_transpose_y_0, x = var_2043_cast_fp16, y = var_2045_cast_fp16)[name = tensor("mh_w_53_cast_fp16")]; + tensor obj_125_cast_fp16 = softmax(axis = var_1892, x = mh_w_53_cast_fp16)[name = tensor("obj_125_cast_fp16")]; + tensor var_2049 = const()[name = tensor("op_2049"), val = tensor([1, 20, 64, -1])]; + tensor var_2050_cast_fp16 = reshape(shape = var_2049, x = value_35_cast_fp16)[name = tensor("op_2050_cast_fp16")]; + tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; + tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(true)]; + tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = var_2050_cast_fp16, y = obj_125_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_2053 = const()[name = tensor("op_2053"), val = tensor([1, 1280, 1, -1])]; + tensor input_83_cast_fp16 = reshape(shape = var_2053, x = attn_35_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor var_2057 = const()[name = tensor("op_2057"), val = tensor([1, 1])]; + tensor var_2059 = const()[name = tensor("op_2059"), val = tensor([1, 1])]; + tensor obj_123_pad_type_0 = const()[name = tensor("obj_123_pad_type_0"), val = tensor("custom")]; + tensor obj_123_pad_0 = const()[name = tensor("obj_123_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(576681792)))]; + tensor layers_8_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579958656)))]; + tensor obj_123_cast_fp16 = conv(bias = layers_8_encoder_attn_o_proj_bias_to_fp16, dilations = var_2059, groups = var_1899, pad = obj_123_pad_0, pad_type = obj_123_pad_type_0, strides = var_2057, weight = layers_8_encoder_attn_o_proj_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("obj_123_cast_fp16")]; + tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = obj_123_cast_fp16)[name = tensor("inputs_53_cast_fp16")]; + tensor var_2065 = const()[name = tensor("op_2065"), val = tensor([1])]; + tensor channels_mean_53_cast_fp16 = reduce_mean(axes = var_2065, keep_dims = var_1900, x = inputs_53_cast_fp16)[name = tensor("channels_mean_53_cast_fp16")]; + tensor zero_mean_53_cast_fp16 = sub(x = inputs_53_cast_fp16, y = channels_mean_53_cast_fp16)[name = tensor("zero_mean_53_cast_fp16")]; + tensor zero_mean_sq_53_cast_fp16 = mul(x = zero_mean_53_cast_fp16, y = zero_mean_53_cast_fp16)[name = tensor("zero_mean_sq_53_cast_fp16")]; + tensor var_2069 = const()[name = tensor("op_2069"), val = tensor([1])]; + tensor var_2070_cast_fp16 = reduce_mean(axes = var_2069, keep_dims = var_1900, x = zero_mean_sq_53_cast_fp16)[name = tensor("op_2070_cast_fp16")]; + tensor var_2071_to_fp16 = const()[name = tensor("op_2071_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2072_cast_fp16 = add(x = var_2070_cast_fp16, y = var_2071_to_fp16)[name = tensor("op_2072_cast_fp16")]; + tensor denom_53_epsilon_0_to_fp16 = const()[name = tensor("denom_53_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_53_cast_fp16 = rsqrt(epsilon = denom_53_epsilon_0_to_fp16, x = var_2072_cast_fp16)[name = tensor("denom_53_cast_fp16")]; + tensor out_53_cast_fp16 = mul(x = zero_mean_53_cast_fp16, y = denom_53_cast_fp16)[name = tensor("out_53_cast_fp16")]; + tensor input_85_gamma_0_to_fp16 = const()[name = tensor("input_85_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579961280)))]; + tensor input_85_beta_0_to_fp16 = const()[name = tensor("input_85_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579963904)))]; + tensor input_85_epsilon_0_to_fp16 = const()[name = tensor("input_85_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_85_cast_fp16 = batch_norm(beta = input_85_beta_0_to_fp16, epsilon = input_85_epsilon_0_to_fp16, gamma = input_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_53_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor var_2083 = const()[name = tensor("op_2083"), val = tensor([1, 1])]; + tensor var_2085 = const()[name = tensor("op_2085"), val = tensor([1, 1])]; + tensor input_87_pad_type_0 = const()[name = tensor("input_87_pad_type_0"), val = tensor("custom")]; + tensor input_87_pad_0 = const()[name = tensor("input_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_fc1_weight_to_fp16 = const()[name = tensor("layers_8_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579966528)))]; + tensor layers_8_fc1_bias_to_fp16 = const()[name = tensor("layers_8_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(593073792)))]; + tensor input_87_cast_fp16 = conv(bias = layers_8_fc1_bias_to_fp16, dilations = var_2085, groups = var_1899, pad = input_87_pad_0, pad_type = input_87_pad_type_0, strides = var_2083, weight = layers_8_fc1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_mode_0 = const()[name = tensor("input_89_mode_0"), val = tensor("EXACT")]; + tensor input_89_cast_fp16 = gelu(mode = input_89_mode_0, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor var_2091 = const()[name = tensor("op_2091"), val = tensor([1, 1])]; + tensor var_2093 = const()[name = tensor("op_2093"), val = tensor([1, 1])]; + tensor hidden_states_19_pad_type_0 = const()[name = tensor("hidden_states_19_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_19_pad_0 = const()[name = tensor("hidden_states_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_fc2_weight_to_fp16 = const()[name = tensor("layers_8_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(593084096)))]; + tensor layers_8_fc2_bias_to_fp16 = const()[name = tensor("layers_8_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(606191360)))]; + tensor hidden_states_19_cast_fp16 = conv(bias = layers_8_fc2_bias_to_fp16, dilations = var_2093, groups = var_1899, pad = hidden_states_19_pad_0, pad_type = hidden_states_19_pad_type_0, strides = var_2091, weight = layers_8_fc2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("hidden_states_19_cast_fp16")]; + tensor inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = hidden_states_19_cast_fp16)[name = tensor("inputs_55_cast_fp16")]; + tensor var_2106 = const()[name = tensor("op_2106"), val = tensor(3)]; + tensor var_2113 = const()[name = tensor("op_2113"), val = tensor(1)]; + tensor var_2114 = const()[name = tensor("op_2114"), val = tensor(true)]; + tensor var_2126 = const()[name = tensor("op_2126"), val = tensor([1])]; + tensor channels_mean_55_cast_fp16 = reduce_mean(axes = var_2126, keep_dims = var_2114, x = inputs_55_cast_fp16)[name = tensor("channels_mean_55_cast_fp16")]; + tensor zero_mean_55_cast_fp16 = sub(x = inputs_55_cast_fp16, y = channels_mean_55_cast_fp16)[name = tensor("zero_mean_55_cast_fp16")]; + tensor zero_mean_sq_55_cast_fp16 = mul(x = zero_mean_55_cast_fp16, y = zero_mean_55_cast_fp16)[name = tensor("zero_mean_sq_55_cast_fp16")]; + tensor var_2130 = const()[name = tensor("op_2130"), val = tensor([1])]; + tensor var_2131_cast_fp16 = reduce_mean(axes = var_2130, keep_dims = var_2114, x = zero_mean_sq_55_cast_fp16)[name = tensor("op_2131_cast_fp16")]; + tensor var_2132_to_fp16 = const()[name = tensor("op_2132_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2133_cast_fp16 = add(x = var_2131_cast_fp16, y = var_2132_to_fp16)[name = tensor("op_2133_cast_fp16")]; + tensor denom_55_epsilon_0_to_fp16 = const()[name = tensor("denom_55_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_55_cast_fp16 = rsqrt(epsilon = denom_55_epsilon_0_to_fp16, x = var_2133_cast_fp16)[name = tensor("denom_55_cast_fp16")]; + tensor out_55_cast_fp16 = mul(x = zero_mean_55_cast_fp16, y = denom_55_cast_fp16)[name = tensor("out_55_cast_fp16")]; + tensor obj_127_gamma_0_to_fp16 = const()[name = tensor("obj_127_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(606193984)))]; + tensor obj_127_beta_0_to_fp16 = const()[name = tensor("obj_127_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(606196608)))]; + tensor obj_127_epsilon_0_to_fp16 = const()[name = tensor("obj_127_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_127_cast_fp16 = batch_norm(beta = obj_127_beta_0_to_fp16, epsilon = obj_127_epsilon_0_to_fp16, gamma = obj_127_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_55_cast_fp16)[name = tensor("obj_127_cast_fp16")]; + tensor var_2148 = const()[name = tensor("op_2148"), val = tensor([1, 1])]; + tensor var_2150 = const()[name = tensor("op_2150"), val = tensor([1, 1])]; + tensor query_37_pad_type_0 = const()[name = tensor("query_37_pad_type_0"), val = tensor("custom")]; + tensor query_37_pad_0 = const()[name = tensor("query_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(606199232)))]; + tensor layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(609476096)))]; + tensor query_37_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_bias_to_fp16, dilations = var_2150, groups = var_2113, pad = query_37_pad_0, pad_type = query_37_pad_type_0, strides = var_2148, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("query_37_cast_fp16")]; + tensor var_2154 = const()[name = tensor("op_2154"), val = tensor([1, 1])]; + tensor var_2156 = const()[name = tensor("op_2156"), val = tensor([1, 1])]; + tensor current_key_19_pad_type_0 = const()[name = tensor("current_key_19_pad_type_0"), val = tensor("custom")]; + tensor current_key_19_pad_0 = const()[name = tensor("current_key_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(609478720)))]; + tensor current_key_19_cast_fp16 = conv(dilations = var_2156, groups = var_2113, pad = current_key_19_pad_0, pad_type = current_key_19_pad_type_0, strides = var_2154, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("current_key_19_cast_fp16")]; + tensor var_2161 = const()[name = tensor("op_2161"), val = tensor([1, 1])]; + tensor var_2163 = const()[name = tensor("op_2163"), val = tensor([1, 1])]; + tensor current_value_19_pad_type_0 = const()[name = tensor("current_value_19_pad_type_0"), val = tensor("custom")]; + tensor current_value_19_pad_0 = const()[name = tensor("current_value_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(612755584)))]; + tensor layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(616032448)))]; + tensor current_value_19_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_bias_to_fp16, dilations = var_2163, groups = var_2113, pad = current_value_19_pad_0, pad_type = current_value_19_pad_type_0, strides = var_2161, weight = layers_9_self_attn_v_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("current_value_19_cast_fp16")]; + tensor var_2170_cast_fp16 = mul(x = current_key_19_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_2170_cast_fp16")]; + tensor var_2172_cast_fp16 = mul(x = var_103_cast_fp16_9, y = var_241_cast_fp16)[name = tensor("op_2172_cast_fp16")]; + tensor key_37_cast_fp16 = add(x = var_2170_cast_fp16, y = var_2172_cast_fp16)[name = tensor("key_37_cast_fp16")]; + tensor var_2174_cast_fp16 = mul(x = current_value_19_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_2174_cast_fp16")]; + tensor var_2176_cast_fp16 = mul(x = var_138_cast_fp16_9, y = var_241_cast_fp16)[name = tensor("op_2176_cast_fp16")]; + tensor value_37_cast_fp16 = add(x = var_2174_cast_fp16, y = var_2176_cast_fp16)[name = tensor("value_37_cast_fp16")]; + tensor var_2179 = const()[name = tensor("op_2179"), val = tensor([1, 20, 64, -1])]; + tensor var_2180_cast_fp16 = reshape(shape = var_2179, x = query_37_cast_fp16)[name = tensor("op_2180_cast_fp16")]; + tensor var_2181_to_fp16 = const()[name = tensor("op_2181_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2182_cast_fp16 = mul(x = var_2180_cast_fp16, y = var_2181_to_fp16)[name = tensor("op_2182_cast_fp16")]; + tensor var_2183 = const()[name = tensor("op_2183"), val = tensor([1, 20, 64, -1])]; + tensor var_2184_cast_fp16 = reshape(shape = var_2183, x = key_37_cast_fp16)[name = tensor("op_2184_cast_fp16")]; + tensor mh_w_55_transpose_x_0 = const()[name = tensor("mh_w_55_transpose_x_0"), val = tensor(true)]; + tensor mh_w_55_transpose_y_0 = const()[name = tensor("mh_w_55_transpose_y_0"), val = tensor(false)]; + tensor mh_w_55_cast_fp16 = matmul(transpose_x = mh_w_55_transpose_x_0, transpose_y = mh_w_55_transpose_y_0, x = var_2182_cast_fp16, y = var_2184_cast_fp16)[name = tensor("mh_w_55_cast_fp16")]; + tensor mh_w_57_cast_fp16 = add(x = mh_w_55_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_57_cast_fp16")]; + tensor var_2192_cast_fp16 = softmax(axis = var_2106, x = mh_w_57_cast_fp16)[name = tensor("op_2192_cast_fp16")]; + tensor var_2193 = const()[name = tensor("op_2193"), val = tensor([1, 20, 64, -1])]; + tensor var_2194_cast_fp16 = reshape(shape = var_2193, x = value_37_cast_fp16)[name = tensor("op_2194_cast_fp16")]; + tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; + tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(true)]; + tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = var_2194_cast_fp16, y = var_2192_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_2197 = const()[name = tensor("op_2197"), val = tensor([1, 1280, 1, -1])]; + tensor input_91_cast_fp16 = reshape(shape = var_2197, x = attn_37_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor var_2201 = const()[name = tensor("op_2201"), val = tensor([1, 1])]; + tensor var_2203 = const()[name = tensor("op_2203"), val = tensor([1, 1])]; + tensor obj_133_pad_type_0 = const()[name = tensor("obj_133_pad_type_0"), val = tensor("custom")]; + tensor obj_133_pad_0 = const()[name = tensor("obj_133_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(616035072)))]; + tensor layers_9_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(619311936)))]; + tensor obj_133_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_bias_to_fp16, dilations = var_2203, groups = var_2113, pad = obj_133_pad_0, pad_type = obj_133_pad_type_0, strides = var_2201, weight = layers_9_self_attn_o_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("obj_133_cast_fp16")]; + tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = obj_133_cast_fp16)[name = tensor("inputs_57_cast_fp16")]; + tensor var_2213 = const()[name = tensor("op_2213"), val = tensor([1])]; + tensor channels_mean_57_cast_fp16 = reduce_mean(axes = var_2213, keep_dims = var_2114, x = inputs_57_cast_fp16)[name = tensor("channels_mean_57_cast_fp16")]; + tensor zero_mean_57_cast_fp16 = sub(x = inputs_57_cast_fp16, y = channels_mean_57_cast_fp16)[name = tensor("zero_mean_57_cast_fp16")]; + tensor zero_mean_sq_57_cast_fp16 = mul(x = zero_mean_57_cast_fp16, y = zero_mean_57_cast_fp16)[name = tensor("zero_mean_sq_57_cast_fp16")]; + tensor var_2217 = const()[name = tensor("op_2217"), val = tensor([1])]; + tensor var_2218_cast_fp16 = reduce_mean(axes = var_2217, keep_dims = var_2114, x = zero_mean_sq_57_cast_fp16)[name = tensor("op_2218_cast_fp16")]; + tensor var_2219_to_fp16 = const()[name = tensor("op_2219_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2220_cast_fp16 = add(x = var_2218_cast_fp16, y = var_2219_to_fp16)[name = tensor("op_2220_cast_fp16")]; + tensor denom_57_epsilon_0_to_fp16 = const()[name = tensor("denom_57_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_57_cast_fp16 = rsqrt(epsilon = denom_57_epsilon_0_to_fp16, x = var_2220_cast_fp16)[name = tensor("denom_57_cast_fp16")]; + tensor out_57_cast_fp16 = mul(x = zero_mean_57_cast_fp16, y = denom_57_cast_fp16)[name = tensor("out_57_cast_fp16")]; + tensor obj_135_gamma_0_to_fp16 = const()[name = tensor("obj_135_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(619314560)))]; + tensor obj_135_beta_0_to_fp16 = const()[name = tensor("obj_135_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(619317184)))]; + tensor obj_135_epsilon_0_to_fp16 = const()[name = tensor("obj_135_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_135_cast_fp16 = batch_norm(beta = obj_135_beta_0_to_fp16, epsilon = obj_135_epsilon_0_to_fp16, gamma = obj_135_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_57_cast_fp16)[name = tensor("obj_135_cast_fp16")]; + tensor var_2235 = const()[name = tensor("op_2235"), val = tensor([1, 1])]; + tensor var_2237 = const()[name = tensor("op_2237"), val = tensor([1, 1])]; + tensor query_39_pad_type_0 = const()[name = tensor("query_39_pad_type_0"), val = tensor("custom")]; + tensor query_39_pad_0 = const()[name = tensor("query_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(619319808)))]; + tensor layers_9_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(622596672)))]; + tensor query_39_cast_fp16 = conv(bias = layers_9_encoder_attn_q_proj_bias_to_fp16, dilations = var_2237, groups = var_2113, pad = query_39_pad_0, pad_type = query_39_pad_type_0, strides = var_2235, weight = layers_9_encoder_attn_q_proj_weight_to_fp16, x = obj_135_cast_fp16)[name = tensor("query_39_cast_fp16")]; + tensor var_2241 = const()[name = tensor("op_2241"), val = tensor([1, 1])]; + tensor var_2243 = const()[name = tensor("op_2243"), val = tensor([1, 1])]; + tensor key_39_pad_type_0 = const()[name = tensor("key_39_pad_type_0"), val = tensor("custom")]; + tensor key_39_pad_0 = const()[name = tensor("key_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(622599296)))]; + tensor key_39_cast_fp16 = conv(dilations = var_2243, groups = var_2113, pad = key_39_pad_0, pad_type = key_39_pad_type_0, strides = var_2241, weight = layers_9_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_39_cast_fp16")]; + tensor var_2248 = const()[name = tensor("op_2248"), val = tensor([1, 1])]; + tensor var_2250 = const()[name = tensor("op_2250"), val = tensor([1, 1])]; + tensor value_39_pad_type_0 = const()[name = tensor("value_39_pad_type_0"), val = tensor("custom")]; + tensor value_39_pad_0 = const()[name = tensor("value_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(625876160)))]; + tensor layers_9_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(629153024)))]; + tensor value_39_cast_fp16 = conv(bias = layers_9_encoder_attn_v_proj_bias_to_fp16, dilations = var_2250, groups = var_2113, pad = value_39_pad_0, pad_type = value_39_pad_type_0, strides = var_2248, weight = layers_9_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_39_cast_fp16")]; + tensor var_2254 = const()[name = tensor("op_2254"), val = tensor([1, 20, 64, -1])]; + tensor var_2255_cast_fp16 = reshape(shape = var_2254, x = query_39_cast_fp16)[name = tensor("op_2255_cast_fp16")]; + tensor var_2256_to_fp16 = const()[name = tensor("op_2256_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2257_cast_fp16 = mul(x = var_2255_cast_fp16, y = var_2256_to_fp16)[name = tensor("op_2257_cast_fp16")]; + tensor var_2258 = const()[name = tensor("op_2258"), val = tensor([1, 20, 64, -1])]; + tensor var_2259_cast_fp16 = reshape(shape = var_2258, x = key_39_cast_fp16)[name = tensor("op_2259_cast_fp16")]; + tensor mh_w_59_transpose_x_0 = const()[name = tensor("mh_w_59_transpose_x_0"), val = tensor(true)]; + tensor mh_w_59_transpose_y_0 = const()[name = tensor("mh_w_59_transpose_y_0"), val = tensor(false)]; + tensor mh_w_59_cast_fp16 = matmul(transpose_x = mh_w_59_transpose_x_0, transpose_y = mh_w_59_transpose_y_0, x = var_2257_cast_fp16, y = var_2259_cast_fp16)[name = tensor("mh_w_59_cast_fp16")]; + tensor obj_139_cast_fp16 = softmax(axis = var_2106, x = mh_w_59_cast_fp16)[name = tensor("obj_139_cast_fp16")]; + tensor var_2263 = const()[name = tensor("op_2263"), val = tensor([1, 20, 64, -1])]; + tensor var_2264_cast_fp16 = reshape(shape = var_2263, x = value_39_cast_fp16)[name = tensor("op_2264_cast_fp16")]; + tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; + tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(true)]; + tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = var_2264_cast_fp16, y = obj_139_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_2267 = const()[name = tensor("op_2267"), val = tensor([1, 1280, 1, -1])]; + tensor input_93_cast_fp16 = reshape(shape = var_2267, x = attn_39_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor var_2271 = const()[name = tensor("op_2271"), val = tensor([1, 1])]; + tensor var_2273 = const()[name = tensor("op_2273"), val = tensor([1, 1])]; + tensor obj_137_pad_type_0 = const()[name = tensor("obj_137_pad_type_0"), val = tensor("custom")]; + tensor obj_137_pad_0 = const()[name = tensor("obj_137_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(629155648)))]; + tensor layers_9_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(632432512)))]; + tensor obj_137_cast_fp16 = conv(bias = layers_9_encoder_attn_o_proj_bias_to_fp16, dilations = var_2273, groups = var_2113, pad = obj_137_pad_0, pad_type = obj_137_pad_type_0, strides = var_2271, weight = layers_9_encoder_attn_o_proj_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("obj_137_cast_fp16")]; + tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_137_cast_fp16)[name = tensor("inputs_59_cast_fp16")]; + tensor var_2279 = const()[name = tensor("op_2279"), val = tensor([1])]; + tensor channels_mean_59_cast_fp16 = reduce_mean(axes = var_2279, keep_dims = var_2114, x = inputs_59_cast_fp16)[name = tensor("channels_mean_59_cast_fp16")]; + tensor zero_mean_59_cast_fp16 = sub(x = inputs_59_cast_fp16, y = channels_mean_59_cast_fp16)[name = tensor("zero_mean_59_cast_fp16")]; + tensor zero_mean_sq_59_cast_fp16 = mul(x = zero_mean_59_cast_fp16, y = zero_mean_59_cast_fp16)[name = tensor("zero_mean_sq_59_cast_fp16")]; + tensor var_2283 = const()[name = tensor("op_2283"), val = tensor([1])]; + tensor var_2284_cast_fp16 = reduce_mean(axes = var_2283, keep_dims = var_2114, x = zero_mean_sq_59_cast_fp16)[name = tensor("op_2284_cast_fp16")]; + tensor var_2285_to_fp16 = const()[name = tensor("op_2285_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2286_cast_fp16 = add(x = var_2284_cast_fp16, y = var_2285_to_fp16)[name = tensor("op_2286_cast_fp16")]; + tensor denom_59_epsilon_0_to_fp16 = const()[name = tensor("denom_59_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_59_cast_fp16 = rsqrt(epsilon = denom_59_epsilon_0_to_fp16, x = var_2286_cast_fp16)[name = tensor("denom_59_cast_fp16")]; + tensor out_59_cast_fp16 = mul(x = zero_mean_59_cast_fp16, y = denom_59_cast_fp16)[name = tensor("out_59_cast_fp16")]; + tensor input_95_gamma_0_to_fp16 = const()[name = tensor("input_95_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(632435136)))]; + tensor input_95_beta_0_to_fp16 = const()[name = tensor("input_95_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(632437760)))]; + tensor input_95_epsilon_0_to_fp16 = const()[name = tensor("input_95_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_95_cast_fp16 = batch_norm(beta = input_95_beta_0_to_fp16, epsilon = input_95_epsilon_0_to_fp16, gamma = input_95_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_59_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor var_2297 = const()[name = tensor("op_2297"), val = tensor([1, 1])]; + tensor var_2299 = const()[name = tensor("op_2299"), val = tensor([1, 1])]; + tensor input_97_pad_type_0 = const()[name = tensor("input_97_pad_type_0"), val = tensor("custom")]; + tensor input_97_pad_0 = const()[name = tensor("input_97_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_fc1_weight_to_fp16 = const()[name = tensor("layers_9_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(632440384)))]; + tensor layers_9_fc1_bias_to_fp16 = const()[name = tensor("layers_9_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(645547648)))]; + tensor input_97_cast_fp16 = conv(bias = layers_9_fc1_bias_to_fp16, dilations = var_2299, groups = var_2113, pad = input_97_pad_0, pad_type = input_97_pad_type_0, strides = var_2297, weight = layers_9_fc1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_mode_0 = const()[name = tensor("input_99_mode_0"), val = tensor("EXACT")]; + tensor input_99_cast_fp16 = gelu(mode = input_99_mode_0, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor var_2305 = const()[name = tensor("op_2305"), val = tensor([1, 1])]; + tensor var_2307 = const()[name = tensor("op_2307"), val = tensor([1, 1])]; + tensor hidden_states_21_pad_type_0 = const()[name = tensor("hidden_states_21_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_21_pad_0 = const()[name = tensor("hidden_states_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_fc2_weight_to_fp16 = const()[name = tensor("layers_9_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(645557952)))]; + tensor layers_9_fc2_bias_to_fp16 = const()[name = tensor("layers_9_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(658665216)))]; + tensor hidden_states_21_cast_fp16 = conv(bias = layers_9_fc2_bias_to_fp16, dilations = var_2307, groups = var_2113, pad = hidden_states_21_pad_0, pad_type = hidden_states_21_pad_type_0, strides = var_2305, weight = layers_9_fc2_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("hidden_states_21_cast_fp16")]; + tensor inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = hidden_states_21_cast_fp16)[name = tensor("inputs_61_cast_fp16")]; + tensor var_2320 = const()[name = tensor("op_2320"), val = tensor(3)]; + tensor var_2327 = const()[name = tensor("op_2327"), val = tensor(1)]; + tensor var_2328 = const()[name = tensor("op_2328"), val = tensor(true)]; + tensor var_2340 = const()[name = tensor("op_2340"), val = tensor([1])]; + tensor channels_mean_61_cast_fp16 = reduce_mean(axes = var_2340, keep_dims = var_2328, x = inputs_61_cast_fp16)[name = tensor("channels_mean_61_cast_fp16")]; + tensor zero_mean_61_cast_fp16 = sub(x = inputs_61_cast_fp16, y = channels_mean_61_cast_fp16)[name = tensor("zero_mean_61_cast_fp16")]; + tensor zero_mean_sq_61_cast_fp16 = mul(x = zero_mean_61_cast_fp16, y = zero_mean_61_cast_fp16)[name = tensor("zero_mean_sq_61_cast_fp16")]; + tensor var_2344 = const()[name = tensor("op_2344"), val = tensor([1])]; + tensor var_2345_cast_fp16 = reduce_mean(axes = var_2344, keep_dims = var_2328, x = zero_mean_sq_61_cast_fp16)[name = tensor("op_2345_cast_fp16")]; + tensor var_2346_to_fp16 = const()[name = tensor("op_2346_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2347_cast_fp16 = add(x = var_2345_cast_fp16, y = var_2346_to_fp16)[name = tensor("op_2347_cast_fp16")]; + tensor denom_61_epsilon_0_to_fp16 = const()[name = tensor("denom_61_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_61_cast_fp16 = rsqrt(epsilon = denom_61_epsilon_0_to_fp16, x = var_2347_cast_fp16)[name = tensor("denom_61_cast_fp16")]; + tensor out_61_cast_fp16 = mul(x = zero_mean_61_cast_fp16, y = denom_61_cast_fp16)[name = tensor("out_61_cast_fp16")]; + tensor obj_141_gamma_0_to_fp16 = const()[name = tensor("obj_141_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(658667840)))]; + tensor obj_141_beta_0_to_fp16 = const()[name = tensor("obj_141_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(658670464)))]; + tensor obj_141_epsilon_0_to_fp16 = const()[name = tensor("obj_141_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_141_cast_fp16 = batch_norm(beta = obj_141_beta_0_to_fp16, epsilon = obj_141_epsilon_0_to_fp16, gamma = obj_141_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_61_cast_fp16)[name = tensor("obj_141_cast_fp16")]; + tensor var_2362 = const()[name = tensor("op_2362"), val = tensor([1, 1])]; + tensor var_2364 = const()[name = tensor("op_2364"), val = tensor([1, 1])]; + tensor query_41_pad_type_0 = const()[name = tensor("query_41_pad_type_0"), val = tensor("custom")]; + tensor query_41_pad_0 = const()[name = tensor("query_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(658673088)))]; + tensor layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(661949952)))]; + tensor query_41_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_bias_to_fp16, dilations = var_2364, groups = var_2327, pad = query_41_pad_0, pad_type = query_41_pad_type_0, strides = var_2362, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("query_41_cast_fp16")]; + tensor var_2368 = const()[name = tensor("op_2368"), val = tensor([1, 1])]; + tensor var_2370 = const()[name = tensor("op_2370"), val = tensor([1, 1])]; + tensor current_key_21_pad_type_0 = const()[name = tensor("current_key_21_pad_type_0"), val = tensor("custom")]; + tensor current_key_21_pad_0 = const()[name = tensor("current_key_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(661952576)))]; + tensor current_key_21_cast_fp16 = conv(dilations = var_2370, groups = var_2327, pad = current_key_21_pad_0, pad_type = current_key_21_pad_type_0, strides = var_2368, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("current_key_21_cast_fp16")]; + tensor var_2375 = const()[name = tensor("op_2375"), val = tensor([1, 1])]; + tensor var_2377 = const()[name = tensor("op_2377"), val = tensor([1, 1])]; + tensor current_value_21_pad_type_0 = const()[name = tensor("current_value_21_pad_type_0"), val = tensor("custom")]; + tensor current_value_21_pad_0 = const()[name = tensor("current_value_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(665229440)))]; + tensor layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(668506304)))]; + tensor current_value_21_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_bias_to_fp16, dilations = var_2377, groups = var_2327, pad = current_value_21_pad_0, pad_type = current_value_21_pad_type_0, strides = var_2375, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("current_value_21_cast_fp16")]; + tensor var_2384_cast_fp16 = mul(x = current_key_21_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_2384_cast_fp16")]; + tensor var_2386_cast_fp16 = mul(x = var_103_cast_fp16_10, y = var_241_cast_fp16)[name = tensor("op_2386_cast_fp16")]; + tensor key_41_cast_fp16 = add(x = var_2384_cast_fp16, y = var_2386_cast_fp16)[name = tensor("key_41_cast_fp16")]; + tensor var_2388_cast_fp16 = mul(x = current_value_21_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_2388_cast_fp16")]; + tensor var_2390_cast_fp16 = mul(x = var_138_cast_fp16_10, y = var_241_cast_fp16)[name = tensor("op_2390_cast_fp16")]; + tensor value_41_cast_fp16 = add(x = var_2388_cast_fp16, y = var_2390_cast_fp16)[name = tensor("value_41_cast_fp16")]; + tensor var_2393 = const()[name = tensor("op_2393"), val = tensor([1, 20, 64, -1])]; + tensor var_2394_cast_fp16 = reshape(shape = var_2393, x = query_41_cast_fp16)[name = tensor("op_2394_cast_fp16")]; + tensor var_2395_to_fp16 = const()[name = tensor("op_2395_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2396_cast_fp16 = mul(x = var_2394_cast_fp16, y = var_2395_to_fp16)[name = tensor("op_2396_cast_fp16")]; + tensor var_2397 = const()[name = tensor("op_2397"), val = tensor([1, 20, 64, -1])]; + tensor var_2398_cast_fp16 = reshape(shape = var_2397, x = key_41_cast_fp16)[name = tensor("op_2398_cast_fp16")]; + tensor mh_w_61_transpose_x_0 = const()[name = tensor("mh_w_61_transpose_x_0"), val = tensor(true)]; + tensor mh_w_61_transpose_y_0 = const()[name = tensor("mh_w_61_transpose_y_0"), val = tensor(false)]; + tensor mh_w_61_cast_fp16 = matmul(transpose_x = mh_w_61_transpose_x_0, transpose_y = mh_w_61_transpose_y_0, x = var_2396_cast_fp16, y = var_2398_cast_fp16)[name = tensor("mh_w_61_cast_fp16")]; + tensor mh_w_63_cast_fp16 = add(x = mh_w_61_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_63_cast_fp16")]; + tensor var_2406_cast_fp16 = softmax(axis = var_2320, x = mh_w_63_cast_fp16)[name = tensor("op_2406_cast_fp16")]; + tensor var_2407 = const()[name = tensor("op_2407"), val = tensor([1, 20, 64, -1])]; + tensor var_2408_cast_fp16 = reshape(shape = var_2407, x = value_41_cast_fp16)[name = tensor("op_2408_cast_fp16")]; + tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; + tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(true)]; + tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = var_2408_cast_fp16, y = var_2406_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_2411 = const()[name = tensor("op_2411"), val = tensor([1, 1280, 1, -1])]; + tensor input_101_cast_fp16 = reshape(shape = var_2411, x = attn_41_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor var_2415 = const()[name = tensor("op_2415"), val = tensor([1, 1])]; + tensor var_2417 = const()[name = tensor("op_2417"), val = tensor([1, 1])]; + tensor obj_147_pad_type_0 = const()[name = tensor("obj_147_pad_type_0"), val = tensor("custom")]; + tensor obj_147_pad_0 = const()[name = tensor("obj_147_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(668508928)))]; + tensor layers_10_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(671785792)))]; + tensor obj_147_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_bias_to_fp16, dilations = var_2417, groups = var_2327, pad = obj_147_pad_0, pad_type = obj_147_pad_type_0, strides = var_2415, weight = layers_10_self_attn_o_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("obj_147_cast_fp16")]; + tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_147_cast_fp16)[name = tensor("inputs_63_cast_fp16")]; + tensor var_2427 = const()[name = tensor("op_2427"), val = tensor([1])]; + tensor channels_mean_63_cast_fp16 = reduce_mean(axes = var_2427, keep_dims = var_2328, x = inputs_63_cast_fp16)[name = tensor("channels_mean_63_cast_fp16")]; + tensor zero_mean_63_cast_fp16 = sub(x = inputs_63_cast_fp16, y = channels_mean_63_cast_fp16)[name = tensor("zero_mean_63_cast_fp16")]; + tensor zero_mean_sq_63_cast_fp16 = mul(x = zero_mean_63_cast_fp16, y = zero_mean_63_cast_fp16)[name = tensor("zero_mean_sq_63_cast_fp16")]; + tensor var_2431 = const()[name = tensor("op_2431"), val = tensor([1])]; + tensor var_2432_cast_fp16 = reduce_mean(axes = var_2431, keep_dims = var_2328, x = zero_mean_sq_63_cast_fp16)[name = tensor("op_2432_cast_fp16")]; + tensor var_2433_to_fp16 = const()[name = tensor("op_2433_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2434_cast_fp16 = add(x = var_2432_cast_fp16, y = var_2433_to_fp16)[name = tensor("op_2434_cast_fp16")]; + tensor denom_63_epsilon_0_to_fp16 = const()[name = tensor("denom_63_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_63_cast_fp16 = rsqrt(epsilon = denom_63_epsilon_0_to_fp16, x = var_2434_cast_fp16)[name = tensor("denom_63_cast_fp16")]; + tensor out_63_cast_fp16 = mul(x = zero_mean_63_cast_fp16, y = denom_63_cast_fp16)[name = tensor("out_63_cast_fp16")]; + tensor obj_149_gamma_0_to_fp16 = const()[name = tensor("obj_149_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(671788416)))]; + tensor obj_149_beta_0_to_fp16 = const()[name = tensor("obj_149_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(671791040)))]; + tensor obj_149_epsilon_0_to_fp16 = const()[name = tensor("obj_149_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_149_cast_fp16 = batch_norm(beta = obj_149_beta_0_to_fp16, epsilon = obj_149_epsilon_0_to_fp16, gamma = obj_149_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_63_cast_fp16)[name = tensor("obj_149_cast_fp16")]; + tensor var_2449 = const()[name = tensor("op_2449"), val = tensor([1, 1])]; + tensor var_2451 = const()[name = tensor("op_2451"), val = tensor([1, 1])]; + tensor query_43_pad_type_0 = const()[name = tensor("query_43_pad_type_0"), val = tensor("custom")]; + tensor query_43_pad_0 = const()[name = tensor("query_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(671793664)))]; + tensor layers_10_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(675070528)))]; + tensor query_43_cast_fp16 = conv(bias = layers_10_encoder_attn_q_proj_bias_to_fp16, dilations = var_2451, groups = var_2327, pad = query_43_pad_0, pad_type = query_43_pad_type_0, strides = var_2449, weight = layers_10_encoder_attn_q_proj_weight_to_fp16, x = obj_149_cast_fp16)[name = tensor("query_43_cast_fp16")]; + tensor var_2455 = const()[name = tensor("op_2455"), val = tensor([1, 1])]; + tensor var_2457 = const()[name = tensor("op_2457"), val = tensor([1, 1])]; + tensor key_43_pad_type_0 = const()[name = tensor("key_43_pad_type_0"), val = tensor("custom")]; + tensor key_43_pad_0 = const()[name = tensor("key_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(675073152)))]; + tensor key_43_cast_fp16 = conv(dilations = var_2457, groups = var_2327, pad = key_43_pad_0, pad_type = key_43_pad_type_0, strides = var_2455, weight = layers_10_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_43_cast_fp16")]; + tensor var_2462 = const()[name = tensor("op_2462"), val = tensor([1, 1])]; + tensor var_2464 = const()[name = tensor("op_2464"), val = tensor([1, 1])]; + tensor value_43_pad_type_0 = const()[name = tensor("value_43_pad_type_0"), val = tensor("custom")]; + tensor value_43_pad_0 = const()[name = tensor("value_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(678350016)))]; + tensor layers_10_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(681626880)))]; + tensor value_43_cast_fp16 = conv(bias = layers_10_encoder_attn_v_proj_bias_to_fp16, dilations = var_2464, groups = var_2327, pad = value_43_pad_0, pad_type = value_43_pad_type_0, strides = var_2462, weight = layers_10_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_43_cast_fp16")]; + tensor var_2468 = const()[name = tensor("op_2468"), val = tensor([1, 20, 64, -1])]; + tensor var_2469_cast_fp16 = reshape(shape = var_2468, x = query_43_cast_fp16)[name = tensor("op_2469_cast_fp16")]; + tensor var_2470_to_fp16 = const()[name = tensor("op_2470_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2471_cast_fp16 = mul(x = var_2469_cast_fp16, y = var_2470_to_fp16)[name = tensor("op_2471_cast_fp16")]; + tensor var_2472 = const()[name = tensor("op_2472"), val = tensor([1, 20, 64, -1])]; + tensor var_2473_cast_fp16 = reshape(shape = var_2472, x = key_43_cast_fp16)[name = tensor("op_2473_cast_fp16")]; + tensor mh_w_65_transpose_x_0 = const()[name = tensor("mh_w_65_transpose_x_0"), val = tensor(true)]; + tensor mh_w_65_transpose_y_0 = const()[name = tensor("mh_w_65_transpose_y_0"), val = tensor(false)]; + tensor mh_w_65_cast_fp16 = matmul(transpose_x = mh_w_65_transpose_x_0, transpose_y = mh_w_65_transpose_y_0, x = var_2471_cast_fp16, y = var_2473_cast_fp16)[name = tensor("mh_w_65_cast_fp16")]; + tensor obj_153_cast_fp16 = softmax(axis = var_2320, x = mh_w_65_cast_fp16)[name = tensor("obj_153_cast_fp16")]; + tensor var_2477 = const()[name = tensor("op_2477"), val = tensor([1, 20, 64, -1])]; + tensor var_2478_cast_fp16 = reshape(shape = var_2477, x = value_43_cast_fp16)[name = tensor("op_2478_cast_fp16")]; + tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; + tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(true)]; + tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = var_2478_cast_fp16, y = obj_153_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_2481 = const()[name = tensor("op_2481"), val = tensor([1, 1280, 1, -1])]; + tensor input_103_cast_fp16 = reshape(shape = var_2481, x = attn_43_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor var_2485 = const()[name = tensor("op_2485"), val = tensor([1, 1])]; + tensor var_2487 = const()[name = tensor("op_2487"), val = tensor([1, 1])]; + tensor obj_151_pad_type_0 = const()[name = tensor("obj_151_pad_type_0"), val = tensor("custom")]; + tensor obj_151_pad_0 = const()[name = tensor("obj_151_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(681629504)))]; + tensor layers_10_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(684906368)))]; + tensor obj_151_cast_fp16 = conv(bias = layers_10_encoder_attn_o_proj_bias_to_fp16, dilations = var_2487, groups = var_2327, pad = obj_151_pad_0, pad_type = obj_151_pad_type_0, strides = var_2485, weight = layers_10_encoder_attn_o_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("obj_151_cast_fp16")]; + tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = obj_151_cast_fp16)[name = tensor("inputs_65_cast_fp16")]; + tensor var_2496 = const()[name = tensor("op_2496"), val = tensor([1])]; + tensor channels_mean_65_cast_fp16 = reduce_mean(axes = var_2496, keep_dims = var_2328, x = inputs_65_cast_fp16)[name = tensor("channels_mean_65_cast_fp16")]; + tensor zero_mean_65_cast_fp16 = sub(x = inputs_65_cast_fp16, y = channels_mean_65_cast_fp16)[name = tensor("zero_mean_65_cast_fp16")]; + tensor zero_mean_sq_65_cast_fp16 = mul(x = zero_mean_65_cast_fp16, y = zero_mean_65_cast_fp16)[name = tensor("zero_mean_sq_65_cast_fp16")]; + tensor var_2500 = const()[name = tensor("op_2500"), val = tensor([1])]; + tensor var_2501_cast_fp16 = reduce_mean(axes = var_2500, keep_dims = var_2328, x = zero_mean_sq_65_cast_fp16)[name = tensor("op_2501_cast_fp16")]; + tensor var_2502_to_fp16 = const()[name = tensor("op_2502_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2503_cast_fp16 = add(x = var_2501_cast_fp16, y = var_2502_to_fp16)[name = tensor("op_2503_cast_fp16")]; + tensor denom_65_epsilon_0_to_fp16 = const()[name = tensor("denom_65_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_65_cast_fp16 = rsqrt(epsilon = denom_65_epsilon_0_to_fp16, x = var_2503_cast_fp16)[name = tensor("denom_65_cast_fp16")]; + tensor out_65_cast_fp16 = mul(x = zero_mean_65_cast_fp16, y = denom_65_cast_fp16)[name = tensor("out_65_cast_fp16")]; + tensor input_105_gamma_0_to_fp16 = const()[name = tensor("input_105_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(684908992)))]; + tensor input_105_beta_0_to_fp16 = const()[name = tensor("input_105_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(684911616)))]; + tensor input_105_epsilon_0_to_fp16 = const()[name = tensor("input_105_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_105_cast_fp16 = batch_norm(beta = input_105_beta_0_to_fp16, epsilon = input_105_epsilon_0_to_fp16, gamma = input_105_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_65_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor var_2514 = const()[name = tensor("op_2514"), val = tensor([1, 1])]; + tensor var_2516 = const()[name = tensor("op_2516"), val = tensor([1, 1])]; + tensor input_107_pad_type_0 = const()[name = tensor("input_107_pad_type_0"), val = tensor("custom")]; + tensor input_107_pad_0 = const()[name = tensor("input_107_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_fc1_weight_to_fp16 = const()[name = tensor("layers_10_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(684914240)))]; + tensor layers_10_fc1_bias_to_fp16 = const()[name = tensor("layers_10_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(698021504)))]; + tensor input_107_cast_fp16 = conv(bias = layers_10_fc1_bias_to_fp16, dilations = var_2516, groups = var_2327, pad = input_107_pad_0, pad_type = input_107_pad_type_0, strides = var_2514, weight = layers_10_fc1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_mode_0 = const()[name = tensor("input_109_mode_0"), val = tensor("EXACT")]; + tensor input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor var_2522 = const()[name = tensor("op_2522"), val = tensor([1, 1])]; + tensor var_2524 = const()[name = tensor("op_2524"), val = tensor([1, 1])]; + tensor hidden_states_23_pad_type_0 = const()[name = tensor("hidden_states_23_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_23_pad_0 = const()[name = tensor("hidden_states_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_fc2_weight_to_fp16 = const()[name = tensor("layers_10_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(698031808)))]; + tensor layers_10_fc2_bias_to_fp16 = const()[name = tensor("layers_10_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(711139072)))]; + tensor hidden_states_23_cast_fp16 = conv(bias = layers_10_fc2_bias_to_fp16, dilations = var_2524, groups = var_2327, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = var_2522, weight = layers_10_fc2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("hidden_states_23_cast_fp16")]; + tensor inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = hidden_states_23_cast_fp16)[name = tensor("inputs_67_cast_fp16")]; + tensor var_2538 = const()[name = tensor("op_2538"), val = tensor(3)]; + tensor var_2545 = const()[name = tensor("op_2545"), val = tensor(1)]; + tensor var_2546 = const()[name = tensor("op_2546"), val = tensor(true)]; + tensor var_2558 = const()[name = tensor("op_2558"), val = tensor([1])]; + tensor channels_mean_67_cast_fp16 = reduce_mean(axes = var_2558, keep_dims = var_2546, x = inputs_67_cast_fp16)[name = tensor("channels_mean_67_cast_fp16")]; + tensor zero_mean_67_cast_fp16 = sub(x = inputs_67_cast_fp16, y = channels_mean_67_cast_fp16)[name = tensor("zero_mean_67_cast_fp16")]; + tensor zero_mean_sq_67_cast_fp16 = mul(x = zero_mean_67_cast_fp16, y = zero_mean_67_cast_fp16)[name = tensor("zero_mean_sq_67_cast_fp16")]; + tensor var_2562 = const()[name = tensor("op_2562"), val = tensor([1])]; + tensor var_2563_cast_fp16 = reduce_mean(axes = var_2562, keep_dims = var_2546, x = zero_mean_sq_67_cast_fp16)[name = tensor("op_2563_cast_fp16")]; + tensor var_2564_to_fp16 = const()[name = tensor("op_2564_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2565_cast_fp16 = add(x = var_2563_cast_fp16, y = var_2564_to_fp16)[name = tensor("op_2565_cast_fp16")]; + tensor denom_67_epsilon_0_to_fp16 = const()[name = tensor("denom_67_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_67_cast_fp16 = rsqrt(epsilon = denom_67_epsilon_0_to_fp16, x = var_2565_cast_fp16)[name = tensor("denom_67_cast_fp16")]; + tensor out_67_cast_fp16 = mul(x = zero_mean_67_cast_fp16, y = denom_67_cast_fp16)[name = tensor("out_67_cast_fp16")]; + tensor obj_155_gamma_0_to_fp16 = const()[name = tensor("obj_155_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(711141696)))]; + tensor obj_155_beta_0_to_fp16 = const()[name = tensor("obj_155_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(711144320)))]; + tensor obj_155_epsilon_0_to_fp16 = const()[name = tensor("obj_155_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_155_cast_fp16 = batch_norm(beta = obj_155_beta_0_to_fp16, epsilon = obj_155_epsilon_0_to_fp16, gamma = obj_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_67_cast_fp16)[name = tensor("obj_155_cast_fp16")]; + tensor var_2580 = const()[name = tensor("op_2580"), val = tensor([1, 1])]; + tensor var_2582 = const()[name = tensor("op_2582"), val = tensor([1, 1])]; + tensor query_45_pad_type_0 = const()[name = tensor("query_45_pad_type_0"), val = tensor("custom")]; + tensor query_45_pad_0 = const()[name = tensor("query_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(711146944)))]; + tensor layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(714423808)))]; + tensor query_45_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_bias_to_fp16, dilations = var_2582, groups = var_2545, pad = query_45_pad_0, pad_type = query_45_pad_type_0, strides = var_2580, weight = layers_11_self_attn_q_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("query_45_cast_fp16")]; + tensor var_2586 = const()[name = tensor("op_2586"), val = tensor([1, 1])]; + tensor var_2588 = const()[name = tensor("op_2588"), val = tensor([1, 1])]; + tensor current_key_23_pad_type_0 = const()[name = tensor("current_key_23_pad_type_0"), val = tensor("custom")]; + tensor current_key_23_pad_0 = const()[name = tensor("current_key_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(714426432)))]; + tensor current_key_23_cast_fp16 = conv(dilations = var_2588, groups = var_2545, pad = current_key_23_pad_0, pad_type = current_key_23_pad_type_0, strides = var_2586, weight = layers_11_self_attn_k_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("current_key_23_cast_fp16")]; + tensor var_2593 = const()[name = tensor("op_2593"), val = tensor([1, 1])]; + tensor var_2595 = const()[name = tensor("op_2595"), val = tensor([1, 1])]; + tensor current_value_23_pad_type_0 = const()[name = tensor("current_value_23_pad_type_0"), val = tensor("custom")]; + tensor current_value_23_pad_0 = const()[name = tensor("current_value_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(717703296)))]; + tensor layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(720980160)))]; + tensor current_value_23_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_bias_to_fp16, dilations = var_2595, groups = var_2545, pad = current_value_23_pad_0, pad_type = current_value_23_pad_type_0, strides = var_2593, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("current_value_23_cast_fp16")]; + tensor var_2602_cast_fp16 = mul(x = current_key_23_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_2602_cast_fp16")]; + tensor var_2604_cast_fp16 = mul(x = var_103_cast_fp16_11, y = var_241_cast_fp16)[name = tensor("op_2604_cast_fp16")]; + tensor key_45_cast_fp16 = add(x = var_2602_cast_fp16, y = var_2604_cast_fp16)[name = tensor("key_45_cast_fp16")]; + tensor var_2606_cast_fp16 = mul(x = current_value_23_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_2606_cast_fp16")]; + tensor var_2608_cast_fp16 = mul(x = var_138_cast_fp16_11, y = var_241_cast_fp16)[name = tensor("op_2608_cast_fp16")]; + tensor value_45_cast_fp16 = add(x = var_2606_cast_fp16, y = var_2608_cast_fp16)[name = tensor("value_45_cast_fp16")]; + tensor var_2611 = const()[name = tensor("op_2611"), val = tensor([1, 20, 64, -1])]; + tensor var_2612_cast_fp16 = reshape(shape = var_2611, x = query_45_cast_fp16)[name = tensor("op_2612_cast_fp16")]; + tensor var_2613_to_fp16 = const()[name = tensor("op_2613_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2614_cast_fp16 = mul(x = var_2612_cast_fp16, y = var_2613_to_fp16)[name = tensor("op_2614_cast_fp16")]; + tensor var_2615 = const()[name = tensor("op_2615"), val = tensor([1, 20, 64, -1])]; + tensor var_2616_cast_fp16 = reshape(shape = var_2615, x = key_45_cast_fp16)[name = tensor("op_2616_cast_fp16")]; + tensor mh_w_67_transpose_x_0 = const()[name = tensor("mh_w_67_transpose_x_0"), val = tensor(true)]; + tensor mh_w_67_transpose_y_0 = const()[name = tensor("mh_w_67_transpose_y_0"), val = tensor(false)]; + tensor mh_w_67_cast_fp16 = matmul(transpose_x = mh_w_67_transpose_x_0, transpose_y = mh_w_67_transpose_y_0, x = var_2614_cast_fp16, y = var_2616_cast_fp16)[name = tensor("mh_w_67_cast_fp16")]; + tensor mh_w_69_cast_fp16 = add(x = mh_w_67_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_69_cast_fp16")]; + tensor var_2624_cast_fp16 = softmax(axis = var_2538, x = mh_w_69_cast_fp16)[name = tensor("op_2624_cast_fp16")]; + tensor var_2625 = const()[name = tensor("op_2625"), val = tensor([1, 20, 64, -1])]; + tensor var_2626_cast_fp16 = reshape(shape = var_2625, x = value_45_cast_fp16)[name = tensor("op_2626_cast_fp16")]; + tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; + tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(true)]; + tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = var_2626_cast_fp16, y = var_2624_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_2629 = const()[name = tensor("op_2629"), val = tensor([1, 1280, 1, -1])]; + tensor input_111_cast_fp16 = reshape(shape = var_2629, x = attn_45_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor var_2633 = const()[name = tensor("op_2633"), val = tensor([1, 1])]; + tensor var_2635 = const()[name = tensor("op_2635"), val = tensor([1, 1])]; + tensor obj_161_pad_type_0 = const()[name = tensor("obj_161_pad_type_0"), val = tensor("custom")]; + tensor obj_161_pad_0 = const()[name = tensor("obj_161_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(720982784)))]; + tensor layers_11_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(724259648)))]; + tensor obj_161_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_bias_to_fp16, dilations = var_2635, groups = var_2545, pad = obj_161_pad_0, pad_type = obj_161_pad_type_0, strides = var_2633, weight = layers_11_self_attn_o_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("obj_161_cast_fp16")]; + tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = obj_161_cast_fp16)[name = tensor("inputs_69_cast_fp16")]; + tensor var_2645 = const()[name = tensor("op_2645"), val = tensor([1])]; + tensor channels_mean_69_cast_fp16 = reduce_mean(axes = var_2645, keep_dims = var_2546, x = inputs_69_cast_fp16)[name = tensor("channels_mean_69_cast_fp16")]; + tensor zero_mean_69_cast_fp16 = sub(x = inputs_69_cast_fp16, y = channels_mean_69_cast_fp16)[name = tensor("zero_mean_69_cast_fp16")]; + tensor zero_mean_sq_69_cast_fp16 = mul(x = zero_mean_69_cast_fp16, y = zero_mean_69_cast_fp16)[name = tensor("zero_mean_sq_69_cast_fp16")]; + tensor var_2649 = const()[name = tensor("op_2649"), val = tensor([1])]; + tensor var_2650_cast_fp16 = reduce_mean(axes = var_2649, keep_dims = var_2546, x = zero_mean_sq_69_cast_fp16)[name = tensor("op_2650_cast_fp16")]; + tensor var_2651_to_fp16 = const()[name = tensor("op_2651_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2652_cast_fp16 = add(x = var_2650_cast_fp16, y = var_2651_to_fp16)[name = tensor("op_2652_cast_fp16")]; + tensor denom_69_epsilon_0_to_fp16 = const()[name = tensor("denom_69_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_69_cast_fp16 = rsqrt(epsilon = denom_69_epsilon_0_to_fp16, x = var_2652_cast_fp16)[name = tensor("denom_69_cast_fp16")]; + tensor out_69_cast_fp16 = mul(x = zero_mean_69_cast_fp16, y = denom_69_cast_fp16)[name = tensor("out_69_cast_fp16")]; + tensor obj_163_gamma_0_to_fp16 = const()[name = tensor("obj_163_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(724262272)))]; + tensor obj_163_beta_0_to_fp16 = const()[name = tensor("obj_163_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(724264896)))]; + tensor obj_163_epsilon_0_to_fp16 = const()[name = tensor("obj_163_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_163_cast_fp16 = batch_norm(beta = obj_163_beta_0_to_fp16, epsilon = obj_163_epsilon_0_to_fp16, gamma = obj_163_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_69_cast_fp16)[name = tensor("obj_163_cast_fp16")]; + tensor var_2667 = const()[name = tensor("op_2667"), val = tensor([1, 1])]; + tensor var_2669 = const()[name = tensor("op_2669"), val = tensor([1, 1])]; + tensor query_47_pad_type_0 = const()[name = tensor("query_47_pad_type_0"), val = tensor("custom")]; + tensor query_47_pad_0 = const()[name = tensor("query_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(724267520)))]; + tensor layers_11_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(727544384)))]; + tensor query_47_cast_fp16 = conv(bias = layers_11_encoder_attn_q_proj_bias_to_fp16, dilations = var_2669, groups = var_2545, pad = query_47_pad_0, pad_type = query_47_pad_type_0, strides = var_2667, weight = layers_11_encoder_attn_q_proj_weight_to_fp16, x = obj_163_cast_fp16)[name = tensor("query_47_cast_fp16")]; + tensor var_2673 = const()[name = tensor("op_2673"), val = tensor([1, 1])]; + tensor var_2675 = const()[name = tensor("op_2675"), val = tensor([1, 1])]; + tensor key_47_pad_type_0 = const()[name = tensor("key_47_pad_type_0"), val = tensor("custom")]; + tensor key_47_pad_0 = const()[name = tensor("key_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(727547008)))]; + tensor key_47_cast_fp16 = conv(dilations = var_2675, groups = var_2545, pad = key_47_pad_0, pad_type = key_47_pad_type_0, strides = var_2673, weight = layers_11_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_47_cast_fp16")]; + tensor var_2680 = const()[name = tensor("op_2680"), val = tensor([1, 1])]; + tensor var_2682 = const()[name = tensor("op_2682"), val = tensor([1, 1])]; + tensor value_47_pad_type_0 = const()[name = tensor("value_47_pad_type_0"), val = tensor("custom")]; + tensor value_47_pad_0 = const()[name = tensor("value_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(730823872)))]; + tensor layers_11_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(734100736)))]; + tensor value_47_cast_fp16 = conv(bias = layers_11_encoder_attn_v_proj_bias_to_fp16, dilations = var_2682, groups = var_2545, pad = value_47_pad_0, pad_type = value_47_pad_type_0, strides = var_2680, weight = layers_11_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_47_cast_fp16")]; + tensor var_2686 = const()[name = tensor("op_2686"), val = tensor([1, 20, 64, -1])]; + tensor var_2687_cast_fp16 = reshape(shape = var_2686, x = query_47_cast_fp16)[name = tensor("op_2687_cast_fp16")]; + tensor var_2688_to_fp16 = const()[name = tensor("op_2688_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2689_cast_fp16 = mul(x = var_2687_cast_fp16, y = var_2688_to_fp16)[name = tensor("op_2689_cast_fp16")]; + tensor var_2690 = const()[name = tensor("op_2690"), val = tensor([1, 20, 64, -1])]; + tensor var_2691_cast_fp16 = reshape(shape = var_2690, x = key_47_cast_fp16)[name = tensor("op_2691_cast_fp16")]; + tensor mh_w_71_transpose_x_0 = const()[name = tensor("mh_w_71_transpose_x_0"), val = tensor(true)]; + tensor mh_w_71_transpose_y_0 = const()[name = tensor("mh_w_71_transpose_y_0"), val = tensor(false)]; + tensor mh_w_71_cast_fp16 = matmul(transpose_x = mh_w_71_transpose_x_0, transpose_y = mh_w_71_transpose_y_0, x = var_2689_cast_fp16, y = var_2691_cast_fp16)[name = tensor("mh_w_71_cast_fp16")]; + tensor obj_167_cast_fp16 = softmax(axis = var_2538, x = mh_w_71_cast_fp16)[name = tensor("obj_167_cast_fp16")]; + tensor var_2695 = const()[name = tensor("op_2695"), val = tensor([1, 20, 64, -1])]; + tensor var_2696_cast_fp16 = reshape(shape = var_2695, x = value_47_cast_fp16)[name = tensor("op_2696_cast_fp16")]; + tensor attn_47_transpose_x_0 = const()[name = tensor("attn_47_transpose_x_0"), val = tensor(false)]; + tensor attn_47_transpose_y_0 = const()[name = tensor("attn_47_transpose_y_0"), val = tensor(true)]; + tensor attn_47_cast_fp16 = matmul(transpose_x = attn_47_transpose_x_0, transpose_y = attn_47_transpose_y_0, x = var_2696_cast_fp16, y = obj_167_cast_fp16)[name = tensor("attn_47_cast_fp16")]; + tensor var_2699 = const()[name = tensor("op_2699"), val = tensor([1, 1280, 1, -1])]; + tensor input_113_cast_fp16 = reshape(shape = var_2699, x = attn_47_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor var_2703 = const()[name = tensor("op_2703"), val = tensor([1, 1])]; + tensor var_2705 = const()[name = tensor("op_2705"), val = tensor([1, 1])]; + tensor obj_165_pad_type_0 = const()[name = tensor("obj_165_pad_type_0"), val = tensor("custom")]; + tensor obj_165_pad_0 = const()[name = tensor("obj_165_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(734103360)))]; + tensor layers_11_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(737380224)))]; + tensor obj_165_cast_fp16 = conv(bias = layers_11_encoder_attn_o_proj_bias_to_fp16, dilations = var_2705, groups = var_2545, pad = obj_165_pad_0, pad_type = obj_165_pad_type_0, strides = var_2703, weight = layers_11_encoder_attn_o_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("obj_165_cast_fp16")]; + tensor inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_165_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; + tensor var_2711 = const()[name = tensor("op_2711"), val = tensor([1])]; + tensor channels_mean_71_cast_fp16 = reduce_mean(axes = var_2711, keep_dims = var_2546, x = inputs_71_cast_fp16)[name = tensor("channels_mean_71_cast_fp16")]; + tensor zero_mean_71_cast_fp16 = sub(x = inputs_71_cast_fp16, y = channels_mean_71_cast_fp16)[name = tensor("zero_mean_71_cast_fp16")]; + tensor zero_mean_sq_71_cast_fp16 = mul(x = zero_mean_71_cast_fp16, y = zero_mean_71_cast_fp16)[name = tensor("zero_mean_sq_71_cast_fp16")]; + tensor var_2715 = const()[name = tensor("op_2715"), val = tensor([1])]; + tensor var_2716_cast_fp16 = reduce_mean(axes = var_2715, keep_dims = var_2546, x = zero_mean_sq_71_cast_fp16)[name = tensor("op_2716_cast_fp16")]; + tensor var_2717_to_fp16 = const()[name = tensor("op_2717_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2718_cast_fp16 = add(x = var_2716_cast_fp16, y = var_2717_to_fp16)[name = tensor("op_2718_cast_fp16")]; + tensor denom_71_epsilon_0_to_fp16 = const()[name = tensor("denom_71_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_71_cast_fp16 = rsqrt(epsilon = denom_71_epsilon_0_to_fp16, x = var_2718_cast_fp16)[name = tensor("denom_71_cast_fp16")]; + tensor out_71_cast_fp16 = mul(x = zero_mean_71_cast_fp16, y = denom_71_cast_fp16)[name = tensor("out_71_cast_fp16")]; + tensor input_115_gamma_0_to_fp16 = const()[name = tensor("input_115_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(737382848)))]; + tensor input_115_beta_0_to_fp16 = const()[name = tensor("input_115_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(737385472)))]; + tensor input_115_epsilon_0_to_fp16 = const()[name = tensor("input_115_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_115_cast_fp16 = batch_norm(beta = input_115_beta_0_to_fp16, epsilon = input_115_epsilon_0_to_fp16, gamma = input_115_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_71_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor var_2729 = const()[name = tensor("op_2729"), val = tensor([1, 1])]; + tensor var_2731 = const()[name = tensor("op_2731"), val = tensor([1, 1])]; + tensor input_117_pad_type_0 = const()[name = tensor("input_117_pad_type_0"), val = tensor("custom")]; + tensor input_117_pad_0 = const()[name = tensor("input_117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_fc1_weight_to_fp16 = const()[name = tensor("layers_11_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(737388096)))]; + tensor layers_11_fc1_bias_to_fp16 = const()[name = tensor("layers_11_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(750495360)))]; + tensor input_117_cast_fp16 = conv(bias = layers_11_fc1_bias_to_fp16, dilations = var_2731, groups = var_2545, pad = input_117_pad_0, pad_type = input_117_pad_type_0, strides = var_2729, weight = layers_11_fc1_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_mode_0 = const()[name = tensor("input_119_mode_0"), val = tensor("EXACT")]; + tensor input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor var_2737 = const()[name = tensor("op_2737"), val = tensor([1, 1])]; + tensor var_2739 = const()[name = tensor("op_2739"), val = tensor([1, 1])]; + tensor hidden_states_25_pad_type_0 = const()[name = tensor("hidden_states_25_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_25_pad_0 = const()[name = tensor("hidden_states_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_fc2_weight_to_fp16 = const()[name = tensor("layers_11_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(750505664)))]; + tensor layers_11_fc2_bias_to_fp16 = const()[name = tensor("layers_11_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(763612928)))]; + tensor hidden_states_25_cast_fp16 = conv(bias = layers_11_fc2_bias_to_fp16, dilations = var_2739, groups = var_2545, pad = hidden_states_25_pad_0, pad_type = hidden_states_25_pad_type_0, strides = var_2737, weight = layers_11_fc2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("hidden_states_25_cast_fp16")]; + tensor inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = hidden_states_25_cast_fp16)[name = tensor("inputs_73_cast_fp16")]; + tensor var_2752 = const()[name = tensor("op_2752"), val = tensor(3)]; + tensor var_2759 = const()[name = tensor("op_2759"), val = tensor(1)]; + tensor var_2760 = const()[name = tensor("op_2760"), val = tensor(true)]; + tensor var_2772 = const()[name = tensor("op_2772"), val = tensor([1])]; + tensor channels_mean_73_cast_fp16 = reduce_mean(axes = var_2772, keep_dims = var_2760, x = inputs_73_cast_fp16)[name = tensor("channels_mean_73_cast_fp16")]; + tensor zero_mean_73_cast_fp16 = sub(x = inputs_73_cast_fp16, y = channels_mean_73_cast_fp16)[name = tensor("zero_mean_73_cast_fp16")]; + tensor zero_mean_sq_73_cast_fp16 = mul(x = zero_mean_73_cast_fp16, y = zero_mean_73_cast_fp16)[name = tensor("zero_mean_sq_73_cast_fp16")]; + tensor var_2776 = const()[name = tensor("op_2776"), val = tensor([1])]; + tensor var_2777_cast_fp16 = reduce_mean(axes = var_2776, keep_dims = var_2760, x = zero_mean_sq_73_cast_fp16)[name = tensor("op_2777_cast_fp16")]; + tensor var_2778_to_fp16 = const()[name = tensor("op_2778_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2779_cast_fp16 = add(x = var_2777_cast_fp16, y = var_2778_to_fp16)[name = tensor("op_2779_cast_fp16")]; + tensor denom_73_epsilon_0_to_fp16 = const()[name = tensor("denom_73_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_73_cast_fp16 = rsqrt(epsilon = denom_73_epsilon_0_to_fp16, x = var_2779_cast_fp16)[name = tensor("denom_73_cast_fp16")]; + tensor out_73_cast_fp16 = mul(x = zero_mean_73_cast_fp16, y = denom_73_cast_fp16)[name = tensor("out_73_cast_fp16")]; + tensor obj_169_gamma_0_to_fp16 = const()[name = tensor("obj_169_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(763615552)))]; + tensor obj_169_beta_0_to_fp16 = const()[name = tensor("obj_169_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(763618176)))]; + tensor obj_169_epsilon_0_to_fp16 = const()[name = tensor("obj_169_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_169_cast_fp16 = batch_norm(beta = obj_169_beta_0_to_fp16, epsilon = obj_169_epsilon_0_to_fp16, gamma = obj_169_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_73_cast_fp16)[name = tensor("obj_169_cast_fp16")]; + tensor var_2794 = const()[name = tensor("op_2794"), val = tensor([1, 1])]; + tensor var_2796 = const()[name = tensor("op_2796"), val = tensor([1, 1])]; + tensor query_49_pad_type_0 = const()[name = tensor("query_49_pad_type_0"), val = tensor("custom")]; + tensor query_49_pad_0 = const()[name = tensor("query_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(763620800)))]; + tensor layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(766897664)))]; + tensor query_49_cast_fp16 = conv(bias = layers_12_self_attn_q_proj_bias_to_fp16, dilations = var_2796, groups = var_2759, pad = query_49_pad_0, pad_type = query_49_pad_type_0, strides = var_2794, weight = layers_12_self_attn_q_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("query_49_cast_fp16")]; + tensor var_2800 = const()[name = tensor("op_2800"), val = tensor([1, 1])]; + tensor var_2802 = const()[name = tensor("op_2802"), val = tensor([1, 1])]; + tensor current_key_25_pad_type_0 = const()[name = tensor("current_key_25_pad_type_0"), val = tensor("custom")]; + tensor current_key_25_pad_0 = const()[name = tensor("current_key_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(766900288)))]; + tensor current_key_25_cast_fp16 = conv(dilations = var_2802, groups = var_2759, pad = current_key_25_pad_0, pad_type = current_key_25_pad_type_0, strides = var_2800, weight = layers_12_self_attn_k_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("current_key_25_cast_fp16")]; + tensor var_2807 = const()[name = tensor("op_2807"), val = tensor([1, 1])]; + tensor var_2809 = const()[name = tensor("op_2809"), val = tensor([1, 1])]; + tensor current_value_25_pad_type_0 = const()[name = tensor("current_value_25_pad_type_0"), val = tensor("custom")]; + tensor current_value_25_pad_0 = const()[name = tensor("current_value_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(770177152)))]; + tensor layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(773454016)))]; + tensor current_value_25_cast_fp16 = conv(bias = layers_12_self_attn_v_proj_bias_to_fp16, dilations = var_2809, groups = var_2759, pad = current_value_25_pad_0, pad_type = current_value_25_pad_type_0, strides = var_2807, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("current_value_25_cast_fp16")]; + tensor var_2816_cast_fp16 = mul(x = current_key_25_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_2816_cast_fp16")]; + tensor var_2818_cast_fp16 = mul(x = var_103_cast_fp16_12, y = var_241_cast_fp16)[name = tensor("op_2818_cast_fp16")]; + tensor key_49_cast_fp16 = add(x = var_2816_cast_fp16, y = var_2818_cast_fp16)[name = tensor("key_49_cast_fp16")]; + tensor var_2820_cast_fp16 = mul(x = current_value_25_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_2820_cast_fp16")]; + tensor var_2822_cast_fp16 = mul(x = var_138_cast_fp16_12, y = var_241_cast_fp16)[name = tensor("op_2822_cast_fp16")]; + tensor value_49_cast_fp16 = add(x = var_2820_cast_fp16, y = var_2822_cast_fp16)[name = tensor("value_49_cast_fp16")]; + tensor var_2825 = const()[name = tensor("op_2825"), val = tensor([1, 20, 64, -1])]; + tensor var_2826_cast_fp16 = reshape(shape = var_2825, x = query_49_cast_fp16)[name = tensor("op_2826_cast_fp16")]; + tensor var_2827_to_fp16 = const()[name = tensor("op_2827_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2828_cast_fp16 = mul(x = var_2826_cast_fp16, y = var_2827_to_fp16)[name = tensor("op_2828_cast_fp16")]; + tensor var_2829 = const()[name = tensor("op_2829"), val = tensor([1, 20, 64, -1])]; + tensor var_2830_cast_fp16 = reshape(shape = var_2829, x = key_49_cast_fp16)[name = tensor("op_2830_cast_fp16")]; + tensor mh_w_73_transpose_x_0 = const()[name = tensor("mh_w_73_transpose_x_0"), val = tensor(true)]; + tensor mh_w_73_transpose_y_0 = const()[name = tensor("mh_w_73_transpose_y_0"), val = tensor(false)]; + tensor mh_w_73_cast_fp16 = matmul(transpose_x = mh_w_73_transpose_x_0, transpose_y = mh_w_73_transpose_y_0, x = var_2828_cast_fp16, y = var_2830_cast_fp16)[name = tensor("mh_w_73_cast_fp16")]; + tensor mh_w_75_cast_fp16 = add(x = mh_w_73_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_75_cast_fp16")]; + tensor var_2838_cast_fp16 = softmax(axis = var_2752, x = mh_w_75_cast_fp16)[name = tensor("op_2838_cast_fp16")]; + tensor var_2839 = const()[name = tensor("op_2839"), val = tensor([1, 20, 64, -1])]; + tensor var_2840_cast_fp16 = reshape(shape = var_2839, x = value_49_cast_fp16)[name = tensor("op_2840_cast_fp16")]; + tensor attn_49_transpose_x_0 = const()[name = tensor("attn_49_transpose_x_0"), val = tensor(false)]; + tensor attn_49_transpose_y_0 = const()[name = tensor("attn_49_transpose_y_0"), val = tensor(true)]; + tensor attn_49_cast_fp16 = matmul(transpose_x = attn_49_transpose_x_0, transpose_y = attn_49_transpose_y_0, x = var_2840_cast_fp16, y = var_2838_cast_fp16)[name = tensor("attn_49_cast_fp16")]; + tensor var_2843 = const()[name = tensor("op_2843"), val = tensor([1, 1280, 1, -1])]; + tensor input_121_cast_fp16 = reshape(shape = var_2843, x = attn_49_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor var_2847 = const()[name = tensor("op_2847"), val = tensor([1, 1])]; + tensor var_2849 = const()[name = tensor("op_2849"), val = tensor([1, 1])]; + tensor obj_175_pad_type_0 = const()[name = tensor("obj_175_pad_type_0"), val = tensor("custom")]; + tensor obj_175_pad_0 = const()[name = tensor("obj_175_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(773456640)))]; + tensor layers_12_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(776733504)))]; + tensor obj_175_cast_fp16 = conv(bias = layers_12_self_attn_o_proj_bias_to_fp16, dilations = var_2849, groups = var_2759, pad = obj_175_pad_0, pad_type = obj_175_pad_type_0, strides = var_2847, weight = layers_12_self_attn_o_proj_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("obj_175_cast_fp16")]; + tensor inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = obj_175_cast_fp16)[name = tensor("inputs_75_cast_fp16")]; + tensor var_2859 = const()[name = tensor("op_2859"), val = tensor([1])]; + tensor channels_mean_75_cast_fp16 = reduce_mean(axes = var_2859, keep_dims = var_2760, x = inputs_75_cast_fp16)[name = tensor("channels_mean_75_cast_fp16")]; + tensor zero_mean_75_cast_fp16 = sub(x = inputs_75_cast_fp16, y = channels_mean_75_cast_fp16)[name = tensor("zero_mean_75_cast_fp16")]; + tensor zero_mean_sq_75_cast_fp16 = mul(x = zero_mean_75_cast_fp16, y = zero_mean_75_cast_fp16)[name = tensor("zero_mean_sq_75_cast_fp16")]; + tensor var_2863 = const()[name = tensor("op_2863"), val = tensor([1])]; + tensor var_2864_cast_fp16 = reduce_mean(axes = var_2863, keep_dims = var_2760, x = zero_mean_sq_75_cast_fp16)[name = tensor("op_2864_cast_fp16")]; + tensor var_2865_to_fp16 = const()[name = tensor("op_2865_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2866_cast_fp16 = add(x = var_2864_cast_fp16, y = var_2865_to_fp16)[name = tensor("op_2866_cast_fp16")]; + tensor denom_75_epsilon_0_to_fp16 = const()[name = tensor("denom_75_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_75_cast_fp16 = rsqrt(epsilon = denom_75_epsilon_0_to_fp16, x = var_2866_cast_fp16)[name = tensor("denom_75_cast_fp16")]; + tensor out_75_cast_fp16 = mul(x = zero_mean_75_cast_fp16, y = denom_75_cast_fp16)[name = tensor("out_75_cast_fp16")]; + tensor obj_177_gamma_0_to_fp16 = const()[name = tensor("obj_177_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(776736128)))]; + tensor obj_177_beta_0_to_fp16 = const()[name = tensor("obj_177_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(776738752)))]; + tensor obj_177_epsilon_0_to_fp16 = const()[name = tensor("obj_177_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_177_cast_fp16 = batch_norm(beta = obj_177_beta_0_to_fp16, epsilon = obj_177_epsilon_0_to_fp16, gamma = obj_177_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_75_cast_fp16)[name = tensor("obj_177_cast_fp16")]; + tensor var_2881 = const()[name = tensor("op_2881"), val = tensor([1, 1])]; + tensor var_2883 = const()[name = tensor("op_2883"), val = tensor([1, 1])]; + tensor query_51_pad_type_0 = const()[name = tensor("query_51_pad_type_0"), val = tensor("custom")]; + tensor query_51_pad_0 = const()[name = tensor("query_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(776741376)))]; + tensor layers_12_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(780018240)))]; + tensor query_51_cast_fp16 = conv(bias = layers_12_encoder_attn_q_proj_bias_to_fp16, dilations = var_2883, groups = var_2759, pad = query_51_pad_0, pad_type = query_51_pad_type_0, strides = var_2881, weight = layers_12_encoder_attn_q_proj_weight_to_fp16, x = obj_177_cast_fp16)[name = tensor("query_51_cast_fp16")]; + tensor var_2887 = const()[name = tensor("op_2887"), val = tensor([1, 1])]; + tensor var_2889 = const()[name = tensor("op_2889"), val = tensor([1, 1])]; + tensor key_51_pad_type_0 = const()[name = tensor("key_51_pad_type_0"), val = tensor("custom")]; + tensor key_51_pad_0 = const()[name = tensor("key_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(780020864)))]; + tensor key_51_cast_fp16 = conv(dilations = var_2889, groups = var_2759, pad = key_51_pad_0, pad_type = key_51_pad_type_0, strides = var_2887, weight = layers_12_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_51_cast_fp16")]; + tensor var_2894 = const()[name = tensor("op_2894"), val = tensor([1, 1])]; + tensor var_2896 = const()[name = tensor("op_2896"), val = tensor([1, 1])]; + tensor value_51_pad_type_0 = const()[name = tensor("value_51_pad_type_0"), val = tensor("custom")]; + tensor value_51_pad_0 = const()[name = tensor("value_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(783297728)))]; + tensor layers_12_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(786574592)))]; + tensor value_51_cast_fp16 = conv(bias = layers_12_encoder_attn_v_proj_bias_to_fp16, dilations = var_2896, groups = var_2759, pad = value_51_pad_0, pad_type = value_51_pad_type_0, strides = var_2894, weight = layers_12_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_51_cast_fp16")]; + tensor var_2900 = const()[name = tensor("op_2900"), val = tensor([1, 20, 64, -1])]; + tensor var_2901_cast_fp16 = reshape(shape = var_2900, x = query_51_cast_fp16)[name = tensor("op_2901_cast_fp16")]; + tensor var_2902_to_fp16 = const()[name = tensor("op_2902_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2903_cast_fp16 = mul(x = var_2901_cast_fp16, y = var_2902_to_fp16)[name = tensor("op_2903_cast_fp16")]; + tensor var_2904 = const()[name = tensor("op_2904"), val = tensor([1, 20, 64, -1])]; + tensor var_2905_cast_fp16 = reshape(shape = var_2904, x = key_51_cast_fp16)[name = tensor("op_2905_cast_fp16")]; + tensor mh_w_77_transpose_x_0 = const()[name = tensor("mh_w_77_transpose_x_0"), val = tensor(true)]; + tensor mh_w_77_transpose_y_0 = const()[name = tensor("mh_w_77_transpose_y_0"), val = tensor(false)]; + tensor mh_w_77_cast_fp16 = matmul(transpose_x = mh_w_77_transpose_x_0, transpose_y = mh_w_77_transpose_y_0, x = var_2903_cast_fp16, y = var_2905_cast_fp16)[name = tensor("mh_w_77_cast_fp16")]; + tensor obj_181_cast_fp16 = softmax(axis = var_2752, x = mh_w_77_cast_fp16)[name = tensor("obj_181_cast_fp16")]; + tensor var_2909 = const()[name = tensor("op_2909"), val = tensor([1, 20, 64, -1])]; + tensor var_2910_cast_fp16 = reshape(shape = var_2909, x = value_51_cast_fp16)[name = tensor("op_2910_cast_fp16")]; + tensor attn_51_transpose_x_0 = const()[name = tensor("attn_51_transpose_x_0"), val = tensor(false)]; + tensor attn_51_transpose_y_0 = const()[name = tensor("attn_51_transpose_y_0"), val = tensor(true)]; + tensor attn_51_cast_fp16 = matmul(transpose_x = attn_51_transpose_x_0, transpose_y = attn_51_transpose_y_0, x = var_2910_cast_fp16, y = obj_181_cast_fp16)[name = tensor("attn_51_cast_fp16")]; + tensor var_2913 = const()[name = tensor("op_2913"), val = tensor([1, 1280, 1, -1])]; + tensor input_123_cast_fp16 = reshape(shape = var_2913, x = attn_51_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor var_2917 = const()[name = tensor("op_2917"), val = tensor([1, 1])]; + tensor var_2919 = const()[name = tensor("op_2919"), val = tensor([1, 1])]; + tensor obj_179_pad_type_0 = const()[name = tensor("obj_179_pad_type_0"), val = tensor("custom")]; + tensor obj_179_pad_0 = const()[name = tensor("obj_179_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(786577216)))]; + tensor layers_12_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(789854080)))]; + tensor obj_179_cast_fp16 = conv(bias = layers_12_encoder_attn_o_proj_bias_to_fp16, dilations = var_2919, groups = var_2759, pad = obj_179_pad_0, pad_type = obj_179_pad_type_0, strides = var_2917, weight = layers_12_encoder_attn_o_proj_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("obj_179_cast_fp16")]; + tensor inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = obj_179_cast_fp16)[name = tensor("inputs_77_cast_fp16")]; + tensor var_2928 = const()[name = tensor("op_2928"), val = tensor([1])]; + tensor channels_mean_77_cast_fp16 = reduce_mean(axes = var_2928, keep_dims = var_2760, x = inputs_77_cast_fp16)[name = tensor("channels_mean_77_cast_fp16")]; + tensor zero_mean_77_cast_fp16 = sub(x = inputs_77_cast_fp16, y = channels_mean_77_cast_fp16)[name = tensor("zero_mean_77_cast_fp16")]; + tensor zero_mean_sq_77_cast_fp16 = mul(x = zero_mean_77_cast_fp16, y = zero_mean_77_cast_fp16)[name = tensor("zero_mean_sq_77_cast_fp16")]; + tensor var_2932 = const()[name = tensor("op_2932"), val = tensor([1])]; + tensor var_2933_cast_fp16 = reduce_mean(axes = var_2932, keep_dims = var_2760, x = zero_mean_sq_77_cast_fp16)[name = tensor("op_2933_cast_fp16")]; + tensor var_2934_to_fp16 = const()[name = tensor("op_2934_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2935_cast_fp16 = add(x = var_2933_cast_fp16, y = var_2934_to_fp16)[name = tensor("op_2935_cast_fp16")]; + tensor denom_77_epsilon_0_to_fp16 = const()[name = tensor("denom_77_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_77_cast_fp16 = rsqrt(epsilon = denom_77_epsilon_0_to_fp16, x = var_2935_cast_fp16)[name = tensor("denom_77_cast_fp16")]; + tensor out_77_cast_fp16 = mul(x = zero_mean_77_cast_fp16, y = denom_77_cast_fp16)[name = tensor("out_77_cast_fp16")]; + tensor input_125_gamma_0_to_fp16 = const()[name = tensor("input_125_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(789856704)))]; + tensor input_125_beta_0_to_fp16 = const()[name = tensor("input_125_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(789859328)))]; + tensor input_125_epsilon_0_to_fp16 = const()[name = tensor("input_125_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_125_cast_fp16 = batch_norm(beta = input_125_beta_0_to_fp16, epsilon = input_125_epsilon_0_to_fp16, gamma = input_125_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_77_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor var_2946 = const()[name = tensor("op_2946"), val = tensor([1, 1])]; + tensor var_2948 = const()[name = tensor("op_2948"), val = tensor([1, 1])]; + tensor input_127_pad_type_0 = const()[name = tensor("input_127_pad_type_0"), val = tensor("custom")]; + tensor input_127_pad_0 = const()[name = tensor("input_127_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_fc1_weight_to_fp16 = const()[name = tensor("layers_12_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(789861952)))]; + tensor layers_12_fc1_bias_to_fp16 = const()[name = tensor("layers_12_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(802969216)))]; + tensor input_127_cast_fp16 = conv(bias = layers_12_fc1_bias_to_fp16, dilations = var_2948, groups = var_2759, pad = input_127_pad_0, pad_type = input_127_pad_type_0, strides = var_2946, weight = layers_12_fc1_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_mode_0 = const()[name = tensor("input_129_mode_0"), val = tensor("EXACT")]; + tensor input_129_cast_fp16 = gelu(mode = input_129_mode_0, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor var_2954 = const()[name = tensor("op_2954"), val = tensor([1, 1])]; + tensor var_2956 = const()[name = tensor("op_2956"), val = tensor([1, 1])]; + tensor hidden_states_27_pad_type_0 = const()[name = tensor("hidden_states_27_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_27_pad_0 = const()[name = tensor("hidden_states_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_fc2_weight_to_fp16 = const()[name = tensor("layers_12_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(802979520)))]; + tensor layers_12_fc2_bias_to_fp16 = const()[name = tensor("layers_12_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(816086784)))]; + tensor hidden_states_27_cast_fp16 = conv(bias = layers_12_fc2_bias_to_fp16, dilations = var_2956, groups = var_2759, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = var_2954, weight = layers_12_fc2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("hidden_states_27_cast_fp16")]; + tensor inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = hidden_states_27_cast_fp16)[name = tensor("inputs_79_cast_fp16")]; + tensor var_2970 = const()[name = tensor("op_2970"), val = tensor(3)]; + tensor var_2977 = const()[name = tensor("op_2977"), val = tensor(1)]; + tensor var_2978 = const()[name = tensor("op_2978"), val = tensor(true)]; + tensor var_2990 = const()[name = tensor("op_2990"), val = tensor([1])]; + tensor channels_mean_79_cast_fp16 = reduce_mean(axes = var_2990, keep_dims = var_2978, x = inputs_79_cast_fp16)[name = tensor("channels_mean_79_cast_fp16")]; + tensor zero_mean_79_cast_fp16 = sub(x = inputs_79_cast_fp16, y = channels_mean_79_cast_fp16)[name = tensor("zero_mean_79_cast_fp16")]; + tensor zero_mean_sq_79_cast_fp16 = mul(x = zero_mean_79_cast_fp16, y = zero_mean_79_cast_fp16)[name = tensor("zero_mean_sq_79_cast_fp16")]; + tensor var_2994 = const()[name = tensor("op_2994"), val = tensor([1])]; + tensor var_2995_cast_fp16 = reduce_mean(axes = var_2994, keep_dims = var_2978, x = zero_mean_sq_79_cast_fp16)[name = tensor("op_2995_cast_fp16")]; + tensor var_2996_to_fp16 = const()[name = tensor("op_2996_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2997_cast_fp16 = add(x = var_2995_cast_fp16, y = var_2996_to_fp16)[name = tensor("op_2997_cast_fp16")]; + tensor denom_79_epsilon_0_to_fp16 = const()[name = tensor("denom_79_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_79_cast_fp16 = rsqrt(epsilon = denom_79_epsilon_0_to_fp16, x = var_2997_cast_fp16)[name = tensor("denom_79_cast_fp16")]; + tensor out_79_cast_fp16 = mul(x = zero_mean_79_cast_fp16, y = denom_79_cast_fp16)[name = tensor("out_79_cast_fp16")]; + tensor obj_183_gamma_0_to_fp16 = const()[name = tensor("obj_183_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(816089408)))]; + tensor obj_183_beta_0_to_fp16 = const()[name = tensor("obj_183_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(816092032)))]; + tensor obj_183_epsilon_0_to_fp16 = const()[name = tensor("obj_183_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_183_cast_fp16 = batch_norm(beta = obj_183_beta_0_to_fp16, epsilon = obj_183_epsilon_0_to_fp16, gamma = obj_183_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_79_cast_fp16)[name = tensor("obj_183_cast_fp16")]; + tensor var_3012 = const()[name = tensor("op_3012"), val = tensor([1, 1])]; + tensor var_3014 = const()[name = tensor("op_3014"), val = tensor([1, 1])]; + tensor query_53_pad_type_0 = const()[name = tensor("query_53_pad_type_0"), val = tensor("custom")]; + tensor query_53_pad_0 = const()[name = tensor("query_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(816094656)))]; + tensor layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(819371520)))]; + tensor query_53_cast_fp16 = conv(bias = layers_13_self_attn_q_proj_bias_to_fp16, dilations = var_3014, groups = var_2977, pad = query_53_pad_0, pad_type = query_53_pad_type_0, strides = var_3012, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor("query_53_cast_fp16")]; + tensor var_3018 = const()[name = tensor("op_3018"), val = tensor([1, 1])]; + tensor var_3020 = const()[name = tensor("op_3020"), val = tensor([1, 1])]; + tensor current_key_27_pad_type_0 = const()[name = tensor("current_key_27_pad_type_0"), val = tensor("custom")]; + tensor current_key_27_pad_0 = const()[name = tensor("current_key_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(819374144)))]; + tensor current_key_27_cast_fp16 = conv(dilations = var_3020, groups = var_2977, pad = current_key_27_pad_0, pad_type = current_key_27_pad_type_0, strides = var_3018, weight = layers_13_self_attn_k_proj_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor("current_key_27_cast_fp16")]; + tensor var_3025 = const()[name = tensor("op_3025"), val = tensor([1, 1])]; + tensor var_3027 = const()[name = tensor("op_3027"), val = tensor([1, 1])]; + tensor current_value_27_pad_type_0 = const()[name = tensor("current_value_27_pad_type_0"), val = tensor("custom")]; + tensor current_value_27_pad_0 = const()[name = tensor("current_value_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(822651008)))]; + tensor layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(825927872)))]; + tensor current_value_27_cast_fp16 = conv(bias = layers_13_self_attn_v_proj_bias_to_fp16, dilations = var_3027, groups = var_2977, pad = current_value_27_pad_0, pad_type = current_value_27_pad_type_0, strides = var_3025, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor("current_value_27_cast_fp16")]; + tensor var_3034_cast_fp16 = mul(x = current_key_27_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3034_cast_fp16")]; + tensor var_3036_cast_fp16 = mul(x = var_103_cast_fp16_13, y = var_241_cast_fp16)[name = tensor("op_3036_cast_fp16")]; + tensor key_53_cast_fp16 = add(x = var_3034_cast_fp16, y = var_3036_cast_fp16)[name = tensor("key_53_cast_fp16")]; + tensor var_3038_cast_fp16 = mul(x = current_value_27_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3038_cast_fp16")]; + tensor var_3040_cast_fp16 = mul(x = var_138_cast_fp16_13, y = var_241_cast_fp16)[name = tensor("op_3040_cast_fp16")]; + tensor value_53_cast_fp16 = add(x = var_3038_cast_fp16, y = var_3040_cast_fp16)[name = tensor("value_53_cast_fp16")]; + tensor var_3043 = const()[name = tensor("op_3043"), val = tensor([1, 20, 64, -1])]; + tensor var_3044_cast_fp16 = reshape(shape = var_3043, x = query_53_cast_fp16)[name = tensor("op_3044_cast_fp16")]; + tensor var_3045_to_fp16 = const()[name = tensor("op_3045_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3046_cast_fp16 = mul(x = var_3044_cast_fp16, y = var_3045_to_fp16)[name = tensor("op_3046_cast_fp16")]; + tensor var_3047 = const()[name = tensor("op_3047"), val = tensor([1, 20, 64, -1])]; + tensor var_3048_cast_fp16 = reshape(shape = var_3047, x = key_53_cast_fp16)[name = tensor("op_3048_cast_fp16")]; + tensor mh_w_79_transpose_x_0 = const()[name = tensor("mh_w_79_transpose_x_0"), val = tensor(true)]; + tensor mh_w_79_transpose_y_0 = const()[name = tensor("mh_w_79_transpose_y_0"), val = tensor(false)]; + tensor mh_w_79_cast_fp16 = matmul(transpose_x = mh_w_79_transpose_x_0, transpose_y = mh_w_79_transpose_y_0, x = var_3046_cast_fp16, y = var_3048_cast_fp16)[name = tensor("mh_w_79_cast_fp16")]; + tensor mh_w_81_cast_fp16 = add(x = mh_w_79_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_81_cast_fp16")]; + tensor var_3056_cast_fp16 = softmax(axis = var_2970, x = mh_w_81_cast_fp16)[name = tensor("op_3056_cast_fp16")]; + tensor var_3057 = const()[name = tensor("op_3057"), val = tensor([1, 20, 64, -1])]; + tensor var_3058_cast_fp16 = reshape(shape = var_3057, x = value_53_cast_fp16)[name = tensor("op_3058_cast_fp16")]; + tensor attn_53_transpose_x_0 = const()[name = tensor("attn_53_transpose_x_0"), val = tensor(false)]; + tensor attn_53_transpose_y_0 = const()[name = tensor("attn_53_transpose_y_0"), val = tensor(true)]; + tensor attn_53_cast_fp16 = matmul(transpose_x = attn_53_transpose_x_0, transpose_y = attn_53_transpose_y_0, x = var_3058_cast_fp16, y = var_3056_cast_fp16)[name = tensor("attn_53_cast_fp16")]; + tensor var_3061 = const()[name = tensor("op_3061"), val = tensor([1, 1280, 1, -1])]; + tensor input_131_cast_fp16 = reshape(shape = var_3061, x = attn_53_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor var_3065 = const()[name = tensor("op_3065"), val = tensor([1, 1])]; + tensor var_3067 = const()[name = tensor("op_3067"), val = tensor([1, 1])]; + tensor obj_189_pad_type_0 = const()[name = tensor("obj_189_pad_type_0"), val = tensor("custom")]; + tensor obj_189_pad_0 = const()[name = tensor("obj_189_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(825930496)))]; + tensor layers_13_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(829207360)))]; + tensor obj_189_cast_fp16 = conv(bias = layers_13_self_attn_o_proj_bias_to_fp16, dilations = var_3067, groups = var_2977, pad = obj_189_pad_0, pad_type = obj_189_pad_type_0, strides = var_3065, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("obj_189_cast_fp16")]; + tensor inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = obj_189_cast_fp16)[name = tensor("inputs_81_cast_fp16")]; + tensor var_3077 = const()[name = tensor("op_3077"), val = tensor([1])]; + tensor channels_mean_81_cast_fp16 = reduce_mean(axes = var_3077, keep_dims = var_2978, x = inputs_81_cast_fp16)[name = tensor("channels_mean_81_cast_fp16")]; + tensor zero_mean_81_cast_fp16 = sub(x = inputs_81_cast_fp16, y = channels_mean_81_cast_fp16)[name = tensor("zero_mean_81_cast_fp16")]; + tensor zero_mean_sq_81_cast_fp16 = mul(x = zero_mean_81_cast_fp16, y = zero_mean_81_cast_fp16)[name = tensor("zero_mean_sq_81_cast_fp16")]; + tensor var_3081 = const()[name = tensor("op_3081"), val = tensor([1])]; + tensor var_3082_cast_fp16 = reduce_mean(axes = var_3081, keep_dims = var_2978, x = zero_mean_sq_81_cast_fp16)[name = tensor("op_3082_cast_fp16")]; + tensor var_3083_to_fp16 = const()[name = tensor("op_3083_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3084_cast_fp16 = add(x = var_3082_cast_fp16, y = var_3083_to_fp16)[name = tensor("op_3084_cast_fp16")]; + tensor denom_81_epsilon_0_to_fp16 = const()[name = tensor("denom_81_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_81_cast_fp16 = rsqrt(epsilon = denom_81_epsilon_0_to_fp16, x = var_3084_cast_fp16)[name = tensor("denom_81_cast_fp16")]; + tensor out_81_cast_fp16 = mul(x = zero_mean_81_cast_fp16, y = denom_81_cast_fp16)[name = tensor("out_81_cast_fp16")]; + tensor obj_191_gamma_0_to_fp16 = const()[name = tensor("obj_191_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(829209984)))]; + tensor obj_191_beta_0_to_fp16 = const()[name = tensor("obj_191_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(829212608)))]; + tensor obj_191_epsilon_0_to_fp16 = const()[name = tensor("obj_191_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_191_cast_fp16 = batch_norm(beta = obj_191_beta_0_to_fp16, epsilon = obj_191_epsilon_0_to_fp16, gamma = obj_191_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_81_cast_fp16)[name = tensor("obj_191_cast_fp16")]; + tensor var_3099 = const()[name = tensor("op_3099"), val = tensor([1, 1])]; + tensor var_3101 = const()[name = tensor("op_3101"), val = tensor([1, 1])]; + tensor query_55_pad_type_0 = const()[name = tensor("query_55_pad_type_0"), val = tensor("custom")]; + tensor query_55_pad_0 = const()[name = tensor("query_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(829215232)))]; + tensor layers_13_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(832492096)))]; + tensor query_55_cast_fp16 = conv(bias = layers_13_encoder_attn_q_proj_bias_to_fp16, dilations = var_3101, groups = var_2977, pad = query_55_pad_0, pad_type = query_55_pad_type_0, strides = var_3099, weight = layers_13_encoder_attn_q_proj_weight_to_fp16, x = obj_191_cast_fp16)[name = tensor("query_55_cast_fp16")]; + tensor var_3105 = const()[name = tensor("op_3105"), val = tensor([1, 1])]; + tensor var_3107 = const()[name = tensor("op_3107"), val = tensor([1, 1])]; + tensor key_55_pad_type_0 = const()[name = tensor("key_55_pad_type_0"), val = tensor("custom")]; + tensor key_55_pad_0 = const()[name = tensor("key_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(832494720)))]; + tensor key_55_cast_fp16 = conv(dilations = var_3107, groups = var_2977, pad = key_55_pad_0, pad_type = key_55_pad_type_0, strides = var_3105, weight = layers_13_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_55_cast_fp16")]; + tensor var_3112 = const()[name = tensor("op_3112"), val = tensor([1, 1])]; + tensor var_3114 = const()[name = tensor("op_3114"), val = tensor([1, 1])]; + tensor value_55_pad_type_0 = const()[name = tensor("value_55_pad_type_0"), val = tensor("custom")]; + tensor value_55_pad_0 = const()[name = tensor("value_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(835771584)))]; + tensor layers_13_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(839048448)))]; + tensor value_55_cast_fp16 = conv(bias = layers_13_encoder_attn_v_proj_bias_to_fp16, dilations = var_3114, groups = var_2977, pad = value_55_pad_0, pad_type = value_55_pad_type_0, strides = var_3112, weight = layers_13_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_55_cast_fp16")]; + tensor var_3118 = const()[name = tensor("op_3118"), val = tensor([1, 20, 64, -1])]; + tensor var_3119_cast_fp16 = reshape(shape = var_3118, x = query_55_cast_fp16)[name = tensor("op_3119_cast_fp16")]; + tensor var_3120_to_fp16 = const()[name = tensor("op_3120_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3121_cast_fp16 = mul(x = var_3119_cast_fp16, y = var_3120_to_fp16)[name = tensor("op_3121_cast_fp16")]; + tensor var_3122 = const()[name = tensor("op_3122"), val = tensor([1, 20, 64, -1])]; + tensor var_3123_cast_fp16 = reshape(shape = var_3122, x = key_55_cast_fp16)[name = tensor("op_3123_cast_fp16")]; + tensor mh_w_83_transpose_x_0 = const()[name = tensor("mh_w_83_transpose_x_0"), val = tensor(true)]; + tensor mh_w_83_transpose_y_0 = const()[name = tensor("mh_w_83_transpose_y_0"), val = tensor(false)]; + tensor mh_w_83_cast_fp16 = matmul(transpose_x = mh_w_83_transpose_x_0, transpose_y = mh_w_83_transpose_y_0, x = var_3121_cast_fp16, y = var_3123_cast_fp16)[name = tensor("mh_w_83_cast_fp16")]; + tensor obj_195_cast_fp16 = softmax(axis = var_2970, x = mh_w_83_cast_fp16)[name = tensor("obj_195_cast_fp16")]; + tensor var_3127 = const()[name = tensor("op_3127"), val = tensor([1, 20, 64, -1])]; + tensor var_3128_cast_fp16 = reshape(shape = var_3127, x = value_55_cast_fp16)[name = tensor("op_3128_cast_fp16")]; + tensor attn_55_transpose_x_0 = const()[name = tensor("attn_55_transpose_x_0"), val = tensor(false)]; + tensor attn_55_transpose_y_0 = const()[name = tensor("attn_55_transpose_y_0"), val = tensor(true)]; + tensor attn_55_cast_fp16 = matmul(transpose_x = attn_55_transpose_x_0, transpose_y = attn_55_transpose_y_0, x = var_3128_cast_fp16, y = obj_195_cast_fp16)[name = tensor("attn_55_cast_fp16")]; + tensor var_3131 = const()[name = tensor("op_3131"), val = tensor([1, 1280, 1, -1])]; + tensor input_133_cast_fp16 = reshape(shape = var_3131, x = attn_55_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor var_3135 = const()[name = tensor("op_3135"), val = tensor([1, 1])]; + tensor var_3137 = const()[name = tensor("op_3137"), val = tensor([1, 1])]; + tensor obj_193_pad_type_0 = const()[name = tensor("obj_193_pad_type_0"), val = tensor("custom")]; + tensor obj_193_pad_0 = const()[name = tensor("obj_193_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(839051072)))]; + tensor layers_13_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(842327936)))]; + tensor obj_193_cast_fp16 = conv(bias = layers_13_encoder_attn_o_proj_bias_to_fp16, dilations = var_3137, groups = var_2977, pad = obj_193_pad_0, pad_type = obj_193_pad_type_0, strides = var_3135, weight = layers_13_encoder_attn_o_proj_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("obj_193_cast_fp16")]; + tensor inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = obj_193_cast_fp16)[name = tensor("inputs_83_cast_fp16")]; + tensor var_3146 = const()[name = tensor("op_3146"), val = tensor([1])]; + tensor channels_mean_83_cast_fp16 = reduce_mean(axes = var_3146, keep_dims = var_2978, x = inputs_83_cast_fp16)[name = tensor("channels_mean_83_cast_fp16")]; + tensor zero_mean_83_cast_fp16 = sub(x = inputs_83_cast_fp16, y = channels_mean_83_cast_fp16)[name = tensor("zero_mean_83_cast_fp16")]; + tensor zero_mean_sq_83_cast_fp16 = mul(x = zero_mean_83_cast_fp16, y = zero_mean_83_cast_fp16)[name = tensor("zero_mean_sq_83_cast_fp16")]; + tensor var_3150 = const()[name = tensor("op_3150"), val = tensor([1])]; + tensor var_3151_cast_fp16 = reduce_mean(axes = var_3150, keep_dims = var_2978, x = zero_mean_sq_83_cast_fp16)[name = tensor("op_3151_cast_fp16")]; + tensor var_3152_to_fp16 = const()[name = tensor("op_3152_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3153_cast_fp16 = add(x = var_3151_cast_fp16, y = var_3152_to_fp16)[name = tensor("op_3153_cast_fp16")]; + tensor denom_83_epsilon_0_to_fp16 = const()[name = tensor("denom_83_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_83_cast_fp16 = rsqrt(epsilon = denom_83_epsilon_0_to_fp16, x = var_3153_cast_fp16)[name = tensor("denom_83_cast_fp16")]; + tensor out_83_cast_fp16 = mul(x = zero_mean_83_cast_fp16, y = denom_83_cast_fp16)[name = tensor("out_83_cast_fp16")]; + tensor input_135_gamma_0_to_fp16 = const()[name = tensor("input_135_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(842330560)))]; + tensor input_135_beta_0_to_fp16 = const()[name = tensor("input_135_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(842333184)))]; + tensor input_135_epsilon_0_to_fp16 = const()[name = tensor("input_135_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_135_cast_fp16 = batch_norm(beta = input_135_beta_0_to_fp16, epsilon = input_135_epsilon_0_to_fp16, gamma = input_135_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_83_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor var_3164 = const()[name = tensor("op_3164"), val = tensor([1, 1])]; + tensor var_3166 = const()[name = tensor("op_3166"), val = tensor([1, 1])]; + tensor input_137_pad_type_0 = const()[name = tensor("input_137_pad_type_0"), val = tensor("custom")]; + tensor input_137_pad_0 = const()[name = tensor("input_137_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_fc1_weight_to_fp16 = const()[name = tensor("layers_13_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(842335808)))]; + tensor layers_13_fc1_bias_to_fp16 = const()[name = tensor("layers_13_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(855443072)))]; + tensor input_137_cast_fp16 = conv(bias = layers_13_fc1_bias_to_fp16, dilations = var_3166, groups = var_2977, pad = input_137_pad_0, pad_type = input_137_pad_type_0, strides = var_3164, weight = layers_13_fc1_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor input_139_mode_0 = const()[name = tensor("input_139_mode_0"), val = tensor("EXACT")]; + tensor input_139_cast_fp16 = gelu(mode = input_139_mode_0, x = input_137_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor var_3172 = const()[name = tensor("op_3172"), val = tensor([1, 1])]; + tensor var_3174 = const()[name = tensor("op_3174"), val = tensor([1, 1])]; + tensor hidden_states_29_pad_type_0 = const()[name = tensor("hidden_states_29_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_29_pad_0 = const()[name = tensor("hidden_states_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_fc2_weight_to_fp16 = const()[name = tensor("layers_13_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(855453376)))]; + tensor layers_13_fc2_bias_to_fp16 = const()[name = tensor("layers_13_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(868560640)))]; + tensor hidden_states_29_cast_fp16 = conv(bias = layers_13_fc2_bias_to_fp16, dilations = var_3174, groups = var_2977, pad = hidden_states_29_pad_0, pad_type = hidden_states_29_pad_type_0, strides = var_3172, weight = layers_13_fc2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("hidden_states_29_cast_fp16")]; + tensor inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = hidden_states_29_cast_fp16)[name = tensor("inputs_85_cast_fp16")]; + tensor var_3188 = const()[name = tensor("op_3188"), val = tensor(3)]; + tensor var_3195 = const()[name = tensor("op_3195"), val = tensor(1)]; + tensor var_3196 = const()[name = tensor("op_3196"), val = tensor(true)]; + tensor var_3208 = const()[name = tensor("op_3208"), val = tensor([1])]; + tensor channels_mean_85_cast_fp16 = reduce_mean(axes = var_3208, keep_dims = var_3196, x = inputs_85_cast_fp16)[name = tensor("channels_mean_85_cast_fp16")]; + tensor zero_mean_85_cast_fp16 = sub(x = inputs_85_cast_fp16, y = channels_mean_85_cast_fp16)[name = tensor("zero_mean_85_cast_fp16")]; + tensor zero_mean_sq_85_cast_fp16 = mul(x = zero_mean_85_cast_fp16, y = zero_mean_85_cast_fp16)[name = tensor("zero_mean_sq_85_cast_fp16")]; + tensor var_3212 = const()[name = tensor("op_3212"), val = tensor([1])]; + tensor var_3213_cast_fp16 = reduce_mean(axes = var_3212, keep_dims = var_3196, x = zero_mean_sq_85_cast_fp16)[name = tensor("op_3213_cast_fp16")]; + tensor var_3214_to_fp16 = const()[name = tensor("op_3214_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3215_cast_fp16 = add(x = var_3213_cast_fp16, y = var_3214_to_fp16)[name = tensor("op_3215_cast_fp16")]; + tensor denom_85_epsilon_0_to_fp16 = const()[name = tensor("denom_85_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_85_cast_fp16 = rsqrt(epsilon = denom_85_epsilon_0_to_fp16, x = var_3215_cast_fp16)[name = tensor("denom_85_cast_fp16")]; + tensor out_85_cast_fp16 = mul(x = zero_mean_85_cast_fp16, y = denom_85_cast_fp16)[name = tensor("out_85_cast_fp16")]; + tensor obj_197_gamma_0_to_fp16 = const()[name = tensor("obj_197_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(868563264)))]; + tensor obj_197_beta_0_to_fp16 = const()[name = tensor("obj_197_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(868565888)))]; + tensor obj_197_epsilon_0_to_fp16 = const()[name = tensor("obj_197_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_197_cast_fp16 = batch_norm(beta = obj_197_beta_0_to_fp16, epsilon = obj_197_epsilon_0_to_fp16, gamma = obj_197_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_85_cast_fp16)[name = tensor("obj_197_cast_fp16")]; + tensor var_3230 = const()[name = tensor("op_3230"), val = tensor([1, 1])]; + tensor var_3232 = const()[name = tensor("op_3232"), val = tensor([1, 1])]; + tensor query_57_pad_type_0 = const()[name = tensor("query_57_pad_type_0"), val = tensor("custom")]; + tensor query_57_pad_0 = const()[name = tensor("query_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(868568512)))]; + tensor layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(871845376)))]; + tensor query_57_cast_fp16 = conv(bias = layers_14_self_attn_q_proj_bias_to_fp16, dilations = var_3232, groups = var_3195, pad = query_57_pad_0, pad_type = query_57_pad_type_0, strides = var_3230, weight = layers_14_self_attn_q_proj_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor("query_57_cast_fp16")]; + tensor var_3236 = const()[name = tensor("op_3236"), val = tensor([1, 1])]; + tensor var_3238 = const()[name = tensor("op_3238"), val = tensor([1, 1])]; + tensor current_key_29_pad_type_0 = const()[name = tensor("current_key_29_pad_type_0"), val = tensor("custom")]; + tensor current_key_29_pad_0 = const()[name = tensor("current_key_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(871848000)))]; + tensor current_key_29_cast_fp16 = conv(dilations = var_3238, groups = var_3195, pad = current_key_29_pad_0, pad_type = current_key_29_pad_type_0, strides = var_3236, weight = layers_14_self_attn_k_proj_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor("current_key_29_cast_fp16")]; + tensor var_3243 = const()[name = tensor("op_3243"), val = tensor([1, 1])]; + tensor var_3245 = const()[name = tensor("op_3245"), val = tensor([1, 1])]; + tensor current_value_29_pad_type_0 = const()[name = tensor("current_value_29_pad_type_0"), val = tensor("custom")]; + tensor current_value_29_pad_0 = const()[name = tensor("current_value_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(875124864)))]; + tensor layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(878401728)))]; + tensor current_value_29_cast_fp16 = conv(bias = layers_14_self_attn_v_proj_bias_to_fp16, dilations = var_3245, groups = var_3195, pad = current_value_29_pad_0, pad_type = current_value_29_pad_type_0, strides = var_3243, weight = layers_14_self_attn_v_proj_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor("current_value_29_cast_fp16")]; + tensor var_3252_cast_fp16 = mul(x = current_key_29_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3252_cast_fp16")]; + tensor var_3254_cast_fp16 = mul(x = var_103_cast_fp16_14, y = var_241_cast_fp16)[name = tensor("op_3254_cast_fp16")]; + tensor key_57_cast_fp16 = add(x = var_3252_cast_fp16, y = var_3254_cast_fp16)[name = tensor("key_57_cast_fp16")]; + tensor var_3256_cast_fp16 = mul(x = current_value_29_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3256_cast_fp16")]; + tensor var_3258_cast_fp16 = mul(x = var_138_cast_fp16_14, y = var_241_cast_fp16)[name = tensor("op_3258_cast_fp16")]; + tensor value_57_cast_fp16 = add(x = var_3256_cast_fp16, y = var_3258_cast_fp16)[name = tensor("value_57_cast_fp16")]; + tensor var_3261 = const()[name = tensor("op_3261"), val = tensor([1, 20, 64, -1])]; + tensor var_3262_cast_fp16 = reshape(shape = var_3261, x = query_57_cast_fp16)[name = tensor("op_3262_cast_fp16")]; + tensor var_3263_to_fp16 = const()[name = tensor("op_3263_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3264_cast_fp16 = mul(x = var_3262_cast_fp16, y = var_3263_to_fp16)[name = tensor("op_3264_cast_fp16")]; + tensor var_3265 = const()[name = tensor("op_3265"), val = tensor([1, 20, 64, -1])]; + tensor var_3266_cast_fp16 = reshape(shape = var_3265, x = key_57_cast_fp16)[name = tensor("op_3266_cast_fp16")]; + tensor mh_w_85_transpose_x_0 = const()[name = tensor("mh_w_85_transpose_x_0"), val = tensor(true)]; + tensor mh_w_85_transpose_y_0 = const()[name = tensor("mh_w_85_transpose_y_0"), val = tensor(false)]; + tensor mh_w_85_cast_fp16 = matmul(transpose_x = mh_w_85_transpose_x_0, transpose_y = mh_w_85_transpose_y_0, x = var_3264_cast_fp16, y = var_3266_cast_fp16)[name = tensor("mh_w_85_cast_fp16")]; + tensor mh_w_87_cast_fp16 = add(x = mh_w_85_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_87_cast_fp16")]; + tensor var_3274_cast_fp16 = softmax(axis = var_3188, x = mh_w_87_cast_fp16)[name = tensor("op_3274_cast_fp16")]; + tensor var_3275 = const()[name = tensor("op_3275"), val = tensor([1, 20, 64, -1])]; + tensor var_3276_cast_fp16 = reshape(shape = var_3275, x = value_57_cast_fp16)[name = tensor("op_3276_cast_fp16")]; + tensor attn_57_transpose_x_0 = const()[name = tensor("attn_57_transpose_x_0"), val = tensor(false)]; + tensor attn_57_transpose_y_0 = const()[name = tensor("attn_57_transpose_y_0"), val = tensor(true)]; + tensor attn_57_cast_fp16 = matmul(transpose_x = attn_57_transpose_x_0, transpose_y = attn_57_transpose_y_0, x = var_3276_cast_fp16, y = var_3274_cast_fp16)[name = tensor("attn_57_cast_fp16")]; + tensor var_3279 = const()[name = tensor("op_3279"), val = tensor([1, 1280, 1, -1])]; + tensor input_141_cast_fp16 = reshape(shape = var_3279, x = attn_57_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor var_3283 = const()[name = tensor("op_3283"), val = tensor([1, 1])]; + tensor var_3285 = const()[name = tensor("op_3285"), val = tensor([1, 1])]; + tensor obj_203_pad_type_0 = const()[name = tensor("obj_203_pad_type_0"), val = tensor("custom")]; + tensor obj_203_pad_0 = const()[name = tensor("obj_203_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(878404352)))]; + tensor layers_14_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(881681216)))]; + tensor obj_203_cast_fp16 = conv(bias = layers_14_self_attn_o_proj_bias_to_fp16, dilations = var_3285, groups = var_3195, pad = obj_203_pad_0, pad_type = obj_203_pad_type_0, strides = var_3283, weight = layers_14_self_attn_o_proj_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("obj_203_cast_fp16")]; + tensor inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = obj_203_cast_fp16)[name = tensor("inputs_87_cast_fp16")]; + tensor var_3295 = const()[name = tensor("op_3295"), val = tensor([1])]; + tensor channels_mean_87_cast_fp16 = reduce_mean(axes = var_3295, keep_dims = var_3196, x = inputs_87_cast_fp16)[name = tensor("channels_mean_87_cast_fp16")]; + tensor zero_mean_87_cast_fp16 = sub(x = inputs_87_cast_fp16, y = channels_mean_87_cast_fp16)[name = tensor("zero_mean_87_cast_fp16")]; + tensor zero_mean_sq_87_cast_fp16 = mul(x = zero_mean_87_cast_fp16, y = zero_mean_87_cast_fp16)[name = tensor("zero_mean_sq_87_cast_fp16")]; + tensor var_3299 = const()[name = tensor("op_3299"), val = tensor([1])]; + tensor var_3300_cast_fp16 = reduce_mean(axes = var_3299, keep_dims = var_3196, x = zero_mean_sq_87_cast_fp16)[name = tensor("op_3300_cast_fp16")]; + tensor var_3301_to_fp16 = const()[name = tensor("op_3301_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3302_cast_fp16 = add(x = var_3300_cast_fp16, y = var_3301_to_fp16)[name = tensor("op_3302_cast_fp16")]; + tensor denom_87_epsilon_0_to_fp16 = const()[name = tensor("denom_87_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_87_cast_fp16 = rsqrt(epsilon = denom_87_epsilon_0_to_fp16, x = var_3302_cast_fp16)[name = tensor("denom_87_cast_fp16")]; + tensor out_87_cast_fp16 = mul(x = zero_mean_87_cast_fp16, y = denom_87_cast_fp16)[name = tensor("out_87_cast_fp16")]; + tensor obj_205_gamma_0_to_fp16 = const()[name = tensor("obj_205_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(881683840)))]; + tensor obj_205_beta_0_to_fp16 = const()[name = tensor("obj_205_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(881686464)))]; + tensor obj_205_epsilon_0_to_fp16 = const()[name = tensor("obj_205_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_205_cast_fp16 = batch_norm(beta = obj_205_beta_0_to_fp16, epsilon = obj_205_epsilon_0_to_fp16, gamma = obj_205_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_87_cast_fp16)[name = tensor("obj_205_cast_fp16")]; + tensor var_3317 = const()[name = tensor("op_3317"), val = tensor([1, 1])]; + tensor var_3319 = const()[name = tensor("op_3319"), val = tensor([1, 1])]; + tensor query_59_pad_type_0 = const()[name = tensor("query_59_pad_type_0"), val = tensor("custom")]; + tensor query_59_pad_0 = const()[name = tensor("query_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(881689088)))]; + tensor layers_14_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(884965952)))]; + tensor query_59_cast_fp16 = conv(bias = layers_14_encoder_attn_q_proj_bias_to_fp16, dilations = var_3319, groups = var_3195, pad = query_59_pad_0, pad_type = query_59_pad_type_0, strides = var_3317, weight = layers_14_encoder_attn_q_proj_weight_to_fp16, x = obj_205_cast_fp16)[name = tensor("query_59_cast_fp16")]; + tensor var_3323 = const()[name = tensor("op_3323"), val = tensor([1, 1])]; + tensor var_3325 = const()[name = tensor("op_3325"), val = tensor([1, 1])]; + tensor key_59_pad_type_0 = const()[name = tensor("key_59_pad_type_0"), val = tensor("custom")]; + tensor key_59_pad_0 = const()[name = tensor("key_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(884968576)))]; + tensor key_59_cast_fp16 = conv(dilations = var_3325, groups = var_3195, pad = key_59_pad_0, pad_type = key_59_pad_type_0, strides = var_3323, weight = layers_14_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_59_cast_fp16")]; + tensor var_3330 = const()[name = tensor("op_3330"), val = tensor([1, 1])]; + tensor var_3332 = const()[name = tensor("op_3332"), val = tensor([1, 1])]; + tensor value_59_pad_type_0 = const()[name = tensor("value_59_pad_type_0"), val = tensor("custom")]; + tensor value_59_pad_0 = const()[name = tensor("value_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(888245440)))]; + tensor layers_14_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(891522304)))]; + tensor value_59_cast_fp16 = conv(bias = layers_14_encoder_attn_v_proj_bias_to_fp16, dilations = var_3332, groups = var_3195, pad = value_59_pad_0, pad_type = value_59_pad_type_0, strides = var_3330, weight = layers_14_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_59_cast_fp16")]; + tensor var_3336 = const()[name = tensor("op_3336"), val = tensor([1, 20, 64, -1])]; + tensor var_3337_cast_fp16 = reshape(shape = var_3336, x = query_59_cast_fp16)[name = tensor("op_3337_cast_fp16")]; + tensor var_3338_to_fp16 = const()[name = tensor("op_3338_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3339_cast_fp16 = mul(x = var_3337_cast_fp16, y = var_3338_to_fp16)[name = tensor("op_3339_cast_fp16")]; + tensor var_3340 = const()[name = tensor("op_3340"), val = tensor([1, 20, 64, -1])]; + tensor var_3341_cast_fp16 = reshape(shape = var_3340, x = key_59_cast_fp16)[name = tensor("op_3341_cast_fp16")]; + tensor mh_w_89_transpose_x_0 = const()[name = tensor("mh_w_89_transpose_x_0"), val = tensor(true)]; + tensor mh_w_89_transpose_y_0 = const()[name = tensor("mh_w_89_transpose_y_0"), val = tensor(false)]; + tensor mh_w_89_cast_fp16 = matmul(transpose_x = mh_w_89_transpose_x_0, transpose_y = mh_w_89_transpose_y_0, x = var_3339_cast_fp16, y = var_3341_cast_fp16)[name = tensor("mh_w_89_cast_fp16")]; + tensor obj_209_cast_fp16 = softmax(axis = var_3188, x = mh_w_89_cast_fp16)[name = tensor("obj_209_cast_fp16")]; + tensor var_3345 = const()[name = tensor("op_3345"), val = tensor([1, 20, 64, -1])]; + tensor var_3346_cast_fp16 = reshape(shape = var_3345, x = value_59_cast_fp16)[name = tensor("op_3346_cast_fp16")]; + tensor attn_59_transpose_x_0 = const()[name = tensor("attn_59_transpose_x_0"), val = tensor(false)]; + tensor attn_59_transpose_y_0 = const()[name = tensor("attn_59_transpose_y_0"), val = tensor(true)]; + tensor attn_59_cast_fp16 = matmul(transpose_x = attn_59_transpose_x_0, transpose_y = attn_59_transpose_y_0, x = var_3346_cast_fp16, y = obj_209_cast_fp16)[name = tensor("attn_59_cast_fp16")]; + tensor var_3349 = const()[name = tensor("op_3349"), val = tensor([1, 1280, 1, -1])]; + tensor input_143_cast_fp16 = reshape(shape = var_3349, x = attn_59_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor var_3353 = const()[name = tensor("op_3353"), val = tensor([1, 1])]; + tensor var_3355 = const()[name = tensor("op_3355"), val = tensor([1, 1])]; + tensor obj_207_pad_type_0 = const()[name = tensor("obj_207_pad_type_0"), val = tensor("custom")]; + tensor obj_207_pad_0 = const()[name = tensor("obj_207_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(891524928)))]; + tensor layers_14_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(894801792)))]; + tensor obj_207_cast_fp16 = conv(bias = layers_14_encoder_attn_o_proj_bias_to_fp16, dilations = var_3355, groups = var_3195, pad = obj_207_pad_0, pad_type = obj_207_pad_type_0, strides = var_3353, weight = layers_14_encoder_attn_o_proj_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("obj_207_cast_fp16")]; + tensor inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = obj_207_cast_fp16)[name = tensor("inputs_89_cast_fp16")]; + tensor var_3361 = const()[name = tensor("op_3361"), val = tensor([1])]; + tensor channels_mean_89_cast_fp16 = reduce_mean(axes = var_3361, keep_dims = var_3196, x = inputs_89_cast_fp16)[name = tensor("channels_mean_89_cast_fp16")]; + tensor zero_mean_89_cast_fp16 = sub(x = inputs_89_cast_fp16, y = channels_mean_89_cast_fp16)[name = tensor("zero_mean_89_cast_fp16")]; + tensor zero_mean_sq_89_cast_fp16 = mul(x = zero_mean_89_cast_fp16, y = zero_mean_89_cast_fp16)[name = tensor("zero_mean_sq_89_cast_fp16")]; + tensor var_3365 = const()[name = tensor("op_3365"), val = tensor([1])]; + tensor var_3366_cast_fp16 = reduce_mean(axes = var_3365, keep_dims = var_3196, x = zero_mean_sq_89_cast_fp16)[name = tensor("op_3366_cast_fp16")]; + tensor var_3367_to_fp16 = const()[name = tensor("op_3367_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3368_cast_fp16 = add(x = var_3366_cast_fp16, y = var_3367_to_fp16)[name = tensor("op_3368_cast_fp16")]; + tensor denom_89_epsilon_0_to_fp16 = const()[name = tensor("denom_89_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_89_cast_fp16 = rsqrt(epsilon = denom_89_epsilon_0_to_fp16, x = var_3368_cast_fp16)[name = tensor("denom_89_cast_fp16")]; + tensor out_89_cast_fp16 = mul(x = zero_mean_89_cast_fp16, y = denom_89_cast_fp16)[name = tensor("out_89_cast_fp16")]; + tensor input_145_gamma_0_to_fp16 = const()[name = tensor("input_145_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(894804416)))]; + tensor input_145_beta_0_to_fp16 = const()[name = tensor("input_145_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(894807040)))]; + tensor input_145_epsilon_0_to_fp16 = const()[name = tensor("input_145_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_145_cast_fp16 = batch_norm(beta = input_145_beta_0_to_fp16, epsilon = input_145_epsilon_0_to_fp16, gamma = input_145_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_89_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor var_3379 = const()[name = tensor("op_3379"), val = tensor([1, 1])]; + tensor var_3381 = const()[name = tensor("op_3381"), val = tensor([1, 1])]; + tensor input_147_pad_type_0 = const()[name = tensor("input_147_pad_type_0"), val = tensor("custom")]; + tensor input_147_pad_0 = const()[name = tensor("input_147_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_fc1_weight_to_fp16 = const()[name = tensor("layers_14_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(894809664)))]; + tensor layers_14_fc1_bias_to_fp16 = const()[name = tensor("layers_14_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(907916928)))]; + tensor input_147_cast_fp16 = conv(bias = layers_14_fc1_bias_to_fp16, dilations = var_3381, groups = var_3195, pad = input_147_pad_0, pad_type = input_147_pad_type_0, strides = var_3379, weight = layers_14_fc1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_mode_0 = const()[name = tensor("input_149_mode_0"), val = tensor("EXACT")]; + tensor input_149_cast_fp16 = gelu(mode = input_149_mode_0, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor var_3387 = const()[name = tensor("op_3387"), val = tensor([1, 1])]; + tensor var_3389 = const()[name = tensor("op_3389"), val = tensor([1, 1])]; + tensor hidden_states_31_pad_type_0 = const()[name = tensor("hidden_states_31_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_31_pad_0 = const()[name = tensor("hidden_states_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_fc2_weight_to_fp16 = const()[name = tensor("layers_14_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(907927232)))]; + tensor layers_14_fc2_bias_to_fp16 = const()[name = tensor("layers_14_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(921034496)))]; + tensor hidden_states_31_cast_fp16 = conv(bias = layers_14_fc2_bias_to_fp16, dilations = var_3389, groups = var_3195, pad = hidden_states_31_pad_0, pad_type = hidden_states_31_pad_type_0, strides = var_3387, weight = layers_14_fc2_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("hidden_states_31_cast_fp16")]; + tensor inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = hidden_states_31_cast_fp16)[name = tensor("inputs_91_cast_fp16")]; + tensor var_3402 = const()[name = tensor("op_3402"), val = tensor(3)]; + tensor var_3409 = const()[name = tensor("op_3409"), val = tensor(1)]; + tensor var_3410 = const()[name = tensor("op_3410"), val = tensor(true)]; + tensor var_3422 = const()[name = tensor("op_3422"), val = tensor([1])]; + tensor channels_mean_91_cast_fp16 = reduce_mean(axes = var_3422, keep_dims = var_3410, x = inputs_91_cast_fp16)[name = tensor("channels_mean_91_cast_fp16")]; + tensor zero_mean_91_cast_fp16 = sub(x = inputs_91_cast_fp16, y = channels_mean_91_cast_fp16)[name = tensor("zero_mean_91_cast_fp16")]; + tensor zero_mean_sq_91_cast_fp16 = mul(x = zero_mean_91_cast_fp16, y = zero_mean_91_cast_fp16)[name = tensor("zero_mean_sq_91_cast_fp16")]; + tensor var_3426 = const()[name = tensor("op_3426"), val = tensor([1])]; + tensor var_3427_cast_fp16 = reduce_mean(axes = var_3426, keep_dims = var_3410, x = zero_mean_sq_91_cast_fp16)[name = tensor("op_3427_cast_fp16")]; + tensor var_3428_to_fp16 = const()[name = tensor("op_3428_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3429_cast_fp16 = add(x = var_3427_cast_fp16, y = var_3428_to_fp16)[name = tensor("op_3429_cast_fp16")]; + tensor denom_91_epsilon_0_to_fp16 = const()[name = tensor("denom_91_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_91_cast_fp16 = rsqrt(epsilon = denom_91_epsilon_0_to_fp16, x = var_3429_cast_fp16)[name = tensor("denom_91_cast_fp16")]; + tensor out_91_cast_fp16 = mul(x = zero_mean_91_cast_fp16, y = denom_91_cast_fp16)[name = tensor("out_91_cast_fp16")]; + tensor obj_211_gamma_0_to_fp16 = const()[name = tensor("obj_211_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(921037120)))]; + tensor obj_211_beta_0_to_fp16 = const()[name = tensor("obj_211_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(921039744)))]; + tensor obj_211_epsilon_0_to_fp16 = const()[name = tensor("obj_211_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_211_cast_fp16 = batch_norm(beta = obj_211_beta_0_to_fp16, epsilon = obj_211_epsilon_0_to_fp16, gamma = obj_211_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_91_cast_fp16)[name = tensor("obj_211_cast_fp16")]; + tensor var_3444 = const()[name = tensor("op_3444"), val = tensor([1, 1])]; + tensor var_3446 = const()[name = tensor("op_3446"), val = tensor([1, 1])]; + tensor query_61_pad_type_0 = const()[name = tensor("query_61_pad_type_0"), val = tensor("custom")]; + tensor query_61_pad_0 = const()[name = tensor("query_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(921042368)))]; + tensor layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(924319232)))]; + tensor query_61_cast_fp16 = conv(bias = layers_15_self_attn_q_proj_bias_to_fp16, dilations = var_3446, groups = var_3409, pad = query_61_pad_0, pad_type = query_61_pad_type_0, strides = var_3444, weight = layers_15_self_attn_q_proj_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor("query_61_cast_fp16")]; + tensor var_3450 = const()[name = tensor("op_3450"), val = tensor([1, 1])]; + tensor var_3452 = const()[name = tensor("op_3452"), val = tensor([1, 1])]; + tensor current_key_31_pad_type_0 = const()[name = tensor("current_key_31_pad_type_0"), val = tensor("custom")]; + tensor current_key_31_pad_0 = const()[name = tensor("current_key_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(924321856)))]; + tensor current_key_31_cast_fp16 = conv(dilations = var_3452, groups = var_3409, pad = current_key_31_pad_0, pad_type = current_key_31_pad_type_0, strides = var_3450, weight = layers_15_self_attn_k_proj_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor("current_key_31_cast_fp16")]; + tensor var_3457 = const()[name = tensor("op_3457"), val = tensor([1, 1])]; + tensor var_3459 = const()[name = tensor("op_3459"), val = tensor([1, 1])]; + tensor current_value_31_pad_type_0 = const()[name = tensor("current_value_31_pad_type_0"), val = tensor("custom")]; + tensor current_value_31_pad_0 = const()[name = tensor("current_value_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(927598720)))]; + tensor layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(930875584)))]; + tensor current_value_31_cast_fp16 = conv(bias = layers_15_self_attn_v_proj_bias_to_fp16, dilations = var_3459, groups = var_3409, pad = current_value_31_pad_0, pad_type = current_value_31_pad_type_0, strides = var_3457, weight = layers_15_self_attn_v_proj_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor("current_value_31_cast_fp16")]; + tensor var_3466_cast_fp16 = mul(x = current_key_31_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3466_cast_fp16")]; + tensor var_3468_cast_fp16 = mul(x = var_103_cast_fp16_15, y = var_241_cast_fp16)[name = tensor("op_3468_cast_fp16")]; + tensor key_61_cast_fp16 = add(x = var_3466_cast_fp16, y = var_3468_cast_fp16)[name = tensor("key_61_cast_fp16")]; + tensor var_3470_cast_fp16 = mul(x = current_value_31_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3470_cast_fp16")]; + tensor var_3472_cast_fp16 = mul(x = var_138_cast_fp16_15, y = var_241_cast_fp16)[name = tensor("op_3472_cast_fp16")]; + tensor value_61_cast_fp16 = add(x = var_3470_cast_fp16, y = var_3472_cast_fp16)[name = tensor("value_61_cast_fp16")]; + tensor var_3475 = const()[name = tensor("op_3475"), val = tensor([1, 20, 64, -1])]; + tensor var_3476_cast_fp16 = reshape(shape = var_3475, x = query_61_cast_fp16)[name = tensor("op_3476_cast_fp16")]; + tensor var_3477_to_fp16 = const()[name = tensor("op_3477_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3478_cast_fp16 = mul(x = var_3476_cast_fp16, y = var_3477_to_fp16)[name = tensor("op_3478_cast_fp16")]; + tensor var_3479 = const()[name = tensor("op_3479"), val = tensor([1, 20, 64, -1])]; + tensor var_3480_cast_fp16 = reshape(shape = var_3479, x = key_61_cast_fp16)[name = tensor("op_3480_cast_fp16")]; + tensor mh_w_91_transpose_x_0 = const()[name = tensor("mh_w_91_transpose_x_0"), val = tensor(true)]; + tensor mh_w_91_transpose_y_0 = const()[name = tensor("mh_w_91_transpose_y_0"), val = tensor(false)]; + tensor mh_w_91_cast_fp16 = matmul(transpose_x = mh_w_91_transpose_x_0, transpose_y = mh_w_91_transpose_y_0, x = var_3478_cast_fp16, y = var_3480_cast_fp16)[name = tensor("mh_w_91_cast_fp16")]; + tensor mh_w_93_cast_fp16 = add(x = mh_w_91_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_93_cast_fp16")]; + tensor var_3488_cast_fp16 = softmax(axis = var_3402, x = mh_w_93_cast_fp16)[name = tensor("op_3488_cast_fp16")]; + tensor var_3489 = const()[name = tensor("op_3489"), val = tensor([1, 20, 64, -1])]; + tensor var_3490_cast_fp16 = reshape(shape = var_3489, x = value_61_cast_fp16)[name = tensor("op_3490_cast_fp16")]; + tensor attn_61_transpose_x_0 = const()[name = tensor("attn_61_transpose_x_0"), val = tensor(false)]; + tensor attn_61_transpose_y_0 = const()[name = tensor("attn_61_transpose_y_0"), val = tensor(true)]; + tensor attn_61_cast_fp16 = matmul(transpose_x = attn_61_transpose_x_0, transpose_y = attn_61_transpose_y_0, x = var_3490_cast_fp16, y = var_3488_cast_fp16)[name = tensor("attn_61_cast_fp16")]; + tensor var_3493 = const()[name = tensor("op_3493"), val = tensor([1, 1280, 1, -1])]; + tensor input_151_cast_fp16 = reshape(shape = var_3493, x = attn_61_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor var_3497 = const()[name = tensor("op_3497"), val = tensor([1, 1])]; + tensor var_3499 = const()[name = tensor("op_3499"), val = tensor([1, 1])]; + tensor obj_217_pad_type_0 = const()[name = tensor("obj_217_pad_type_0"), val = tensor("custom")]; + tensor obj_217_pad_0 = const()[name = tensor("obj_217_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(930878208)))]; + tensor layers_15_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(934155072)))]; + tensor obj_217_cast_fp16 = conv(bias = layers_15_self_attn_o_proj_bias_to_fp16, dilations = var_3499, groups = var_3409, pad = obj_217_pad_0, pad_type = obj_217_pad_type_0, strides = var_3497, weight = layers_15_self_attn_o_proj_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("obj_217_cast_fp16")]; + tensor inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = obj_217_cast_fp16)[name = tensor("inputs_93_cast_fp16")]; + tensor var_3509 = const()[name = tensor("op_3509"), val = tensor([1])]; + tensor channels_mean_93_cast_fp16 = reduce_mean(axes = var_3509, keep_dims = var_3410, x = inputs_93_cast_fp16)[name = tensor("channels_mean_93_cast_fp16")]; + tensor zero_mean_93_cast_fp16 = sub(x = inputs_93_cast_fp16, y = channels_mean_93_cast_fp16)[name = tensor("zero_mean_93_cast_fp16")]; + tensor zero_mean_sq_93_cast_fp16 = mul(x = zero_mean_93_cast_fp16, y = zero_mean_93_cast_fp16)[name = tensor("zero_mean_sq_93_cast_fp16")]; + tensor var_3513 = const()[name = tensor("op_3513"), val = tensor([1])]; + tensor var_3514_cast_fp16 = reduce_mean(axes = var_3513, keep_dims = var_3410, x = zero_mean_sq_93_cast_fp16)[name = tensor("op_3514_cast_fp16")]; + tensor var_3515_to_fp16 = const()[name = tensor("op_3515_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3516_cast_fp16 = add(x = var_3514_cast_fp16, y = var_3515_to_fp16)[name = tensor("op_3516_cast_fp16")]; + tensor denom_93_epsilon_0_to_fp16 = const()[name = tensor("denom_93_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_93_cast_fp16 = rsqrt(epsilon = denom_93_epsilon_0_to_fp16, x = var_3516_cast_fp16)[name = tensor("denom_93_cast_fp16")]; + tensor out_93_cast_fp16 = mul(x = zero_mean_93_cast_fp16, y = denom_93_cast_fp16)[name = tensor("out_93_cast_fp16")]; + tensor obj_219_gamma_0_to_fp16 = const()[name = tensor("obj_219_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(934157696)))]; + tensor obj_219_beta_0_to_fp16 = const()[name = tensor("obj_219_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(934160320)))]; + tensor obj_219_epsilon_0_to_fp16 = const()[name = tensor("obj_219_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_219_cast_fp16 = batch_norm(beta = obj_219_beta_0_to_fp16, epsilon = obj_219_epsilon_0_to_fp16, gamma = obj_219_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_93_cast_fp16)[name = tensor("obj_219_cast_fp16")]; + tensor var_3531 = const()[name = tensor("op_3531"), val = tensor([1, 1])]; + tensor var_3533 = const()[name = tensor("op_3533"), val = tensor([1, 1])]; + tensor query_63_pad_type_0 = const()[name = tensor("query_63_pad_type_0"), val = tensor("custom")]; + tensor query_63_pad_0 = const()[name = tensor("query_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(934162944)))]; + tensor layers_15_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(937439808)))]; + tensor query_63_cast_fp16 = conv(bias = layers_15_encoder_attn_q_proj_bias_to_fp16, dilations = var_3533, groups = var_3409, pad = query_63_pad_0, pad_type = query_63_pad_type_0, strides = var_3531, weight = layers_15_encoder_attn_q_proj_weight_to_fp16, x = obj_219_cast_fp16)[name = tensor("query_63_cast_fp16")]; + tensor var_3537 = const()[name = tensor("op_3537"), val = tensor([1, 1])]; + tensor var_3539 = const()[name = tensor("op_3539"), val = tensor([1, 1])]; + tensor key_63_pad_type_0 = const()[name = tensor("key_63_pad_type_0"), val = tensor("custom")]; + tensor key_63_pad_0 = const()[name = tensor("key_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(937442432)))]; + tensor key_63_cast_fp16 = conv(dilations = var_3539, groups = var_3409, pad = key_63_pad_0, pad_type = key_63_pad_type_0, strides = var_3537, weight = layers_15_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_63_cast_fp16")]; + tensor var_3544 = const()[name = tensor("op_3544"), val = tensor([1, 1])]; + tensor var_3546 = const()[name = tensor("op_3546"), val = tensor([1, 1])]; + tensor value_63_pad_type_0 = const()[name = tensor("value_63_pad_type_0"), val = tensor("custom")]; + tensor value_63_pad_0 = const()[name = tensor("value_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(940719296)))]; + tensor layers_15_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(943996160)))]; + tensor value_63_cast_fp16 = conv(bias = layers_15_encoder_attn_v_proj_bias_to_fp16, dilations = var_3546, groups = var_3409, pad = value_63_pad_0, pad_type = value_63_pad_type_0, strides = var_3544, weight = layers_15_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_63_cast_fp16")]; + tensor var_3550 = const()[name = tensor("op_3550"), val = tensor([1, 20, 64, -1])]; + tensor var_3551_cast_fp16 = reshape(shape = var_3550, x = query_63_cast_fp16)[name = tensor("op_3551_cast_fp16")]; + tensor var_3552_to_fp16 = const()[name = tensor("op_3552_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3553_cast_fp16 = mul(x = var_3551_cast_fp16, y = var_3552_to_fp16)[name = tensor("op_3553_cast_fp16")]; + tensor var_3554 = const()[name = tensor("op_3554"), val = tensor([1, 20, 64, -1])]; + tensor var_3555_cast_fp16 = reshape(shape = var_3554, x = key_63_cast_fp16)[name = tensor("op_3555_cast_fp16")]; + tensor mh_w_95_transpose_x_0 = const()[name = tensor("mh_w_95_transpose_x_0"), val = tensor(true)]; + tensor mh_w_95_transpose_y_0 = const()[name = tensor("mh_w_95_transpose_y_0"), val = tensor(false)]; + tensor mh_w_95_cast_fp16 = matmul(transpose_x = mh_w_95_transpose_x_0, transpose_y = mh_w_95_transpose_y_0, x = var_3553_cast_fp16, y = var_3555_cast_fp16)[name = tensor("mh_w_95_cast_fp16")]; + tensor obj_223_cast_fp16 = softmax(axis = var_3402, x = mh_w_95_cast_fp16)[name = tensor("obj_223_cast_fp16")]; + tensor var_3559 = const()[name = tensor("op_3559"), val = tensor([1, 20, 64, -1])]; + tensor var_3560_cast_fp16 = reshape(shape = var_3559, x = value_63_cast_fp16)[name = tensor("op_3560_cast_fp16")]; + tensor attn_63_transpose_x_0 = const()[name = tensor("attn_63_transpose_x_0"), val = tensor(false)]; + tensor attn_63_transpose_y_0 = const()[name = tensor("attn_63_transpose_y_0"), val = tensor(true)]; + tensor attn_63_cast_fp16 = matmul(transpose_x = attn_63_transpose_x_0, transpose_y = attn_63_transpose_y_0, x = var_3560_cast_fp16, y = obj_223_cast_fp16)[name = tensor("attn_63_cast_fp16")]; + tensor var_3563 = const()[name = tensor("op_3563"), val = tensor([1, 1280, 1, -1])]; + tensor input_153_cast_fp16 = reshape(shape = var_3563, x = attn_63_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor var_3567 = const()[name = tensor("op_3567"), val = tensor([1, 1])]; + tensor var_3569 = const()[name = tensor("op_3569"), val = tensor([1, 1])]; + tensor obj_221_pad_type_0 = const()[name = tensor("obj_221_pad_type_0"), val = tensor("custom")]; + tensor obj_221_pad_0 = const()[name = tensor("obj_221_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(943998784)))]; + tensor layers_15_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(947275648)))]; + tensor obj_221_cast_fp16 = conv(bias = layers_15_encoder_attn_o_proj_bias_to_fp16, dilations = var_3569, groups = var_3409, pad = obj_221_pad_0, pad_type = obj_221_pad_type_0, strides = var_3567, weight = layers_15_encoder_attn_o_proj_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("obj_221_cast_fp16")]; + tensor inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = obj_221_cast_fp16)[name = tensor("inputs_95_cast_fp16")]; + tensor var_3575 = const()[name = tensor("op_3575"), val = tensor([1])]; + tensor channels_mean_95_cast_fp16 = reduce_mean(axes = var_3575, keep_dims = var_3410, x = inputs_95_cast_fp16)[name = tensor("channels_mean_95_cast_fp16")]; + tensor zero_mean_95_cast_fp16 = sub(x = inputs_95_cast_fp16, y = channels_mean_95_cast_fp16)[name = tensor("zero_mean_95_cast_fp16")]; + tensor zero_mean_sq_95_cast_fp16 = mul(x = zero_mean_95_cast_fp16, y = zero_mean_95_cast_fp16)[name = tensor("zero_mean_sq_95_cast_fp16")]; + tensor var_3579 = const()[name = tensor("op_3579"), val = tensor([1])]; + tensor var_3580_cast_fp16 = reduce_mean(axes = var_3579, keep_dims = var_3410, x = zero_mean_sq_95_cast_fp16)[name = tensor("op_3580_cast_fp16")]; + tensor var_3581_to_fp16 = const()[name = tensor("op_3581_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3582_cast_fp16 = add(x = var_3580_cast_fp16, y = var_3581_to_fp16)[name = tensor("op_3582_cast_fp16")]; + tensor denom_95_epsilon_0_to_fp16 = const()[name = tensor("denom_95_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_95_cast_fp16 = rsqrt(epsilon = denom_95_epsilon_0_to_fp16, x = var_3582_cast_fp16)[name = tensor("denom_95_cast_fp16")]; + tensor out_95_cast_fp16 = mul(x = zero_mean_95_cast_fp16, y = denom_95_cast_fp16)[name = tensor("out_95_cast_fp16")]; + tensor input_155_gamma_0_to_fp16 = const()[name = tensor("input_155_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(947278272)))]; + tensor input_155_beta_0_to_fp16 = const()[name = tensor("input_155_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(947280896)))]; + tensor input_155_epsilon_0_to_fp16 = const()[name = tensor("input_155_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_155_cast_fp16 = batch_norm(beta = input_155_beta_0_to_fp16, epsilon = input_155_epsilon_0_to_fp16, gamma = input_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_95_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor var_3593 = const()[name = tensor("op_3593"), val = tensor([1, 1])]; + tensor var_3595 = const()[name = tensor("op_3595"), val = tensor([1, 1])]; + tensor input_157_pad_type_0 = const()[name = tensor("input_157_pad_type_0"), val = tensor("custom")]; + tensor input_157_pad_0 = const()[name = tensor("input_157_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_fc1_weight_to_fp16 = const()[name = tensor("layers_15_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(947283520)))]; + tensor layers_15_fc1_bias_to_fp16 = const()[name = tensor("layers_15_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(960390784)))]; + tensor input_157_cast_fp16 = conv(bias = layers_15_fc1_bias_to_fp16, dilations = var_3595, groups = var_3409, pad = input_157_pad_0, pad_type = input_157_pad_type_0, strides = var_3593, weight = layers_15_fc1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_mode_0 = const()[name = tensor("input_159_mode_0"), val = tensor("EXACT")]; + tensor input_159_cast_fp16 = gelu(mode = input_159_mode_0, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor var_3601 = const()[name = tensor("op_3601"), val = tensor([1, 1])]; + tensor var_3603 = const()[name = tensor("op_3603"), val = tensor([1, 1])]; + tensor hidden_states_33_pad_type_0 = const()[name = tensor("hidden_states_33_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_33_pad_0 = const()[name = tensor("hidden_states_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_fc2_weight_to_fp16 = const()[name = tensor("layers_15_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(960401088)))]; + tensor layers_15_fc2_bias_to_fp16 = const()[name = tensor("layers_15_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(973508352)))]; + tensor hidden_states_33_cast_fp16 = conv(bias = layers_15_fc2_bias_to_fp16, dilations = var_3603, groups = var_3409, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = var_3601, weight = layers_15_fc2_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("hidden_states_33_cast_fp16")]; + tensor inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = hidden_states_33_cast_fp16)[name = tensor("inputs_97_cast_fp16")]; + tensor var_3616 = const()[name = tensor("op_3616"), val = tensor(3)]; + tensor var_3623 = const()[name = tensor("op_3623"), val = tensor(1)]; + tensor var_3624 = const()[name = tensor("op_3624"), val = tensor(true)]; + tensor var_3636 = const()[name = tensor("op_3636"), val = tensor([1])]; + tensor channels_mean_97_cast_fp16 = reduce_mean(axes = var_3636, keep_dims = var_3624, x = inputs_97_cast_fp16)[name = tensor("channels_mean_97_cast_fp16")]; + tensor zero_mean_97_cast_fp16 = sub(x = inputs_97_cast_fp16, y = channels_mean_97_cast_fp16)[name = tensor("zero_mean_97_cast_fp16")]; + tensor zero_mean_sq_97_cast_fp16 = mul(x = zero_mean_97_cast_fp16, y = zero_mean_97_cast_fp16)[name = tensor("zero_mean_sq_97_cast_fp16")]; + tensor var_3640 = const()[name = tensor("op_3640"), val = tensor([1])]; + tensor var_3641_cast_fp16 = reduce_mean(axes = var_3640, keep_dims = var_3624, x = zero_mean_sq_97_cast_fp16)[name = tensor("op_3641_cast_fp16")]; + tensor var_3642_to_fp16 = const()[name = tensor("op_3642_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3643_cast_fp16 = add(x = var_3641_cast_fp16, y = var_3642_to_fp16)[name = tensor("op_3643_cast_fp16")]; + tensor denom_97_epsilon_0_to_fp16 = const()[name = tensor("denom_97_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_97_cast_fp16 = rsqrt(epsilon = denom_97_epsilon_0_to_fp16, x = var_3643_cast_fp16)[name = tensor("denom_97_cast_fp16")]; + tensor out_97_cast_fp16 = mul(x = zero_mean_97_cast_fp16, y = denom_97_cast_fp16)[name = tensor("out_97_cast_fp16")]; + tensor obj_225_gamma_0_to_fp16 = const()[name = tensor("obj_225_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(973510976)))]; + tensor obj_225_beta_0_to_fp16 = const()[name = tensor("obj_225_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(973513600)))]; + tensor obj_225_epsilon_0_to_fp16 = const()[name = tensor("obj_225_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_225_cast_fp16 = batch_norm(beta = obj_225_beta_0_to_fp16, epsilon = obj_225_epsilon_0_to_fp16, gamma = obj_225_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_97_cast_fp16)[name = tensor("obj_225_cast_fp16")]; + tensor var_3658 = const()[name = tensor("op_3658"), val = tensor([1, 1])]; + tensor var_3660 = const()[name = tensor("op_3660"), val = tensor([1, 1])]; + tensor query_65_pad_type_0 = const()[name = tensor("query_65_pad_type_0"), val = tensor("custom")]; + tensor query_65_pad_0 = const()[name = tensor("query_65_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(973516224)))]; + tensor layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(976793088)))]; + tensor query_65_cast_fp16 = conv(bias = layers_16_self_attn_q_proj_bias_to_fp16, dilations = var_3660, groups = var_3623, pad = query_65_pad_0, pad_type = query_65_pad_type_0, strides = var_3658, weight = layers_16_self_attn_q_proj_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("query_65_cast_fp16")]; + tensor var_3664 = const()[name = tensor("op_3664"), val = tensor([1, 1])]; + tensor var_3666 = const()[name = tensor("op_3666"), val = tensor([1, 1])]; + tensor current_key_33_pad_type_0 = const()[name = tensor("current_key_33_pad_type_0"), val = tensor("custom")]; + tensor current_key_33_pad_0 = const()[name = tensor("current_key_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(976795712)))]; + tensor current_key_33_cast_fp16 = conv(dilations = var_3666, groups = var_3623, pad = current_key_33_pad_0, pad_type = current_key_33_pad_type_0, strides = var_3664, weight = layers_16_self_attn_k_proj_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("current_key_33_cast_fp16")]; + tensor var_3671 = const()[name = tensor("op_3671"), val = tensor([1, 1])]; + tensor var_3673 = const()[name = tensor("op_3673"), val = tensor([1, 1])]; + tensor current_value_33_pad_type_0 = const()[name = tensor("current_value_33_pad_type_0"), val = tensor("custom")]; + tensor current_value_33_pad_0 = const()[name = tensor("current_value_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(980072576)))]; + tensor layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(983349440)))]; + tensor current_value_33_cast_fp16 = conv(bias = layers_16_self_attn_v_proj_bias_to_fp16, dilations = var_3673, groups = var_3623, pad = current_value_33_pad_0, pad_type = current_value_33_pad_type_0, strides = var_3671, weight = layers_16_self_attn_v_proj_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("current_value_33_cast_fp16")]; + tensor var_3680_cast_fp16 = mul(x = current_key_33_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3680_cast_fp16")]; + tensor var_3682_cast_fp16 = mul(x = var_103_cast_fp16_16, y = var_241_cast_fp16)[name = tensor("op_3682_cast_fp16")]; + tensor key_65_cast_fp16 = add(x = var_3680_cast_fp16, y = var_3682_cast_fp16)[name = tensor("key_65_cast_fp16")]; + tensor var_3684_cast_fp16 = mul(x = current_value_33_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3684_cast_fp16")]; + tensor var_3686_cast_fp16 = mul(x = var_138_cast_fp16_16, y = var_241_cast_fp16)[name = tensor("op_3686_cast_fp16")]; + tensor value_65_cast_fp16 = add(x = var_3684_cast_fp16, y = var_3686_cast_fp16)[name = tensor("value_65_cast_fp16")]; + tensor var_3689 = const()[name = tensor("op_3689"), val = tensor([1, 20, 64, -1])]; + tensor var_3690_cast_fp16 = reshape(shape = var_3689, x = query_65_cast_fp16)[name = tensor("op_3690_cast_fp16")]; + tensor var_3691_to_fp16 = const()[name = tensor("op_3691_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3692_cast_fp16 = mul(x = var_3690_cast_fp16, y = var_3691_to_fp16)[name = tensor("op_3692_cast_fp16")]; + tensor var_3693 = const()[name = tensor("op_3693"), val = tensor([1, 20, 64, -1])]; + tensor var_3694_cast_fp16 = reshape(shape = var_3693, x = key_65_cast_fp16)[name = tensor("op_3694_cast_fp16")]; + tensor mh_w_97_transpose_x_0 = const()[name = tensor("mh_w_97_transpose_x_0"), val = tensor(true)]; + tensor mh_w_97_transpose_y_0 = const()[name = tensor("mh_w_97_transpose_y_0"), val = tensor(false)]; + tensor mh_w_97_cast_fp16 = matmul(transpose_x = mh_w_97_transpose_x_0, transpose_y = mh_w_97_transpose_y_0, x = var_3692_cast_fp16, y = var_3694_cast_fp16)[name = tensor("mh_w_97_cast_fp16")]; + tensor mh_w_99_cast_fp16 = add(x = mh_w_97_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_99_cast_fp16")]; + tensor var_3702_cast_fp16 = softmax(axis = var_3616, x = mh_w_99_cast_fp16)[name = tensor("op_3702_cast_fp16")]; + tensor var_3703 = const()[name = tensor("op_3703"), val = tensor([1, 20, 64, -1])]; + tensor var_3704_cast_fp16 = reshape(shape = var_3703, x = value_65_cast_fp16)[name = tensor("op_3704_cast_fp16")]; + tensor attn_65_transpose_x_0 = const()[name = tensor("attn_65_transpose_x_0"), val = tensor(false)]; + tensor attn_65_transpose_y_0 = const()[name = tensor("attn_65_transpose_y_0"), val = tensor(true)]; + tensor attn_65_cast_fp16 = matmul(transpose_x = attn_65_transpose_x_0, transpose_y = attn_65_transpose_y_0, x = var_3704_cast_fp16, y = var_3702_cast_fp16)[name = tensor("attn_65_cast_fp16")]; + tensor var_3707 = const()[name = tensor("op_3707"), val = tensor([1, 1280, 1, -1])]; + tensor input_161_cast_fp16 = reshape(shape = var_3707, x = attn_65_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor var_3711 = const()[name = tensor("op_3711"), val = tensor([1, 1])]; + tensor var_3713 = const()[name = tensor("op_3713"), val = tensor([1, 1])]; + tensor obj_231_pad_type_0 = const()[name = tensor("obj_231_pad_type_0"), val = tensor("custom")]; + tensor obj_231_pad_0 = const()[name = tensor("obj_231_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(983352064)))]; + tensor layers_16_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(986628928)))]; + tensor obj_231_cast_fp16 = conv(bias = layers_16_self_attn_o_proj_bias_to_fp16, dilations = var_3713, groups = var_3623, pad = obj_231_pad_0, pad_type = obj_231_pad_type_0, strides = var_3711, weight = layers_16_self_attn_o_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("obj_231_cast_fp16")]; + tensor inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = obj_231_cast_fp16)[name = tensor("inputs_99_cast_fp16")]; + tensor var_3723 = const()[name = tensor("op_3723"), val = tensor([1])]; + tensor channels_mean_99_cast_fp16 = reduce_mean(axes = var_3723, keep_dims = var_3624, x = inputs_99_cast_fp16)[name = tensor("channels_mean_99_cast_fp16")]; + tensor zero_mean_99_cast_fp16 = sub(x = inputs_99_cast_fp16, y = channels_mean_99_cast_fp16)[name = tensor("zero_mean_99_cast_fp16")]; + tensor zero_mean_sq_99_cast_fp16 = mul(x = zero_mean_99_cast_fp16, y = zero_mean_99_cast_fp16)[name = tensor("zero_mean_sq_99_cast_fp16")]; + tensor var_3727 = const()[name = tensor("op_3727"), val = tensor([1])]; + tensor var_3728_cast_fp16 = reduce_mean(axes = var_3727, keep_dims = var_3624, x = zero_mean_sq_99_cast_fp16)[name = tensor("op_3728_cast_fp16")]; + tensor var_3729_to_fp16 = const()[name = tensor("op_3729_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3730_cast_fp16 = add(x = var_3728_cast_fp16, y = var_3729_to_fp16)[name = tensor("op_3730_cast_fp16")]; + tensor denom_99_epsilon_0_to_fp16 = const()[name = tensor("denom_99_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_99_cast_fp16 = rsqrt(epsilon = denom_99_epsilon_0_to_fp16, x = var_3730_cast_fp16)[name = tensor("denom_99_cast_fp16")]; + tensor out_99_cast_fp16 = mul(x = zero_mean_99_cast_fp16, y = denom_99_cast_fp16)[name = tensor("out_99_cast_fp16")]; + tensor obj_233_gamma_0_to_fp16 = const()[name = tensor("obj_233_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(986631552)))]; + tensor obj_233_beta_0_to_fp16 = const()[name = tensor("obj_233_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(986634176)))]; + tensor obj_233_epsilon_0_to_fp16 = const()[name = tensor("obj_233_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_233_cast_fp16 = batch_norm(beta = obj_233_beta_0_to_fp16, epsilon = obj_233_epsilon_0_to_fp16, gamma = obj_233_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_99_cast_fp16)[name = tensor("obj_233_cast_fp16")]; + tensor var_3745 = const()[name = tensor("op_3745"), val = tensor([1, 1])]; + tensor var_3747 = const()[name = tensor("op_3747"), val = tensor([1, 1])]; + tensor query_67_pad_type_0 = const()[name = tensor("query_67_pad_type_0"), val = tensor("custom")]; + tensor query_67_pad_0 = const()[name = tensor("query_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(986636800)))]; + tensor layers_16_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(989913664)))]; + tensor query_67_cast_fp16 = conv(bias = layers_16_encoder_attn_q_proj_bias_to_fp16, dilations = var_3747, groups = var_3623, pad = query_67_pad_0, pad_type = query_67_pad_type_0, strides = var_3745, weight = layers_16_encoder_attn_q_proj_weight_to_fp16, x = obj_233_cast_fp16)[name = tensor("query_67_cast_fp16")]; + tensor var_3751 = const()[name = tensor("op_3751"), val = tensor([1, 1])]; + tensor var_3753 = const()[name = tensor("op_3753"), val = tensor([1, 1])]; + tensor key_67_pad_type_0 = const()[name = tensor("key_67_pad_type_0"), val = tensor("custom")]; + tensor key_67_pad_0 = const()[name = tensor("key_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(989916288)))]; + tensor key_67_cast_fp16 = conv(dilations = var_3753, groups = var_3623, pad = key_67_pad_0, pad_type = key_67_pad_type_0, strides = var_3751, weight = layers_16_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_67_cast_fp16")]; + tensor var_3758 = const()[name = tensor("op_3758"), val = tensor([1, 1])]; + tensor var_3760 = const()[name = tensor("op_3760"), val = tensor([1, 1])]; + tensor value_67_pad_type_0 = const()[name = tensor("value_67_pad_type_0"), val = tensor("custom")]; + tensor value_67_pad_0 = const()[name = tensor("value_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(993193152)))]; + tensor layers_16_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(996470016)))]; + tensor value_67_cast_fp16 = conv(bias = layers_16_encoder_attn_v_proj_bias_to_fp16, dilations = var_3760, groups = var_3623, pad = value_67_pad_0, pad_type = value_67_pad_type_0, strides = var_3758, weight = layers_16_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_67_cast_fp16")]; + tensor var_3764 = const()[name = tensor("op_3764"), val = tensor([1, 20, 64, -1])]; + tensor var_3765_cast_fp16 = reshape(shape = var_3764, x = query_67_cast_fp16)[name = tensor("op_3765_cast_fp16")]; + tensor var_3766_to_fp16 = const()[name = tensor("op_3766_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3767_cast_fp16 = mul(x = var_3765_cast_fp16, y = var_3766_to_fp16)[name = tensor("op_3767_cast_fp16")]; + tensor var_3768 = const()[name = tensor("op_3768"), val = tensor([1, 20, 64, -1])]; + tensor var_3769_cast_fp16 = reshape(shape = var_3768, x = key_67_cast_fp16)[name = tensor("op_3769_cast_fp16")]; + tensor mh_w_101_transpose_x_0 = const()[name = tensor("mh_w_101_transpose_x_0"), val = tensor(true)]; + tensor mh_w_101_transpose_y_0 = const()[name = tensor("mh_w_101_transpose_y_0"), val = tensor(false)]; + tensor mh_w_101_cast_fp16 = matmul(transpose_x = mh_w_101_transpose_x_0, transpose_y = mh_w_101_transpose_y_0, x = var_3767_cast_fp16, y = var_3769_cast_fp16)[name = tensor("mh_w_101_cast_fp16")]; + tensor obj_237_cast_fp16 = softmax(axis = var_3616, x = mh_w_101_cast_fp16)[name = tensor("obj_237_cast_fp16")]; + tensor var_3773 = const()[name = tensor("op_3773"), val = tensor([1, 20, 64, -1])]; + tensor var_3774_cast_fp16 = reshape(shape = var_3773, x = value_67_cast_fp16)[name = tensor("op_3774_cast_fp16")]; + tensor attn_67_transpose_x_0 = const()[name = tensor("attn_67_transpose_x_0"), val = tensor(false)]; + tensor attn_67_transpose_y_0 = const()[name = tensor("attn_67_transpose_y_0"), val = tensor(true)]; + tensor attn_67_cast_fp16 = matmul(transpose_x = attn_67_transpose_x_0, transpose_y = attn_67_transpose_y_0, x = var_3774_cast_fp16, y = obj_237_cast_fp16)[name = tensor("attn_67_cast_fp16")]; + tensor var_3777 = const()[name = tensor("op_3777"), val = tensor([1, 1280, 1, -1])]; + tensor input_163_cast_fp16 = reshape(shape = var_3777, x = attn_67_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor var_3781 = const()[name = tensor("op_3781"), val = tensor([1, 1])]; + tensor var_3783 = const()[name = tensor("op_3783"), val = tensor([1, 1])]; + tensor obj_235_pad_type_0 = const()[name = tensor("obj_235_pad_type_0"), val = tensor("custom")]; + tensor obj_235_pad_0 = const()[name = tensor("obj_235_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(996472640)))]; + tensor layers_16_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(999749504)))]; + tensor obj_235_cast_fp16 = conv(bias = layers_16_encoder_attn_o_proj_bias_to_fp16, dilations = var_3783, groups = var_3623, pad = obj_235_pad_0, pad_type = obj_235_pad_type_0, strides = var_3781, weight = layers_16_encoder_attn_o_proj_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("obj_235_cast_fp16")]; + tensor inputs_101_cast_fp16 = add(x = inputs_99_cast_fp16, y = obj_235_cast_fp16)[name = tensor("inputs_101_cast_fp16")]; + tensor var_3792 = const()[name = tensor("op_3792"), val = tensor([1])]; + tensor channels_mean_101_cast_fp16 = reduce_mean(axes = var_3792, keep_dims = var_3624, x = inputs_101_cast_fp16)[name = tensor("channels_mean_101_cast_fp16")]; + tensor zero_mean_101_cast_fp16 = sub(x = inputs_101_cast_fp16, y = channels_mean_101_cast_fp16)[name = tensor("zero_mean_101_cast_fp16")]; + tensor zero_mean_sq_101_cast_fp16 = mul(x = zero_mean_101_cast_fp16, y = zero_mean_101_cast_fp16)[name = tensor("zero_mean_sq_101_cast_fp16")]; + tensor var_3796 = const()[name = tensor("op_3796"), val = tensor([1])]; + tensor var_3797_cast_fp16 = reduce_mean(axes = var_3796, keep_dims = var_3624, x = zero_mean_sq_101_cast_fp16)[name = tensor("op_3797_cast_fp16")]; + tensor var_3798_to_fp16 = const()[name = tensor("op_3798_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3799_cast_fp16 = add(x = var_3797_cast_fp16, y = var_3798_to_fp16)[name = tensor("op_3799_cast_fp16")]; + tensor denom_101_epsilon_0_to_fp16 = const()[name = tensor("denom_101_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_101_cast_fp16 = rsqrt(epsilon = denom_101_epsilon_0_to_fp16, x = var_3799_cast_fp16)[name = tensor("denom_101_cast_fp16")]; + tensor out_101_cast_fp16 = mul(x = zero_mean_101_cast_fp16, y = denom_101_cast_fp16)[name = tensor("out_101_cast_fp16")]; + tensor input_165_gamma_0_to_fp16 = const()[name = tensor("input_165_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(999752128)))]; + tensor input_165_beta_0_to_fp16 = const()[name = tensor("input_165_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(999754752)))]; + tensor input_165_epsilon_0_to_fp16 = const()[name = tensor("input_165_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_165_cast_fp16 = batch_norm(beta = input_165_beta_0_to_fp16, epsilon = input_165_epsilon_0_to_fp16, gamma = input_165_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_101_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor var_3810 = const()[name = tensor("op_3810"), val = tensor([1, 1])]; + tensor var_3812 = const()[name = tensor("op_3812"), val = tensor([1, 1])]; + tensor input_167_pad_type_0 = const()[name = tensor("input_167_pad_type_0"), val = tensor("custom")]; + tensor input_167_pad_0 = const()[name = tensor("input_167_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_fc1_weight_to_fp16 = const()[name = tensor("layers_16_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(999757376)))]; + tensor layers_16_fc1_bias_to_fp16 = const()[name = tensor("layers_16_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1012864640)))]; + tensor input_167_cast_fp16 = conv(bias = layers_16_fc1_bias_to_fp16, dilations = var_3812, groups = var_3623, pad = input_167_pad_0, pad_type = input_167_pad_type_0, strides = var_3810, weight = layers_16_fc1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor input_169_mode_0 = const()[name = tensor("input_169_mode_0"), val = tensor("EXACT")]; + tensor input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor var_3818 = const()[name = tensor("op_3818"), val = tensor([1, 1])]; + tensor var_3820 = const()[name = tensor("op_3820"), val = tensor([1, 1])]; + tensor hidden_states_35_pad_type_0 = const()[name = tensor("hidden_states_35_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_35_pad_0 = const()[name = tensor("hidden_states_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_fc2_weight_to_fp16 = const()[name = tensor("layers_16_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1012874944)))]; + tensor layers_16_fc2_bias_to_fp16 = const()[name = tensor("layers_16_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1025982208)))]; + tensor hidden_states_35_cast_fp16 = conv(bias = layers_16_fc2_bias_to_fp16, dilations = var_3820, groups = var_3623, pad = hidden_states_35_pad_0, pad_type = hidden_states_35_pad_type_0, strides = var_3818, weight = layers_16_fc2_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("hidden_states_35_cast_fp16")]; + tensor inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = hidden_states_35_cast_fp16)[name = tensor("inputs_103_cast_fp16")]; + tensor var_3834 = const()[name = tensor("op_3834"), val = tensor(3)]; + tensor var_3841 = const()[name = tensor("op_3841"), val = tensor(1)]; + tensor var_3842 = const()[name = tensor("op_3842"), val = tensor(true)]; + tensor var_3854 = const()[name = tensor("op_3854"), val = tensor([1])]; + tensor channels_mean_103_cast_fp16 = reduce_mean(axes = var_3854, keep_dims = var_3842, x = inputs_103_cast_fp16)[name = tensor("channels_mean_103_cast_fp16")]; + tensor zero_mean_103_cast_fp16 = sub(x = inputs_103_cast_fp16, y = channels_mean_103_cast_fp16)[name = tensor("zero_mean_103_cast_fp16")]; + tensor zero_mean_sq_103_cast_fp16 = mul(x = zero_mean_103_cast_fp16, y = zero_mean_103_cast_fp16)[name = tensor("zero_mean_sq_103_cast_fp16")]; + tensor var_3858 = const()[name = tensor("op_3858"), val = tensor([1])]; + tensor var_3859_cast_fp16 = reduce_mean(axes = var_3858, keep_dims = var_3842, x = zero_mean_sq_103_cast_fp16)[name = tensor("op_3859_cast_fp16")]; + tensor var_3860_to_fp16 = const()[name = tensor("op_3860_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3861_cast_fp16 = add(x = var_3859_cast_fp16, y = var_3860_to_fp16)[name = tensor("op_3861_cast_fp16")]; + tensor denom_103_epsilon_0_to_fp16 = const()[name = tensor("denom_103_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_103_cast_fp16 = rsqrt(epsilon = denom_103_epsilon_0_to_fp16, x = var_3861_cast_fp16)[name = tensor("denom_103_cast_fp16")]; + tensor out_103_cast_fp16 = mul(x = zero_mean_103_cast_fp16, y = denom_103_cast_fp16)[name = tensor("out_103_cast_fp16")]; + tensor obj_239_gamma_0_to_fp16 = const()[name = tensor("obj_239_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1025984832)))]; + tensor obj_239_beta_0_to_fp16 = const()[name = tensor("obj_239_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1025987456)))]; + tensor obj_239_epsilon_0_to_fp16 = const()[name = tensor("obj_239_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_239_cast_fp16 = batch_norm(beta = obj_239_beta_0_to_fp16, epsilon = obj_239_epsilon_0_to_fp16, gamma = obj_239_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_103_cast_fp16)[name = tensor("obj_239_cast_fp16")]; + tensor var_3876 = const()[name = tensor("op_3876"), val = tensor([1, 1])]; + tensor var_3878 = const()[name = tensor("op_3878"), val = tensor([1, 1])]; + tensor query_69_pad_type_0 = const()[name = tensor("query_69_pad_type_0"), val = tensor("custom")]; + tensor query_69_pad_0 = const()[name = tensor("query_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1025990080)))]; + tensor layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1029266944)))]; + tensor query_69_cast_fp16 = conv(bias = layers_17_self_attn_q_proj_bias_to_fp16, dilations = var_3878, groups = var_3841, pad = query_69_pad_0, pad_type = query_69_pad_type_0, strides = var_3876, weight = layers_17_self_attn_q_proj_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor("query_69_cast_fp16")]; + tensor var_3882 = const()[name = tensor("op_3882"), val = tensor([1, 1])]; + tensor var_3884 = const()[name = tensor("op_3884"), val = tensor([1, 1])]; + tensor current_key_35_pad_type_0 = const()[name = tensor("current_key_35_pad_type_0"), val = tensor("custom")]; + tensor current_key_35_pad_0 = const()[name = tensor("current_key_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1029269568)))]; + tensor current_key_35_cast_fp16 = conv(dilations = var_3884, groups = var_3841, pad = current_key_35_pad_0, pad_type = current_key_35_pad_type_0, strides = var_3882, weight = layers_17_self_attn_k_proj_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor("current_key_35_cast_fp16")]; + tensor var_3889 = const()[name = tensor("op_3889"), val = tensor([1, 1])]; + tensor var_3891 = const()[name = tensor("op_3891"), val = tensor([1, 1])]; + tensor current_value_35_pad_type_0 = const()[name = tensor("current_value_35_pad_type_0"), val = tensor("custom")]; + tensor current_value_35_pad_0 = const()[name = tensor("current_value_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1032546432)))]; + tensor layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1035823296)))]; + tensor current_value_35_cast_fp16 = conv(bias = layers_17_self_attn_v_proj_bias_to_fp16, dilations = var_3891, groups = var_3841, pad = current_value_35_pad_0, pad_type = current_value_35_pad_type_0, strides = var_3889, weight = layers_17_self_attn_v_proj_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor("current_value_35_cast_fp16")]; + tensor var_3898_cast_fp16 = mul(x = current_key_35_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3898_cast_fp16")]; + tensor var_3900_cast_fp16 = mul(x = var_103_cast_fp16_17, y = var_241_cast_fp16)[name = tensor("op_3900_cast_fp16")]; + tensor key_69_cast_fp16 = add(x = var_3898_cast_fp16, y = var_3900_cast_fp16)[name = tensor("key_69_cast_fp16")]; + tensor var_3902_cast_fp16 = mul(x = current_value_35_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3902_cast_fp16")]; + tensor var_3904_cast_fp16 = mul(x = var_138_cast_fp16_17, y = var_241_cast_fp16)[name = tensor("op_3904_cast_fp16")]; + tensor value_69_cast_fp16 = add(x = var_3902_cast_fp16, y = var_3904_cast_fp16)[name = tensor("value_69_cast_fp16")]; + tensor var_3907 = const()[name = tensor("op_3907"), val = tensor([1, 20, 64, -1])]; + tensor var_3908_cast_fp16 = reshape(shape = var_3907, x = query_69_cast_fp16)[name = tensor("op_3908_cast_fp16")]; + tensor var_3909_to_fp16 = const()[name = tensor("op_3909_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3910_cast_fp16 = mul(x = var_3908_cast_fp16, y = var_3909_to_fp16)[name = tensor("op_3910_cast_fp16")]; + tensor var_3911 = const()[name = tensor("op_3911"), val = tensor([1, 20, 64, -1])]; + tensor var_3912_cast_fp16 = reshape(shape = var_3911, x = key_69_cast_fp16)[name = tensor("op_3912_cast_fp16")]; + tensor mh_w_103_transpose_x_0 = const()[name = tensor("mh_w_103_transpose_x_0"), val = tensor(true)]; + tensor mh_w_103_transpose_y_0 = const()[name = tensor("mh_w_103_transpose_y_0"), val = tensor(false)]; + tensor mh_w_103_cast_fp16 = matmul(transpose_x = mh_w_103_transpose_x_0, transpose_y = mh_w_103_transpose_y_0, x = var_3910_cast_fp16, y = var_3912_cast_fp16)[name = tensor("mh_w_103_cast_fp16")]; + tensor mh_w_105_cast_fp16 = add(x = mh_w_103_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_105_cast_fp16")]; + tensor var_3920_cast_fp16 = softmax(axis = var_3834, x = mh_w_105_cast_fp16)[name = tensor("op_3920_cast_fp16")]; + tensor var_3921 = const()[name = tensor("op_3921"), val = tensor([1, 20, 64, -1])]; + tensor var_3922_cast_fp16 = reshape(shape = var_3921, x = value_69_cast_fp16)[name = tensor("op_3922_cast_fp16")]; + tensor attn_69_transpose_x_0 = const()[name = tensor("attn_69_transpose_x_0"), val = tensor(false)]; + tensor attn_69_transpose_y_0 = const()[name = tensor("attn_69_transpose_y_0"), val = tensor(true)]; + tensor attn_69_cast_fp16 = matmul(transpose_x = attn_69_transpose_x_0, transpose_y = attn_69_transpose_y_0, x = var_3922_cast_fp16, y = var_3920_cast_fp16)[name = tensor("attn_69_cast_fp16")]; + tensor var_3925 = const()[name = tensor("op_3925"), val = tensor([1, 1280, 1, -1])]; + tensor input_171_cast_fp16 = reshape(shape = var_3925, x = attn_69_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor var_3929 = const()[name = tensor("op_3929"), val = tensor([1, 1])]; + tensor var_3931 = const()[name = tensor("op_3931"), val = tensor([1, 1])]; + tensor obj_245_pad_type_0 = const()[name = tensor("obj_245_pad_type_0"), val = tensor("custom")]; + tensor obj_245_pad_0 = const()[name = tensor("obj_245_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1035825920)))]; + tensor layers_17_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1039102784)))]; + tensor obj_245_cast_fp16 = conv(bias = layers_17_self_attn_o_proj_bias_to_fp16, dilations = var_3931, groups = var_3841, pad = obj_245_pad_0, pad_type = obj_245_pad_type_0, strides = var_3929, weight = layers_17_self_attn_o_proj_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("obj_245_cast_fp16")]; + tensor inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = obj_245_cast_fp16)[name = tensor("inputs_105_cast_fp16")]; + tensor var_3941 = const()[name = tensor("op_3941"), val = tensor([1])]; + tensor channels_mean_105_cast_fp16 = reduce_mean(axes = var_3941, keep_dims = var_3842, x = inputs_105_cast_fp16)[name = tensor("channels_mean_105_cast_fp16")]; + tensor zero_mean_105_cast_fp16 = sub(x = inputs_105_cast_fp16, y = channels_mean_105_cast_fp16)[name = tensor("zero_mean_105_cast_fp16")]; + tensor zero_mean_sq_105_cast_fp16 = mul(x = zero_mean_105_cast_fp16, y = zero_mean_105_cast_fp16)[name = tensor("zero_mean_sq_105_cast_fp16")]; + tensor var_3945 = const()[name = tensor("op_3945"), val = tensor([1])]; + tensor var_3946_cast_fp16 = reduce_mean(axes = var_3945, keep_dims = var_3842, x = zero_mean_sq_105_cast_fp16)[name = tensor("op_3946_cast_fp16")]; + tensor var_3947_to_fp16 = const()[name = tensor("op_3947_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3948_cast_fp16 = add(x = var_3946_cast_fp16, y = var_3947_to_fp16)[name = tensor("op_3948_cast_fp16")]; + tensor denom_105_epsilon_0_to_fp16 = const()[name = tensor("denom_105_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_105_cast_fp16 = rsqrt(epsilon = denom_105_epsilon_0_to_fp16, x = var_3948_cast_fp16)[name = tensor("denom_105_cast_fp16")]; + tensor out_105_cast_fp16 = mul(x = zero_mean_105_cast_fp16, y = denom_105_cast_fp16)[name = tensor("out_105_cast_fp16")]; + tensor obj_247_gamma_0_to_fp16 = const()[name = tensor("obj_247_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1039105408)))]; + tensor obj_247_beta_0_to_fp16 = const()[name = tensor("obj_247_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1039108032)))]; + tensor obj_247_epsilon_0_to_fp16 = const()[name = tensor("obj_247_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_247_cast_fp16 = batch_norm(beta = obj_247_beta_0_to_fp16, epsilon = obj_247_epsilon_0_to_fp16, gamma = obj_247_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_105_cast_fp16)[name = tensor("obj_247_cast_fp16")]; + tensor var_3963 = const()[name = tensor("op_3963"), val = tensor([1, 1])]; + tensor var_3965 = const()[name = tensor("op_3965"), val = tensor([1, 1])]; + tensor query_71_pad_type_0 = const()[name = tensor("query_71_pad_type_0"), val = tensor("custom")]; + tensor query_71_pad_0 = const()[name = tensor("query_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1039110656)))]; + tensor layers_17_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1042387520)))]; + tensor query_71_cast_fp16 = conv(bias = layers_17_encoder_attn_q_proj_bias_to_fp16, dilations = var_3965, groups = var_3841, pad = query_71_pad_0, pad_type = query_71_pad_type_0, strides = var_3963, weight = layers_17_encoder_attn_q_proj_weight_to_fp16, x = obj_247_cast_fp16)[name = tensor("query_71_cast_fp16")]; + tensor var_3969 = const()[name = tensor("op_3969"), val = tensor([1, 1])]; + tensor var_3971 = const()[name = tensor("op_3971"), val = tensor([1, 1])]; + tensor key_71_pad_type_0 = const()[name = tensor("key_71_pad_type_0"), val = tensor("custom")]; + tensor key_71_pad_0 = const()[name = tensor("key_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1042390144)))]; + tensor key_71_cast_fp16 = conv(dilations = var_3971, groups = var_3841, pad = key_71_pad_0, pad_type = key_71_pad_type_0, strides = var_3969, weight = layers_17_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_71_cast_fp16")]; + tensor var_3976 = const()[name = tensor("op_3976"), val = tensor([1, 1])]; + tensor var_3978 = const()[name = tensor("op_3978"), val = tensor([1, 1])]; + tensor value_71_pad_type_0 = const()[name = tensor("value_71_pad_type_0"), val = tensor("custom")]; + tensor value_71_pad_0 = const()[name = tensor("value_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1045667008)))]; + tensor layers_17_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1048943872)))]; + tensor value_71_cast_fp16 = conv(bias = layers_17_encoder_attn_v_proj_bias_to_fp16, dilations = var_3978, groups = var_3841, pad = value_71_pad_0, pad_type = value_71_pad_type_0, strides = var_3976, weight = layers_17_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_71_cast_fp16")]; + tensor var_3982 = const()[name = tensor("op_3982"), val = tensor([1, 20, 64, -1])]; + tensor var_3983_cast_fp16 = reshape(shape = var_3982, x = query_71_cast_fp16)[name = tensor("op_3983_cast_fp16")]; + tensor var_3984_to_fp16 = const()[name = tensor("op_3984_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3985_cast_fp16 = mul(x = var_3983_cast_fp16, y = var_3984_to_fp16)[name = tensor("op_3985_cast_fp16")]; + tensor var_3986 = const()[name = tensor("op_3986"), val = tensor([1, 20, 64, -1])]; + tensor var_3987_cast_fp16 = reshape(shape = var_3986, x = key_71_cast_fp16)[name = tensor("op_3987_cast_fp16")]; + tensor mh_w_107_transpose_x_0 = const()[name = tensor("mh_w_107_transpose_x_0"), val = tensor(true)]; + tensor mh_w_107_transpose_y_0 = const()[name = tensor("mh_w_107_transpose_y_0"), val = tensor(false)]; + tensor mh_w_107_cast_fp16 = matmul(transpose_x = mh_w_107_transpose_x_0, transpose_y = mh_w_107_transpose_y_0, x = var_3985_cast_fp16, y = var_3987_cast_fp16)[name = tensor("mh_w_107_cast_fp16")]; + tensor obj_251_cast_fp16 = softmax(axis = var_3834, x = mh_w_107_cast_fp16)[name = tensor("obj_251_cast_fp16")]; + tensor var_3991 = const()[name = tensor("op_3991"), val = tensor([1, 20, 64, -1])]; + tensor var_3992_cast_fp16 = reshape(shape = var_3991, x = value_71_cast_fp16)[name = tensor("op_3992_cast_fp16")]; + tensor attn_71_transpose_x_0 = const()[name = tensor("attn_71_transpose_x_0"), val = tensor(false)]; + tensor attn_71_transpose_y_0 = const()[name = tensor("attn_71_transpose_y_0"), val = tensor(true)]; + tensor attn_71_cast_fp16 = matmul(transpose_x = attn_71_transpose_x_0, transpose_y = attn_71_transpose_y_0, x = var_3992_cast_fp16, y = obj_251_cast_fp16)[name = tensor("attn_71_cast_fp16")]; + tensor var_3995 = const()[name = tensor("op_3995"), val = tensor([1, 1280, 1, -1])]; + tensor input_173_cast_fp16 = reshape(shape = var_3995, x = attn_71_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor var_3999 = const()[name = tensor("op_3999"), val = tensor([1, 1])]; + tensor var_4001 = const()[name = tensor("op_4001"), val = tensor([1, 1])]; + tensor obj_249_pad_type_0 = const()[name = tensor("obj_249_pad_type_0"), val = tensor("custom")]; + tensor obj_249_pad_0 = const()[name = tensor("obj_249_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1048946496)))]; + tensor layers_17_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1052223360)))]; + tensor obj_249_cast_fp16 = conv(bias = layers_17_encoder_attn_o_proj_bias_to_fp16, dilations = var_4001, groups = var_3841, pad = obj_249_pad_0, pad_type = obj_249_pad_type_0, strides = var_3999, weight = layers_17_encoder_attn_o_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("obj_249_cast_fp16")]; + tensor inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = obj_249_cast_fp16)[name = tensor("inputs_107_cast_fp16")]; + tensor var_4010 = const()[name = tensor("op_4010"), val = tensor([1])]; + tensor channels_mean_107_cast_fp16 = reduce_mean(axes = var_4010, keep_dims = var_3842, x = inputs_107_cast_fp16)[name = tensor("channels_mean_107_cast_fp16")]; + tensor zero_mean_107_cast_fp16 = sub(x = inputs_107_cast_fp16, y = channels_mean_107_cast_fp16)[name = tensor("zero_mean_107_cast_fp16")]; + tensor zero_mean_sq_107_cast_fp16 = mul(x = zero_mean_107_cast_fp16, y = zero_mean_107_cast_fp16)[name = tensor("zero_mean_sq_107_cast_fp16")]; + tensor var_4014 = const()[name = tensor("op_4014"), val = tensor([1])]; + tensor var_4015_cast_fp16 = reduce_mean(axes = var_4014, keep_dims = var_3842, x = zero_mean_sq_107_cast_fp16)[name = tensor("op_4015_cast_fp16")]; + tensor var_4016_to_fp16 = const()[name = tensor("op_4016_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4017_cast_fp16 = add(x = var_4015_cast_fp16, y = var_4016_to_fp16)[name = tensor("op_4017_cast_fp16")]; + tensor denom_107_epsilon_0_to_fp16 = const()[name = tensor("denom_107_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_107_cast_fp16 = rsqrt(epsilon = denom_107_epsilon_0_to_fp16, x = var_4017_cast_fp16)[name = tensor("denom_107_cast_fp16")]; + tensor out_107_cast_fp16 = mul(x = zero_mean_107_cast_fp16, y = denom_107_cast_fp16)[name = tensor("out_107_cast_fp16")]; + tensor input_175_gamma_0_to_fp16 = const()[name = tensor("input_175_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1052225984)))]; + tensor input_175_beta_0_to_fp16 = const()[name = tensor("input_175_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1052228608)))]; + tensor input_175_epsilon_0_to_fp16 = const()[name = tensor("input_175_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_175_cast_fp16 = batch_norm(beta = input_175_beta_0_to_fp16, epsilon = input_175_epsilon_0_to_fp16, gamma = input_175_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_107_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor var_4028 = const()[name = tensor("op_4028"), val = tensor([1, 1])]; + tensor var_4030 = const()[name = tensor("op_4030"), val = tensor([1, 1])]; + tensor input_177_pad_type_0 = const()[name = tensor("input_177_pad_type_0"), val = tensor("custom")]; + tensor input_177_pad_0 = const()[name = tensor("input_177_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_fc1_weight_to_fp16 = const()[name = tensor("layers_17_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1052231232)))]; + tensor layers_17_fc1_bias_to_fp16 = const()[name = tensor("layers_17_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1065338496)))]; + tensor input_177_cast_fp16 = conv(bias = layers_17_fc1_bias_to_fp16, dilations = var_4030, groups = var_3841, pad = input_177_pad_0, pad_type = input_177_pad_type_0, strides = var_4028, weight = layers_17_fc1_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_mode_0 = const()[name = tensor("input_179_mode_0"), val = tensor("EXACT")]; + tensor input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor var_4036 = const()[name = tensor("op_4036"), val = tensor([1, 1])]; + tensor var_4038 = const()[name = tensor("op_4038"), val = tensor([1, 1])]; + tensor hidden_states_37_pad_type_0 = const()[name = tensor("hidden_states_37_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_37_pad_0 = const()[name = tensor("hidden_states_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_fc2_weight_to_fp16 = const()[name = tensor("layers_17_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1065348800)))]; + tensor layers_17_fc2_bias_to_fp16 = const()[name = tensor("layers_17_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1078456064)))]; + tensor hidden_states_37_cast_fp16 = conv(bias = layers_17_fc2_bias_to_fp16, dilations = var_4038, groups = var_3841, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = var_4036, weight = layers_17_fc2_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("hidden_states_37_cast_fp16")]; + tensor inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = hidden_states_37_cast_fp16)[name = tensor("inputs_109_cast_fp16")]; + tensor var_4052 = const()[name = tensor("op_4052"), val = tensor(3)]; + tensor var_4059 = const()[name = tensor("op_4059"), val = tensor(1)]; + tensor var_4060 = const()[name = tensor("op_4060"), val = tensor(true)]; + tensor var_4072 = const()[name = tensor("op_4072"), val = tensor([1])]; + tensor channels_mean_109_cast_fp16 = reduce_mean(axes = var_4072, keep_dims = var_4060, x = inputs_109_cast_fp16)[name = tensor("channels_mean_109_cast_fp16")]; + tensor zero_mean_109_cast_fp16 = sub(x = inputs_109_cast_fp16, y = channels_mean_109_cast_fp16)[name = tensor("zero_mean_109_cast_fp16")]; + tensor zero_mean_sq_109_cast_fp16 = mul(x = zero_mean_109_cast_fp16, y = zero_mean_109_cast_fp16)[name = tensor("zero_mean_sq_109_cast_fp16")]; + tensor var_4076 = const()[name = tensor("op_4076"), val = tensor([1])]; + tensor var_4077_cast_fp16 = reduce_mean(axes = var_4076, keep_dims = var_4060, x = zero_mean_sq_109_cast_fp16)[name = tensor("op_4077_cast_fp16")]; + tensor var_4078_to_fp16 = const()[name = tensor("op_4078_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4079_cast_fp16 = add(x = var_4077_cast_fp16, y = var_4078_to_fp16)[name = tensor("op_4079_cast_fp16")]; + tensor denom_109_epsilon_0_to_fp16 = const()[name = tensor("denom_109_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_109_cast_fp16 = rsqrt(epsilon = denom_109_epsilon_0_to_fp16, x = var_4079_cast_fp16)[name = tensor("denom_109_cast_fp16")]; + tensor out_109_cast_fp16 = mul(x = zero_mean_109_cast_fp16, y = denom_109_cast_fp16)[name = tensor("out_109_cast_fp16")]; + tensor obj_253_gamma_0_to_fp16 = const()[name = tensor("obj_253_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1078458688)))]; + tensor obj_253_beta_0_to_fp16 = const()[name = tensor("obj_253_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1078461312)))]; + tensor obj_253_epsilon_0_to_fp16 = const()[name = tensor("obj_253_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_253_cast_fp16 = batch_norm(beta = obj_253_beta_0_to_fp16, epsilon = obj_253_epsilon_0_to_fp16, gamma = obj_253_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_109_cast_fp16)[name = tensor("obj_253_cast_fp16")]; + tensor var_4094 = const()[name = tensor("op_4094"), val = tensor([1, 1])]; + tensor var_4096 = const()[name = tensor("op_4096"), val = tensor([1, 1])]; + tensor query_73_pad_type_0 = const()[name = tensor("query_73_pad_type_0"), val = tensor("custom")]; + tensor query_73_pad_0 = const()[name = tensor("query_73_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1078463936)))]; + tensor layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1081740800)))]; + tensor query_73_cast_fp16 = conv(bias = layers_18_self_attn_q_proj_bias_to_fp16, dilations = var_4096, groups = var_4059, pad = query_73_pad_0, pad_type = query_73_pad_type_0, strides = var_4094, weight = layers_18_self_attn_q_proj_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("query_73_cast_fp16")]; + tensor var_4100 = const()[name = tensor("op_4100"), val = tensor([1, 1])]; + tensor var_4102 = const()[name = tensor("op_4102"), val = tensor([1, 1])]; + tensor current_key_37_pad_type_0 = const()[name = tensor("current_key_37_pad_type_0"), val = tensor("custom")]; + tensor current_key_37_pad_0 = const()[name = tensor("current_key_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1081743424)))]; + tensor current_key_37_cast_fp16 = conv(dilations = var_4102, groups = var_4059, pad = current_key_37_pad_0, pad_type = current_key_37_pad_type_0, strides = var_4100, weight = layers_18_self_attn_k_proj_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("current_key_37_cast_fp16")]; + tensor var_4107 = const()[name = tensor("op_4107"), val = tensor([1, 1])]; + tensor var_4109 = const()[name = tensor("op_4109"), val = tensor([1, 1])]; + tensor current_value_37_pad_type_0 = const()[name = tensor("current_value_37_pad_type_0"), val = tensor("custom")]; + tensor current_value_37_pad_0 = const()[name = tensor("current_value_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1085020288)))]; + tensor layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1088297152)))]; + tensor current_value_37_cast_fp16 = conv(bias = layers_18_self_attn_v_proj_bias_to_fp16, dilations = var_4109, groups = var_4059, pad = current_value_37_pad_0, pad_type = current_value_37_pad_type_0, strides = var_4107, weight = layers_18_self_attn_v_proj_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("current_value_37_cast_fp16")]; + tensor var_4116_cast_fp16 = mul(x = current_key_37_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4116_cast_fp16")]; + tensor var_4118_cast_fp16 = mul(x = var_103_cast_fp16_18, y = var_241_cast_fp16)[name = tensor("op_4118_cast_fp16")]; + tensor key_73_cast_fp16 = add(x = var_4116_cast_fp16, y = var_4118_cast_fp16)[name = tensor("key_73_cast_fp16")]; + tensor var_4120_cast_fp16 = mul(x = current_value_37_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4120_cast_fp16")]; + tensor var_4122_cast_fp16 = mul(x = var_138_cast_fp16_18, y = var_241_cast_fp16)[name = tensor("op_4122_cast_fp16")]; + tensor value_73_cast_fp16 = add(x = var_4120_cast_fp16, y = var_4122_cast_fp16)[name = tensor("value_73_cast_fp16")]; + tensor var_4125 = const()[name = tensor("op_4125"), val = tensor([1, 20, 64, -1])]; + tensor var_4126_cast_fp16 = reshape(shape = var_4125, x = query_73_cast_fp16)[name = tensor("op_4126_cast_fp16")]; + tensor var_4127_to_fp16 = const()[name = tensor("op_4127_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4128_cast_fp16 = mul(x = var_4126_cast_fp16, y = var_4127_to_fp16)[name = tensor("op_4128_cast_fp16")]; + tensor var_4129 = const()[name = tensor("op_4129"), val = tensor([1, 20, 64, -1])]; + tensor var_4130_cast_fp16 = reshape(shape = var_4129, x = key_73_cast_fp16)[name = tensor("op_4130_cast_fp16")]; + tensor mh_w_109_transpose_x_0 = const()[name = tensor("mh_w_109_transpose_x_0"), val = tensor(true)]; + tensor mh_w_109_transpose_y_0 = const()[name = tensor("mh_w_109_transpose_y_0"), val = tensor(false)]; + tensor mh_w_109_cast_fp16 = matmul(transpose_x = mh_w_109_transpose_x_0, transpose_y = mh_w_109_transpose_y_0, x = var_4128_cast_fp16, y = var_4130_cast_fp16)[name = tensor("mh_w_109_cast_fp16")]; + tensor mh_w_111_cast_fp16 = add(x = mh_w_109_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_111_cast_fp16")]; + tensor var_4138_cast_fp16 = softmax(axis = var_4052, x = mh_w_111_cast_fp16)[name = tensor("op_4138_cast_fp16")]; + tensor var_4139 = const()[name = tensor("op_4139"), val = tensor([1, 20, 64, -1])]; + tensor var_4140_cast_fp16 = reshape(shape = var_4139, x = value_73_cast_fp16)[name = tensor("op_4140_cast_fp16")]; + tensor attn_73_transpose_x_0 = const()[name = tensor("attn_73_transpose_x_0"), val = tensor(false)]; + tensor attn_73_transpose_y_0 = const()[name = tensor("attn_73_transpose_y_0"), val = tensor(true)]; + tensor attn_73_cast_fp16 = matmul(transpose_x = attn_73_transpose_x_0, transpose_y = attn_73_transpose_y_0, x = var_4140_cast_fp16, y = var_4138_cast_fp16)[name = tensor("attn_73_cast_fp16")]; + tensor var_4143 = const()[name = tensor("op_4143"), val = tensor([1, 1280, 1, -1])]; + tensor input_181_cast_fp16 = reshape(shape = var_4143, x = attn_73_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor var_4147 = const()[name = tensor("op_4147"), val = tensor([1, 1])]; + tensor var_4149 = const()[name = tensor("op_4149"), val = tensor([1, 1])]; + tensor obj_259_pad_type_0 = const()[name = tensor("obj_259_pad_type_0"), val = tensor("custom")]; + tensor obj_259_pad_0 = const()[name = tensor("obj_259_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1088299776)))]; + tensor layers_18_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1091576640)))]; + tensor obj_259_cast_fp16 = conv(bias = layers_18_self_attn_o_proj_bias_to_fp16, dilations = var_4149, groups = var_4059, pad = obj_259_pad_0, pad_type = obj_259_pad_type_0, strides = var_4147, weight = layers_18_self_attn_o_proj_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("obj_259_cast_fp16")]; + tensor inputs_111_cast_fp16 = add(x = inputs_109_cast_fp16, y = obj_259_cast_fp16)[name = tensor("inputs_111_cast_fp16")]; + tensor var_4159 = const()[name = tensor("op_4159"), val = tensor([1])]; + tensor channels_mean_111_cast_fp16 = reduce_mean(axes = var_4159, keep_dims = var_4060, x = inputs_111_cast_fp16)[name = tensor("channels_mean_111_cast_fp16")]; + tensor zero_mean_111_cast_fp16 = sub(x = inputs_111_cast_fp16, y = channels_mean_111_cast_fp16)[name = tensor("zero_mean_111_cast_fp16")]; + tensor zero_mean_sq_111_cast_fp16 = mul(x = zero_mean_111_cast_fp16, y = zero_mean_111_cast_fp16)[name = tensor("zero_mean_sq_111_cast_fp16")]; + tensor var_4163 = const()[name = tensor("op_4163"), val = tensor([1])]; + tensor var_4164_cast_fp16 = reduce_mean(axes = var_4163, keep_dims = var_4060, x = zero_mean_sq_111_cast_fp16)[name = tensor("op_4164_cast_fp16")]; + tensor var_4165_to_fp16 = const()[name = tensor("op_4165_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4166_cast_fp16 = add(x = var_4164_cast_fp16, y = var_4165_to_fp16)[name = tensor("op_4166_cast_fp16")]; + tensor denom_111_epsilon_0_to_fp16 = const()[name = tensor("denom_111_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_111_cast_fp16 = rsqrt(epsilon = denom_111_epsilon_0_to_fp16, x = var_4166_cast_fp16)[name = tensor("denom_111_cast_fp16")]; + tensor out_111_cast_fp16 = mul(x = zero_mean_111_cast_fp16, y = denom_111_cast_fp16)[name = tensor("out_111_cast_fp16")]; + tensor obj_261_gamma_0_to_fp16 = const()[name = tensor("obj_261_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1091579264)))]; + tensor obj_261_beta_0_to_fp16 = const()[name = tensor("obj_261_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1091581888)))]; + tensor obj_261_epsilon_0_to_fp16 = const()[name = tensor("obj_261_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_261_cast_fp16 = batch_norm(beta = obj_261_beta_0_to_fp16, epsilon = obj_261_epsilon_0_to_fp16, gamma = obj_261_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_111_cast_fp16)[name = tensor("obj_261_cast_fp16")]; + tensor var_4181 = const()[name = tensor("op_4181"), val = tensor([1, 1])]; + tensor var_4183 = const()[name = tensor("op_4183"), val = tensor([1, 1])]; + tensor query_75_pad_type_0 = const()[name = tensor("query_75_pad_type_0"), val = tensor("custom")]; + tensor query_75_pad_0 = const()[name = tensor("query_75_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1091584512)))]; + tensor layers_18_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1094861376)))]; + tensor query_75_cast_fp16 = conv(bias = layers_18_encoder_attn_q_proj_bias_to_fp16, dilations = var_4183, groups = var_4059, pad = query_75_pad_0, pad_type = query_75_pad_type_0, strides = var_4181, weight = layers_18_encoder_attn_q_proj_weight_to_fp16, x = obj_261_cast_fp16)[name = tensor("query_75_cast_fp16")]; + tensor var_4187 = const()[name = tensor("op_4187"), val = tensor([1, 1])]; + tensor var_4189 = const()[name = tensor("op_4189"), val = tensor([1, 1])]; + tensor key_75_pad_type_0 = const()[name = tensor("key_75_pad_type_0"), val = tensor("custom")]; + tensor key_75_pad_0 = const()[name = tensor("key_75_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1094864000)))]; + tensor key_75_cast_fp16 = conv(dilations = var_4189, groups = var_4059, pad = key_75_pad_0, pad_type = key_75_pad_type_0, strides = var_4187, weight = layers_18_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_75_cast_fp16")]; + tensor var_4194 = const()[name = tensor("op_4194"), val = tensor([1, 1])]; + tensor var_4196 = const()[name = tensor("op_4196"), val = tensor([1, 1])]; + tensor value_75_pad_type_0 = const()[name = tensor("value_75_pad_type_0"), val = tensor("custom")]; + tensor value_75_pad_0 = const()[name = tensor("value_75_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1098140864)))]; + tensor layers_18_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1101417728)))]; + tensor value_75_cast_fp16 = conv(bias = layers_18_encoder_attn_v_proj_bias_to_fp16, dilations = var_4196, groups = var_4059, pad = value_75_pad_0, pad_type = value_75_pad_type_0, strides = var_4194, weight = layers_18_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_75_cast_fp16")]; + tensor var_4200 = const()[name = tensor("op_4200"), val = tensor([1, 20, 64, -1])]; + tensor var_4201_cast_fp16 = reshape(shape = var_4200, x = query_75_cast_fp16)[name = tensor("op_4201_cast_fp16")]; + tensor var_4202_to_fp16 = const()[name = tensor("op_4202_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4203_cast_fp16 = mul(x = var_4201_cast_fp16, y = var_4202_to_fp16)[name = tensor("op_4203_cast_fp16")]; + tensor var_4204 = const()[name = tensor("op_4204"), val = tensor([1, 20, 64, -1])]; + tensor var_4205_cast_fp16 = reshape(shape = var_4204, x = key_75_cast_fp16)[name = tensor("op_4205_cast_fp16")]; + tensor mh_w_113_transpose_x_0 = const()[name = tensor("mh_w_113_transpose_x_0"), val = tensor(true)]; + tensor mh_w_113_transpose_y_0 = const()[name = tensor("mh_w_113_transpose_y_0"), val = tensor(false)]; + tensor mh_w_113_cast_fp16 = matmul(transpose_x = mh_w_113_transpose_x_0, transpose_y = mh_w_113_transpose_y_0, x = var_4203_cast_fp16, y = var_4205_cast_fp16)[name = tensor("mh_w_113_cast_fp16")]; + tensor obj_265_cast_fp16 = softmax(axis = var_4052, x = mh_w_113_cast_fp16)[name = tensor("obj_265_cast_fp16")]; + tensor var_4209 = const()[name = tensor("op_4209"), val = tensor([1, 20, 64, -1])]; + tensor var_4210_cast_fp16 = reshape(shape = var_4209, x = value_75_cast_fp16)[name = tensor("op_4210_cast_fp16")]; + tensor attn_75_transpose_x_0 = const()[name = tensor("attn_75_transpose_x_0"), val = tensor(false)]; + tensor attn_75_transpose_y_0 = const()[name = tensor("attn_75_transpose_y_0"), val = tensor(true)]; + tensor attn_75_cast_fp16 = matmul(transpose_x = attn_75_transpose_x_0, transpose_y = attn_75_transpose_y_0, x = var_4210_cast_fp16, y = obj_265_cast_fp16)[name = tensor("attn_75_cast_fp16")]; + tensor var_4213 = const()[name = tensor("op_4213"), val = tensor([1, 1280, 1, -1])]; + tensor input_183_cast_fp16 = reshape(shape = var_4213, x = attn_75_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor var_4217 = const()[name = tensor("op_4217"), val = tensor([1, 1])]; + tensor var_4219 = const()[name = tensor("op_4219"), val = tensor([1, 1])]; + tensor obj_263_pad_type_0 = const()[name = tensor("obj_263_pad_type_0"), val = tensor("custom")]; + tensor obj_263_pad_0 = const()[name = tensor("obj_263_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1101420352)))]; + tensor layers_18_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1104697216)))]; + tensor obj_263_cast_fp16 = conv(bias = layers_18_encoder_attn_o_proj_bias_to_fp16, dilations = var_4219, groups = var_4059, pad = obj_263_pad_0, pad_type = obj_263_pad_type_0, strides = var_4217, weight = layers_18_encoder_attn_o_proj_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("obj_263_cast_fp16")]; + tensor inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = obj_263_cast_fp16)[name = tensor("inputs_113_cast_fp16")]; + tensor var_4225 = const()[name = tensor("op_4225"), val = tensor([1])]; + tensor channels_mean_113_cast_fp16 = reduce_mean(axes = var_4225, keep_dims = var_4060, x = inputs_113_cast_fp16)[name = tensor("channels_mean_113_cast_fp16")]; + tensor zero_mean_113_cast_fp16 = sub(x = inputs_113_cast_fp16, y = channels_mean_113_cast_fp16)[name = tensor("zero_mean_113_cast_fp16")]; + tensor zero_mean_sq_113_cast_fp16 = mul(x = zero_mean_113_cast_fp16, y = zero_mean_113_cast_fp16)[name = tensor("zero_mean_sq_113_cast_fp16")]; + tensor var_4229 = const()[name = tensor("op_4229"), val = tensor([1])]; + tensor var_4230_cast_fp16 = reduce_mean(axes = var_4229, keep_dims = var_4060, x = zero_mean_sq_113_cast_fp16)[name = tensor("op_4230_cast_fp16")]; + tensor var_4231_to_fp16 = const()[name = tensor("op_4231_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4232_cast_fp16 = add(x = var_4230_cast_fp16, y = var_4231_to_fp16)[name = tensor("op_4232_cast_fp16")]; + tensor denom_113_epsilon_0_to_fp16 = const()[name = tensor("denom_113_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_113_cast_fp16 = rsqrt(epsilon = denom_113_epsilon_0_to_fp16, x = var_4232_cast_fp16)[name = tensor("denom_113_cast_fp16")]; + tensor out_113_cast_fp16 = mul(x = zero_mean_113_cast_fp16, y = denom_113_cast_fp16)[name = tensor("out_113_cast_fp16")]; + tensor input_185_gamma_0_to_fp16 = const()[name = tensor("input_185_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1104699840)))]; + tensor input_185_beta_0_to_fp16 = const()[name = tensor("input_185_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1104702464)))]; + tensor input_185_epsilon_0_to_fp16 = const()[name = tensor("input_185_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_185_cast_fp16 = batch_norm(beta = input_185_beta_0_to_fp16, epsilon = input_185_epsilon_0_to_fp16, gamma = input_185_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_113_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor var_4243 = const()[name = tensor("op_4243"), val = tensor([1, 1])]; + tensor var_4245 = const()[name = tensor("op_4245"), val = tensor([1, 1])]; + tensor input_187_pad_type_0 = const()[name = tensor("input_187_pad_type_0"), val = tensor("custom")]; + tensor input_187_pad_0 = const()[name = tensor("input_187_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_fc1_weight_to_fp16 = const()[name = tensor("layers_18_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1104705088)))]; + tensor layers_18_fc1_bias_to_fp16 = const()[name = tensor("layers_18_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1117812352)))]; + tensor input_187_cast_fp16 = conv(bias = layers_18_fc1_bias_to_fp16, dilations = var_4245, groups = var_4059, pad = input_187_pad_0, pad_type = input_187_pad_type_0, strides = var_4243, weight = layers_18_fc1_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_mode_0 = const()[name = tensor("input_189_mode_0"), val = tensor("EXACT")]; + tensor input_189_cast_fp16 = gelu(mode = input_189_mode_0, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor var_4251 = const()[name = tensor("op_4251"), val = tensor([1, 1])]; + tensor var_4253 = const()[name = tensor("op_4253"), val = tensor([1, 1])]; + tensor hidden_states_39_pad_type_0 = const()[name = tensor("hidden_states_39_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_39_pad_0 = const()[name = tensor("hidden_states_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_fc2_weight_to_fp16 = const()[name = tensor("layers_18_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1117822656)))]; + tensor layers_18_fc2_bias_to_fp16 = const()[name = tensor("layers_18_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1130929920)))]; + tensor hidden_states_39_cast_fp16 = conv(bias = layers_18_fc2_bias_to_fp16, dilations = var_4253, groups = var_4059, pad = hidden_states_39_pad_0, pad_type = hidden_states_39_pad_type_0, strides = var_4251, weight = layers_18_fc2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("hidden_states_39_cast_fp16")]; + tensor inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = hidden_states_39_cast_fp16)[name = tensor("inputs_115_cast_fp16")]; + tensor var_4266 = const()[name = tensor("op_4266"), val = tensor(3)]; + tensor var_4273 = const()[name = tensor("op_4273"), val = tensor(1)]; + tensor var_4274 = const()[name = tensor("op_4274"), val = tensor(true)]; + tensor var_4286 = const()[name = tensor("op_4286"), val = tensor([1])]; + tensor channels_mean_115_cast_fp16 = reduce_mean(axes = var_4286, keep_dims = var_4274, x = inputs_115_cast_fp16)[name = tensor("channels_mean_115_cast_fp16")]; + tensor zero_mean_115_cast_fp16 = sub(x = inputs_115_cast_fp16, y = channels_mean_115_cast_fp16)[name = tensor("zero_mean_115_cast_fp16")]; + tensor zero_mean_sq_115_cast_fp16 = mul(x = zero_mean_115_cast_fp16, y = zero_mean_115_cast_fp16)[name = tensor("zero_mean_sq_115_cast_fp16")]; + tensor var_4290 = const()[name = tensor("op_4290"), val = tensor([1])]; + tensor var_4291_cast_fp16 = reduce_mean(axes = var_4290, keep_dims = var_4274, x = zero_mean_sq_115_cast_fp16)[name = tensor("op_4291_cast_fp16")]; + tensor var_4292_to_fp16 = const()[name = tensor("op_4292_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4293_cast_fp16 = add(x = var_4291_cast_fp16, y = var_4292_to_fp16)[name = tensor("op_4293_cast_fp16")]; + tensor denom_115_epsilon_0_to_fp16 = const()[name = tensor("denom_115_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_115_cast_fp16 = rsqrt(epsilon = denom_115_epsilon_0_to_fp16, x = var_4293_cast_fp16)[name = tensor("denom_115_cast_fp16")]; + tensor out_115_cast_fp16 = mul(x = zero_mean_115_cast_fp16, y = denom_115_cast_fp16)[name = tensor("out_115_cast_fp16")]; + tensor obj_267_gamma_0_to_fp16 = const()[name = tensor("obj_267_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1130932544)))]; + tensor obj_267_beta_0_to_fp16 = const()[name = tensor("obj_267_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1130935168)))]; + tensor obj_267_epsilon_0_to_fp16 = const()[name = tensor("obj_267_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_267_cast_fp16 = batch_norm(beta = obj_267_beta_0_to_fp16, epsilon = obj_267_epsilon_0_to_fp16, gamma = obj_267_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_115_cast_fp16)[name = tensor("obj_267_cast_fp16")]; + tensor var_4308 = const()[name = tensor("op_4308"), val = tensor([1, 1])]; + tensor var_4310 = const()[name = tensor("op_4310"), val = tensor([1, 1])]; + tensor query_77_pad_type_0 = const()[name = tensor("query_77_pad_type_0"), val = tensor("custom")]; + tensor query_77_pad_0 = const()[name = tensor("query_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1130937792)))]; + tensor layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1134214656)))]; + tensor query_77_cast_fp16 = conv(bias = layers_19_self_attn_q_proj_bias_to_fp16, dilations = var_4310, groups = var_4273, pad = query_77_pad_0, pad_type = query_77_pad_type_0, strides = var_4308, weight = layers_19_self_attn_q_proj_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor("query_77_cast_fp16")]; + tensor var_4314 = const()[name = tensor("op_4314"), val = tensor([1, 1])]; + tensor var_4316 = const()[name = tensor("op_4316"), val = tensor([1, 1])]; + tensor current_key_39_pad_type_0 = const()[name = tensor("current_key_39_pad_type_0"), val = tensor("custom")]; + tensor current_key_39_pad_0 = const()[name = tensor("current_key_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1134217280)))]; + tensor current_key_39_cast_fp16 = conv(dilations = var_4316, groups = var_4273, pad = current_key_39_pad_0, pad_type = current_key_39_pad_type_0, strides = var_4314, weight = layers_19_self_attn_k_proj_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor("current_key_39_cast_fp16")]; + tensor var_4321 = const()[name = tensor("op_4321"), val = tensor([1, 1])]; + tensor var_4323 = const()[name = tensor("op_4323"), val = tensor([1, 1])]; + tensor current_value_39_pad_type_0 = const()[name = tensor("current_value_39_pad_type_0"), val = tensor("custom")]; + tensor current_value_39_pad_0 = const()[name = tensor("current_value_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1137494144)))]; + tensor layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1140771008)))]; + tensor current_value_39_cast_fp16 = conv(bias = layers_19_self_attn_v_proj_bias_to_fp16, dilations = var_4323, groups = var_4273, pad = current_value_39_pad_0, pad_type = current_value_39_pad_type_0, strides = var_4321, weight = layers_19_self_attn_v_proj_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor("current_value_39_cast_fp16")]; + tensor var_4330_cast_fp16 = mul(x = current_key_39_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4330_cast_fp16")]; + tensor var_4332_cast_fp16 = mul(x = var_103_cast_fp16_19, y = var_241_cast_fp16)[name = tensor("op_4332_cast_fp16")]; + tensor key_77_cast_fp16 = add(x = var_4330_cast_fp16, y = var_4332_cast_fp16)[name = tensor("key_77_cast_fp16")]; + tensor var_4334_cast_fp16 = mul(x = current_value_39_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4334_cast_fp16")]; + tensor var_4336_cast_fp16 = mul(x = var_138_cast_fp16_19, y = var_241_cast_fp16)[name = tensor("op_4336_cast_fp16")]; + tensor value_77_cast_fp16 = add(x = var_4334_cast_fp16, y = var_4336_cast_fp16)[name = tensor("value_77_cast_fp16")]; + tensor var_4339 = const()[name = tensor("op_4339"), val = tensor([1, 20, 64, -1])]; + tensor var_4340_cast_fp16 = reshape(shape = var_4339, x = query_77_cast_fp16)[name = tensor("op_4340_cast_fp16")]; + tensor var_4341_to_fp16 = const()[name = tensor("op_4341_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4342_cast_fp16 = mul(x = var_4340_cast_fp16, y = var_4341_to_fp16)[name = tensor("op_4342_cast_fp16")]; + tensor var_4343 = const()[name = tensor("op_4343"), val = tensor([1, 20, 64, -1])]; + tensor var_4344_cast_fp16 = reshape(shape = var_4343, x = key_77_cast_fp16)[name = tensor("op_4344_cast_fp16")]; + tensor mh_w_115_transpose_x_0 = const()[name = tensor("mh_w_115_transpose_x_0"), val = tensor(true)]; + tensor mh_w_115_transpose_y_0 = const()[name = tensor("mh_w_115_transpose_y_0"), val = tensor(false)]; + tensor mh_w_115_cast_fp16 = matmul(transpose_x = mh_w_115_transpose_x_0, transpose_y = mh_w_115_transpose_y_0, x = var_4342_cast_fp16, y = var_4344_cast_fp16)[name = tensor("mh_w_115_cast_fp16")]; + tensor mh_w_117_cast_fp16 = add(x = mh_w_115_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_117_cast_fp16")]; + tensor var_4352_cast_fp16 = softmax(axis = var_4266, x = mh_w_117_cast_fp16)[name = tensor("op_4352_cast_fp16")]; + tensor var_4353 = const()[name = tensor("op_4353"), val = tensor([1, 20, 64, -1])]; + tensor var_4354_cast_fp16 = reshape(shape = var_4353, x = value_77_cast_fp16)[name = tensor("op_4354_cast_fp16")]; + tensor attn_77_transpose_x_0 = const()[name = tensor("attn_77_transpose_x_0"), val = tensor(false)]; + tensor attn_77_transpose_y_0 = const()[name = tensor("attn_77_transpose_y_0"), val = tensor(true)]; + tensor attn_77_cast_fp16 = matmul(transpose_x = attn_77_transpose_x_0, transpose_y = attn_77_transpose_y_0, x = var_4354_cast_fp16, y = var_4352_cast_fp16)[name = tensor("attn_77_cast_fp16")]; + tensor var_4357 = const()[name = tensor("op_4357"), val = tensor([1, 1280, 1, -1])]; + tensor input_191_cast_fp16 = reshape(shape = var_4357, x = attn_77_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor var_4361 = const()[name = tensor("op_4361"), val = tensor([1, 1])]; + tensor var_4363 = const()[name = tensor("op_4363"), val = tensor([1, 1])]; + tensor obj_273_pad_type_0 = const()[name = tensor("obj_273_pad_type_0"), val = tensor("custom")]; + tensor obj_273_pad_0 = const()[name = tensor("obj_273_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1140773632)))]; + tensor layers_19_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1144050496)))]; + tensor obj_273_cast_fp16 = conv(bias = layers_19_self_attn_o_proj_bias_to_fp16, dilations = var_4363, groups = var_4273, pad = obj_273_pad_0, pad_type = obj_273_pad_type_0, strides = var_4361, weight = layers_19_self_attn_o_proj_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("obj_273_cast_fp16")]; + tensor inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = obj_273_cast_fp16)[name = tensor("inputs_117_cast_fp16")]; + tensor var_4373 = const()[name = tensor("op_4373"), val = tensor([1])]; + tensor channels_mean_117_cast_fp16 = reduce_mean(axes = var_4373, keep_dims = var_4274, x = inputs_117_cast_fp16)[name = tensor("channels_mean_117_cast_fp16")]; + tensor zero_mean_117_cast_fp16 = sub(x = inputs_117_cast_fp16, y = channels_mean_117_cast_fp16)[name = tensor("zero_mean_117_cast_fp16")]; + tensor zero_mean_sq_117_cast_fp16 = mul(x = zero_mean_117_cast_fp16, y = zero_mean_117_cast_fp16)[name = tensor("zero_mean_sq_117_cast_fp16")]; + tensor var_4377 = const()[name = tensor("op_4377"), val = tensor([1])]; + tensor var_4378_cast_fp16 = reduce_mean(axes = var_4377, keep_dims = var_4274, x = zero_mean_sq_117_cast_fp16)[name = tensor("op_4378_cast_fp16")]; + tensor var_4379_to_fp16 = const()[name = tensor("op_4379_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4380_cast_fp16 = add(x = var_4378_cast_fp16, y = var_4379_to_fp16)[name = tensor("op_4380_cast_fp16")]; + tensor denom_117_epsilon_0_to_fp16 = const()[name = tensor("denom_117_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_117_cast_fp16 = rsqrt(epsilon = denom_117_epsilon_0_to_fp16, x = var_4380_cast_fp16)[name = tensor("denom_117_cast_fp16")]; + tensor out_117_cast_fp16 = mul(x = zero_mean_117_cast_fp16, y = denom_117_cast_fp16)[name = tensor("out_117_cast_fp16")]; + tensor obj_275_gamma_0_to_fp16 = const()[name = tensor("obj_275_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1144053120)))]; + tensor obj_275_beta_0_to_fp16 = const()[name = tensor("obj_275_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1144055744)))]; + tensor obj_275_epsilon_0_to_fp16 = const()[name = tensor("obj_275_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_275_cast_fp16 = batch_norm(beta = obj_275_beta_0_to_fp16, epsilon = obj_275_epsilon_0_to_fp16, gamma = obj_275_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_117_cast_fp16)[name = tensor("obj_275_cast_fp16")]; + tensor var_4395 = const()[name = tensor("op_4395"), val = tensor([1, 1])]; + tensor var_4397 = const()[name = tensor("op_4397"), val = tensor([1, 1])]; + tensor query_79_pad_type_0 = const()[name = tensor("query_79_pad_type_0"), val = tensor("custom")]; + tensor query_79_pad_0 = const()[name = tensor("query_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1144058368)))]; + tensor layers_19_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1147335232)))]; + tensor query_79_cast_fp16 = conv(bias = layers_19_encoder_attn_q_proj_bias_to_fp16, dilations = var_4397, groups = var_4273, pad = query_79_pad_0, pad_type = query_79_pad_type_0, strides = var_4395, weight = layers_19_encoder_attn_q_proj_weight_to_fp16, x = obj_275_cast_fp16)[name = tensor("query_79_cast_fp16")]; + tensor var_4401 = const()[name = tensor("op_4401"), val = tensor([1, 1])]; + tensor var_4403 = const()[name = tensor("op_4403"), val = tensor([1, 1])]; + tensor key_79_pad_type_0 = const()[name = tensor("key_79_pad_type_0"), val = tensor("custom")]; + tensor key_79_pad_0 = const()[name = tensor("key_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1147337856)))]; + tensor key_79_cast_fp16 = conv(dilations = var_4403, groups = var_4273, pad = key_79_pad_0, pad_type = key_79_pad_type_0, strides = var_4401, weight = layers_19_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_79_cast_fp16")]; + tensor var_4408 = const()[name = tensor("op_4408"), val = tensor([1, 1])]; + tensor var_4410 = const()[name = tensor("op_4410"), val = tensor([1, 1])]; + tensor value_79_pad_type_0 = const()[name = tensor("value_79_pad_type_0"), val = tensor("custom")]; + tensor value_79_pad_0 = const()[name = tensor("value_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1150614720)))]; + tensor layers_19_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1153891584)))]; + tensor value_79_cast_fp16 = conv(bias = layers_19_encoder_attn_v_proj_bias_to_fp16, dilations = var_4410, groups = var_4273, pad = value_79_pad_0, pad_type = value_79_pad_type_0, strides = var_4408, weight = layers_19_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_79_cast_fp16")]; + tensor var_4414 = const()[name = tensor("op_4414"), val = tensor([1, 20, 64, -1])]; + tensor var_4415_cast_fp16 = reshape(shape = var_4414, x = query_79_cast_fp16)[name = tensor("op_4415_cast_fp16")]; + tensor var_4416_to_fp16 = const()[name = tensor("op_4416_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4417_cast_fp16 = mul(x = var_4415_cast_fp16, y = var_4416_to_fp16)[name = tensor("op_4417_cast_fp16")]; + tensor var_4418 = const()[name = tensor("op_4418"), val = tensor([1, 20, 64, -1])]; + tensor var_4419_cast_fp16 = reshape(shape = var_4418, x = key_79_cast_fp16)[name = tensor("op_4419_cast_fp16")]; + tensor mh_w_119_transpose_x_0 = const()[name = tensor("mh_w_119_transpose_x_0"), val = tensor(true)]; + tensor mh_w_119_transpose_y_0 = const()[name = tensor("mh_w_119_transpose_y_0"), val = tensor(false)]; + tensor mh_w_119_cast_fp16 = matmul(transpose_x = mh_w_119_transpose_x_0, transpose_y = mh_w_119_transpose_y_0, x = var_4417_cast_fp16, y = var_4419_cast_fp16)[name = tensor("mh_w_119_cast_fp16")]; + tensor obj_279_cast_fp16 = softmax(axis = var_4266, x = mh_w_119_cast_fp16)[name = tensor("obj_279_cast_fp16")]; + tensor var_4423 = const()[name = tensor("op_4423"), val = tensor([1, 20, 64, -1])]; + tensor var_4424_cast_fp16 = reshape(shape = var_4423, x = value_79_cast_fp16)[name = tensor("op_4424_cast_fp16")]; + tensor attn_79_transpose_x_0 = const()[name = tensor("attn_79_transpose_x_0"), val = tensor(false)]; + tensor attn_79_transpose_y_0 = const()[name = tensor("attn_79_transpose_y_0"), val = tensor(true)]; + tensor attn_79_cast_fp16 = matmul(transpose_x = attn_79_transpose_x_0, transpose_y = attn_79_transpose_y_0, x = var_4424_cast_fp16, y = obj_279_cast_fp16)[name = tensor("attn_79_cast_fp16")]; + tensor var_4427 = const()[name = tensor("op_4427"), val = tensor([1, 1280, 1, -1])]; + tensor input_193_cast_fp16 = reshape(shape = var_4427, x = attn_79_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor var_4431 = const()[name = tensor("op_4431"), val = tensor([1, 1])]; + tensor var_4433 = const()[name = tensor("op_4433"), val = tensor([1, 1])]; + tensor obj_277_pad_type_0 = const()[name = tensor("obj_277_pad_type_0"), val = tensor("custom")]; + tensor obj_277_pad_0 = const()[name = tensor("obj_277_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1153894208)))]; + tensor layers_19_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1157171072)))]; + tensor obj_277_cast_fp16 = conv(bias = layers_19_encoder_attn_o_proj_bias_to_fp16, dilations = var_4433, groups = var_4273, pad = obj_277_pad_0, pad_type = obj_277_pad_type_0, strides = var_4431, weight = layers_19_encoder_attn_o_proj_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("obj_277_cast_fp16")]; + tensor inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = obj_277_cast_fp16)[name = tensor("inputs_119_cast_fp16")]; + tensor var_4442 = const()[name = tensor("op_4442"), val = tensor([1])]; + tensor channels_mean_119_cast_fp16 = reduce_mean(axes = var_4442, keep_dims = var_4274, x = inputs_119_cast_fp16)[name = tensor("channels_mean_119_cast_fp16")]; + tensor zero_mean_119_cast_fp16 = sub(x = inputs_119_cast_fp16, y = channels_mean_119_cast_fp16)[name = tensor("zero_mean_119_cast_fp16")]; + tensor zero_mean_sq_119_cast_fp16 = mul(x = zero_mean_119_cast_fp16, y = zero_mean_119_cast_fp16)[name = tensor("zero_mean_sq_119_cast_fp16")]; + tensor var_4446 = const()[name = tensor("op_4446"), val = tensor([1])]; + tensor var_4447_cast_fp16 = reduce_mean(axes = var_4446, keep_dims = var_4274, x = zero_mean_sq_119_cast_fp16)[name = tensor("op_4447_cast_fp16")]; + tensor var_4448_to_fp16 = const()[name = tensor("op_4448_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4449_cast_fp16 = add(x = var_4447_cast_fp16, y = var_4448_to_fp16)[name = tensor("op_4449_cast_fp16")]; + tensor denom_119_epsilon_0_to_fp16 = const()[name = tensor("denom_119_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_119_cast_fp16 = rsqrt(epsilon = denom_119_epsilon_0_to_fp16, x = var_4449_cast_fp16)[name = tensor("denom_119_cast_fp16")]; + tensor out_119_cast_fp16 = mul(x = zero_mean_119_cast_fp16, y = denom_119_cast_fp16)[name = tensor("out_119_cast_fp16")]; + tensor input_195_gamma_0_to_fp16 = const()[name = tensor("input_195_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1157173696)))]; + tensor input_195_beta_0_to_fp16 = const()[name = tensor("input_195_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1157176320)))]; + tensor input_195_epsilon_0_to_fp16 = const()[name = tensor("input_195_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_195_cast_fp16 = batch_norm(beta = input_195_beta_0_to_fp16, epsilon = input_195_epsilon_0_to_fp16, gamma = input_195_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_119_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor var_4460 = const()[name = tensor("op_4460"), val = tensor([1, 1])]; + tensor var_4462 = const()[name = tensor("op_4462"), val = tensor([1, 1])]; + tensor input_197_pad_type_0 = const()[name = tensor("input_197_pad_type_0"), val = tensor("custom")]; + tensor input_197_pad_0 = const()[name = tensor("input_197_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_fc1_weight_to_fp16 = const()[name = tensor("layers_19_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1157178944)))]; + tensor layers_19_fc1_bias_to_fp16 = const()[name = tensor("layers_19_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1170286208)))]; + tensor input_197_cast_fp16 = conv(bias = layers_19_fc1_bias_to_fp16, dilations = var_4462, groups = var_4273, pad = input_197_pad_0, pad_type = input_197_pad_type_0, strides = var_4460, weight = layers_19_fc1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor input_199_mode_0 = const()[name = tensor("input_199_mode_0"), val = tensor("EXACT")]; + tensor input_199_cast_fp16 = gelu(mode = input_199_mode_0, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor var_4468 = const()[name = tensor("op_4468"), val = tensor([1, 1])]; + tensor var_4470 = const()[name = tensor("op_4470"), val = tensor([1, 1])]; + tensor hidden_states_41_pad_type_0 = const()[name = tensor("hidden_states_41_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_41_pad_0 = const()[name = tensor("hidden_states_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_fc2_weight_to_fp16 = const()[name = tensor("layers_19_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1170296512)))]; + tensor layers_19_fc2_bias_to_fp16 = const()[name = tensor("layers_19_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1183403776)))]; + tensor hidden_states_41_cast_fp16 = conv(bias = layers_19_fc2_bias_to_fp16, dilations = var_4470, groups = var_4273, pad = hidden_states_41_pad_0, pad_type = hidden_states_41_pad_type_0, strides = var_4468, weight = layers_19_fc2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("hidden_states_41_cast_fp16")]; + tensor inputs_121_cast_fp16 = add(x = inputs_119_cast_fp16, y = hidden_states_41_cast_fp16)[name = tensor("inputs_121_cast_fp16")]; + tensor var_4484 = const()[name = tensor("op_4484"), val = tensor(3)]; + tensor var_4491 = const()[name = tensor("op_4491"), val = tensor(1)]; + tensor var_4492 = const()[name = tensor("op_4492"), val = tensor(true)]; + tensor var_4504 = const()[name = tensor("op_4504"), val = tensor([1])]; + tensor channels_mean_121_cast_fp16 = reduce_mean(axes = var_4504, keep_dims = var_4492, x = inputs_121_cast_fp16)[name = tensor("channels_mean_121_cast_fp16")]; + tensor zero_mean_121_cast_fp16 = sub(x = inputs_121_cast_fp16, y = channels_mean_121_cast_fp16)[name = tensor("zero_mean_121_cast_fp16")]; + tensor zero_mean_sq_121_cast_fp16 = mul(x = zero_mean_121_cast_fp16, y = zero_mean_121_cast_fp16)[name = tensor("zero_mean_sq_121_cast_fp16")]; + tensor var_4508 = const()[name = tensor("op_4508"), val = tensor([1])]; + tensor var_4509_cast_fp16 = reduce_mean(axes = var_4508, keep_dims = var_4492, x = zero_mean_sq_121_cast_fp16)[name = tensor("op_4509_cast_fp16")]; + tensor var_4510_to_fp16 = const()[name = tensor("op_4510_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4511_cast_fp16 = add(x = var_4509_cast_fp16, y = var_4510_to_fp16)[name = tensor("op_4511_cast_fp16")]; + tensor denom_121_epsilon_0_to_fp16 = const()[name = tensor("denom_121_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_121_cast_fp16 = rsqrt(epsilon = denom_121_epsilon_0_to_fp16, x = var_4511_cast_fp16)[name = tensor("denom_121_cast_fp16")]; + tensor out_121_cast_fp16 = mul(x = zero_mean_121_cast_fp16, y = denom_121_cast_fp16)[name = tensor("out_121_cast_fp16")]; + tensor obj_281_gamma_0_to_fp16 = const()[name = tensor("obj_281_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1183406400)))]; + tensor obj_281_beta_0_to_fp16 = const()[name = tensor("obj_281_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1183409024)))]; + tensor obj_281_epsilon_0_to_fp16 = const()[name = tensor("obj_281_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_281_cast_fp16 = batch_norm(beta = obj_281_beta_0_to_fp16, epsilon = obj_281_epsilon_0_to_fp16, gamma = obj_281_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_121_cast_fp16)[name = tensor("obj_281_cast_fp16")]; + tensor var_4526 = const()[name = tensor("op_4526"), val = tensor([1, 1])]; + tensor var_4528 = const()[name = tensor("op_4528"), val = tensor([1, 1])]; + tensor query_81_pad_type_0 = const()[name = tensor("query_81_pad_type_0"), val = tensor("custom")]; + tensor query_81_pad_0 = const()[name = tensor("query_81_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1183411648)))]; + tensor layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1186688512)))]; + tensor query_81_cast_fp16 = conv(bias = layers_20_self_attn_q_proj_bias_to_fp16, dilations = var_4528, groups = var_4491, pad = query_81_pad_0, pad_type = query_81_pad_type_0, strides = var_4526, weight = layers_20_self_attn_q_proj_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor("query_81_cast_fp16")]; + tensor var_4532 = const()[name = tensor("op_4532"), val = tensor([1, 1])]; + tensor var_4534 = const()[name = tensor("op_4534"), val = tensor([1, 1])]; + tensor current_key_41_pad_type_0 = const()[name = tensor("current_key_41_pad_type_0"), val = tensor("custom")]; + tensor current_key_41_pad_0 = const()[name = tensor("current_key_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1186691136)))]; + tensor current_key_41_cast_fp16 = conv(dilations = var_4534, groups = var_4491, pad = current_key_41_pad_0, pad_type = current_key_41_pad_type_0, strides = var_4532, weight = layers_20_self_attn_k_proj_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor("current_key_41_cast_fp16")]; + tensor var_4539 = const()[name = tensor("op_4539"), val = tensor([1, 1])]; + tensor var_4541 = const()[name = tensor("op_4541"), val = tensor([1, 1])]; + tensor current_value_41_pad_type_0 = const()[name = tensor("current_value_41_pad_type_0"), val = tensor("custom")]; + tensor current_value_41_pad_0 = const()[name = tensor("current_value_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1189968000)))]; + tensor layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1193244864)))]; + tensor current_value_41_cast_fp16 = conv(bias = layers_20_self_attn_v_proj_bias_to_fp16, dilations = var_4541, groups = var_4491, pad = current_value_41_pad_0, pad_type = current_value_41_pad_type_0, strides = var_4539, weight = layers_20_self_attn_v_proj_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor("current_value_41_cast_fp16")]; + tensor var_4548_cast_fp16 = mul(x = current_key_41_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4548_cast_fp16")]; + tensor var_4550_cast_fp16 = mul(x = var_103_cast_fp16_20, y = var_241_cast_fp16)[name = tensor("op_4550_cast_fp16")]; + tensor key_81_cast_fp16 = add(x = var_4548_cast_fp16, y = var_4550_cast_fp16)[name = tensor("key_81_cast_fp16")]; + tensor var_4552_cast_fp16 = mul(x = current_value_41_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4552_cast_fp16")]; + tensor var_4554_cast_fp16 = mul(x = var_138_cast_fp16_20, y = var_241_cast_fp16)[name = tensor("op_4554_cast_fp16")]; + tensor value_81_cast_fp16 = add(x = var_4552_cast_fp16, y = var_4554_cast_fp16)[name = tensor("value_81_cast_fp16")]; + tensor var_4557 = const()[name = tensor("op_4557"), val = tensor([1, 20, 64, -1])]; + tensor var_4558_cast_fp16 = reshape(shape = var_4557, x = query_81_cast_fp16)[name = tensor("op_4558_cast_fp16")]; + tensor var_4559_to_fp16 = const()[name = tensor("op_4559_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4560_cast_fp16 = mul(x = var_4558_cast_fp16, y = var_4559_to_fp16)[name = tensor("op_4560_cast_fp16")]; + tensor var_4561 = const()[name = tensor("op_4561"), val = tensor([1, 20, 64, -1])]; + tensor var_4562_cast_fp16 = reshape(shape = var_4561, x = key_81_cast_fp16)[name = tensor("op_4562_cast_fp16")]; + tensor mh_w_121_transpose_x_0 = const()[name = tensor("mh_w_121_transpose_x_0"), val = tensor(true)]; + tensor mh_w_121_transpose_y_0 = const()[name = tensor("mh_w_121_transpose_y_0"), val = tensor(false)]; + tensor mh_w_121_cast_fp16 = matmul(transpose_x = mh_w_121_transpose_x_0, transpose_y = mh_w_121_transpose_y_0, x = var_4560_cast_fp16, y = var_4562_cast_fp16)[name = tensor("mh_w_121_cast_fp16")]; + tensor mh_w_123_cast_fp16 = add(x = mh_w_121_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_123_cast_fp16")]; + tensor var_4570_cast_fp16 = softmax(axis = var_4484, x = mh_w_123_cast_fp16)[name = tensor("op_4570_cast_fp16")]; + tensor var_4571 = const()[name = tensor("op_4571"), val = tensor([1, 20, 64, -1])]; + tensor var_4572_cast_fp16 = reshape(shape = var_4571, x = value_81_cast_fp16)[name = tensor("op_4572_cast_fp16")]; + tensor attn_81_transpose_x_0 = const()[name = tensor("attn_81_transpose_x_0"), val = tensor(false)]; + tensor attn_81_transpose_y_0 = const()[name = tensor("attn_81_transpose_y_0"), val = tensor(true)]; + tensor attn_81_cast_fp16 = matmul(transpose_x = attn_81_transpose_x_0, transpose_y = attn_81_transpose_y_0, x = var_4572_cast_fp16, y = var_4570_cast_fp16)[name = tensor("attn_81_cast_fp16")]; + tensor var_4575 = const()[name = tensor("op_4575"), val = tensor([1, 1280, 1, -1])]; + tensor input_201_cast_fp16 = reshape(shape = var_4575, x = attn_81_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor var_4579 = const()[name = tensor("op_4579"), val = tensor([1, 1])]; + tensor var_4581 = const()[name = tensor("op_4581"), val = tensor([1, 1])]; + tensor obj_287_pad_type_0 = const()[name = tensor("obj_287_pad_type_0"), val = tensor("custom")]; + tensor obj_287_pad_0 = const()[name = tensor("obj_287_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1193247488)))]; + tensor layers_20_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1196524352)))]; + tensor obj_287_cast_fp16 = conv(bias = layers_20_self_attn_o_proj_bias_to_fp16, dilations = var_4581, groups = var_4491, pad = obj_287_pad_0, pad_type = obj_287_pad_type_0, strides = var_4579, weight = layers_20_self_attn_o_proj_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("obj_287_cast_fp16")]; + tensor inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = obj_287_cast_fp16)[name = tensor("inputs_123_cast_fp16")]; + tensor var_4591 = const()[name = tensor("op_4591"), val = tensor([1])]; + tensor channels_mean_123_cast_fp16 = reduce_mean(axes = var_4591, keep_dims = var_4492, x = inputs_123_cast_fp16)[name = tensor("channels_mean_123_cast_fp16")]; + tensor zero_mean_123_cast_fp16 = sub(x = inputs_123_cast_fp16, y = channels_mean_123_cast_fp16)[name = tensor("zero_mean_123_cast_fp16")]; + tensor zero_mean_sq_123_cast_fp16 = mul(x = zero_mean_123_cast_fp16, y = zero_mean_123_cast_fp16)[name = tensor("zero_mean_sq_123_cast_fp16")]; + tensor var_4595 = const()[name = tensor("op_4595"), val = tensor([1])]; + tensor var_4596_cast_fp16 = reduce_mean(axes = var_4595, keep_dims = var_4492, x = zero_mean_sq_123_cast_fp16)[name = tensor("op_4596_cast_fp16")]; + tensor var_4597_to_fp16 = const()[name = tensor("op_4597_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4598_cast_fp16 = add(x = var_4596_cast_fp16, y = var_4597_to_fp16)[name = tensor("op_4598_cast_fp16")]; + tensor denom_123_epsilon_0_to_fp16 = const()[name = tensor("denom_123_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_123_cast_fp16 = rsqrt(epsilon = denom_123_epsilon_0_to_fp16, x = var_4598_cast_fp16)[name = tensor("denom_123_cast_fp16")]; + tensor out_123_cast_fp16 = mul(x = zero_mean_123_cast_fp16, y = denom_123_cast_fp16)[name = tensor("out_123_cast_fp16")]; + tensor obj_289_gamma_0_to_fp16 = const()[name = tensor("obj_289_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1196526976)))]; + tensor obj_289_beta_0_to_fp16 = const()[name = tensor("obj_289_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1196529600)))]; + tensor obj_289_epsilon_0_to_fp16 = const()[name = tensor("obj_289_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_289_cast_fp16 = batch_norm(beta = obj_289_beta_0_to_fp16, epsilon = obj_289_epsilon_0_to_fp16, gamma = obj_289_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_123_cast_fp16)[name = tensor("obj_289_cast_fp16")]; + tensor var_4613 = const()[name = tensor("op_4613"), val = tensor([1, 1])]; + tensor var_4615 = const()[name = tensor("op_4615"), val = tensor([1, 1])]; + tensor query_83_pad_type_0 = const()[name = tensor("query_83_pad_type_0"), val = tensor("custom")]; + tensor query_83_pad_0 = const()[name = tensor("query_83_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1196532224)))]; + tensor layers_20_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1199809088)))]; + tensor query_83_cast_fp16 = conv(bias = layers_20_encoder_attn_q_proj_bias_to_fp16, dilations = var_4615, groups = var_4491, pad = query_83_pad_0, pad_type = query_83_pad_type_0, strides = var_4613, weight = layers_20_encoder_attn_q_proj_weight_to_fp16, x = obj_289_cast_fp16)[name = tensor("query_83_cast_fp16")]; + tensor var_4619 = const()[name = tensor("op_4619"), val = tensor([1, 1])]; + tensor var_4621 = const()[name = tensor("op_4621"), val = tensor([1, 1])]; + tensor key_83_pad_type_0 = const()[name = tensor("key_83_pad_type_0"), val = tensor("custom")]; + tensor key_83_pad_0 = const()[name = tensor("key_83_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1199811712)))]; + tensor key_83_cast_fp16 = conv(dilations = var_4621, groups = var_4491, pad = key_83_pad_0, pad_type = key_83_pad_type_0, strides = var_4619, weight = layers_20_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_83_cast_fp16")]; + tensor var_4626 = const()[name = tensor("op_4626"), val = tensor([1, 1])]; + tensor var_4628 = const()[name = tensor("op_4628"), val = tensor([1, 1])]; + tensor value_83_pad_type_0 = const()[name = tensor("value_83_pad_type_0"), val = tensor("custom")]; + tensor value_83_pad_0 = const()[name = tensor("value_83_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1203088576)))]; + tensor layers_20_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1206365440)))]; + tensor value_83_cast_fp16 = conv(bias = layers_20_encoder_attn_v_proj_bias_to_fp16, dilations = var_4628, groups = var_4491, pad = value_83_pad_0, pad_type = value_83_pad_type_0, strides = var_4626, weight = layers_20_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_83_cast_fp16")]; + tensor var_4632 = const()[name = tensor("op_4632"), val = tensor([1, 20, 64, -1])]; + tensor var_4633_cast_fp16 = reshape(shape = var_4632, x = query_83_cast_fp16)[name = tensor("op_4633_cast_fp16")]; + tensor var_4634_to_fp16 = const()[name = tensor("op_4634_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4635_cast_fp16 = mul(x = var_4633_cast_fp16, y = var_4634_to_fp16)[name = tensor("op_4635_cast_fp16")]; + tensor var_4636 = const()[name = tensor("op_4636"), val = tensor([1, 20, 64, -1])]; + tensor var_4637_cast_fp16 = reshape(shape = var_4636, x = key_83_cast_fp16)[name = tensor("op_4637_cast_fp16")]; + tensor mh_w_125_transpose_x_0 = const()[name = tensor("mh_w_125_transpose_x_0"), val = tensor(true)]; + tensor mh_w_125_transpose_y_0 = const()[name = tensor("mh_w_125_transpose_y_0"), val = tensor(false)]; + tensor mh_w_125_cast_fp16 = matmul(transpose_x = mh_w_125_transpose_x_0, transpose_y = mh_w_125_transpose_y_0, x = var_4635_cast_fp16, y = var_4637_cast_fp16)[name = tensor("mh_w_125_cast_fp16")]; + tensor obj_293_cast_fp16 = softmax(axis = var_4484, x = mh_w_125_cast_fp16)[name = tensor("obj_293_cast_fp16")]; + tensor var_4641 = const()[name = tensor("op_4641"), val = tensor([1, 20, 64, -1])]; + tensor var_4642_cast_fp16 = reshape(shape = var_4641, x = value_83_cast_fp16)[name = tensor("op_4642_cast_fp16")]; + tensor attn_83_transpose_x_0 = const()[name = tensor("attn_83_transpose_x_0"), val = tensor(false)]; + tensor attn_83_transpose_y_0 = const()[name = tensor("attn_83_transpose_y_0"), val = tensor(true)]; + tensor attn_83_cast_fp16 = matmul(transpose_x = attn_83_transpose_x_0, transpose_y = attn_83_transpose_y_0, x = var_4642_cast_fp16, y = obj_293_cast_fp16)[name = tensor("attn_83_cast_fp16")]; + tensor var_4645 = const()[name = tensor("op_4645"), val = tensor([1, 1280, 1, -1])]; + tensor input_203_cast_fp16 = reshape(shape = var_4645, x = attn_83_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor var_4649 = const()[name = tensor("op_4649"), val = tensor([1, 1])]; + tensor var_4651 = const()[name = tensor("op_4651"), val = tensor([1, 1])]; + tensor obj_291_pad_type_0 = const()[name = tensor("obj_291_pad_type_0"), val = tensor("custom")]; + tensor obj_291_pad_0 = const()[name = tensor("obj_291_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1206368064)))]; + tensor layers_20_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1209644928)))]; + tensor obj_291_cast_fp16 = conv(bias = layers_20_encoder_attn_o_proj_bias_to_fp16, dilations = var_4651, groups = var_4491, pad = obj_291_pad_0, pad_type = obj_291_pad_type_0, strides = var_4649, weight = layers_20_encoder_attn_o_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("obj_291_cast_fp16")]; + tensor inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = obj_291_cast_fp16)[name = tensor("inputs_125_cast_fp16")]; + tensor var_4657 = const()[name = tensor("op_4657"), val = tensor([1])]; + tensor channels_mean_125_cast_fp16 = reduce_mean(axes = var_4657, keep_dims = var_4492, x = inputs_125_cast_fp16)[name = tensor("channels_mean_125_cast_fp16")]; + tensor zero_mean_125_cast_fp16 = sub(x = inputs_125_cast_fp16, y = channels_mean_125_cast_fp16)[name = tensor("zero_mean_125_cast_fp16")]; + tensor zero_mean_sq_125_cast_fp16 = mul(x = zero_mean_125_cast_fp16, y = zero_mean_125_cast_fp16)[name = tensor("zero_mean_sq_125_cast_fp16")]; + tensor var_4661 = const()[name = tensor("op_4661"), val = tensor([1])]; + tensor var_4662_cast_fp16 = reduce_mean(axes = var_4661, keep_dims = var_4492, x = zero_mean_sq_125_cast_fp16)[name = tensor("op_4662_cast_fp16")]; + tensor var_4663_to_fp16 = const()[name = tensor("op_4663_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4664_cast_fp16 = add(x = var_4662_cast_fp16, y = var_4663_to_fp16)[name = tensor("op_4664_cast_fp16")]; + tensor denom_125_epsilon_0_to_fp16 = const()[name = tensor("denom_125_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_125_cast_fp16 = rsqrt(epsilon = denom_125_epsilon_0_to_fp16, x = var_4664_cast_fp16)[name = tensor("denom_125_cast_fp16")]; + tensor out_125_cast_fp16 = mul(x = zero_mean_125_cast_fp16, y = denom_125_cast_fp16)[name = tensor("out_125_cast_fp16")]; + tensor input_205_gamma_0_to_fp16 = const()[name = tensor("input_205_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1209647552)))]; + tensor input_205_beta_0_to_fp16 = const()[name = tensor("input_205_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1209650176)))]; + tensor input_205_epsilon_0_to_fp16 = const()[name = tensor("input_205_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_205_cast_fp16 = batch_norm(beta = input_205_beta_0_to_fp16, epsilon = input_205_epsilon_0_to_fp16, gamma = input_205_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_125_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor var_4675 = const()[name = tensor("op_4675"), val = tensor([1, 1])]; + tensor var_4677 = const()[name = tensor("op_4677"), val = tensor([1, 1])]; + tensor input_207_pad_type_0 = const()[name = tensor("input_207_pad_type_0"), val = tensor("custom")]; + tensor input_207_pad_0 = const()[name = tensor("input_207_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_fc1_weight_to_fp16 = const()[name = tensor("layers_20_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1209652800)))]; + tensor layers_20_fc1_bias_to_fp16 = const()[name = tensor("layers_20_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1222760064)))]; + tensor input_207_cast_fp16 = conv(bias = layers_20_fc1_bias_to_fp16, dilations = var_4677, groups = var_4491, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = var_4675, weight = layers_20_fc1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_mode_0 = const()[name = tensor("input_209_mode_0"), val = tensor("EXACT")]; + tensor input_209_cast_fp16 = gelu(mode = input_209_mode_0, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor var_4683 = const()[name = tensor("op_4683"), val = tensor([1, 1])]; + tensor var_4685 = const()[name = tensor("op_4685"), val = tensor([1, 1])]; + tensor hidden_states_43_pad_type_0 = const()[name = tensor("hidden_states_43_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_43_pad_0 = const()[name = tensor("hidden_states_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_fc2_weight_to_fp16 = const()[name = tensor("layers_20_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1222770368)))]; + tensor layers_20_fc2_bias_to_fp16 = const()[name = tensor("layers_20_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1235877632)))]; + tensor hidden_states_43_cast_fp16 = conv(bias = layers_20_fc2_bias_to_fp16, dilations = var_4685, groups = var_4491, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = var_4683, weight = layers_20_fc2_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("hidden_states_43_cast_fp16")]; + tensor inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = hidden_states_43_cast_fp16)[name = tensor("inputs_127_cast_fp16")]; + tensor var_4698 = const()[name = tensor("op_4698"), val = tensor(3)]; + tensor var_4705 = const()[name = tensor("op_4705"), val = tensor(1)]; + tensor var_4706 = const()[name = tensor("op_4706"), val = tensor(true)]; + tensor var_4718 = const()[name = tensor("op_4718"), val = tensor([1])]; + tensor channels_mean_127_cast_fp16 = reduce_mean(axes = var_4718, keep_dims = var_4706, x = inputs_127_cast_fp16)[name = tensor("channels_mean_127_cast_fp16")]; + tensor zero_mean_127_cast_fp16 = sub(x = inputs_127_cast_fp16, y = channels_mean_127_cast_fp16)[name = tensor("zero_mean_127_cast_fp16")]; + tensor zero_mean_sq_127_cast_fp16 = mul(x = zero_mean_127_cast_fp16, y = zero_mean_127_cast_fp16)[name = tensor("zero_mean_sq_127_cast_fp16")]; + tensor var_4722 = const()[name = tensor("op_4722"), val = tensor([1])]; + tensor var_4723_cast_fp16 = reduce_mean(axes = var_4722, keep_dims = var_4706, x = zero_mean_sq_127_cast_fp16)[name = tensor("op_4723_cast_fp16")]; + tensor var_4724_to_fp16 = const()[name = tensor("op_4724_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4725_cast_fp16 = add(x = var_4723_cast_fp16, y = var_4724_to_fp16)[name = tensor("op_4725_cast_fp16")]; + tensor denom_127_epsilon_0_to_fp16 = const()[name = tensor("denom_127_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_127_cast_fp16 = rsqrt(epsilon = denom_127_epsilon_0_to_fp16, x = var_4725_cast_fp16)[name = tensor("denom_127_cast_fp16")]; + tensor out_127_cast_fp16 = mul(x = zero_mean_127_cast_fp16, y = denom_127_cast_fp16)[name = tensor("out_127_cast_fp16")]; + tensor obj_295_gamma_0_to_fp16 = const()[name = tensor("obj_295_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1235880256)))]; + tensor obj_295_beta_0_to_fp16 = const()[name = tensor("obj_295_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1235882880)))]; + tensor obj_295_epsilon_0_to_fp16 = const()[name = tensor("obj_295_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_295_cast_fp16 = batch_norm(beta = obj_295_beta_0_to_fp16, epsilon = obj_295_epsilon_0_to_fp16, gamma = obj_295_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_127_cast_fp16)[name = tensor("obj_295_cast_fp16")]; + tensor var_4740 = const()[name = tensor("op_4740"), val = tensor([1, 1])]; + tensor var_4742 = const()[name = tensor("op_4742"), val = tensor([1, 1])]; + tensor query_85_pad_type_0 = const()[name = tensor("query_85_pad_type_0"), val = tensor("custom")]; + tensor query_85_pad_0 = const()[name = tensor("query_85_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1235885504)))]; + tensor layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1239162368)))]; + tensor query_85_cast_fp16 = conv(bias = layers_21_self_attn_q_proj_bias_to_fp16, dilations = var_4742, groups = var_4705, pad = query_85_pad_0, pad_type = query_85_pad_type_0, strides = var_4740, weight = layers_21_self_attn_q_proj_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor("query_85_cast_fp16")]; + tensor var_4746 = const()[name = tensor("op_4746"), val = tensor([1, 1])]; + tensor var_4748 = const()[name = tensor("op_4748"), val = tensor([1, 1])]; + tensor current_key_43_pad_type_0 = const()[name = tensor("current_key_43_pad_type_0"), val = tensor("custom")]; + tensor current_key_43_pad_0 = const()[name = tensor("current_key_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1239164992)))]; + tensor current_key_43_cast_fp16 = conv(dilations = var_4748, groups = var_4705, pad = current_key_43_pad_0, pad_type = current_key_43_pad_type_0, strides = var_4746, weight = layers_21_self_attn_k_proj_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor("current_key_43_cast_fp16")]; + tensor var_4753 = const()[name = tensor("op_4753"), val = tensor([1, 1])]; + tensor var_4755 = const()[name = tensor("op_4755"), val = tensor([1, 1])]; + tensor current_value_43_pad_type_0 = const()[name = tensor("current_value_43_pad_type_0"), val = tensor("custom")]; + tensor current_value_43_pad_0 = const()[name = tensor("current_value_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1242441856)))]; + tensor layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1245718720)))]; + tensor current_value_43_cast_fp16 = conv(bias = layers_21_self_attn_v_proj_bias_to_fp16, dilations = var_4755, groups = var_4705, pad = current_value_43_pad_0, pad_type = current_value_43_pad_type_0, strides = var_4753, weight = layers_21_self_attn_v_proj_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor("current_value_43_cast_fp16")]; + tensor var_4762_cast_fp16 = mul(x = current_key_43_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4762_cast_fp16")]; + tensor var_4764_cast_fp16 = mul(x = var_103_cast_fp16_21, y = var_241_cast_fp16)[name = tensor("op_4764_cast_fp16")]; + tensor key_85_cast_fp16 = add(x = var_4762_cast_fp16, y = var_4764_cast_fp16)[name = tensor("key_85_cast_fp16")]; + tensor var_4766_cast_fp16 = mul(x = current_value_43_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4766_cast_fp16")]; + tensor var_4768_cast_fp16 = mul(x = var_138_cast_fp16_21, y = var_241_cast_fp16)[name = tensor("op_4768_cast_fp16")]; + tensor value_85_cast_fp16 = add(x = var_4766_cast_fp16, y = var_4768_cast_fp16)[name = tensor("value_85_cast_fp16")]; + tensor var_4771 = const()[name = tensor("op_4771"), val = tensor([1, 20, 64, -1])]; + tensor var_4772_cast_fp16 = reshape(shape = var_4771, x = query_85_cast_fp16)[name = tensor("op_4772_cast_fp16")]; + tensor var_4773_to_fp16 = const()[name = tensor("op_4773_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4774_cast_fp16 = mul(x = var_4772_cast_fp16, y = var_4773_to_fp16)[name = tensor("op_4774_cast_fp16")]; + tensor var_4775 = const()[name = tensor("op_4775"), val = tensor([1, 20, 64, -1])]; + tensor var_4776_cast_fp16 = reshape(shape = var_4775, x = key_85_cast_fp16)[name = tensor("op_4776_cast_fp16")]; + tensor mh_w_127_transpose_x_0 = const()[name = tensor("mh_w_127_transpose_x_0"), val = tensor(true)]; + tensor mh_w_127_transpose_y_0 = const()[name = tensor("mh_w_127_transpose_y_0"), val = tensor(false)]; + tensor mh_w_127_cast_fp16 = matmul(transpose_x = mh_w_127_transpose_x_0, transpose_y = mh_w_127_transpose_y_0, x = var_4774_cast_fp16, y = var_4776_cast_fp16)[name = tensor("mh_w_127_cast_fp16")]; + tensor mh_w_129_cast_fp16 = add(x = mh_w_127_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_129_cast_fp16")]; + tensor var_4784_cast_fp16 = softmax(axis = var_4698, x = mh_w_129_cast_fp16)[name = tensor("op_4784_cast_fp16")]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([1, 20, 64, -1])]; + tensor var_4786_cast_fp16 = reshape(shape = var_4785, x = value_85_cast_fp16)[name = tensor("op_4786_cast_fp16")]; + tensor attn_85_transpose_x_0 = const()[name = tensor("attn_85_transpose_x_0"), val = tensor(false)]; + tensor attn_85_transpose_y_0 = const()[name = tensor("attn_85_transpose_y_0"), val = tensor(true)]; + tensor attn_85_cast_fp16 = matmul(transpose_x = attn_85_transpose_x_0, transpose_y = attn_85_transpose_y_0, x = var_4786_cast_fp16, y = var_4784_cast_fp16)[name = tensor("attn_85_cast_fp16")]; + tensor var_4789 = const()[name = tensor("op_4789"), val = tensor([1, 1280, 1, -1])]; + tensor input_211_cast_fp16 = reshape(shape = var_4789, x = attn_85_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor var_4793 = const()[name = tensor("op_4793"), val = tensor([1, 1])]; + tensor var_4795 = const()[name = tensor("op_4795"), val = tensor([1, 1])]; + tensor obj_301_pad_type_0 = const()[name = tensor("obj_301_pad_type_0"), val = tensor("custom")]; + tensor obj_301_pad_0 = const()[name = tensor("obj_301_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1245721344)))]; + tensor layers_21_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1248998208)))]; + tensor obj_301_cast_fp16 = conv(bias = layers_21_self_attn_o_proj_bias_to_fp16, dilations = var_4795, groups = var_4705, pad = obj_301_pad_0, pad_type = obj_301_pad_type_0, strides = var_4793, weight = layers_21_self_attn_o_proj_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("obj_301_cast_fp16")]; + tensor inputs_129_cast_fp16 = add(x = inputs_127_cast_fp16, y = obj_301_cast_fp16)[name = tensor("inputs_129_cast_fp16")]; + tensor var_4805 = const()[name = tensor("op_4805"), val = tensor([1])]; + tensor channels_mean_129_cast_fp16 = reduce_mean(axes = var_4805, keep_dims = var_4706, x = inputs_129_cast_fp16)[name = tensor("channels_mean_129_cast_fp16")]; + tensor zero_mean_129_cast_fp16 = sub(x = inputs_129_cast_fp16, y = channels_mean_129_cast_fp16)[name = tensor("zero_mean_129_cast_fp16")]; + tensor zero_mean_sq_129_cast_fp16 = mul(x = zero_mean_129_cast_fp16, y = zero_mean_129_cast_fp16)[name = tensor("zero_mean_sq_129_cast_fp16")]; + tensor var_4809 = const()[name = tensor("op_4809"), val = tensor([1])]; + tensor var_4810_cast_fp16 = reduce_mean(axes = var_4809, keep_dims = var_4706, x = zero_mean_sq_129_cast_fp16)[name = tensor("op_4810_cast_fp16")]; + tensor var_4811_to_fp16 = const()[name = tensor("op_4811_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4812_cast_fp16 = add(x = var_4810_cast_fp16, y = var_4811_to_fp16)[name = tensor("op_4812_cast_fp16")]; + tensor denom_129_epsilon_0_to_fp16 = const()[name = tensor("denom_129_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_129_cast_fp16 = rsqrt(epsilon = denom_129_epsilon_0_to_fp16, x = var_4812_cast_fp16)[name = tensor("denom_129_cast_fp16")]; + tensor out_129_cast_fp16 = mul(x = zero_mean_129_cast_fp16, y = denom_129_cast_fp16)[name = tensor("out_129_cast_fp16")]; + tensor obj_303_gamma_0_to_fp16 = const()[name = tensor("obj_303_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1249000832)))]; + tensor obj_303_beta_0_to_fp16 = const()[name = tensor("obj_303_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1249003456)))]; + tensor obj_303_epsilon_0_to_fp16 = const()[name = tensor("obj_303_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_303_cast_fp16 = batch_norm(beta = obj_303_beta_0_to_fp16, epsilon = obj_303_epsilon_0_to_fp16, gamma = obj_303_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_129_cast_fp16)[name = tensor("obj_303_cast_fp16")]; + tensor var_4827 = const()[name = tensor("op_4827"), val = tensor([1, 1])]; + tensor var_4829 = const()[name = tensor("op_4829"), val = tensor([1, 1])]; + tensor query_87_pad_type_0 = const()[name = tensor("query_87_pad_type_0"), val = tensor("custom")]; + tensor query_87_pad_0 = const()[name = tensor("query_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1249006080)))]; + tensor layers_21_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1252282944)))]; + tensor query_87_cast_fp16 = conv(bias = layers_21_encoder_attn_q_proj_bias_to_fp16, dilations = var_4829, groups = var_4705, pad = query_87_pad_0, pad_type = query_87_pad_type_0, strides = var_4827, weight = layers_21_encoder_attn_q_proj_weight_to_fp16, x = obj_303_cast_fp16)[name = tensor("query_87_cast_fp16")]; + tensor var_4833 = const()[name = tensor("op_4833"), val = tensor([1, 1])]; + tensor var_4835 = const()[name = tensor("op_4835"), val = tensor([1, 1])]; + tensor key_87_pad_type_0 = const()[name = tensor("key_87_pad_type_0"), val = tensor("custom")]; + tensor key_87_pad_0 = const()[name = tensor("key_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1252285568)))]; + tensor key_87_cast_fp16 = conv(dilations = var_4835, groups = var_4705, pad = key_87_pad_0, pad_type = key_87_pad_type_0, strides = var_4833, weight = layers_21_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_87_cast_fp16")]; + tensor var_4840 = const()[name = tensor("op_4840"), val = tensor([1, 1])]; + tensor var_4842 = const()[name = tensor("op_4842"), val = tensor([1, 1])]; + tensor value_87_pad_type_0 = const()[name = tensor("value_87_pad_type_0"), val = tensor("custom")]; + tensor value_87_pad_0 = const()[name = tensor("value_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1255562432)))]; + tensor layers_21_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1258839296)))]; + tensor value_87_cast_fp16 = conv(bias = layers_21_encoder_attn_v_proj_bias_to_fp16, dilations = var_4842, groups = var_4705, pad = value_87_pad_0, pad_type = value_87_pad_type_0, strides = var_4840, weight = layers_21_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_87_cast_fp16")]; + tensor var_4846 = const()[name = tensor("op_4846"), val = tensor([1, 20, 64, -1])]; + tensor var_4847_cast_fp16 = reshape(shape = var_4846, x = query_87_cast_fp16)[name = tensor("op_4847_cast_fp16")]; + tensor var_4848_to_fp16 = const()[name = tensor("op_4848_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4849_cast_fp16 = mul(x = var_4847_cast_fp16, y = var_4848_to_fp16)[name = tensor("op_4849_cast_fp16")]; + tensor var_4850 = const()[name = tensor("op_4850"), val = tensor([1, 20, 64, -1])]; + tensor var_4851_cast_fp16 = reshape(shape = var_4850, x = key_87_cast_fp16)[name = tensor("op_4851_cast_fp16")]; + tensor mh_w_131_transpose_x_0 = const()[name = tensor("mh_w_131_transpose_x_0"), val = tensor(true)]; + tensor mh_w_131_transpose_y_0 = const()[name = tensor("mh_w_131_transpose_y_0"), val = tensor(false)]; + tensor mh_w_131_cast_fp16 = matmul(transpose_x = mh_w_131_transpose_x_0, transpose_y = mh_w_131_transpose_y_0, x = var_4849_cast_fp16, y = var_4851_cast_fp16)[name = tensor("mh_w_131_cast_fp16")]; + tensor obj_307_cast_fp16 = softmax(axis = var_4698, x = mh_w_131_cast_fp16)[name = tensor("obj_307_cast_fp16")]; + tensor var_4855 = const()[name = tensor("op_4855"), val = tensor([1, 20, 64, -1])]; + tensor var_4856_cast_fp16 = reshape(shape = var_4855, x = value_87_cast_fp16)[name = tensor("op_4856_cast_fp16")]; + tensor attn_87_transpose_x_0 = const()[name = tensor("attn_87_transpose_x_0"), val = tensor(false)]; + tensor attn_87_transpose_y_0 = const()[name = tensor("attn_87_transpose_y_0"), val = tensor(true)]; + tensor attn_87_cast_fp16 = matmul(transpose_x = attn_87_transpose_x_0, transpose_y = attn_87_transpose_y_0, x = var_4856_cast_fp16, y = obj_307_cast_fp16)[name = tensor("attn_87_cast_fp16")]; + tensor var_4859 = const()[name = tensor("op_4859"), val = tensor([1, 1280, 1, -1])]; + tensor input_213_cast_fp16 = reshape(shape = var_4859, x = attn_87_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor var_4863 = const()[name = tensor("op_4863"), val = tensor([1, 1])]; + tensor var_4865 = const()[name = tensor("op_4865"), val = tensor([1, 1])]; + tensor obj_305_pad_type_0 = const()[name = tensor("obj_305_pad_type_0"), val = tensor("custom")]; + tensor obj_305_pad_0 = const()[name = tensor("obj_305_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1258841920)))]; + tensor layers_21_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1262118784)))]; + tensor obj_305_cast_fp16 = conv(bias = layers_21_encoder_attn_o_proj_bias_to_fp16, dilations = var_4865, groups = var_4705, pad = obj_305_pad_0, pad_type = obj_305_pad_type_0, strides = var_4863, weight = layers_21_encoder_attn_o_proj_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("obj_305_cast_fp16")]; + tensor inputs_131_cast_fp16 = add(x = inputs_129_cast_fp16, y = obj_305_cast_fp16)[name = tensor("inputs_131_cast_fp16")]; + tensor var_4874 = const()[name = tensor("op_4874"), val = tensor([1])]; + tensor channels_mean_131_cast_fp16 = reduce_mean(axes = var_4874, keep_dims = var_4706, x = inputs_131_cast_fp16)[name = tensor("channels_mean_131_cast_fp16")]; + tensor zero_mean_131_cast_fp16 = sub(x = inputs_131_cast_fp16, y = channels_mean_131_cast_fp16)[name = tensor("zero_mean_131_cast_fp16")]; + tensor zero_mean_sq_131_cast_fp16 = mul(x = zero_mean_131_cast_fp16, y = zero_mean_131_cast_fp16)[name = tensor("zero_mean_sq_131_cast_fp16")]; + tensor var_4878 = const()[name = tensor("op_4878"), val = tensor([1])]; + tensor var_4879_cast_fp16 = reduce_mean(axes = var_4878, keep_dims = var_4706, x = zero_mean_sq_131_cast_fp16)[name = tensor("op_4879_cast_fp16")]; + tensor var_4880_to_fp16 = const()[name = tensor("op_4880_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4881_cast_fp16 = add(x = var_4879_cast_fp16, y = var_4880_to_fp16)[name = tensor("op_4881_cast_fp16")]; + tensor denom_131_epsilon_0_to_fp16 = const()[name = tensor("denom_131_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_131_cast_fp16 = rsqrt(epsilon = denom_131_epsilon_0_to_fp16, x = var_4881_cast_fp16)[name = tensor("denom_131_cast_fp16")]; + tensor out_131_cast_fp16 = mul(x = zero_mean_131_cast_fp16, y = denom_131_cast_fp16)[name = tensor("out_131_cast_fp16")]; + tensor input_215_gamma_0_to_fp16 = const()[name = tensor("input_215_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1262121408)))]; + tensor input_215_beta_0_to_fp16 = const()[name = tensor("input_215_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1262124032)))]; + tensor input_215_epsilon_0_to_fp16 = const()[name = tensor("input_215_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_215_cast_fp16 = batch_norm(beta = input_215_beta_0_to_fp16, epsilon = input_215_epsilon_0_to_fp16, gamma = input_215_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_131_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor var_4892 = const()[name = tensor("op_4892"), val = tensor([1, 1])]; + tensor var_4894 = const()[name = tensor("op_4894"), val = tensor([1, 1])]; + tensor input_217_pad_type_0 = const()[name = tensor("input_217_pad_type_0"), val = tensor("custom")]; + tensor input_217_pad_0 = const()[name = tensor("input_217_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_fc1_weight_to_fp16 = const()[name = tensor("layers_21_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1262126656)))]; + tensor layers_21_fc1_bias_to_fp16 = const()[name = tensor("layers_21_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1275233920)))]; + tensor input_217_cast_fp16 = conv(bias = layers_21_fc1_bias_to_fp16, dilations = var_4894, groups = var_4705, pad = input_217_pad_0, pad_type = input_217_pad_type_0, strides = var_4892, weight = layers_21_fc1_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_mode_0 = const()[name = tensor("input_219_mode_0"), val = tensor("EXACT")]; + tensor input_219_cast_fp16 = gelu(mode = input_219_mode_0, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor var_4900 = const()[name = tensor("op_4900"), val = tensor([1, 1])]; + tensor var_4902 = const()[name = tensor("op_4902"), val = tensor([1, 1])]; + tensor hidden_states_45_pad_type_0 = const()[name = tensor("hidden_states_45_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_45_pad_0 = const()[name = tensor("hidden_states_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_fc2_weight_to_fp16 = const()[name = tensor("layers_21_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1275244224)))]; + tensor layers_21_fc2_bias_to_fp16 = const()[name = tensor("layers_21_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1288351488)))]; + tensor hidden_states_45_cast_fp16 = conv(bias = layers_21_fc2_bias_to_fp16, dilations = var_4902, groups = var_4705, pad = hidden_states_45_pad_0, pad_type = hidden_states_45_pad_type_0, strides = var_4900, weight = layers_21_fc2_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("hidden_states_45_cast_fp16")]; + tensor inputs_133_cast_fp16 = add(x = inputs_131_cast_fp16, y = hidden_states_45_cast_fp16)[name = tensor("inputs_133_cast_fp16")]; + tensor var_4916 = const()[name = tensor("op_4916"), val = tensor(3)]; + tensor var_4923 = const()[name = tensor("op_4923"), val = tensor(1)]; + tensor var_4924 = const()[name = tensor("op_4924"), val = tensor(true)]; + tensor var_4936 = const()[name = tensor("op_4936"), val = tensor([1])]; + tensor channels_mean_133_cast_fp16 = reduce_mean(axes = var_4936, keep_dims = var_4924, x = inputs_133_cast_fp16)[name = tensor("channels_mean_133_cast_fp16")]; + tensor zero_mean_133_cast_fp16 = sub(x = inputs_133_cast_fp16, y = channels_mean_133_cast_fp16)[name = tensor("zero_mean_133_cast_fp16")]; + tensor zero_mean_sq_133_cast_fp16 = mul(x = zero_mean_133_cast_fp16, y = zero_mean_133_cast_fp16)[name = tensor("zero_mean_sq_133_cast_fp16")]; + tensor var_4940 = const()[name = tensor("op_4940"), val = tensor([1])]; + tensor var_4941_cast_fp16 = reduce_mean(axes = var_4940, keep_dims = var_4924, x = zero_mean_sq_133_cast_fp16)[name = tensor("op_4941_cast_fp16")]; + tensor var_4942_to_fp16 = const()[name = tensor("op_4942_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4943_cast_fp16 = add(x = var_4941_cast_fp16, y = var_4942_to_fp16)[name = tensor("op_4943_cast_fp16")]; + tensor denom_133_epsilon_0_to_fp16 = const()[name = tensor("denom_133_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_133_cast_fp16 = rsqrt(epsilon = denom_133_epsilon_0_to_fp16, x = var_4943_cast_fp16)[name = tensor("denom_133_cast_fp16")]; + tensor out_133_cast_fp16 = mul(x = zero_mean_133_cast_fp16, y = denom_133_cast_fp16)[name = tensor("out_133_cast_fp16")]; + tensor obj_309_gamma_0_to_fp16 = const()[name = tensor("obj_309_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1288354112)))]; + tensor obj_309_beta_0_to_fp16 = const()[name = tensor("obj_309_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1288356736)))]; + tensor obj_309_epsilon_0_to_fp16 = const()[name = tensor("obj_309_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_309_cast_fp16 = batch_norm(beta = obj_309_beta_0_to_fp16, epsilon = obj_309_epsilon_0_to_fp16, gamma = obj_309_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_133_cast_fp16)[name = tensor("obj_309_cast_fp16")]; + tensor var_4958 = const()[name = tensor("op_4958"), val = tensor([1, 1])]; + tensor var_4960 = const()[name = tensor("op_4960"), val = tensor([1, 1])]; + tensor query_89_pad_type_0 = const()[name = tensor("query_89_pad_type_0"), val = tensor("custom")]; + tensor query_89_pad_0 = const()[name = tensor("query_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1288359360)))]; + tensor layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1291636224)))]; + tensor query_89_cast_fp16 = conv(bias = layers_22_self_attn_q_proj_bias_to_fp16, dilations = var_4960, groups = var_4923, pad = query_89_pad_0, pad_type = query_89_pad_type_0, strides = var_4958, weight = layers_22_self_attn_q_proj_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("query_89_cast_fp16")]; + tensor var_4964 = const()[name = tensor("op_4964"), val = tensor([1, 1])]; + tensor var_4966 = const()[name = tensor("op_4966"), val = tensor([1, 1])]; + tensor current_key_45_pad_type_0 = const()[name = tensor("current_key_45_pad_type_0"), val = tensor("custom")]; + tensor current_key_45_pad_0 = const()[name = tensor("current_key_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1291638848)))]; + tensor current_key_45_cast_fp16 = conv(dilations = var_4966, groups = var_4923, pad = current_key_45_pad_0, pad_type = current_key_45_pad_type_0, strides = var_4964, weight = layers_22_self_attn_k_proj_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("current_key_45_cast_fp16")]; + tensor var_4971 = const()[name = tensor("op_4971"), val = tensor([1, 1])]; + tensor var_4973 = const()[name = tensor("op_4973"), val = tensor([1, 1])]; + tensor current_value_45_pad_type_0 = const()[name = tensor("current_value_45_pad_type_0"), val = tensor("custom")]; + tensor current_value_45_pad_0 = const()[name = tensor("current_value_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1294915712)))]; + tensor layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1298192576)))]; + tensor current_value_45_cast_fp16 = conv(bias = layers_22_self_attn_v_proj_bias_to_fp16, dilations = var_4973, groups = var_4923, pad = current_value_45_pad_0, pad_type = current_value_45_pad_type_0, strides = var_4971, weight = layers_22_self_attn_v_proj_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("current_value_45_cast_fp16")]; + tensor var_4980_cast_fp16 = mul(x = current_key_45_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4980_cast_fp16")]; + tensor var_4982_cast_fp16 = mul(x = var_103_cast_fp16_22, y = var_241_cast_fp16)[name = tensor("op_4982_cast_fp16")]; + tensor key_89_cast_fp16 = add(x = var_4980_cast_fp16, y = var_4982_cast_fp16)[name = tensor("key_89_cast_fp16")]; + tensor var_4984_cast_fp16 = mul(x = current_value_45_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4984_cast_fp16")]; + tensor var_4986_cast_fp16 = mul(x = var_138_cast_fp16_22, y = var_241_cast_fp16)[name = tensor("op_4986_cast_fp16")]; + tensor value_89_cast_fp16 = add(x = var_4984_cast_fp16, y = var_4986_cast_fp16)[name = tensor("value_89_cast_fp16")]; + tensor var_4989 = const()[name = tensor("op_4989"), val = tensor([1, 20, 64, -1])]; + tensor var_4990_cast_fp16 = reshape(shape = var_4989, x = query_89_cast_fp16)[name = tensor("op_4990_cast_fp16")]; + tensor var_4991_to_fp16 = const()[name = tensor("op_4991_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4992_cast_fp16 = mul(x = var_4990_cast_fp16, y = var_4991_to_fp16)[name = tensor("op_4992_cast_fp16")]; + tensor var_4993 = const()[name = tensor("op_4993"), val = tensor([1, 20, 64, -1])]; + tensor var_4994_cast_fp16 = reshape(shape = var_4993, x = key_89_cast_fp16)[name = tensor("op_4994_cast_fp16")]; + tensor mh_w_133_transpose_x_0 = const()[name = tensor("mh_w_133_transpose_x_0"), val = tensor(true)]; + tensor mh_w_133_transpose_y_0 = const()[name = tensor("mh_w_133_transpose_y_0"), val = tensor(false)]; + tensor mh_w_133_cast_fp16 = matmul(transpose_x = mh_w_133_transpose_x_0, transpose_y = mh_w_133_transpose_y_0, x = var_4992_cast_fp16, y = var_4994_cast_fp16)[name = tensor("mh_w_133_cast_fp16")]; + tensor mh_w_135_cast_fp16 = add(x = mh_w_133_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_135_cast_fp16")]; + tensor var_5002_cast_fp16 = softmax(axis = var_4916, x = mh_w_135_cast_fp16)[name = tensor("op_5002_cast_fp16")]; + tensor var_5003 = const()[name = tensor("op_5003"), val = tensor([1, 20, 64, -1])]; + tensor var_5004_cast_fp16 = reshape(shape = var_5003, x = value_89_cast_fp16)[name = tensor("op_5004_cast_fp16")]; + tensor attn_89_transpose_x_0 = const()[name = tensor("attn_89_transpose_x_0"), val = tensor(false)]; + tensor attn_89_transpose_y_0 = const()[name = tensor("attn_89_transpose_y_0"), val = tensor(true)]; + tensor attn_89_cast_fp16 = matmul(transpose_x = attn_89_transpose_x_0, transpose_y = attn_89_transpose_y_0, x = var_5004_cast_fp16, y = var_5002_cast_fp16)[name = tensor("attn_89_cast_fp16")]; + tensor var_5007 = const()[name = tensor("op_5007"), val = tensor([1, 1280, 1, -1])]; + tensor input_221_cast_fp16 = reshape(shape = var_5007, x = attn_89_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor var_5011 = const()[name = tensor("op_5011"), val = tensor([1, 1])]; + tensor var_5013 = const()[name = tensor("op_5013"), val = tensor([1, 1])]; + tensor obj_315_pad_type_0 = const()[name = tensor("obj_315_pad_type_0"), val = tensor("custom")]; + tensor obj_315_pad_0 = const()[name = tensor("obj_315_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1298195200)))]; + tensor layers_22_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1301472064)))]; + tensor obj_315_cast_fp16 = conv(bias = layers_22_self_attn_o_proj_bias_to_fp16, dilations = var_5013, groups = var_4923, pad = obj_315_pad_0, pad_type = obj_315_pad_type_0, strides = var_5011, weight = layers_22_self_attn_o_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("obj_315_cast_fp16")]; + tensor inputs_135_cast_fp16 = add(x = inputs_133_cast_fp16, y = obj_315_cast_fp16)[name = tensor("inputs_135_cast_fp16")]; + tensor var_5023 = const()[name = tensor("op_5023"), val = tensor([1])]; + tensor channels_mean_135_cast_fp16 = reduce_mean(axes = var_5023, keep_dims = var_4924, x = inputs_135_cast_fp16)[name = tensor("channels_mean_135_cast_fp16")]; + tensor zero_mean_135_cast_fp16 = sub(x = inputs_135_cast_fp16, y = channels_mean_135_cast_fp16)[name = tensor("zero_mean_135_cast_fp16")]; + tensor zero_mean_sq_135_cast_fp16 = mul(x = zero_mean_135_cast_fp16, y = zero_mean_135_cast_fp16)[name = tensor("zero_mean_sq_135_cast_fp16")]; + tensor var_5027 = const()[name = tensor("op_5027"), val = tensor([1])]; + tensor var_5028_cast_fp16 = reduce_mean(axes = var_5027, keep_dims = var_4924, x = zero_mean_sq_135_cast_fp16)[name = tensor("op_5028_cast_fp16")]; + tensor var_5029_to_fp16 = const()[name = tensor("op_5029_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5030_cast_fp16 = add(x = var_5028_cast_fp16, y = var_5029_to_fp16)[name = tensor("op_5030_cast_fp16")]; + tensor denom_135_epsilon_0_to_fp16 = const()[name = tensor("denom_135_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_135_cast_fp16 = rsqrt(epsilon = denom_135_epsilon_0_to_fp16, x = var_5030_cast_fp16)[name = tensor("denom_135_cast_fp16")]; + tensor out_135_cast_fp16 = mul(x = zero_mean_135_cast_fp16, y = denom_135_cast_fp16)[name = tensor("out_135_cast_fp16")]; + tensor obj_317_gamma_0_to_fp16 = const()[name = tensor("obj_317_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1301474688)))]; + tensor obj_317_beta_0_to_fp16 = const()[name = tensor("obj_317_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1301477312)))]; + tensor obj_317_epsilon_0_to_fp16 = const()[name = tensor("obj_317_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_317_cast_fp16 = batch_norm(beta = obj_317_beta_0_to_fp16, epsilon = obj_317_epsilon_0_to_fp16, gamma = obj_317_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_135_cast_fp16)[name = tensor("obj_317_cast_fp16")]; + tensor var_5045 = const()[name = tensor("op_5045"), val = tensor([1, 1])]; + tensor var_5047 = const()[name = tensor("op_5047"), val = tensor([1, 1])]; + tensor query_91_pad_type_0 = const()[name = tensor("query_91_pad_type_0"), val = tensor("custom")]; + tensor query_91_pad_0 = const()[name = tensor("query_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1301479936)))]; + tensor layers_22_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1304756800)))]; + tensor query_91_cast_fp16 = conv(bias = layers_22_encoder_attn_q_proj_bias_to_fp16, dilations = var_5047, groups = var_4923, pad = query_91_pad_0, pad_type = query_91_pad_type_0, strides = var_5045, weight = layers_22_encoder_attn_q_proj_weight_to_fp16, x = obj_317_cast_fp16)[name = tensor("query_91_cast_fp16")]; + tensor var_5051 = const()[name = tensor("op_5051"), val = tensor([1, 1])]; + tensor var_5053 = const()[name = tensor("op_5053"), val = tensor([1, 1])]; + tensor key_91_pad_type_0 = const()[name = tensor("key_91_pad_type_0"), val = tensor("custom")]; + tensor key_91_pad_0 = const()[name = tensor("key_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1304759424)))]; + tensor key_91_cast_fp16 = conv(dilations = var_5053, groups = var_4923, pad = key_91_pad_0, pad_type = key_91_pad_type_0, strides = var_5051, weight = layers_22_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_91_cast_fp16")]; + tensor var_5058 = const()[name = tensor("op_5058"), val = tensor([1, 1])]; + tensor var_5060 = const()[name = tensor("op_5060"), val = tensor([1, 1])]; + tensor value_91_pad_type_0 = const()[name = tensor("value_91_pad_type_0"), val = tensor("custom")]; + tensor value_91_pad_0 = const()[name = tensor("value_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1308036288)))]; + tensor layers_22_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1311313152)))]; + tensor value_91_cast_fp16 = conv(bias = layers_22_encoder_attn_v_proj_bias_to_fp16, dilations = var_5060, groups = var_4923, pad = value_91_pad_0, pad_type = value_91_pad_type_0, strides = var_5058, weight = layers_22_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_91_cast_fp16")]; + tensor var_5064 = const()[name = tensor("op_5064"), val = tensor([1, 20, 64, -1])]; + tensor var_5065_cast_fp16 = reshape(shape = var_5064, x = query_91_cast_fp16)[name = tensor("op_5065_cast_fp16")]; + tensor var_5066_to_fp16 = const()[name = tensor("op_5066_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5067_cast_fp16 = mul(x = var_5065_cast_fp16, y = var_5066_to_fp16)[name = tensor("op_5067_cast_fp16")]; + tensor var_5068 = const()[name = tensor("op_5068"), val = tensor([1, 20, 64, -1])]; + tensor var_5069_cast_fp16 = reshape(shape = var_5068, x = key_91_cast_fp16)[name = tensor("op_5069_cast_fp16")]; + tensor mh_w_137_transpose_x_0 = const()[name = tensor("mh_w_137_transpose_x_0"), val = tensor(true)]; + tensor mh_w_137_transpose_y_0 = const()[name = tensor("mh_w_137_transpose_y_0"), val = tensor(false)]; + tensor mh_w_137_cast_fp16 = matmul(transpose_x = mh_w_137_transpose_x_0, transpose_y = mh_w_137_transpose_y_0, x = var_5067_cast_fp16, y = var_5069_cast_fp16)[name = tensor("mh_w_137_cast_fp16")]; + tensor obj_321_cast_fp16 = softmax(axis = var_4916, x = mh_w_137_cast_fp16)[name = tensor("obj_321_cast_fp16")]; + tensor var_5073 = const()[name = tensor("op_5073"), val = tensor([1, 20, 64, -1])]; + tensor var_5074_cast_fp16 = reshape(shape = var_5073, x = value_91_cast_fp16)[name = tensor("op_5074_cast_fp16")]; + tensor attn_91_transpose_x_0 = const()[name = tensor("attn_91_transpose_x_0"), val = tensor(false)]; + tensor attn_91_transpose_y_0 = const()[name = tensor("attn_91_transpose_y_0"), val = tensor(true)]; + tensor attn_91_cast_fp16 = matmul(transpose_x = attn_91_transpose_x_0, transpose_y = attn_91_transpose_y_0, x = var_5074_cast_fp16, y = obj_321_cast_fp16)[name = tensor("attn_91_cast_fp16")]; + tensor var_5077 = const()[name = tensor("op_5077"), val = tensor([1, 1280, 1, -1])]; + tensor input_223_cast_fp16 = reshape(shape = var_5077, x = attn_91_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor var_5081 = const()[name = tensor("op_5081"), val = tensor([1, 1])]; + tensor var_5083 = const()[name = tensor("op_5083"), val = tensor([1, 1])]; + tensor obj_319_pad_type_0 = const()[name = tensor("obj_319_pad_type_0"), val = tensor("custom")]; + tensor obj_319_pad_0 = const()[name = tensor("obj_319_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1311315776)))]; + tensor layers_22_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1314592640)))]; + tensor obj_319_cast_fp16 = conv(bias = layers_22_encoder_attn_o_proj_bias_to_fp16, dilations = var_5083, groups = var_4923, pad = obj_319_pad_0, pad_type = obj_319_pad_type_0, strides = var_5081, weight = layers_22_encoder_attn_o_proj_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("obj_319_cast_fp16")]; + tensor inputs_137_cast_fp16 = add(x = inputs_135_cast_fp16, y = obj_319_cast_fp16)[name = tensor("inputs_137_cast_fp16")]; + tensor var_5089 = const()[name = tensor("op_5089"), val = tensor([1])]; + tensor channels_mean_137_cast_fp16 = reduce_mean(axes = var_5089, keep_dims = var_4924, x = inputs_137_cast_fp16)[name = tensor("channels_mean_137_cast_fp16")]; + tensor zero_mean_137_cast_fp16 = sub(x = inputs_137_cast_fp16, y = channels_mean_137_cast_fp16)[name = tensor("zero_mean_137_cast_fp16")]; + tensor zero_mean_sq_137_cast_fp16 = mul(x = zero_mean_137_cast_fp16, y = zero_mean_137_cast_fp16)[name = tensor("zero_mean_sq_137_cast_fp16")]; + tensor var_5093 = const()[name = tensor("op_5093"), val = tensor([1])]; + tensor var_5094_cast_fp16 = reduce_mean(axes = var_5093, keep_dims = var_4924, x = zero_mean_sq_137_cast_fp16)[name = tensor("op_5094_cast_fp16")]; + tensor var_5095_to_fp16 = const()[name = tensor("op_5095_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5096_cast_fp16 = add(x = var_5094_cast_fp16, y = var_5095_to_fp16)[name = tensor("op_5096_cast_fp16")]; + tensor denom_137_epsilon_0_to_fp16 = const()[name = tensor("denom_137_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_137_cast_fp16 = rsqrt(epsilon = denom_137_epsilon_0_to_fp16, x = var_5096_cast_fp16)[name = tensor("denom_137_cast_fp16")]; + tensor out_137_cast_fp16 = mul(x = zero_mean_137_cast_fp16, y = denom_137_cast_fp16)[name = tensor("out_137_cast_fp16")]; + tensor input_225_gamma_0_to_fp16 = const()[name = tensor("input_225_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1314595264)))]; + tensor input_225_beta_0_to_fp16 = const()[name = tensor("input_225_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1314597888)))]; + tensor input_225_epsilon_0_to_fp16 = const()[name = tensor("input_225_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_225_cast_fp16 = batch_norm(beta = input_225_beta_0_to_fp16, epsilon = input_225_epsilon_0_to_fp16, gamma = input_225_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_137_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor var_5107 = const()[name = tensor("op_5107"), val = tensor([1, 1])]; + tensor var_5109 = const()[name = tensor("op_5109"), val = tensor([1, 1])]; + tensor input_227_pad_type_0 = const()[name = tensor("input_227_pad_type_0"), val = tensor("custom")]; + tensor input_227_pad_0 = const()[name = tensor("input_227_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_fc1_weight_to_fp16 = const()[name = tensor("layers_22_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1314600512)))]; + tensor layers_22_fc1_bias_to_fp16 = const()[name = tensor("layers_22_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1327707776)))]; + tensor input_227_cast_fp16 = conv(bias = layers_22_fc1_bias_to_fp16, dilations = var_5109, groups = var_4923, pad = input_227_pad_0, pad_type = input_227_pad_type_0, strides = var_5107, weight = layers_22_fc1_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor input_229_mode_0 = const()[name = tensor("input_229_mode_0"), val = tensor("EXACT")]; + tensor input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor var_5115 = const()[name = tensor("op_5115"), val = tensor([1, 1])]; + tensor var_5117 = const()[name = tensor("op_5117"), val = tensor([1, 1])]; + tensor hidden_states_47_pad_type_0 = const()[name = tensor("hidden_states_47_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_47_pad_0 = const()[name = tensor("hidden_states_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_fc2_weight_to_fp16 = const()[name = tensor("layers_22_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1327718080)))]; + tensor layers_22_fc2_bias_to_fp16 = const()[name = tensor("layers_22_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1340825344)))]; + tensor hidden_states_47_cast_fp16 = conv(bias = layers_22_fc2_bias_to_fp16, dilations = var_5117, groups = var_4923, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = var_5115, weight = layers_22_fc2_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("hidden_states_47_cast_fp16")]; + tensor inputs_139_cast_fp16 = add(x = inputs_137_cast_fp16, y = hidden_states_47_cast_fp16)[name = tensor("inputs_139_cast_fp16")]; + tensor var_5130 = const()[name = tensor("op_5130"), val = tensor(3)]; + tensor var_5137 = const()[name = tensor("op_5137"), val = tensor(1)]; + tensor var_5138 = const()[name = tensor("op_5138"), val = tensor(true)]; + tensor var_5150 = const()[name = tensor("op_5150"), val = tensor([1])]; + tensor channels_mean_139_cast_fp16 = reduce_mean(axes = var_5150, keep_dims = var_5138, x = inputs_139_cast_fp16)[name = tensor("channels_mean_139_cast_fp16")]; + tensor zero_mean_139_cast_fp16 = sub(x = inputs_139_cast_fp16, y = channels_mean_139_cast_fp16)[name = tensor("zero_mean_139_cast_fp16")]; + tensor zero_mean_sq_139_cast_fp16 = mul(x = zero_mean_139_cast_fp16, y = zero_mean_139_cast_fp16)[name = tensor("zero_mean_sq_139_cast_fp16")]; + tensor var_5154 = const()[name = tensor("op_5154"), val = tensor([1])]; + tensor var_5155_cast_fp16 = reduce_mean(axes = var_5154, keep_dims = var_5138, x = zero_mean_sq_139_cast_fp16)[name = tensor("op_5155_cast_fp16")]; + tensor var_5156_to_fp16 = const()[name = tensor("op_5156_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5157_cast_fp16 = add(x = var_5155_cast_fp16, y = var_5156_to_fp16)[name = tensor("op_5157_cast_fp16")]; + tensor denom_139_epsilon_0_to_fp16 = const()[name = tensor("denom_139_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_139_cast_fp16 = rsqrt(epsilon = denom_139_epsilon_0_to_fp16, x = var_5157_cast_fp16)[name = tensor("denom_139_cast_fp16")]; + tensor out_139_cast_fp16 = mul(x = zero_mean_139_cast_fp16, y = denom_139_cast_fp16)[name = tensor("out_139_cast_fp16")]; + tensor obj_323_gamma_0_to_fp16 = const()[name = tensor("obj_323_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1340827968)))]; + tensor obj_323_beta_0_to_fp16 = const()[name = tensor("obj_323_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1340830592)))]; + tensor obj_323_epsilon_0_to_fp16 = const()[name = tensor("obj_323_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_323_cast_fp16 = batch_norm(beta = obj_323_beta_0_to_fp16, epsilon = obj_323_epsilon_0_to_fp16, gamma = obj_323_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_139_cast_fp16)[name = tensor("obj_323_cast_fp16")]; + tensor var_5172 = const()[name = tensor("op_5172"), val = tensor([1, 1])]; + tensor var_5174 = const()[name = tensor("op_5174"), val = tensor([1, 1])]; + tensor query_93_pad_type_0 = const()[name = tensor("query_93_pad_type_0"), val = tensor("custom")]; + tensor query_93_pad_0 = const()[name = tensor("query_93_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1340833216)))]; + tensor layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1344110080)))]; + tensor query_93_cast_fp16 = conv(bias = layers_23_self_attn_q_proj_bias_to_fp16, dilations = var_5174, groups = var_5137, pad = query_93_pad_0, pad_type = query_93_pad_type_0, strides = var_5172, weight = layers_23_self_attn_q_proj_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor("query_93_cast_fp16")]; + tensor var_5178 = const()[name = tensor("op_5178"), val = tensor([1, 1])]; + tensor var_5180 = const()[name = tensor("op_5180"), val = tensor([1, 1])]; + tensor current_key_47_pad_type_0 = const()[name = tensor("current_key_47_pad_type_0"), val = tensor("custom")]; + tensor current_key_47_pad_0 = const()[name = tensor("current_key_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1344112704)))]; + tensor current_key_47_cast_fp16 = conv(dilations = var_5180, groups = var_5137, pad = current_key_47_pad_0, pad_type = current_key_47_pad_type_0, strides = var_5178, weight = layers_23_self_attn_k_proj_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor("current_key_47_cast_fp16")]; + tensor var_5185 = const()[name = tensor("op_5185"), val = tensor([1, 1])]; + tensor var_5187 = const()[name = tensor("op_5187"), val = tensor([1, 1])]; + tensor current_value_47_pad_type_0 = const()[name = tensor("current_value_47_pad_type_0"), val = tensor("custom")]; + tensor current_value_47_pad_0 = const()[name = tensor("current_value_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1347389568)))]; + tensor layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1350666432)))]; + tensor current_value_47_cast_fp16 = conv(bias = layers_23_self_attn_v_proj_bias_to_fp16, dilations = var_5187, groups = var_5137, pad = current_value_47_pad_0, pad_type = current_value_47_pad_type_0, strides = var_5185, weight = layers_23_self_attn_v_proj_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor("current_value_47_cast_fp16")]; + tensor var_5194_cast_fp16 = mul(x = current_key_47_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_5194_cast_fp16")]; + tensor var_5196_cast_fp16 = mul(x = var_103_cast_fp16_23, y = var_241_cast_fp16)[name = tensor("op_5196_cast_fp16")]; + tensor key_93_cast_fp16 = add(x = var_5194_cast_fp16, y = var_5196_cast_fp16)[name = tensor("key_93_cast_fp16")]; + tensor var_5198_cast_fp16 = mul(x = current_value_47_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_5198_cast_fp16")]; + tensor var_5200_cast_fp16 = mul(x = var_138_cast_fp16_23, y = var_241_cast_fp16)[name = tensor("op_5200_cast_fp16")]; + tensor value_93_cast_fp16 = add(x = var_5198_cast_fp16, y = var_5200_cast_fp16)[name = tensor("value_93_cast_fp16")]; + tensor var_5203 = const()[name = tensor("op_5203"), val = tensor([1, 20, 64, -1])]; + tensor var_5204_cast_fp16 = reshape(shape = var_5203, x = query_93_cast_fp16)[name = tensor("op_5204_cast_fp16")]; + tensor var_5205_to_fp16 = const()[name = tensor("op_5205_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5206_cast_fp16 = mul(x = var_5204_cast_fp16, y = var_5205_to_fp16)[name = tensor("op_5206_cast_fp16")]; + tensor var_5207 = const()[name = tensor("op_5207"), val = tensor([1, 20, 64, -1])]; + tensor var_5208_cast_fp16 = reshape(shape = var_5207, x = key_93_cast_fp16)[name = tensor("op_5208_cast_fp16")]; + tensor mh_w_139_transpose_x_0 = const()[name = tensor("mh_w_139_transpose_x_0"), val = tensor(true)]; + tensor mh_w_139_transpose_y_0 = const()[name = tensor("mh_w_139_transpose_y_0"), val = tensor(false)]; + tensor mh_w_139_cast_fp16 = matmul(transpose_x = mh_w_139_transpose_x_0, transpose_y = mh_w_139_transpose_y_0, x = var_5206_cast_fp16, y = var_5208_cast_fp16)[name = tensor("mh_w_139_cast_fp16")]; + tensor mh_w_141_cast_fp16 = add(x = mh_w_139_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_141_cast_fp16")]; + tensor var_5216_cast_fp16 = softmax(axis = var_5130, x = mh_w_141_cast_fp16)[name = tensor("op_5216_cast_fp16")]; + tensor var_5217 = const()[name = tensor("op_5217"), val = tensor([1, 20, 64, -1])]; + tensor var_5218_cast_fp16 = reshape(shape = var_5217, x = value_93_cast_fp16)[name = tensor("op_5218_cast_fp16")]; + tensor attn_93_transpose_x_0 = const()[name = tensor("attn_93_transpose_x_0"), val = tensor(false)]; + tensor attn_93_transpose_y_0 = const()[name = tensor("attn_93_transpose_y_0"), val = tensor(true)]; + tensor attn_93_cast_fp16 = matmul(transpose_x = attn_93_transpose_x_0, transpose_y = attn_93_transpose_y_0, x = var_5218_cast_fp16, y = var_5216_cast_fp16)[name = tensor("attn_93_cast_fp16")]; + tensor var_5221 = const()[name = tensor("op_5221"), val = tensor([1, 1280, 1, -1])]; + tensor input_231_cast_fp16 = reshape(shape = var_5221, x = attn_93_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor var_5225 = const()[name = tensor("op_5225"), val = tensor([1, 1])]; + tensor var_5227 = const()[name = tensor("op_5227"), val = tensor([1, 1])]; + tensor obj_329_pad_type_0 = const()[name = tensor("obj_329_pad_type_0"), val = tensor("custom")]; + tensor obj_329_pad_0 = const()[name = tensor("obj_329_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1350669056)))]; + tensor layers_23_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1353945920)))]; + tensor obj_329_cast_fp16 = conv(bias = layers_23_self_attn_o_proj_bias_to_fp16, dilations = var_5227, groups = var_5137, pad = obj_329_pad_0, pad_type = obj_329_pad_type_0, strides = var_5225, weight = layers_23_self_attn_o_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("obj_329_cast_fp16")]; + tensor inputs_141_cast_fp16 = add(x = inputs_139_cast_fp16, y = obj_329_cast_fp16)[name = tensor("inputs_141_cast_fp16")]; + tensor var_5237 = const()[name = tensor("op_5237"), val = tensor([1])]; + tensor channels_mean_141_cast_fp16 = reduce_mean(axes = var_5237, keep_dims = var_5138, x = inputs_141_cast_fp16)[name = tensor("channels_mean_141_cast_fp16")]; + tensor zero_mean_141_cast_fp16 = sub(x = inputs_141_cast_fp16, y = channels_mean_141_cast_fp16)[name = tensor("zero_mean_141_cast_fp16")]; + tensor zero_mean_sq_141_cast_fp16 = mul(x = zero_mean_141_cast_fp16, y = zero_mean_141_cast_fp16)[name = tensor("zero_mean_sq_141_cast_fp16")]; + tensor var_5241 = const()[name = tensor("op_5241"), val = tensor([1])]; + tensor var_5242_cast_fp16 = reduce_mean(axes = var_5241, keep_dims = var_5138, x = zero_mean_sq_141_cast_fp16)[name = tensor("op_5242_cast_fp16")]; + tensor var_5243_to_fp16 = const()[name = tensor("op_5243_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5244_cast_fp16 = add(x = var_5242_cast_fp16, y = var_5243_to_fp16)[name = tensor("op_5244_cast_fp16")]; + tensor denom_141_epsilon_0_to_fp16 = const()[name = tensor("denom_141_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_141_cast_fp16 = rsqrt(epsilon = denom_141_epsilon_0_to_fp16, x = var_5244_cast_fp16)[name = tensor("denom_141_cast_fp16")]; + tensor out_141_cast_fp16 = mul(x = zero_mean_141_cast_fp16, y = denom_141_cast_fp16)[name = tensor("out_141_cast_fp16")]; + tensor obj_331_gamma_0_to_fp16 = const()[name = tensor("obj_331_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1353948544)))]; + tensor obj_331_beta_0_to_fp16 = const()[name = tensor("obj_331_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1353951168)))]; + tensor obj_331_epsilon_0_to_fp16 = const()[name = tensor("obj_331_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_331_cast_fp16 = batch_norm(beta = obj_331_beta_0_to_fp16, epsilon = obj_331_epsilon_0_to_fp16, gamma = obj_331_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_141_cast_fp16)[name = tensor("obj_331_cast_fp16")]; + tensor var_5259 = const()[name = tensor("op_5259"), val = tensor([1, 1])]; + tensor var_5261 = const()[name = tensor("op_5261"), val = tensor([1, 1])]; + tensor query_95_pad_type_0 = const()[name = tensor("query_95_pad_type_0"), val = tensor("custom")]; + tensor query_95_pad_0 = const()[name = tensor("query_95_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1353953792)))]; + tensor layers_23_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1357230656)))]; + tensor query_95_cast_fp16 = conv(bias = layers_23_encoder_attn_q_proj_bias_to_fp16, dilations = var_5261, groups = var_5137, pad = query_95_pad_0, pad_type = query_95_pad_type_0, strides = var_5259, weight = layers_23_encoder_attn_q_proj_weight_to_fp16, x = obj_331_cast_fp16)[name = tensor("query_95_cast_fp16")]; + tensor var_5265 = const()[name = tensor("op_5265"), val = tensor([1, 1])]; + tensor var_5267 = const()[name = tensor("op_5267"), val = tensor([1, 1])]; + tensor key_95_pad_type_0 = const()[name = tensor("key_95_pad_type_0"), val = tensor("custom")]; + tensor key_95_pad_0 = const()[name = tensor("key_95_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1357233280)))]; + tensor key_95_cast_fp16 = conv(dilations = var_5267, groups = var_5137, pad = key_95_pad_0, pad_type = key_95_pad_type_0, strides = var_5265, weight = layers_23_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_95_cast_fp16")]; + tensor var_5272 = const()[name = tensor("op_5272"), val = tensor([1, 1])]; + tensor var_5274 = const()[name = tensor("op_5274"), val = tensor([1, 1])]; + tensor value_95_pad_type_0 = const()[name = tensor("value_95_pad_type_0"), val = tensor("custom")]; + tensor value_95_pad_0 = const()[name = tensor("value_95_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1360510144)))]; + tensor layers_23_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1363787008)))]; + tensor value_95_cast_fp16 = conv(bias = layers_23_encoder_attn_v_proj_bias_to_fp16, dilations = var_5274, groups = var_5137, pad = value_95_pad_0, pad_type = value_95_pad_type_0, strides = var_5272, weight = layers_23_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_95_cast_fp16")]; + tensor var_5278 = const()[name = tensor("op_5278"), val = tensor([1, 20, 64, -1])]; + tensor var_5279_cast_fp16 = reshape(shape = var_5278, x = query_95_cast_fp16)[name = tensor("op_5279_cast_fp16")]; + tensor var_5280_to_fp16 = const()[name = tensor("op_5280_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5281_cast_fp16 = mul(x = var_5279_cast_fp16, y = var_5280_to_fp16)[name = tensor("op_5281_cast_fp16")]; + tensor var_5282 = const()[name = tensor("op_5282"), val = tensor([1, 20, 64, -1])]; + tensor var_5283_cast_fp16 = reshape(shape = var_5282, x = key_95_cast_fp16)[name = tensor("op_5283_cast_fp16")]; + tensor mh_w_143_transpose_x_0 = const()[name = tensor("mh_w_143_transpose_x_0"), val = tensor(true)]; + tensor mh_w_143_transpose_y_0 = const()[name = tensor("mh_w_143_transpose_y_0"), val = tensor(false)]; + tensor mh_w_143_cast_fp16 = matmul(transpose_x = mh_w_143_transpose_x_0, transpose_y = mh_w_143_transpose_y_0, x = var_5281_cast_fp16, y = var_5283_cast_fp16)[name = tensor("mh_w_143_cast_fp16")]; + tensor obj_335_cast_fp16 = softmax(axis = var_5130, x = mh_w_143_cast_fp16)[name = tensor("obj_335_cast_fp16")]; + tensor var_5287 = const()[name = tensor("op_5287"), val = tensor([1, 20, 64, -1])]; + tensor var_5288_cast_fp16 = reshape(shape = var_5287, x = value_95_cast_fp16)[name = tensor("op_5288_cast_fp16")]; + tensor attn_95_transpose_x_0 = const()[name = tensor("attn_95_transpose_x_0"), val = tensor(false)]; + tensor attn_95_transpose_y_0 = const()[name = tensor("attn_95_transpose_y_0"), val = tensor(true)]; + tensor attn_95_cast_fp16 = matmul(transpose_x = attn_95_transpose_x_0, transpose_y = attn_95_transpose_y_0, x = var_5288_cast_fp16, y = obj_335_cast_fp16)[name = tensor("attn_95_cast_fp16")]; + tensor var_5291 = const()[name = tensor("op_5291"), val = tensor([1, 1280, 1, -1])]; + tensor input_233_cast_fp16 = reshape(shape = var_5291, x = attn_95_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor var_5295 = const()[name = tensor("op_5295"), val = tensor([1, 1])]; + tensor var_5297 = const()[name = tensor("op_5297"), val = tensor([1, 1])]; + tensor obj_333_pad_type_0 = const()[name = tensor("obj_333_pad_type_0"), val = tensor("custom")]; + tensor obj_333_pad_0 = const()[name = tensor("obj_333_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1363789632)))]; + tensor layers_23_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1367066496)))]; + tensor obj_333_cast_fp16 = conv(bias = layers_23_encoder_attn_o_proj_bias_to_fp16, dilations = var_5297, groups = var_5137, pad = obj_333_pad_0, pad_type = obj_333_pad_type_0, strides = var_5295, weight = layers_23_encoder_attn_o_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("obj_333_cast_fp16")]; + tensor inputs_143_cast_fp16 = add(x = inputs_141_cast_fp16, y = obj_333_cast_fp16)[name = tensor("inputs_143_cast_fp16")]; + tensor var_5303 = const()[name = tensor("op_5303"), val = tensor([1])]; + tensor channels_mean_143_cast_fp16 = reduce_mean(axes = var_5303, keep_dims = var_5138, x = inputs_143_cast_fp16)[name = tensor("channels_mean_143_cast_fp16")]; + tensor zero_mean_143_cast_fp16 = sub(x = inputs_143_cast_fp16, y = channels_mean_143_cast_fp16)[name = tensor("zero_mean_143_cast_fp16")]; + tensor zero_mean_sq_143_cast_fp16 = mul(x = zero_mean_143_cast_fp16, y = zero_mean_143_cast_fp16)[name = tensor("zero_mean_sq_143_cast_fp16")]; + tensor var_5307 = const()[name = tensor("op_5307"), val = tensor([1])]; + tensor var_5308_cast_fp16 = reduce_mean(axes = var_5307, keep_dims = var_5138, x = zero_mean_sq_143_cast_fp16)[name = tensor("op_5308_cast_fp16")]; + tensor var_5309_to_fp16 = const()[name = tensor("op_5309_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5310_cast_fp16 = add(x = var_5308_cast_fp16, y = var_5309_to_fp16)[name = tensor("op_5310_cast_fp16")]; + tensor denom_143_epsilon_0_to_fp16 = const()[name = tensor("denom_143_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_143_cast_fp16 = rsqrt(epsilon = denom_143_epsilon_0_to_fp16, x = var_5310_cast_fp16)[name = tensor("denom_143_cast_fp16")]; + tensor out_143_cast_fp16 = mul(x = zero_mean_143_cast_fp16, y = denom_143_cast_fp16)[name = tensor("out_143_cast_fp16")]; + tensor input_235_gamma_0_to_fp16 = const()[name = tensor("input_235_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1367069120)))]; + tensor input_235_beta_0_to_fp16 = const()[name = tensor("input_235_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1367071744)))]; + tensor input_235_epsilon_0_to_fp16 = const()[name = tensor("input_235_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_235_cast_fp16 = batch_norm(beta = input_235_beta_0_to_fp16, epsilon = input_235_epsilon_0_to_fp16, gamma = input_235_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_143_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor var_5321 = const()[name = tensor("op_5321"), val = tensor([1, 1])]; + tensor var_5323 = const()[name = tensor("op_5323"), val = tensor([1, 1])]; + tensor input_237_pad_type_0 = const()[name = tensor("input_237_pad_type_0"), val = tensor("custom")]; + tensor input_237_pad_0 = const()[name = tensor("input_237_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_fc1_weight_to_fp16 = const()[name = tensor("layers_23_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1367074368)))]; + tensor layers_23_fc1_bias_to_fp16 = const()[name = tensor("layers_23_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1380181632)))]; + tensor input_237_cast_fp16 = conv(bias = layers_23_fc1_bias_to_fp16, dilations = var_5323, groups = var_5137, pad = input_237_pad_0, pad_type = input_237_pad_type_0, strides = var_5321, weight = layers_23_fc1_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_mode_0 = const()[name = tensor("input_239_mode_0"), val = tensor("EXACT")]; + tensor input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor var_5329 = const()[name = tensor("op_5329"), val = tensor([1, 1])]; + tensor var_5331 = const()[name = tensor("op_5331"), val = tensor([1, 1])]; + tensor hidden_states_49_pad_type_0 = const()[name = tensor("hidden_states_49_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_49_pad_0 = const()[name = tensor("hidden_states_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_fc2_weight_to_fp16 = const()[name = tensor("layers_23_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1380191936)))]; + tensor layers_23_fc2_bias_to_fp16 = const()[name = tensor("layers_23_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1393299200)))]; + tensor hidden_states_49_cast_fp16 = conv(bias = layers_23_fc2_bias_to_fp16, dilations = var_5331, groups = var_5137, pad = hidden_states_49_pad_0, pad_type = hidden_states_49_pad_type_0, strides = var_5329, weight = layers_23_fc2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("hidden_states_49_cast_fp16")]; + tensor inputs_145_cast_fp16 = add(x = inputs_143_cast_fp16, y = hidden_states_49_cast_fp16)[name = tensor("inputs_145_cast_fp16")]; + tensor var_5344 = const()[name = tensor("op_5344"), val = tensor(3)]; + tensor var_5351 = const()[name = tensor("op_5351"), val = tensor(1)]; + tensor var_5352 = const()[name = tensor("op_5352"), val = tensor(true)]; + tensor var_5364 = const()[name = tensor("op_5364"), val = tensor([1])]; + tensor channels_mean_145_cast_fp16 = reduce_mean(axes = var_5364, keep_dims = var_5352, x = inputs_145_cast_fp16)[name = tensor("channels_mean_145_cast_fp16")]; + tensor zero_mean_145_cast_fp16 = sub(x = inputs_145_cast_fp16, y = channels_mean_145_cast_fp16)[name = tensor("zero_mean_145_cast_fp16")]; + tensor zero_mean_sq_145_cast_fp16 = mul(x = zero_mean_145_cast_fp16, y = zero_mean_145_cast_fp16)[name = tensor("zero_mean_sq_145_cast_fp16")]; + tensor var_5368 = const()[name = tensor("op_5368"), val = tensor([1])]; + tensor var_5369_cast_fp16 = reduce_mean(axes = var_5368, keep_dims = var_5352, x = zero_mean_sq_145_cast_fp16)[name = tensor("op_5369_cast_fp16")]; + tensor var_5370_to_fp16 = const()[name = tensor("op_5370_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5371_cast_fp16 = add(x = var_5369_cast_fp16, y = var_5370_to_fp16)[name = tensor("op_5371_cast_fp16")]; + tensor denom_145_epsilon_0_to_fp16 = const()[name = tensor("denom_145_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_145_cast_fp16 = rsqrt(epsilon = denom_145_epsilon_0_to_fp16, x = var_5371_cast_fp16)[name = tensor("denom_145_cast_fp16")]; + tensor out_145_cast_fp16 = mul(x = zero_mean_145_cast_fp16, y = denom_145_cast_fp16)[name = tensor("out_145_cast_fp16")]; + tensor obj_337_gamma_0_to_fp16 = const()[name = tensor("obj_337_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1393301824)))]; + tensor obj_337_beta_0_to_fp16 = const()[name = tensor("obj_337_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1393304448)))]; + tensor obj_337_epsilon_0_to_fp16 = const()[name = tensor("obj_337_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_337_cast_fp16 = batch_norm(beta = obj_337_beta_0_to_fp16, epsilon = obj_337_epsilon_0_to_fp16, gamma = obj_337_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_145_cast_fp16)[name = tensor("obj_337_cast_fp16")]; + tensor var_5386 = const()[name = tensor("op_5386"), val = tensor([1, 1])]; + tensor var_5388 = const()[name = tensor("op_5388"), val = tensor([1, 1])]; + tensor query_97_pad_type_0 = const()[name = tensor("query_97_pad_type_0"), val = tensor("custom")]; + tensor query_97_pad_0 = const()[name = tensor("query_97_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1393307072)))]; + tensor layers_24_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1396583936)))]; + tensor query_97_cast_fp16 = conv(bias = layers_24_self_attn_q_proj_bias_to_fp16, dilations = var_5388, groups = var_5351, pad = query_97_pad_0, pad_type = query_97_pad_type_0, strides = var_5386, weight = layers_24_self_attn_q_proj_weight_to_fp16, x = obj_337_cast_fp16)[name = tensor("query_97_cast_fp16")]; + tensor var_5392 = const()[name = tensor("op_5392"), val = tensor([1, 1])]; + tensor var_5394 = const()[name = tensor("op_5394"), val = tensor([1, 1])]; + tensor current_key_49_pad_type_0 = const()[name = tensor("current_key_49_pad_type_0"), val = tensor("custom")]; + tensor current_key_49_pad_0 = const()[name = tensor("current_key_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1396586560)))]; + tensor current_key_49_cast_fp16 = conv(dilations = var_5394, groups = var_5351, pad = current_key_49_pad_0, pad_type = current_key_49_pad_type_0, strides = var_5392, weight = layers_24_self_attn_k_proj_weight_to_fp16, x = obj_337_cast_fp16)[name = tensor("current_key_49_cast_fp16")]; + tensor var_5399 = const()[name = tensor("op_5399"), val = tensor([1, 1])]; + tensor var_5401 = const()[name = tensor("op_5401"), val = tensor([1, 1])]; + tensor current_value_49_pad_type_0 = const()[name = tensor("current_value_49_pad_type_0"), val = tensor("custom")]; + tensor current_value_49_pad_0 = const()[name = tensor("current_value_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1399863424)))]; + tensor layers_24_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1403140288)))]; + tensor current_value_49_cast_fp16 = conv(bias = layers_24_self_attn_v_proj_bias_to_fp16, dilations = var_5401, groups = var_5351, pad = current_value_49_pad_0, pad_type = current_value_49_pad_type_0, strides = var_5399, weight = layers_24_self_attn_v_proj_weight_to_fp16, x = obj_337_cast_fp16)[name = tensor("current_value_49_cast_fp16")]; + tensor var_5408_cast_fp16 = mul(x = current_key_49_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_5408_cast_fp16")]; + tensor var_5410_cast_fp16 = mul(x = var_103_cast_fp16_24, y = var_241_cast_fp16)[name = tensor("op_5410_cast_fp16")]; + tensor key_97_cast_fp16 = add(x = var_5408_cast_fp16, y = var_5410_cast_fp16)[name = tensor("key_97_cast_fp16")]; + tensor var_5412_cast_fp16 = mul(x = current_value_49_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_5412_cast_fp16")]; + tensor var_5414_cast_fp16 = mul(x = var_138_cast_fp16_24, y = var_241_cast_fp16)[name = tensor("op_5414_cast_fp16")]; + tensor value_97_cast_fp16 = add(x = var_5412_cast_fp16, y = var_5414_cast_fp16)[name = tensor("value_97_cast_fp16")]; + tensor var_5417 = const()[name = tensor("op_5417"), val = tensor([1, 20, 64, -1])]; + tensor var_5418_cast_fp16 = reshape(shape = var_5417, x = query_97_cast_fp16)[name = tensor("op_5418_cast_fp16")]; + tensor var_5419_to_fp16 = const()[name = tensor("op_5419_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5420_cast_fp16 = mul(x = var_5418_cast_fp16, y = var_5419_to_fp16)[name = tensor("op_5420_cast_fp16")]; + tensor var_5421 = const()[name = tensor("op_5421"), val = tensor([1, 20, 64, -1])]; + tensor var_5422_cast_fp16 = reshape(shape = var_5421, x = key_97_cast_fp16)[name = tensor("op_5422_cast_fp16")]; + tensor mh_w_145_transpose_x_0 = const()[name = tensor("mh_w_145_transpose_x_0"), val = tensor(true)]; + tensor mh_w_145_transpose_y_0 = const()[name = tensor("mh_w_145_transpose_y_0"), val = tensor(false)]; + tensor mh_w_145_cast_fp16 = matmul(transpose_x = mh_w_145_transpose_x_0, transpose_y = mh_w_145_transpose_y_0, x = var_5420_cast_fp16, y = var_5422_cast_fp16)[name = tensor("mh_w_145_cast_fp16")]; + tensor mh_w_147_cast_fp16 = add(x = mh_w_145_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_147_cast_fp16")]; + tensor var_5430_cast_fp16 = softmax(axis = var_5344, x = mh_w_147_cast_fp16)[name = tensor("op_5430_cast_fp16")]; + tensor var_5431 = const()[name = tensor("op_5431"), val = tensor([1, 20, 64, -1])]; + tensor var_5432_cast_fp16 = reshape(shape = var_5431, x = value_97_cast_fp16)[name = tensor("op_5432_cast_fp16")]; + tensor attn_97_transpose_x_0 = const()[name = tensor("attn_97_transpose_x_0"), val = tensor(false)]; + tensor attn_97_transpose_y_0 = const()[name = tensor("attn_97_transpose_y_0"), val = tensor(true)]; + tensor attn_97_cast_fp16 = matmul(transpose_x = attn_97_transpose_x_0, transpose_y = attn_97_transpose_y_0, x = var_5432_cast_fp16, y = var_5430_cast_fp16)[name = tensor("attn_97_cast_fp16")]; + tensor var_5435 = const()[name = tensor("op_5435"), val = tensor([1, 1280, 1, -1])]; + tensor input_241_cast_fp16 = reshape(shape = var_5435, x = attn_97_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor var_5439 = const()[name = tensor("op_5439"), val = tensor([1, 1])]; + tensor var_5441 = const()[name = tensor("op_5441"), val = tensor([1, 1])]; + tensor obj_343_pad_type_0 = const()[name = tensor("obj_343_pad_type_0"), val = tensor("custom")]; + tensor obj_343_pad_0 = const()[name = tensor("obj_343_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1403142912)))]; + tensor layers_24_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1406419776)))]; + tensor obj_343_cast_fp16 = conv(bias = layers_24_self_attn_o_proj_bias_to_fp16, dilations = var_5441, groups = var_5351, pad = obj_343_pad_0, pad_type = obj_343_pad_type_0, strides = var_5439, weight = layers_24_self_attn_o_proj_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("obj_343_cast_fp16")]; + tensor inputs_147_cast_fp16 = add(x = inputs_145_cast_fp16, y = obj_343_cast_fp16)[name = tensor("inputs_147_cast_fp16")]; + tensor var_5451 = const()[name = tensor("op_5451"), val = tensor([1])]; + tensor channels_mean_147_cast_fp16 = reduce_mean(axes = var_5451, keep_dims = var_5352, x = inputs_147_cast_fp16)[name = tensor("channels_mean_147_cast_fp16")]; + tensor zero_mean_147_cast_fp16 = sub(x = inputs_147_cast_fp16, y = channels_mean_147_cast_fp16)[name = tensor("zero_mean_147_cast_fp16")]; + tensor zero_mean_sq_147_cast_fp16 = mul(x = zero_mean_147_cast_fp16, y = zero_mean_147_cast_fp16)[name = tensor("zero_mean_sq_147_cast_fp16")]; + tensor var_5455 = const()[name = tensor("op_5455"), val = tensor([1])]; + tensor var_5456_cast_fp16 = reduce_mean(axes = var_5455, keep_dims = var_5352, x = zero_mean_sq_147_cast_fp16)[name = tensor("op_5456_cast_fp16")]; + tensor var_5457_to_fp16 = const()[name = tensor("op_5457_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5458_cast_fp16 = add(x = var_5456_cast_fp16, y = var_5457_to_fp16)[name = tensor("op_5458_cast_fp16")]; + tensor denom_147_epsilon_0_to_fp16 = const()[name = tensor("denom_147_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_147_cast_fp16 = rsqrt(epsilon = denom_147_epsilon_0_to_fp16, x = var_5458_cast_fp16)[name = tensor("denom_147_cast_fp16")]; + tensor out_147_cast_fp16 = mul(x = zero_mean_147_cast_fp16, y = denom_147_cast_fp16)[name = tensor("out_147_cast_fp16")]; + tensor obj_345_gamma_0_to_fp16 = const()[name = tensor("obj_345_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1406422400)))]; + tensor obj_345_beta_0_to_fp16 = const()[name = tensor("obj_345_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1406425024)))]; + tensor obj_345_epsilon_0_to_fp16 = const()[name = tensor("obj_345_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_345_cast_fp16 = batch_norm(beta = obj_345_beta_0_to_fp16, epsilon = obj_345_epsilon_0_to_fp16, gamma = obj_345_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_147_cast_fp16)[name = tensor("obj_345_cast_fp16")]; + tensor var_5473 = const()[name = tensor("op_5473"), val = tensor([1, 1])]; + tensor var_5475 = const()[name = tensor("op_5475"), val = tensor([1, 1])]; + tensor query_99_pad_type_0 = const()[name = tensor("query_99_pad_type_0"), val = tensor("custom")]; + tensor query_99_pad_0 = const()[name = tensor("query_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_24_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1406427648)))]; + tensor layers_24_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_24_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1409704512)))]; + tensor query_99_cast_fp16 = conv(bias = layers_24_encoder_attn_q_proj_bias_to_fp16, dilations = var_5475, groups = var_5351, pad = query_99_pad_0, pad_type = query_99_pad_type_0, strides = var_5473, weight = layers_24_encoder_attn_q_proj_weight_to_fp16, x = obj_345_cast_fp16)[name = tensor("query_99_cast_fp16")]; + tensor var_5479 = const()[name = tensor("op_5479"), val = tensor([1, 1])]; + tensor var_5481 = const()[name = tensor("op_5481"), val = tensor([1, 1])]; + tensor key_99_pad_type_0 = const()[name = tensor("key_99_pad_type_0"), val = tensor("custom")]; + tensor key_99_pad_0 = const()[name = tensor("key_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_24_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1409707136)))]; + tensor key_99_cast_fp16 = conv(dilations = var_5481, groups = var_5351, pad = key_99_pad_0, pad_type = key_99_pad_type_0, strides = var_5479, weight = layers_24_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_99_cast_fp16")]; + tensor var_5486 = const()[name = tensor("op_5486"), val = tensor([1, 1])]; + tensor var_5488 = const()[name = tensor("op_5488"), val = tensor([1, 1])]; + tensor value_99_pad_type_0 = const()[name = tensor("value_99_pad_type_0"), val = tensor("custom")]; + tensor value_99_pad_0 = const()[name = tensor("value_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_24_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1412984000)))]; + tensor layers_24_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_24_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1416260864)))]; + tensor value_99_cast_fp16 = conv(bias = layers_24_encoder_attn_v_proj_bias_to_fp16, dilations = var_5488, groups = var_5351, pad = value_99_pad_0, pad_type = value_99_pad_type_0, strides = var_5486, weight = layers_24_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_99_cast_fp16")]; + tensor var_5492 = const()[name = tensor("op_5492"), val = tensor([1, 20, 64, -1])]; + tensor var_5493_cast_fp16 = reshape(shape = var_5492, x = query_99_cast_fp16)[name = tensor("op_5493_cast_fp16")]; + tensor var_5494_to_fp16 = const()[name = tensor("op_5494_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5495_cast_fp16 = mul(x = var_5493_cast_fp16, y = var_5494_to_fp16)[name = tensor("op_5495_cast_fp16")]; + tensor var_5496 = const()[name = tensor("op_5496"), val = tensor([1, 20, 64, -1])]; + tensor var_5497_cast_fp16 = reshape(shape = var_5496, x = key_99_cast_fp16)[name = tensor("op_5497_cast_fp16")]; + tensor mh_w_149_transpose_x_0 = const()[name = tensor("mh_w_149_transpose_x_0"), val = tensor(true)]; + tensor mh_w_149_transpose_y_0 = const()[name = tensor("mh_w_149_transpose_y_0"), val = tensor(false)]; + tensor mh_w_149_cast_fp16 = matmul(transpose_x = mh_w_149_transpose_x_0, transpose_y = mh_w_149_transpose_y_0, x = var_5495_cast_fp16, y = var_5497_cast_fp16)[name = tensor("mh_w_149_cast_fp16")]; + tensor obj_349_cast_fp16 = softmax(axis = var_5344, x = mh_w_149_cast_fp16)[name = tensor("obj_349_cast_fp16")]; + tensor var_5501 = const()[name = tensor("op_5501"), val = tensor([1, 20, 64, -1])]; + tensor var_5502_cast_fp16 = reshape(shape = var_5501, x = value_99_cast_fp16)[name = tensor("op_5502_cast_fp16")]; + tensor attn_99_transpose_x_0 = const()[name = tensor("attn_99_transpose_x_0"), val = tensor(false)]; + tensor attn_99_transpose_y_0 = const()[name = tensor("attn_99_transpose_y_0"), val = tensor(true)]; + tensor attn_99_cast_fp16 = matmul(transpose_x = attn_99_transpose_x_0, transpose_y = attn_99_transpose_y_0, x = var_5502_cast_fp16, y = obj_349_cast_fp16)[name = tensor("attn_99_cast_fp16")]; + tensor var_5505 = const()[name = tensor("op_5505"), val = tensor([1, 1280, 1, -1])]; + tensor input_243_cast_fp16 = reshape(shape = var_5505, x = attn_99_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor var_5509 = const()[name = tensor("op_5509"), val = tensor([1, 1])]; + tensor var_5511 = const()[name = tensor("op_5511"), val = tensor([1, 1])]; + tensor obj_347_pad_type_0 = const()[name = tensor("obj_347_pad_type_0"), val = tensor("custom")]; + tensor obj_347_pad_0 = const()[name = tensor("obj_347_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_24_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1416263488)))]; + tensor layers_24_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_24_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1419540352)))]; + tensor obj_347_cast_fp16 = conv(bias = layers_24_encoder_attn_o_proj_bias_to_fp16, dilations = var_5511, groups = var_5351, pad = obj_347_pad_0, pad_type = obj_347_pad_type_0, strides = var_5509, weight = layers_24_encoder_attn_o_proj_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("obj_347_cast_fp16")]; + tensor inputs_149_cast_fp16 = add(x = inputs_147_cast_fp16, y = obj_347_cast_fp16)[name = tensor("inputs_149_cast_fp16")]; + tensor var_5520 = const()[name = tensor("op_5520"), val = tensor([1])]; + tensor channels_mean_149_cast_fp16 = reduce_mean(axes = var_5520, keep_dims = var_5352, x = inputs_149_cast_fp16)[name = tensor("channels_mean_149_cast_fp16")]; + tensor zero_mean_149_cast_fp16 = sub(x = inputs_149_cast_fp16, y = channels_mean_149_cast_fp16)[name = tensor("zero_mean_149_cast_fp16")]; + tensor zero_mean_sq_149_cast_fp16 = mul(x = zero_mean_149_cast_fp16, y = zero_mean_149_cast_fp16)[name = tensor("zero_mean_sq_149_cast_fp16")]; + tensor var_5524 = const()[name = tensor("op_5524"), val = tensor([1])]; + tensor var_5525_cast_fp16 = reduce_mean(axes = var_5524, keep_dims = var_5352, x = zero_mean_sq_149_cast_fp16)[name = tensor("op_5525_cast_fp16")]; + tensor var_5526_to_fp16 = const()[name = tensor("op_5526_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5527_cast_fp16 = add(x = var_5525_cast_fp16, y = var_5526_to_fp16)[name = tensor("op_5527_cast_fp16")]; + tensor denom_149_epsilon_0_to_fp16 = const()[name = tensor("denom_149_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_149_cast_fp16 = rsqrt(epsilon = denom_149_epsilon_0_to_fp16, x = var_5527_cast_fp16)[name = tensor("denom_149_cast_fp16")]; + tensor out_149_cast_fp16 = mul(x = zero_mean_149_cast_fp16, y = denom_149_cast_fp16)[name = tensor("out_149_cast_fp16")]; + tensor input_245_gamma_0_to_fp16 = const()[name = tensor("input_245_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1419542976)))]; + tensor input_245_beta_0_to_fp16 = const()[name = tensor("input_245_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1419545600)))]; + tensor input_245_epsilon_0_to_fp16 = const()[name = tensor("input_245_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_245_cast_fp16 = batch_norm(beta = input_245_beta_0_to_fp16, epsilon = input_245_epsilon_0_to_fp16, gamma = input_245_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_149_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor var_5538 = const()[name = tensor("op_5538"), val = tensor([1, 1])]; + tensor var_5540 = const()[name = tensor("op_5540"), val = tensor([1, 1])]; + tensor input_247_pad_type_0 = const()[name = tensor("input_247_pad_type_0"), val = tensor("custom")]; + tensor input_247_pad_0 = const()[name = tensor("input_247_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_fc1_weight_to_fp16 = const()[name = tensor("layers_24_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1419548224)))]; + tensor layers_24_fc1_bias_to_fp16 = const()[name = tensor("layers_24_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1432655488)))]; + tensor input_247_cast_fp16 = conv(bias = layers_24_fc1_bias_to_fp16, dilations = var_5540, groups = var_5351, pad = input_247_pad_0, pad_type = input_247_pad_type_0, strides = var_5538, weight = layers_24_fc1_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor input_249_mode_0 = const()[name = tensor("input_249_mode_0"), val = tensor("EXACT")]; + tensor input_249_cast_fp16 = gelu(mode = input_249_mode_0, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor var_5546 = const()[name = tensor("op_5546"), val = tensor([1, 1])]; + tensor var_5548 = const()[name = tensor("op_5548"), val = tensor([1, 1])]; + tensor hidden_states_51_pad_type_0 = const()[name = tensor("hidden_states_51_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_51_pad_0 = const()[name = tensor("hidden_states_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_fc2_weight_to_fp16 = const()[name = tensor("layers_24_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1432665792)))]; + tensor layers_24_fc2_bias_to_fp16 = const()[name = tensor("layers_24_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1445773056)))]; + tensor hidden_states_51_cast_fp16 = conv(bias = layers_24_fc2_bias_to_fp16, dilations = var_5548, groups = var_5351, pad = hidden_states_51_pad_0, pad_type = hidden_states_51_pad_type_0, strides = var_5546, weight = layers_24_fc2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("hidden_states_51_cast_fp16")]; + tensor inputs_151_cast_fp16 = add(x = inputs_149_cast_fp16, y = hidden_states_51_cast_fp16)[name = tensor("inputs_151_cast_fp16")]; + tensor var_5562 = const()[name = tensor("op_5562"), val = tensor(3)]; + tensor var_5569 = const()[name = tensor("op_5569"), val = tensor(1)]; + tensor var_5570 = const()[name = tensor("op_5570"), val = tensor(true)]; + tensor var_5582 = const()[name = tensor("op_5582"), val = tensor([1])]; + tensor channels_mean_151_cast_fp16 = reduce_mean(axes = var_5582, keep_dims = var_5570, x = inputs_151_cast_fp16)[name = tensor("channels_mean_151_cast_fp16")]; + tensor zero_mean_151_cast_fp16 = sub(x = inputs_151_cast_fp16, y = channels_mean_151_cast_fp16)[name = tensor("zero_mean_151_cast_fp16")]; + tensor zero_mean_sq_151_cast_fp16 = mul(x = zero_mean_151_cast_fp16, y = zero_mean_151_cast_fp16)[name = tensor("zero_mean_sq_151_cast_fp16")]; + tensor var_5586 = const()[name = tensor("op_5586"), val = tensor([1])]; + tensor var_5587_cast_fp16 = reduce_mean(axes = var_5586, keep_dims = var_5570, x = zero_mean_sq_151_cast_fp16)[name = tensor("op_5587_cast_fp16")]; + tensor var_5588_to_fp16 = const()[name = tensor("op_5588_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5589_cast_fp16 = add(x = var_5587_cast_fp16, y = var_5588_to_fp16)[name = tensor("op_5589_cast_fp16")]; + tensor denom_151_epsilon_0_to_fp16 = const()[name = tensor("denom_151_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_151_cast_fp16 = rsqrt(epsilon = denom_151_epsilon_0_to_fp16, x = var_5589_cast_fp16)[name = tensor("denom_151_cast_fp16")]; + tensor out_151_cast_fp16 = mul(x = zero_mean_151_cast_fp16, y = denom_151_cast_fp16)[name = tensor("out_151_cast_fp16")]; + tensor obj_351_gamma_0_to_fp16 = const()[name = tensor("obj_351_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1445775680)))]; + tensor obj_351_beta_0_to_fp16 = const()[name = tensor("obj_351_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1445778304)))]; + tensor obj_351_epsilon_0_to_fp16 = const()[name = tensor("obj_351_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_351_cast_fp16 = batch_norm(beta = obj_351_beta_0_to_fp16, epsilon = obj_351_epsilon_0_to_fp16, gamma = obj_351_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_151_cast_fp16)[name = tensor("obj_351_cast_fp16")]; + tensor var_5604 = const()[name = tensor("op_5604"), val = tensor([1, 1])]; + tensor var_5606 = const()[name = tensor("op_5606"), val = tensor([1, 1])]; + tensor query_101_pad_type_0 = const()[name = tensor("query_101_pad_type_0"), val = tensor("custom")]; + tensor query_101_pad_0 = const()[name = tensor("query_101_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1445780928)))]; + tensor layers_25_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1449057792)))]; + tensor query_101_cast_fp16 = conv(bias = layers_25_self_attn_q_proj_bias_to_fp16, dilations = var_5606, groups = var_5569, pad = query_101_pad_0, pad_type = query_101_pad_type_0, strides = var_5604, weight = layers_25_self_attn_q_proj_weight_to_fp16, x = obj_351_cast_fp16)[name = tensor("query_101_cast_fp16")]; + tensor var_5610 = const()[name = tensor("op_5610"), val = tensor([1, 1])]; + tensor var_5612 = const()[name = tensor("op_5612"), val = tensor([1, 1])]; + tensor current_key_51_pad_type_0 = const()[name = tensor("current_key_51_pad_type_0"), val = tensor("custom")]; + tensor current_key_51_pad_0 = const()[name = tensor("current_key_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1449060416)))]; + tensor current_key_51_cast_fp16 = conv(dilations = var_5612, groups = var_5569, pad = current_key_51_pad_0, pad_type = current_key_51_pad_type_0, strides = var_5610, weight = layers_25_self_attn_k_proj_weight_to_fp16, x = obj_351_cast_fp16)[name = tensor("current_key_51_cast_fp16")]; + tensor var_5617 = const()[name = tensor("op_5617"), val = tensor([1, 1])]; + tensor var_5619 = const()[name = tensor("op_5619"), val = tensor([1, 1])]; + tensor current_value_51_pad_type_0 = const()[name = tensor("current_value_51_pad_type_0"), val = tensor("custom")]; + tensor current_value_51_pad_0 = const()[name = tensor("current_value_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1452337280)))]; + tensor layers_25_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1455614144)))]; + tensor current_value_51_cast_fp16 = conv(bias = layers_25_self_attn_v_proj_bias_to_fp16, dilations = var_5619, groups = var_5569, pad = current_value_51_pad_0, pad_type = current_value_51_pad_type_0, strides = var_5617, weight = layers_25_self_attn_v_proj_weight_to_fp16, x = obj_351_cast_fp16)[name = tensor("current_value_51_cast_fp16")]; + tensor var_5626_cast_fp16 = mul(x = current_key_51_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_5626_cast_fp16")]; + tensor var_5628_cast_fp16 = mul(x = var_103_cast_fp16_25, y = var_241_cast_fp16)[name = tensor("op_5628_cast_fp16")]; + tensor key_101_cast_fp16 = add(x = var_5626_cast_fp16, y = var_5628_cast_fp16)[name = tensor("key_101_cast_fp16")]; + tensor var_5630_cast_fp16 = mul(x = current_value_51_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_5630_cast_fp16")]; + tensor var_5632_cast_fp16 = mul(x = var_138_cast_fp16_25, y = var_241_cast_fp16)[name = tensor("op_5632_cast_fp16")]; + tensor value_101_cast_fp16 = add(x = var_5630_cast_fp16, y = var_5632_cast_fp16)[name = tensor("value_101_cast_fp16")]; + tensor var_5635 = const()[name = tensor("op_5635"), val = tensor([1, 20, 64, -1])]; + tensor var_5636_cast_fp16 = reshape(shape = var_5635, x = query_101_cast_fp16)[name = tensor("op_5636_cast_fp16")]; + tensor var_5637_to_fp16 = const()[name = tensor("op_5637_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5638_cast_fp16 = mul(x = var_5636_cast_fp16, y = var_5637_to_fp16)[name = tensor("op_5638_cast_fp16")]; + tensor var_5639 = const()[name = tensor("op_5639"), val = tensor([1, 20, 64, -1])]; + tensor var_5640_cast_fp16 = reshape(shape = var_5639, x = key_101_cast_fp16)[name = tensor("op_5640_cast_fp16")]; + tensor mh_w_151_transpose_x_0 = const()[name = tensor("mh_w_151_transpose_x_0"), val = tensor(true)]; + tensor mh_w_151_transpose_y_0 = const()[name = tensor("mh_w_151_transpose_y_0"), val = tensor(false)]; + tensor mh_w_151_cast_fp16 = matmul(transpose_x = mh_w_151_transpose_x_0, transpose_y = mh_w_151_transpose_y_0, x = var_5638_cast_fp16, y = var_5640_cast_fp16)[name = tensor("mh_w_151_cast_fp16")]; + tensor mh_w_153_cast_fp16 = add(x = mh_w_151_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_153_cast_fp16")]; + tensor var_5648_cast_fp16 = softmax(axis = var_5562, x = mh_w_153_cast_fp16)[name = tensor("op_5648_cast_fp16")]; + tensor var_5649 = const()[name = tensor("op_5649"), val = tensor([1, 20, 64, -1])]; + tensor var_5650_cast_fp16 = reshape(shape = var_5649, x = value_101_cast_fp16)[name = tensor("op_5650_cast_fp16")]; + tensor attn_101_transpose_x_0 = const()[name = tensor("attn_101_transpose_x_0"), val = tensor(false)]; + tensor attn_101_transpose_y_0 = const()[name = tensor("attn_101_transpose_y_0"), val = tensor(true)]; + tensor attn_101_cast_fp16 = matmul(transpose_x = attn_101_transpose_x_0, transpose_y = attn_101_transpose_y_0, x = var_5650_cast_fp16, y = var_5648_cast_fp16)[name = tensor("attn_101_cast_fp16")]; + tensor var_5653 = const()[name = tensor("op_5653"), val = tensor([1, 1280, 1, -1])]; + tensor input_251_cast_fp16 = reshape(shape = var_5653, x = attn_101_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor var_5657 = const()[name = tensor("op_5657"), val = tensor([1, 1])]; + tensor var_5659 = const()[name = tensor("op_5659"), val = tensor([1, 1])]; + tensor obj_357_pad_type_0 = const()[name = tensor("obj_357_pad_type_0"), val = tensor("custom")]; + tensor obj_357_pad_0 = const()[name = tensor("obj_357_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1455616768)))]; + tensor layers_25_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1458893632)))]; + tensor obj_357_cast_fp16 = conv(bias = layers_25_self_attn_o_proj_bias_to_fp16, dilations = var_5659, groups = var_5569, pad = obj_357_pad_0, pad_type = obj_357_pad_type_0, strides = var_5657, weight = layers_25_self_attn_o_proj_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("obj_357_cast_fp16")]; + tensor inputs_153_cast_fp16 = add(x = inputs_151_cast_fp16, y = obj_357_cast_fp16)[name = tensor("inputs_153_cast_fp16")]; + tensor var_5669 = const()[name = tensor("op_5669"), val = tensor([1])]; + tensor channels_mean_153_cast_fp16 = reduce_mean(axes = var_5669, keep_dims = var_5570, x = inputs_153_cast_fp16)[name = tensor("channels_mean_153_cast_fp16")]; + tensor zero_mean_153_cast_fp16 = sub(x = inputs_153_cast_fp16, y = channels_mean_153_cast_fp16)[name = tensor("zero_mean_153_cast_fp16")]; + tensor zero_mean_sq_153_cast_fp16 = mul(x = zero_mean_153_cast_fp16, y = zero_mean_153_cast_fp16)[name = tensor("zero_mean_sq_153_cast_fp16")]; + tensor var_5673 = const()[name = tensor("op_5673"), val = tensor([1])]; + tensor var_5674_cast_fp16 = reduce_mean(axes = var_5673, keep_dims = var_5570, x = zero_mean_sq_153_cast_fp16)[name = tensor("op_5674_cast_fp16")]; + tensor var_5675_to_fp16 = const()[name = tensor("op_5675_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5676_cast_fp16 = add(x = var_5674_cast_fp16, y = var_5675_to_fp16)[name = tensor("op_5676_cast_fp16")]; + tensor denom_153_epsilon_0_to_fp16 = const()[name = tensor("denom_153_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_153_cast_fp16 = rsqrt(epsilon = denom_153_epsilon_0_to_fp16, x = var_5676_cast_fp16)[name = tensor("denom_153_cast_fp16")]; + tensor out_153_cast_fp16 = mul(x = zero_mean_153_cast_fp16, y = denom_153_cast_fp16)[name = tensor("out_153_cast_fp16")]; + tensor obj_359_gamma_0_to_fp16 = const()[name = tensor("obj_359_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1458896256)))]; + tensor obj_359_beta_0_to_fp16 = const()[name = tensor("obj_359_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1458898880)))]; + tensor obj_359_epsilon_0_to_fp16 = const()[name = tensor("obj_359_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_359_cast_fp16 = batch_norm(beta = obj_359_beta_0_to_fp16, epsilon = obj_359_epsilon_0_to_fp16, gamma = obj_359_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_153_cast_fp16)[name = tensor("obj_359_cast_fp16")]; + tensor var_5691 = const()[name = tensor("op_5691"), val = tensor([1, 1])]; + tensor var_5693 = const()[name = tensor("op_5693"), val = tensor([1, 1])]; + tensor query_103_pad_type_0 = const()[name = tensor("query_103_pad_type_0"), val = tensor("custom")]; + tensor query_103_pad_0 = const()[name = tensor("query_103_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_25_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1458901504)))]; + tensor layers_25_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_25_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1462178368)))]; + tensor query_103_cast_fp16 = conv(bias = layers_25_encoder_attn_q_proj_bias_to_fp16, dilations = var_5693, groups = var_5569, pad = query_103_pad_0, pad_type = query_103_pad_type_0, strides = var_5691, weight = layers_25_encoder_attn_q_proj_weight_to_fp16, x = obj_359_cast_fp16)[name = tensor("query_103_cast_fp16")]; + tensor var_5697 = const()[name = tensor("op_5697"), val = tensor([1, 1])]; + tensor var_5699 = const()[name = tensor("op_5699"), val = tensor([1, 1])]; + tensor key_103_pad_type_0 = const()[name = tensor("key_103_pad_type_0"), val = tensor("custom")]; + tensor key_103_pad_0 = const()[name = tensor("key_103_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_25_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1462180992)))]; + tensor key_103_cast_fp16 = conv(dilations = var_5699, groups = var_5569, pad = key_103_pad_0, pad_type = key_103_pad_type_0, strides = var_5697, weight = layers_25_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_103_cast_fp16")]; + tensor var_5704 = const()[name = tensor("op_5704"), val = tensor([1, 1])]; + tensor var_5706 = const()[name = tensor("op_5706"), val = tensor([1, 1])]; + tensor value_103_pad_type_0 = const()[name = tensor("value_103_pad_type_0"), val = tensor("custom")]; + tensor value_103_pad_0 = const()[name = tensor("value_103_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_25_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1465457856)))]; + tensor layers_25_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_25_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1468734720)))]; + tensor value_103_cast_fp16 = conv(bias = layers_25_encoder_attn_v_proj_bias_to_fp16, dilations = var_5706, groups = var_5569, pad = value_103_pad_0, pad_type = value_103_pad_type_0, strides = var_5704, weight = layers_25_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_103_cast_fp16")]; + tensor var_5710 = const()[name = tensor("op_5710"), val = tensor([1, 20, 64, -1])]; + tensor var_5711_cast_fp16 = reshape(shape = var_5710, x = query_103_cast_fp16)[name = tensor("op_5711_cast_fp16")]; + tensor var_5712_to_fp16 = const()[name = tensor("op_5712_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5713_cast_fp16 = mul(x = var_5711_cast_fp16, y = var_5712_to_fp16)[name = tensor("op_5713_cast_fp16")]; + tensor var_5714 = const()[name = tensor("op_5714"), val = tensor([1, 20, 64, -1])]; + tensor var_5715_cast_fp16 = reshape(shape = var_5714, x = key_103_cast_fp16)[name = tensor("op_5715_cast_fp16")]; + tensor mh_w_155_transpose_x_0 = const()[name = tensor("mh_w_155_transpose_x_0"), val = tensor(true)]; + tensor mh_w_155_transpose_y_0 = const()[name = tensor("mh_w_155_transpose_y_0"), val = tensor(false)]; + tensor mh_w_155_cast_fp16 = matmul(transpose_x = mh_w_155_transpose_x_0, transpose_y = mh_w_155_transpose_y_0, x = var_5713_cast_fp16, y = var_5715_cast_fp16)[name = tensor("mh_w_155_cast_fp16")]; + tensor obj_363_cast_fp16 = softmax(axis = var_5562, x = mh_w_155_cast_fp16)[name = tensor("obj_363_cast_fp16")]; + tensor var_5719 = const()[name = tensor("op_5719"), val = tensor([1, 20, 64, -1])]; + tensor var_5720_cast_fp16 = reshape(shape = var_5719, x = value_103_cast_fp16)[name = tensor("op_5720_cast_fp16")]; + tensor attn_103_transpose_x_0 = const()[name = tensor("attn_103_transpose_x_0"), val = tensor(false)]; + tensor attn_103_transpose_y_0 = const()[name = tensor("attn_103_transpose_y_0"), val = tensor(true)]; + tensor attn_103_cast_fp16 = matmul(transpose_x = attn_103_transpose_x_0, transpose_y = attn_103_transpose_y_0, x = var_5720_cast_fp16, y = obj_363_cast_fp16)[name = tensor("attn_103_cast_fp16")]; + tensor var_5723 = const()[name = tensor("op_5723"), val = tensor([1, 1280, 1, -1])]; + tensor input_253_cast_fp16 = reshape(shape = var_5723, x = attn_103_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor var_5727 = const()[name = tensor("op_5727"), val = tensor([1, 1])]; + tensor var_5729 = const()[name = tensor("op_5729"), val = tensor([1, 1])]; + tensor obj_361_pad_type_0 = const()[name = tensor("obj_361_pad_type_0"), val = tensor("custom")]; + tensor obj_361_pad_0 = const()[name = tensor("obj_361_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_25_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1468737344)))]; + tensor layers_25_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_25_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1472014208)))]; + tensor obj_361_cast_fp16 = conv(bias = layers_25_encoder_attn_o_proj_bias_to_fp16, dilations = var_5729, groups = var_5569, pad = obj_361_pad_0, pad_type = obj_361_pad_type_0, strides = var_5727, weight = layers_25_encoder_attn_o_proj_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("obj_361_cast_fp16")]; + tensor inputs_155_cast_fp16 = add(x = inputs_153_cast_fp16, y = obj_361_cast_fp16)[name = tensor("inputs_155_cast_fp16")]; + tensor var_5738 = const()[name = tensor("op_5738"), val = tensor([1])]; + tensor channels_mean_155_cast_fp16 = reduce_mean(axes = var_5738, keep_dims = var_5570, x = inputs_155_cast_fp16)[name = tensor("channels_mean_155_cast_fp16")]; + tensor zero_mean_155_cast_fp16 = sub(x = inputs_155_cast_fp16, y = channels_mean_155_cast_fp16)[name = tensor("zero_mean_155_cast_fp16")]; + tensor zero_mean_sq_155_cast_fp16 = mul(x = zero_mean_155_cast_fp16, y = zero_mean_155_cast_fp16)[name = tensor("zero_mean_sq_155_cast_fp16")]; + tensor var_5742 = const()[name = tensor("op_5742"), val = tensor([1])]; + tensor var_5743_cast_fp16 = reduce_mean(axes = var_5742, keep_dims = var_5570, x = zero_mean_sq_155_cast_fp16)[name = tensor("op_5743_cast_fp16")]; + tensor var_5744_to_fp16 = const()[name = tensor("op_5744_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5745_cast_fp16 = add(x = var_5743_cast_fp16, y = var_5744_to_fp16)[name = tensor("op_5745_cast_fp16")]; + tensor denom_155_epsilon_0_to_fp16 = const()[name = tensor("denom_155_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_155_cast_fp16 = rsqrt(epsilon = denom_155_epsilon_0_to_fp16, x = var_5745_cast_fp16)[name = tensor("denom_155_cast_fp16")]; + tensor out_155_cast_fp16 = mul(x = zero_mean_155_cast_fp16, y = denom_155_cast_fp16)[name = tensor("out_155_cast_fp16")]; + tensor input_255_gamma_0_to_fp16 = const()[name = tensor("input_255_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1472016832)))]; + tensor input_255_beta_0_to_fp16 = const()[name = tensor("input_255_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1472019456)))]; + tensor input_255_epsilon_0_to_fp16 = const()[name = tensor("input_255_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_255_cast_fp16 = batch_norm(beta = input_255_beta_0_to_fp16, epsilon = input_255_epsilon_0_to_fp16, gamma = input_255_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_155_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor var_5756 = const()[name = tensor("op_5756"), val = tensor([1, 1])]; + tensor var_5758 = const()[name = tensor("op_5758"), val = tensor([1, 1])]; + tensor input_257_pad_type_0 = const()[name = tensor("input_257_pad_type_0"), val = tensor("custom")]; + tensor input_257_pad_0 = const()[name = tensor("input_257_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_fc1_weight_to_fp16 = const()[name = tensor("layers_25_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1472022080)))]; + tensor layers_25_fc1_bias_to_fp16 = const()[name = tensor("layers_25_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1485129344)))]; + tensor input_257_cast_fp16 = conv(bias = layers_25_fc1_bias_to_fp16, dilations = var_5758, groups = var_5569, pad = input_257_pad_0, pad_type = input_257_pad_type_0, strides = var_5756, weight = layers_25_fc1_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor input_259_mode_0 = const()[name = tensor("input_259_mode_0"), val = tensor("EXACT")]; + tensor input_259_cast_fp16 = gelu(mode = input_259_mode_0, x = input_257_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor var_5764 = const()[name = tensor("op_5764"), val = tensor([1, 1])]; + tensor var_5766 = const()[name = tensor("op_5766"), val = tensor([1, 1])]; + tensor hidden_states_53_pad_type_0 = const()[name = tensor("hidden_states_53_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_53_pad_0 = const()[name = tensor("hidden_states_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_fc2_weight_to_fp16 = const()[name = tensor("layers_25_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1485139648)))]; + tensor layers_25_fc2_bias_to_fp16 = const()[name = tensor("layers_25_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1498246912)))]; + tensor hidden_states_53_cast_fp16 = conv(bias = layers_25_fc2_bias_to_fp16, dilations = var_5766, groups = var_5569, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = var_5764, weight = layers_25_fc2_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("hidden_states_53_cast_fp16")]; + tensor inputs_157_cast_fp16 = add(x = inputs_155_cast_fp16, y = hidden_states_53_cast_fp16)[name = tensor("inputs_157_cast_fp16")]; + tensor var_5780 = const()[name = tensor("op_5780"), val = tensor(3)]; + tensor var_5787 = const()[name = tensor("op_5787"), val = tensor(1)]; + tensor var_5788 = const()[name = tensor("op_5788"), val = tensor(true)]; + tensor var_5800 = const()[name = tensor("op_5800"), val = tensor([1])]; + tensor channels_mean_157_cast_fp16 = reduce_mean(axes = var_5800, keep_dims = var_5788, x = inputs_157_cast_fp16)[name = tensor("channels_mean_157_cast_fp16")]; + tensor zero_mean_157_cast_fp16 = sub(x = inputs_157_cast_fp16, y = channels_mean_157_cast_fp16)[name = tensor("zero_mean_157_cast_fp16")]; + tensor zero_mean_sq_157_cast_fp16 = mul(x = zero_mean_157_cast_fp16, y = zero_mean_157_cast_fp16)[name = tensor("zero_mean_sq_157_cast_fp16")]; + tensor var_5804 = const()[name = tensor("op_5804"), val = tensor([1])]; + tensor var_5805_cast_fp16 = reduce_mean(axes = var_5804, keep_dims = var_5788, x = zero_mean_sq_157_cast_fp16)[name = tensor("op_5805_cast_fp16")]; + tensor var_5806_to_fp16 = const()[name = tensor("op_5806_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5807_cast_fp16 = add(x = var_5805_cast_fp16, y = var_5806_to_fp16)[name = tensor("op_5807_cast_fp16")]; + tensor denom_157_epsilon_0_to_fp16 = const()[name = tensor("denom_157_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_157_cast_fp16 = rsqrt(epsilon = denom_157_epsilon_0_to_fp16, x = var_5807_cast_fp16)[name = tensor("denom_157_cast_fp16")]; + tensor out_157_cast_fp16 = mul(x = zero_mean_157_cast_fp16, y = denom_157_cast_fp16)[name = tensor("out_157_cast_fp16")]; + tensor obj_365_gamma_0_to_fp16 = const()[name = tensor("obj_365_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1498249536)))]; + tensor obj_365_beta_0_to_fp16 = const()[name = tensor("obj_365_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1498252160)))]; + tensor obj_365_epsilon_0_to_fp16 = const()[name = tensor("obj_365_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_365_cast_fp16 = batch_norm(beta = obj_365_beta_0_to_fp16, epsilon = obj_365_epsilon_0_to_fp16, gamma = obj_365_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_157_cast_fp16)[name = tensor("obj_365_cast_fp16")]; + tensor var_5822 = const()[name = tensor("op_5822"), val = tensor([1, 1])]; + tensor var_5824 = const()[name = tensor("op_5824"), val = tensor([1, 1])]; + tensor query_105_pad_type_0 = const()[name = tensor("query_105_pad_type_0"), val = tensor("custom")]; + tensor query_105_pad_0 = const()[name = tensor("query_105_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1498254784)))]; + tensor layers_26_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1501531648)))]; + tensor query_105_cast_fp16 = conv(bias = layers_26_self_attn_q_proj_bias_to_fp16, dilations = var_5824, groups = var_5787, pad = query_105_pad_0, pad_type = query_105_pad_type_0, strides = var_5822, weight = layers_26_self_attn_q_proj_weight_to_fp16, x = obj_365_cast_fp16)[name = tensor("query_105_cast_fp16")]; + tensor var_5828 = const()[name = tensor("op_5828"), val = tensor([1, 1])]; + tensor var_5830 = const()[name = tensor("op_5830"), val = tensor([1, 1])]; + tensor current_key_53_pad_type_0 = const()[name = tensor("current_key_53_pad_type_0"), val = tensor("custom")]; + tensor current_key_53_pad_0 = const()[name = tensor("current_key_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1501534272)))]; + tensor current_key_53_cast_fp16 = conv(dilations = var_5830, groups = var_5787, pad = current_key_53_pad_0, pad_type = current_key_53_pad_type_0, strides = var_5828, weight = layers_26_self_attn_k_proj_weight_to_fp16, x = obj_365_cast_fp16)[name = tensor("current_key_53_cast_fp16")]; + tensor var_5835 = const()[name = tensor("op_5835"), val = tensor([1, 1])]; + tensor var_5837 = const()[name = tensor("op_5837"), val = tensor([1, 1])]; + tensor current_value_53_pad_type_0 = const()[name = tensor("current_value_53_pad_type_0"), val = tensor("custom")]; + tensor current_value_53_pad_0 = const()[name = tensor("current_value_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1504811136)))]; + tensor layers_26_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1508088000)))]; + tensor current_value_53_cast_fp16 = conv(bias = layers_26_self_attn_v_proj_bias_to_fp16, dilations = var_5837, groups = var_5787, pad = current_value_53_pad_0, pad_type = current_value_53_pad_type_0, strides = var_5835, weight = layers_26_self_attn_v_proj_weight_to_fp16, x = obj_365_cast_fp16)[name = tensor("current_value_53_cast_fp16")]; + tensor var_5844_cast_fp16 = mul(x = current_key_53_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_5844_cast_fp16")]; + tensor var_5846_cast_fp16 = mul(x = var_103_cast_fp16_26, y = var_241_cast_fp16)[name = tensor("op_5846_cast_fp16")]; + tensor key_105_cast_fp16 = add(x = var_5844_cast_fp16, y = var_5846_cast_fp16)[name = tensor("key_105_cast_fp16")]; + tensor var_5848_cast_fp16 = mul(x = current_value_53_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_5848_cast_fp16")]; + tensor var_5850_cast_fp16 = mul(x = var_138_cast_fp16_26, y = var_241_cast_fp16)[name = tensor("op_5850_cast_fp16")]; + tensor value_105_cast_fp16 = add(x = var_5848_cast_fp16, y = var_5850_cast_fp16)[name = tensor("value_105_cast_fp16")]; + tensor var_5853 = const()[name = tensor("op_5853"), val = tensor([1, 20, 64, -1])]; + tensor var_5854_cast_fp16 = reshape(shape = var_5853, x = query_105_cast_fp16)[name = tensor("op_5854_cast_fp16")]; + tensor var_5855_to_fp16 = const()[name = tensor("op_5855_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5856_cast_fp16 = mul(x = var_5854_cast_fp16, y = var_5855_to_fp16)[name = tensor("op_5856_cast_fp16")]; + tensor var_5857 = const()[name = tensor("op_5857"), val = tensor([1, 20, 64, -1])]; + tensor var_5858_cast_fp16 = reshape(shape = var_5857, x = key_105_cast_fp16)[name = tensor("op_5858_cast_fp16")]; + tensor mh_w_157_transpose_x_0 = const()[name = tensor("mh_w_157_transpose_x_0"), val = tensor(true)]; + tensor mh_w_157_transpose_y_0 = const()[name = tensor("mh_w_157_transpose_y_0"), val = tensor(false)]; + tensor mh_w_157_cast_fp16 = matmul(transpose_x = mh_w_157_transpose_x_0, transpose_y = mh_w_157_transpose_y_0, x = var_5856_cast_fp16, y = var_5858_cast_fp16)[name = tensor("mh_w_157_cast_fp16")]; + tensor mh_w_159_cast_fp16 = add(x = mh_w_157_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_159_cast_fp16")]; + tensor var_5866_cast_fp16 = softmax(axis = var_5780, x = mh_w_159_cast_fp16)[name = tensor("op_5866_cast_fp16")]; + tensor var_5867 = const()[name = tensor("op_5867"), val = tensor([1, 20, 64, -1])]; + tensor var_5868_cast_fp16 = reshape(shape = var_5867, x = value_105_cast_fp16)[name = tensor("op_5868_cast_fp16")]; + tensor attn_105_transpose_x_0 = const()[name = tensor("attn_105_transpose_x_0"), val = tensor(false)]; + tensor attn_105_transpose_y_0 = const()[name = tensor("attn_105_transpose_y_0"), val = tensor(true)]; + tensor attn_105_cast_fp16 = matmul(transpose_x = attn_105_transpose_x_0, transpose_y = attn_105_transpose_y_0, x = var_5868_cast_fp16, y = var_5866_cast_fp16)[name = tensor("attn_105_cast_fp16")]; + tensor var_5871 = const()[name = tensor("op_5871"), val = tensor([1, 1280, 1, -1])]; + tensor input_261_cast_fp16 = reshape(shape = var_5871, x = attn_105_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor var_5875 = const()[name = tensor("op_5875"), val = tensor([1, 1])]; + tensor var_5877 = const()[name = tensor("op_5877"), val = tensor([1, 1])]; + tensor obj_371_pad_type_0 = const()[name = tensor("obj_371_pad_type_0"), val = tensor("custom")]; + tensor obj_371_pad_0 = const()[name = tensor("obj_371_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1508090624)))]; + tensor layers_26_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1511367488)))]; + tensor obj_371_cast_fp16 = conv(bias = layers_26_self_attn_o_proj_bias_to_fp16, dilations = var_5877, groups = var_5787, pad = obj_371_pad_0, pad_type = obj_371_pad_type_0, strides = var_5875, weight = layers_26_self_attn_o_proj_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("obj_371_cast_fp16")]; + tensor inputs_159_cast_fp16 = add(x = inputs_157_cast_fp16, y = obj_371_cast_fp16)[name = tensor("inputs_159_cast_fp16")]; + tensor var_5887 = const()[name = tensor("op_5887"), val = tensor([1])]; + tensor channels_mean_159_cast_fp16 = reduce_mean(axes = var_5887, keep_dims = var_5788, x = inputs_159_cast_fp16)[name = tensor("channels_mean_159_cast_fp16")]; + tensor zero_mean_159_cast_fp16 = sub(x = inputs_159_cast_fp16, y = channels_mean_159_cast_fp16)[name = tensor("zero_mean_159_cast_fp16")]; + tensor zero_mean_sq_159_cast_fp16 = mul(x = zero_mean_159_cast_fp16, y = zero_mean_159_cast_fp16)[name = tensor("zero_mean_sq_159_cast_fp16")]; + tensor var_5891 = const()[name = tensor("op_5891"), val = tensor([1])]; + tensor var_5892_cast_fp16 = reduce_mean(axes = var_5891, keep_dims = var_5788, x = zero_mean_sq_159_cast_fp16)[name = tensor("op_5892_cast_fp16")]; + tensor var_5893_to_fp16 = const()[name = tensor("op_5893_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5894_cast_fp16 = add(x = var_5892_cast_fp16, y = var_5893_to_fp16)[name = tensor("op_5894_cast_fp16")]; + tensor denom_159_epsilon_0_to_fp16 = const()[name = tensor("denom_159_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_159_cast_fp16 = rsqrt(epsilon = denom_159_epsilon_0_to_fp16, x = var_5894_cast_fp16)[name = tensor("denom_159_cast_fp16")]; + tensor out_159_cast_fp16 = mul(x = zero_mean_159_cast_fp16, y = denom_159_cast_fp16)[name = tensor("out_159_cast_fp16")]; + tensor obj_373_gamma_0_to_fp16 = const()[name = tensor("obj_373_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1511370112)))]; + tensor obj_373_beta_0_to_fp16 = const()[name = tensor("obj_373_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1511372736)))]; + tensor obj_373_epsilon_0_to_fp16 = const()[name = tensor("obj_373_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_373_cast_fp16 = batch_norm(beta = obj_373_beta_0_to_fp16, epsilon = obj_373_epsilon_0_to_fp16, gamma = obj_373_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_159_cast_fp16)[name = tensor("obj_373_cast_fp16")]; + tensor var_5909 = const()[name = tensor("op_5909"), val = tensor([1, 1])]; + tensor var_5911 = const()[name = tensor("op_5911"), val = tensor([1, 1])]; + tensor query_107_pad_type_0 = const()[name = tensor("query_107_pad_type_0"), val = tensor("custom")]; + tensor query_107_pad_0 = const()[name = tensor("query_107_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_26_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1511375360)))]; + tensor layers_26_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_26_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1514652224)))]; + tensor query_107_cast_fp16 = conv(bias = layers_26_encoder_attn_q_proj_bias_to_fp16, dilations = var_5911, groups = var_5787, pad = query_107_pad_0, pad_type = query_107_pad_type_0, strides = var_5909, weight = layers_26_encoder_attn_q_proj_weight_to_fp16, x = obj_373_cast_fp16)[name = tensor("query_107_cast_fp16")]; + tensor var_5915 = const()[name = tensor("op_5915"), val = tensor([1, 1])]; + tensor var_5917 = const()[name = tensor("op_5917"), val = tensor([1, 1])]; + tensor key_107_pad_type_0 = const()[name = tensor("key_107_pad_type_0"), val = tensor("custom")]; + tensor key_107_pad_0 = const()[name = tensor("key_107_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_26_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1514654848)))]; + tensor key_107_cast_fp16 = conv(dilations = var_5917, groups = var_5787, pad = key_107_pad_0, pad_type = key_107_pad_type_0, strides = var_5915, weight = layers_26_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_107_cast_fp16")]; + tensor var_5922 = const()[name = tensor("op_5922"), val = tensor([1, 1])]; + tensor var_5924 = const()[name = tensor("op_5924"), val = tensor([1, 1])]; + tensor value_107_pad_type_0 = const()[name = tensor("value_107_pad_type_0"), val = tensor("custom")]; + tensor value_107_pad_0 = const()[name = tensor("value_107_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_26_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1517931712)))]; + tensor layers_26_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_26_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1521208576)))]; + tensor value_107_cast_fp16 = conv(bias = layers_26_encoder_attn_v_proj_bias_to_fp16, dilations = var_5924, groups = var_5787, pad = value_107_pad_0, pad_type = value_107_pad_type_0, strides = var_5922, weight = layers_26_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_107_cast_fp16")]; + tensor var_5928 = const()[name = tensor("op_5928"), val = tensor([1, 20, 64, -1])]; + tensor var_5929_cast_fp16 = reshape(shape = var_5928, x = query_107_cast_fp16)[name = tensor("op_5929_cast_fp16")]; + tensor var_5930_to_fp16 = const()[name = tensor("op_5930_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5931_cast_fp16 = mul(x = var_5929_cast_fp16, y = var_5930_to_fp16)[name = tensor("op_5931_cast_fp16")]; + tensor var_5932 = const()[name = tensor("op_5932"), val = tensor([1, 20, 64, -1])]; + tensor var_5933_cast_fp16 = reshape(shape = var_5932, x = key_107_cast_fp16)[name = tensor("op_5933_cast_fp16")]; + tensor mh_w_161_transpose_x_0 = const()[name = tensor("mh_w_161_transpose_x_0"), val = tensor(true)]; + tensor mh_w_161_transpose_y_0 = const()[name = tensor("mh_w_161_transpose_y_0"), val = tensor(false)]; + tensor mh_w_161_cast_fp16 = matmul(transpose_x = mh_w_161_transpose_x_0, transpose_y = mh_w_161_transpose_y_0, x = var_5931_cast_fp16, y = var_5933_cast_fp16)[name = tensor("mh_w_161_cast_fp16")]; + tensor obj_377_cast_fp16 = softmax(axis = var_5780, x = mh_w_161_cast_fp16)[name = tensor("obj_377_cast_fp16")]; + tensor var_5937 = const()[name = tensor("op_5937"), val = tensor([1, 20, 64, -1])]; + tensor var_5938_cast_fp16 = reshape(shape = var_5937, x = value_107_cast_fp16)[name = tensor("op_5938_cast_fp16")]; + tensor attn_107_transpose_x_0 = const()[name = tensor("attn_107_transpose_x_0"), val = tensor(false)]; + tensor attn_107_transpose_y_0 = const()[name = tensor("attn_107_transpose_y_0"), val = tensor(true)]; + tensor attn_107_cast_fp16 = matmul(transpose_x = attn_107_transpose_x_0, transpose_y = attn_107_transpose_y_0, x = var_5938_cast_fp16, y = obj_377_cast_fp16)[name = tensor("attn_107_cast_fp16")]; + tensor var_5941 = const()[name = tensor("op_5941"), val = tensor([1, 1280, 1, -1])]; + tensor input_263_cast_fp16 = reshape(shape = var_5941, x = attn_107_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor var_5945 = const()[name = tensor("op_5945"), val = tensor([1, 1])]; + tensor var_5947 = const()[name = tensor("op_5947"), val = tensor([1, 1])]; + tensor obj_375_pad_type_0 = const()[name = tensor("obj_375_pad_type_0"), val = tensor("custom")]; + tensor obj_375_pad_0 = const()[name = tensor("obj_375_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_26_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1521211200)))]; + tensor layers_26_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_26_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1524488064)))]; + tensor obj_375_cast_fp16 = conv(bias = layers_26_encoder_attn_o_proj_bias_to_fp16, dilations = var_5947, groups = var_5787, pad = obj_375_pad_0, pad_type = obj_375_pad_type_0, strides = var_5945, weight = layers_26_encoder_attn_o_proj_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("obj_375_cast_fp16")]; + tensor inputs_161_cast_fp16 = add(x = inputs_159_cast_fp16, y = obj_375_cast_fp16)[name = tensor("inputs_161_cast_fp16")]; + tensor var_5953 = const()[name = tensor("op_5953"), val = tensor([1])]; + tensor channels_mean_161_cast_fp16 = reduce_mean(axes = var_5953, keep_dims = var_5788, x = inputs_161_cast_fp16)[name = tensor("channels_mean_161_cast_fp16")]; + tensor zero_mean_161_cast_fp16 = sub(x = inputs_161_cast_fp16, y = channels_mean_161_cast_fp16)[name = tensor("zero_mean_161_cast_fp16")]; + tensor zero_mean_sq_161_cast_fp16 = mul(x = zero_mean_161_cast_fp16, y = zero_mean_161_cast_fp16)[name = tensor("zero_mean_sq_161_cast_fp16")]; + tensor var_5957 = const()[name = tensor("op_5957"), val = tensor([1])]; + tensor var_5958_cast_fp16 = reduce_mean(axes = var_5957, keep_dims = var_5788, x = zero_mean_sq_161_cast_fp16)[name = tensor("op_5958_cast_fp16")]; + tensor var_5959_to_fp16 = const()[name = tensor("op_5959_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5960_cast_fp16 = add(x = var_5958_cast_fp16, y = var_5959_to_fp16)[name = tensor("op_5960_cast_fp16")]; + tensor denom_161_epsilon_0_to_fp16 = const()[name = tensor("denom_161_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_161_cast_fp16 = rsqrt(epsilon = denom_161_epsilon_0_to_fp16, x = var_5960_cast_fp16)[name = tensor("denom_161_cast_fp16")]; + tensor out_161_cast_fp16 = mul(x = zero_mean_161_cast_fp16, y = denom_161_cast_fp16)[name = tensor("out_161_cast_fp16")]; + tensor input_265_gamma_0_to_fp16 = const()[name = tensor("input_265_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1524490688)))]; + tensor input_265_beta_0_to_fp16 = const()[name = tensor("input_265_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1524493312)))]; + tensor input_265_epsilon_0_to_fp16 = const()[name = tensor("input_265_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_265_cast_fp16 = batch_norm(beta = input_265_beta_0_to_fp16, epsilon = input_265_epsilon_0_to_fp16, gamma = input_265_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_161_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor var_5971 = const()[name = tensor("op_5971"), val = tensor([1, 1])]; + tensor var_5973 = const()[name = tensor("op_5973"), val = tensor([1, 1])]; + tensor input_267_pad_type_0 = const()[name = tensor("input_267_pad_type_0"), val = tensor("custom")]; + tensor input_267_pad_0 = const()[name = tensor("input_267_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_fc1_weight_to_fp16 = const()[name = tensor("layers_26_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1524495936)))]; + tensor layers_26_fc1_bias_to_fp16 = const()[name = tensor("layers_26_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1537603200)))]; + tensor input_267_cast_fp16 = conv(bias = layers_26_fc1_bias_to_fp16, dilations = var_5973, groups = var_5787, pad = input_267_pad_0, pad_type = input_267_pad_type_0, strides = var_5971, weight = layers_26_fc1_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor input_269_mode_0 = const()[name = tensor("input_269_mode_0"), val = tensor("EXACT")]; + tensor input_269_cast_fp16 = gelu(mode = input_269_mode_0, x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor var_5979 = const()[name = tensor("op_5979"), val = tensor([1, 1])]; + tensor var_5981 = const()[name = tensor("op_5981"), val = tensor([1, 1])]; + tensor hidden_states_55_pad_type_0 = const()[name = tensor("hidden_states_55_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_55_pad_0 = const()[name = tensor("hidden_states_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_fc2_weight_to_fp16 = const()[name = tensor("layers_26_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1537613504)))]; + tensor layers_26_fc2_bias_to_fp16 = const()[name = tensor("layers_26_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1550720768)))]; + tensor hidden_states_55_cast_fp16 = conv(bias = layers_26_fc2_bias_to_fp16, dilations = var_5981, groups = var_5787, pad = hidden_states_55_pad_0, pad_type = hidden_states_55_pad_type_0, strides = var_5979, weight = layers_26_fc2_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("hidden_states_55_cast_fp16")]; + tensor inputs_163_cast_fp16 = add(x = inputs_161_cast_fp16, y = hidden_states_55_cast_fp16)[name = tensor("inputs_163_cast_fp16")]; + tensor var_5994 = const()[name = tensor("op_5994"), val = tensor(3)]; + tensor var_6001 = const()[name = tensor("op_6001"), val = tensor(1)]; + tensor var_6002 = const()[name = tensor("op_6002"), val = tensor(true)]; + tensor var_6014 = const()[name = tensor("op_6014"), val = tensor([1])]; + tensor channels_mean_163_cast_fp16 = reduce_mean(axes = var_6014, keep_dims = var_6002, x = inputs_163_cast_fp16)[name = tensor("channels_mean_163_cast_fp16")]; + tensor zero_mean_163_cast_fp16 = sub(x = inputs_163_cast_fp16, y = channels_mean_163_cast_fp16)[name = tensor("zero_mean_163_cast_fp16")]; + tensor zero_mean_sq_163_cast_fp16 = mul(x = zero_mean_163_cast_fp16, y = zero_mean_163_cast_fp16)[name = tensor("zero_mean_sq_163_cast_fp16")]; + tensor var_6018 = const()[name = tensor("op_6018"), val = tensor([1])]; + tensor var_6019_cast_fp16 = reduce_mean(axes = var_6018, keep_dims = var_6002, x = zero_mean_sq_163_cast_fp16)[name = tensor("op_6019_cast_fp16")]; + tensor var_6020_to_fp16 = const()[name = tensor("op_6020_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6021_cast_fp16 = add(x = var_6019_cast_fp16, y = var_6020_to_fp16)[name = tensor("op_6021_cast_fp16")]; + tensor denom_163_epsilon_0_to_fp16 = const()[name = tensor("denom_163_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_163_cast_fp16 = rsqrt(epsilon = denom_163_epsilon_0_to_fp16, x = var_6021_cast_fp16)[name = tensor("denom_163_cast_fp16")]; + tensor out_163_cast_fp16 = mul(x = zero_mean_163_cast_fp16, y = denom_163_cast_fp16)[name = tensor("out_163_cast_fp16")]; + tensor obj_379_gamma_0_to_fp16 = const()[name = tensor("obj_379_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1550723392)))]; + tensor obj_379_beta_0_to_fp16 = const()[name = tensor("obj_379_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1550726016)))]; + tensor obj_379_epsilon_0_to_fp16 = const()[name = tensor("obj_379_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_379_cast_fp16 = batch_norm(beta = obj_379_beta_0_to_fp16, epsilon = obj_379_epsilon_0_to_fp16, gamma = obj_379_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_163_cast_fp16)[name = tensor("obj_379_cast_fp16")]; + tensor var_6036 = const()[name = tensor("op_6036"), val = tensor([1, 1])]; + tensor var_6038 = const()[name = tensor("op_6038"), val = tensor([1, 1])]; + tensor query_109_pad_type_0 = const()[name = tensor("query_109_pad_type_0"), val = tensor("custom")]; + tensor query_109_pad_0 = const()[name = tensor("query_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1550728640)))]; + tensor layers_27_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1554005504)))]; + tensor query_109_cast_fp16 = conv(bias = layers_27_self_attn_q_proj_bias_to_fp16, dilations = var_6038, groups = var_6001, pad = query_109_pad_0, pad_type = query_109_pad_type_0, strides = var_6036, weight = layers_27_self_attn_q_proj_weight_to_fp16, x = obj_379_cast_fp16)[name = tensor("query_109_cast_fp16")]; + tensor var_6042 = const()[name = tensor("op_6042"), val = tensor([1, 1])]; + tensor var_6044 = const()[name = tensor("op_6044"), val = tensor([1, 1])]; + tensor current_key_55_pad_type_0 = const()[name = tensor("current_key_55_pad_type_0"), val = tensor("custom")]; + tensor current_key_55_pad_0 = const()[name = tensor("current_key_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1554008128)))]; + tensor current_key_55_cast_fp16 = conv(dilations = var_6044, groups = var_6001, pad = current_key_55_pad_0, pad_type = current_key_55_pad_type_0, strides = var_6042, weight = layers_27_self_attn_k_proj_weight_to_fp16, x = obj_379_cast_fp16)[name = tensor("current_key_55_cast_fp16")]; + tensor var_6049 = const()[name = tensor("op_6049"), val = tensor([1, 1])]; + tensor var_6051 = const()[name = tensor("op_6051"), val = tensor([1, 1])]; + tensor current_value_55_pad_type_0 = const()[name = tensor("current_value_55_pad_type_0"), val = tensor("custom")]; + tensor current_value_55_pad_0 = const()[name = tensor("current_value_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1557284992)))]; + tensor layers_27_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1560561856)))]; + tensor current_value_55_cast_fp16 = conv(bias = layers_27_self_attn_v_proj_bias_to_fp16, dilations = var_6051, groups = var_6001, pad = current_value_55_pad_0, pad_type = current_value_55_pad_type_0, strides = var_6049, weight = layers_27_self_attn_v_proj_weight_to_fp16, x = obj_379_cast_fp16)[name = tensor("current_value_55_cast_fp16")]; + tensor var_6058_cast_fp16 = mul(x = current_key_55_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6058_cast_fp16")]; + tensor var_6060_cast_fp16 = mul(x = var_103_cast_fp16_27, y = var_241_cast_fp16)[name = tensor("op_6060_cast_fp16")]; + tensor key_109_cast_fp16 = add(x = var_6058_cast_fp16, y = var_6060_cast_fp16)[name = tensor("key_109_cast_fp16")]; + tensor var_6062_cast_fp16 = mul(x = current_value_55_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6062_cast_fp16")]; + tensor var_6064_cast_fp16 = mul(x = var_138_cast_fp16_27, y = var_241_cast_fp16)[name = tensor("op_6064_cast_fp16")]; + tensor value_109_cast_fp16 = add(x = var_6062_cast_fp16, y = var_6064_cast_fp16)[name = tensor("value_109_cast_fp16")]; + tensor var_6067 = const()[name = tensor("op_6067"), val = tensor([1, 20, 64, -1])]; + tensor var_6068_cast_fp16 = reshape(shape = var_6067, x = query_109_cast_fp16)[name = tensor("op_6068_cast_fp16")]; + tensor var_6069_to_fp16 = const()[name = tensor("op_6069_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6070_cast_fp16 = mul(x = var_6068_cast_fp16, y = var_6069_to_fp16)[name = tensor("op_6070_cast_fp16")]; + tensor var_6071 = const()[name = tensor("op_6071"), val = tensor([1, 20, 64, -1])]; + tensor var_6072_cast_fp16 = reshape(shape = var_6071, x = key_109_cast_fp16)[name = tensor("op_6072_cast_fp16")]; + tensor mh_w_163_transpose_x_0 = const()[name = tensor("mh_w_163_transpose_x_0"), val = tensor(true)]; + tensor mh_w_163_transpose_y_0 = const()[name = tensor("mh_w_163_transpose_y_0"), val = tensor(false)]; + tensor mh_w_163_cast_fp16 = matmul(transpose_x = mh_w_163_transpose_x_0, transpose_y = mh_w_163_transpose_y_0, x = var_6070_cast_fp16, y = var_6072_cast_fp16)[name = tensor("mh_w_163_cast_fp16")]; + tensor mh_w_165_cast_fp16 = add(x = mh_w_163_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_165_cast_fp16")]; + tensor var_6080_cast_fp16 = softmax(axis = var_5994, x = mh_w_165_cast_fp16)[name = tensor("op_6080_cast_fp16")]; + tensor var_6081 = const()[name = tensor("op_6081"), val = tensor([1, 20, 64, -1])]; + tensor var_6082_cast_fp16 = reshape(shape = var_6081, x = value_109_cast_fp16)[name = tensor("op_6082_cast_fp16")]; + tensor attn_109_transpose_x_0 = const()[name = tensor("attn_109_transpose_x_0"), val = tensor(false)]; + tensor attn_109_transpose_y_0 = const()[name = tensor("attn_109_transpose_y_0"), val = tensor(true)]; + tensor attn_109_cast_fp16 = matmul(transpose_x = attn_109_transpose_x_0, transpose_y = attn_109_transpose_y_0, x = var_6082_cast_fp16, y = var_6080_cast_fp16)[name = tensor("attn_109_cast_fp16")]; + tensor var_6085 = const()[name = tensor("op_6085"), val = tensor([1, 1280, 1, -1])]; + tensor input_271_cast_fp16 = reshape(shape = var_6085, x = attn_109_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor var_6089 = const()[name = tensor("op_6089"), val = tensor([1, 1])]; + tensor var_6091 = const()[name = tensor("op_6091"), val = tensor([1, 1])]; + tensor obj_385_pad_type_0 = const()[name = tensor("obj_385_pad_type_0"), val = tensor("custom")]; + tensor obj_385_pad_0 = const()[name = tensor("obj_385_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1560564480)))]; + tensor layers_27_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1563841344)))]; + tensor obj_385_cast_fp16 = conv(bias = layers_27_self_attn_o_proj_bias_to_fp16, dilations = var_6091, groups = var_6001, pad = obj_385_pad_0, pad_type = obj_385_pad_type_0, strides = var_6089, weight = layers_27_self_attn_o_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("obj_385_cast_fp16")]; + tensor inputs_165_cast_fp16 = add(x = inputs_163_cast_fp16, y = obj_385_cast_fp16)[name = tensor("inputs_165_cast_fp16")]; + tensor var_6101 = const()[name = tensor("op_6101"), val = tensor([1])]; + tensor channels_mean_165_cast_fp16 = reduce_mean(axes = var_6101, keep_dims = var_6002, x = inputs_165_cast_fp16)[name = tensor("channels_mean_165_cast_fp16")]; + tensor zero_mean_165_cast_fp16 = sub(x = inputs_165_cast_fp16, y = channels_mean_165_cast_fp16)[name = tensor("zero_mean_165_cast_fp16")]; + tensor zero_mean_sq_165_cast_fp16 = mul(x = zero_mean_165_cast_fp16, y = zero_mean_165_cast_fp16)[name = tensor("zero_mean_sq_165_cast_fp16")]; + tensor var_6105 = const()[name = tensor("op_6105"), val = tensor([1])]; + tensor var_6106_cast_fp16 = reduce_mean(axes = var_6105, keep_dims = var_6002, x = zero_mean_sq_165_cast_fp16)[name = tensor("op_6106_cast_fp16")]; + tensor var_6107_to_fp16 = const()[name = tensor("op_6107_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6108_cast_fp16 = add(x = var_6106_cast_fp16, y = var_6107_to_fp16)[name = tensor("op_6108_cast_fp16")]; + tensor denom_165_epsilon_0_to_fp16 = const()[name = tensor("denom_165_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_165_cast_fp16 = rsqrt(epsilon = denom_165_epsilon_0_to_fp16, x = var_6108_cast_fp16)[name = tensor("denom_165_cast_fp16")]; + tensor out_165_cast_fp16 = mul(x = zero_mean_165_cast_fp16, y = denom_165_cast_fp16)[name = tensor("out_165_cast_fp16")]; + tensor obj_387_gamma_0_to_fp16 = const()[name = tensor("obj_387_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1563843968)))]; + tensor obj_387_beta_0_to_fp16 = const()[name = tensor("obj_387_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1563846592)))]; + tensor obj_387_epsilon_0_to_fp16 = const()[name = tensor("obj_387_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_387_cast_fp16 = batch_norm(beta = obj_387_beta_0_to_fp16, epsilon = obj_387_epsilon_0_to_fp16, gamma = obj_387_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_165_cast_fp16)[name = tensor("obj_387_cast_fp16")]; + tensor var_6123 = const()[name = tensor("op_6123"), val = tensor([1, 1])]; + tensor var_6125 = const()[name = tensor("op_6125"), val = tensor([1, 1])]; + tensor query_111_pad_type_0 = const()[name = tensor("query_111_pad_type_0"), val = tensor("custom")]; + tensor query_111_pad_0 = const()[name = tensor("query_111_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_27_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1563849216)))]; + tensor layers_27_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_27_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1567126080)))]; + tensor query_111_cast_fp16 = conv(bias = layers_27_encoder_attn_q_proj_bias_to_fp16, dilations = var_6125, groups = var_6001, pad = query_111_pad_0, pad_type = query_111_pad_type_0, strides = var_6123, weight = layers_27_encoder_attn_q_proj_weight_to_fp16, x = obj_387_cast_fp16)[name = tensor("query_111_cast_fp16")]; + tensor var_6129 = const()[name = tensor("op_6129"), val = tensor([1, 1])]; + tensor var_6131 = const()[name = tensor("op_6131"), val = tensor([1, 1])]; + tensor key_111_pad_type_0 = const()[name = tensor("key_111_pad_type_0"), val = tensor("custom")]; + tensor key_111_pad_0 = const()[name = tensor("key_111_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_27_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1567128704)))]; + tensor key_111_cast_fp16 = conv(dilations = var_6131, groups = var_6001, pad = key_111_pad_0, pad_type = key_111_pad_type_0, strides = var_6129, weight = layers_27_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_111_cast_fp16")]; + tensor var_6136 = const()[name = tensor("op_6136"), val = tensor([1, 1])]; + tensor var_6138 = const()[name = tensor("op_6138"), val = tensor([1, 1])]; + tensor value_111_pad_type_0 = const()[name = tensor("value_111_pad_type_0"), val = tensor("custom")]; + tensor value_111_pad_0 = const()[name = tensor("value_111_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_27_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1570405568)))]; + tensor layers_27_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_27_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1573682432)))]; + tensor value_111_cast_fp16 = conv(bias = layers_27_encoder_attn_v_proj_bias_to_fp16, dilations = var_6138, groups = var_6001, pad = value_111_pad_0, pad_type = value_111_pad_type_0, strides = var_6136, weight = layers_27_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_111_cast_fp16")]; + tensor var_6142 = const()[name = tensor("op_6142"), val = tensor([1, 20, 64, -1])]; + tensor var_6143_cast_fp16 = reshape(shape = var_6142, x = query_111_cast_fp16)[name = tensor("op_6143_cast_fp16")]; + tensor var_6144_to_fp16 = const()[name = tensor("op_6144_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6145_cast_fp16 = mul(x = var_6143_cast_fp16, y = var_6144_to_fp16)[name = tensor("op_6145_cast_fp16")]; + tensor var_6146 = const()[name = tensor("op_6146"), val = tensor([1, 20, 64, -1])]; + tensor var_6147_cast_fp16 = reshape(shape = var_6146, x = key_111_cast_fp16)[name = tensor("op_6147_cast_fp16")]; + tensor mh_w_167_transpose_x_0 = const()[name = tensor("mh_w_167_transpose_x_0"), val = tensor(true)]; + tensor mh_w_167_transpose_y_0 = const()[name = tensor("mh_w_167_transpose_y_0"), val = tensor(false)]; + tensor mh_w_167_cast_fp16 = matmul(transpose_x = mh_w_167_transpose_x_0, transpose_y = mh_w_167_transpose_y_0, x = var_6145_cast_fp16, y = var_6147_cast_fp16)[name = tensor("mh_w_167_cast_fp16")]; + tensor obj_391_cast_fp16 = softmax(axis = var_5994, x = mh_w_167_cast_fp16)[name = tensor("obj_391_cast_fp16")]; + tensor var_6151 = const()[name = tensor("op_6151"), val = tensor([1, 20, 64, -1])]; + tensor var_6152_cast_fp16 = reshape(shape = var_6151, x = value_111_cast_fp16)[name = tensor("op_6152_cast_fp16")]; + tensor attn_111_transpose_x_0 = const()[name = tensor("attn_111_transpose_x_0"), val = tensor(false)]; + tensor attn_111_transpose_y_0 = const()[name = tensor("attn_111_transpose_y_0"), val = tensor(true)]; + tensor attn_111_cast_fp16 = matmul(transpose_x = attn_111_transpose_x_0, transpose_y = attn_111_transpose_y_0, x = var_6152_cast_fp16, y = obj_391_cast_fp16)[name = tensor("attn_111_cast_fp16")]; + tensor var_6155 = const()[name = tensor("op_6155"), val = tensor([1, 1280, 1, -1])]; + tensor input_273_cast_fp16 = reshape(shape = var_6155, x = attn_111_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor var_6159 = const()[name = tensor("op_6159"), val = tensor([1, 1])]; + tensor var_6161 = const()[name = tensor("op_6161"), val = tensor([1, 1])]; + tensor obj_389_pad_type_0 = const()[name = tensor("obj_389_pad_type_0"), val = tensor("custom")]; + tensor obj_389_pad_0 = const()[name = tensor("obj_389_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_27_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1573685056)))]; + tensor layers_27_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_27_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1576961920)))]; + tensor obj_389_cast_fp16 = conv(bias = layers_27_encoder_attn_o_proj_bias_to_fp16, dilations = var_6161, groups = var_6001, pad = obj_389_pad_0, pad_type = obj_389_pad_type_0, strides = var_6159, weight = layers_27_encoder_attn_o_proj_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("obj_389_cast_fp16")]; + tensor inputs_167_cast_fp16 = add(x = inputs_165_cast_fp16, y = obj_389_cast_fp16)[name = tensor("inputs_167_cast_fp16")]; + tensor var_6167 = const()[name = tensor("op_6167"), val = tensor([1])]; + tensor channels_mean_167_cast_fp16 = reduce_mean(axes = var_6167, keep_dims = var_6002, x = inputs_167_cast_fp16)[name = tensor("channels_mean_167_cast_fp16")]; + tensor zero_mean_167_cast_fp16 = sub(x = inputs_167_cast_fp16, y = channels_mean_167_cast_fp16)[name = tensor("zero_mean_167_cast_fp16")]; + tensor zero_mean_sq_167_cast_fp16 = mul(x = zero_mean_167_cast_fp16, y = zero_mean_167_cast_fp16)[name = tensor("zero_mean_sq_167_cast_fp16")]; + tensor var_6171 = const()[name = tensor("op_6171"), val = tensor([1])]; + tensor var_6172_cast_fp16 = reduce_mean(axes = var_6171, keep_dims = var_6002, x = zero_mean_sq_167_cast_fp16)[name = tensor("op_6172_cast_fp16")]; + tensor var_6173_to_fp16 = const()[name = tensor("op_6173_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6174_cast_fp16 = add(x = var_6172_cast_fp16, y = var_6173_to_fp16)[name = tensor("op_6174_cast_fp16")]; + tensor denom_167_epsilon_0_to_fp16 = const()[name = tensor("denom_167_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_167_cast_fp16 = rsqrt(epsilon = denom_167_epsilon_0_to_fp16, x = var_6174_cast_fp16)[name = tensor("denom_167_cast_fp16")]; + tensor out_167_cast_fp16 = mul(x = zero_mean_167_cast_fp16, y = denom_167_cast_fp16)[name = tensor("out_167_cast_fp16")]; + tensor input_275_gamma_0_to_fp16 = const()[name = tensor("input_275_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1576964544)))]; + tensor input_275_beta_0_to_fp16 = const()[name = tensor("input_275_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1576967168)))]; + tensor input_275_epsilon_0_to_fp16 = const()[name = tensor("input_275_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_275_cast_fp16 = batch_norm(beta = input_275_beta_0_to_fp16, epsilon = input_275_epsilon_0_to_fp16, gamma = input_275_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_167_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor var_6185 = const()[name = tensor("op_6185"), val = tensor([1, 1])]; + tensor var_6187 = const()[name = tensor("op_6187"), val = tensor([1, 1])]; + tensor input_277_pad_type_0 = const()[name = tensor("input_277_pad_type_0"), val = tensor("custom")]; + tensor input_277_pad_0 = const()[name = tensor("input_277_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_fc1_weight_to_fp16 = const()[name = tensor("layers_27_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1576969792)))]; + tensor layers_27_fc1_bias_to_fp16 = const()[name = tensor("layers_27_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1590077056)))]; + tensor input_277_cast_fp16 = conv(bias = layers_27_fc1_bias_to_fp16, dilations = var_6187, groups = var_6001, pad = input_277_pad_0, pad_type = input_277_pad_type_0, strides = var_6185, weight = layers_27_fc1_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor input_279_mode_0 = const()[name = tensor("input_279_mode_0"), val = tensor("EXACT")]; + tensor input_279_cast_fp16 = gelu(mode = input_279_mode_0, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor var_6193 = const()[name = tensor("op_6193"), val = tensor([1, 1])]; + tensor var_6195 = const()[name = tensor("op_6195"), val = tensor([1, 1])]; + tensor hidden_states_57_pad_type_0 = const()[name = tensor("hidden_states_57_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_57_pad_0 = const()[name = tensor("hidden_states_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_fc2_weight_to_fp16 = const()[name = tensor("layers_27_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1590087360)))]; + tensor layers_27_fc2_bias_to_fp16 = const()[name = tensor("layers_27_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1603194624)))]; + tensor hidden_states_57_cast_fp16 = conv(bias = layers_27_fc2_bias_to_fp16, dilations = var_6195, groups = var_6001, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = var_6193, weight = layers_27_fc2_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("hidden_states_57_cast_fp16")]; + tensor inputs_169_cast_fp16 = add(x = inputs_167_cast_fp16, y = hidden_states_57_cast_fp16)[name = tensor("inputs_169_cast_fp16")]; + tensor var_6208 = const()[name = tensor("op_6208"), val = tensor(3)]; + tensor var_6215 = const()[name = tensor("op_6215"), val = tensor(1)]; + tensor var_6216 = const()[name = tensor("op_6216"), val = tensor(true)]; + tensor var_6228 = const()[name = tensor("op_6228"), val = tensor([1])]; + tensor channels_mean_169_cast_fp16 = reduce_mean(axes = var_6228, keep_dims = var_6216, x = inputs_169_cast_fp16)[name = tensor("channels_mean_169_cast_fp16")]; + tensor zero_mean_169_cast_fp16 = sub(x = inputs_169_cast_fp16, y = channels_mean_169_cast_fp16)[name = tensor("zero_mean_169_cast_fp16")]; + tensor zero_mean_sq_169_cast_fp16 = mul(x = zero_mean_169_cast_fp16, y = zero_mean_169_cast_fp16)[name = tensor("zero_mean_sq_169_cast_fp16")]; + tensor var_6232 = const()[name = tensor("op_6232"), val = tensor([1])]; + tensor var_6233_cast_fp16 = reduce_mean(axes = var_6232, keep_dims = var_6216, x = zero_mean_sq_169_cast_fp16)[name = tensor("op_6233_cast_fp16")]; + tensor var_6234_to_fp16 = const()[name = tensor("op_6234_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6235_cast_fp16 = add(x = var_6233_cast_fp16, y = var_6234_to_fp16)[name = tensor("op_6235_cast_fp16")]; + tensor denom_169_epsilon_0_to_fp16 = const()[name = tensor("denom_169_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_169_cast_fp16 = rsqrt(epsilon = denom_169_epsilon_0_to_fp16, x = var_6235_cast_fp16)[name = tensor("denom_169_cast_fp16")]; + tensor out_169_cast_fp16 = mul(x = zero_mean_169_cast_fp16, y = denom_169_cast_fp16)[name = tensor("out_169_cast_fp16")]; + tensor obj_393_gamma_0_to_fp16 = const()[name = tensor("obj_393_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1603197248)))]; + tensor obj_393_beta_0_to_fp16 = const()[name = tensor("obj_393_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1603199872)))]; + tensor obj_393_epsilon_0_to_fp16 = const()[name = tensor("obj_393_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_393_cast_fp16 = batch_norm(beta = obj_393_beta_0_to_fp16, epsilon = obj_393_epsilon_0_to_fp16, gamma = obj_393_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_169_cast_fp16)[name = tensor("obj_393_cast_fp16")]; + tensor var_6250 = const()[name = tensor("op_6250"), val = tensor([1, 1])]; + tensor var_6252 = const()[name = tensor("op_6252"), val = tensor([1, 1])]; + tensor query_113_pad_type_0 = const()[name = tensor("query_113_pad_type_0"), val = tensor("custom")]; + tensor query_113_pad_0 = const()[name = tensor("query_113_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1603202496)))]; + tensor layers_28_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1606479360)))]; + tensor query_113_cast_fp16 = conv(bias = layers_28_self_attn_q_proj_bias_to_fp16, dilations = var_6252, groups = var_6215, pad = query_113_pad_0, pad_type = query_113_pad_type_0, strides = var_6250, weight = layers_28_self_attn_q_proj_weight_to_fp16, x = obj_393_cast_fp16)[name = tensor("query_113_cast_fp16")]; + tensor var_6256 = const()[name = tensor("op_6256"), val = tensor([1, 1])]; + tensor var_6258 = const()[name = tensor("op_6258"), val = tensor([1, 1])]; + tensor current_key_57_pad_type_0 = const()[name = tensor("current_key_57_pad_type_0"), val = tensor("custom")]; + tensor current_key_57_pad_0 = const()[name = tensor("current_key_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1606481984)))]; + tensor current_key_57_cast_fp16 = conv(dilations = var_6258, groups = var_6215, pad = current_key_57_pad_0, pad_type = current_key_57_pad_type_0, strides = var_6256, weight = layers_28_self_attn_k_proj_weight_to_fp16, x = obj_393_cast_fp16)[name = tensor("current_key_57_cast_fp16")]; + tensor var_6263 = const()[name = tensor("op_6263"), val = tensor([1, 1])]; + tensor var_6265 = const()[name = tensor("op_6265"), val = tensor([1, 1])]; + tensor current_value_57_pad_type_0 = const()[name = tensor("current_value_57_pad_type_0"), val = tensor("custom")]; + tensor current_value_57_pad_0 = const()[name = tensor("current_value_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1609758848)))]; + tensor layers_28_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1613035712)))]; + tensor current_value_57_cast_fp16 = conv(bias = layers_28_self_attn_v_proj_bias_to_fp16, dilations = var_6265, groups = var_6215, pad = current_value_57_pad_0, pad_type = current_value_57_pad_type_0, strides = var_6263, weight = layers_28_self_attn_v_proj_weight_to_fp16, x = obj_393_cast_fp16)[name = tensor("current_value_57_cast_fp16")]; + tensor var_6272_cast_fp16 = mul(x = current_key_57_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6272_cast_fp16")]; + tensor var_6274_cast_fp16 = mul(x = var_103_cast_fp16_28, y = var_241_cast_fp16)[name = tensor("op_6274_cast_fp16")]; + tensor key_113_cast_fp16 = add(x = var_6272_cast_fp16, y = var_6274_cast_fp16)[name = tensor("key_113_cast_fp16")]; + tensor var_6276_cast_fp16 = mul(x = current_value_57_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6276_cast_fp16")]; + tensor var_6278_cast_fp16 = mul(x = var_138_cast_fp16_28, y = var_241_cast_fp16)[name = tensor("op_6278_cast_fp16")]; + tensor value_113_cast_fp16 = add(x = var_6276_cast_fp16, y = var_6278_cast_fp16)[name = tensor("value_113_cast_fp16")]; + tensor var_6281 = const()[name = tensor("op_6281"), val = tensor([1, 20, 64, -1])]; + tensor var_6282_cast_fp16 = reshape(shape = var_6281, x = query_113_cast_fp16)[name = tensor("op_6282_cast_fp16")]; + tensor var_6283_to_fp16 = const()[name = tensor("op_6283_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6284_cast_fp16 = mul(x = var_6282_cast_fp16, y = var_6283_to_fp16)[name = tensor("op_6284_cast_fp16")]; + tensor var_6285 = const()[name = tensor("op_6285"), val = tensor([1, 20, 64, -1])]; + tensor var_6286_cast_fp16 = reshape(shape = var_6285, x = key_113_cast_fp16)[name = tensor("op_6286_cast_fp16")]; + tensor mh_w_169_transpose_x_0 = const()[name = tensor("mh_w_169_transpose_x_0"), val = tensor(true)]; + tensor mh_w_169_transpose_y_0 = const()[name = tensor("mh_w_169_transpose_y_0"), val = tensor(false)]; + tensor mh_w_169_cast_fp16 = matmul(transpose_x = mh_w_169_transpose_x_0, transpose_y = mh_w_169_transpose_y_0, x = var_6284_cast_fp16, y = var_6286_cast_fp16)[name = tensor("mh_w_169_cast_fp16")]; + tensor mh_w_171_cast_fp16 = add(x = mh_w_169_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_171_cast_fp16")]; + tensor var_6294_cast_fp16 = softmax(axis = var_6208, x = mh_w_171_cast_fp16)[name = tensor("op_6294_cast_fp16")]; + tensor var_6295 = const()[name = tensor("op_6295"), val = tensor([1, 20, 64, -1])]; + tensor var_6296_cast_fp16 = reshape(shape = var_6295, x = value_113_cast_fp16)[name = tensor("op_6296_cast_fp16")]; + tensor attn_113_transpose_x_0 = const()[name = tensor("attn_113_transpose_x_0"), val = tensor(false)]; + tensor attn_113_transpose_y_0 = const()[name = tensor("attn_113_transpose_y_0"), val = tensor(true)]; + tensor attn_113_cast_fp16 = matmul(transpose_x = attn_113_transpose_x_0, transpose_y = attn_113_transpose_y_0, x = var_6296_cast_fp16, y = var_6294_cast_fp16)[name = tensor("attn_113_cast_fp16")]; + tensor var_6299 = const()[name = tensor("op_6299"), val = tensor([1, 1280, 1, -1])]; + tensor input_281_cast_fp16 = reshape(shape = var_6299, x = attn_113_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor var_6303 = const()[name = tensor("op_6303"), val = tensor([1, 1])]; + tensor var_6305 = const()[name = tensor("op_6305"), val = tensor([1, 1])]; + tensor obj_399_pad_type_0 = const()[name = tensor("obj_399_pad_type_0"), val = tensor("custom")]; + tensor obj_399_pad_0 = const()[name = tensor("obj_399_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1613038336)))]; + tensor layers_28_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1616315200)))]; + tensor obj_399_cast_fp16 = conv(bias = layers_28_self_attn_o_proj_bias_to_fp16, dilations = var_6305, groups = var_6215, pad = obj_399_pad_0, pad_type = obj_399_pad_type_0, strides = var_6303, weight = layers_28_self_attn_o_proj_weight_to_fp16, x = input_281_cast_fp16)[name = tensor("obj_399_cast_fp16")]; + tensor inputs_171_cast_fp16 = add(x = inputs_169_cast_fp16, y = obj_399_cast_fp16)[name = tensor("inputs_171_cast_fp16")]; + tensor var_6315 = const()[name = tensor("op_6315"), val = tensor([1])]; + tensor channels_mean_171_cast_fp16 = reduce_mean(axes = var_6315, keep_dims = var_6216, x = inputs_171_cast_fp16)[name = tensor("channels_mean_171_cast_fp16")]; + tensor zero_mean_171_cast_fp16 = sub(x = inputs_171_cast_fp16, y = channels_mean_171_cast_fp16)[name = tensor("zero_mean_171_cast_fp16")]; + tensor zero_mean_sq_171_cast_fp16 = mul(x = zero_mean_171_cast_fp16, y = zero_mean_171_cast_fp16)[name = tensor("zero_mean_sq_171_cast_fp16")]; + tensor var_6319 = const()[name = tensor("op_6319"), val = tensor([1])]; + tensor var_6320_cast_fp16 = reduce_mean(axes = var_6319, keep_dims = var_6216, x = zero_mean_sq_171_cast_fp16)[name = tensor("op_6320_cast_fp16")]; + tensor var_6321_to_fp16 = const()[name = tensor("op_6321_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6322_cast_fp16 = add(x = var_6320_cast_fp16, y = var_6321_to_fp16)[name = tensor("op_6322_cast_fp16")]; + tensor denom_171_epsilon_0_to_fp16 = const()[name = tensor("denom_171_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_171_cast_fp16 = rsqrt(epsilon = denom_171_epsilon_0_to_fp16, x = var_6322_cast_fp16)[name = tensor("denom_171_cast_fp16")]; + tensor out_171_cast_fp16 = mul(x = zero_mean_171_cast_fp16, y = denom_171_cast_fp16)[name = tensor("out_171_cast_fp16")]; + tensor obj_401_gamma_0_to_fp16 = const()[name = tensor("obj_401_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1616317824)))]; + tensor obj_401_beta_0_to_fp16 = const()[name = tensor("obj_401_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1616320448)))]; + tensor obj_401_epsilon_0_to_fp16 = const()[name = tensor("obj_401_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_401_cast_fp16 = batch_norm(beta = obj_401_beta_0_to_fp16, epsilon = obj_401_epsilon_0_to_fp16, gamma = obj_401_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_171_cast_fp16)[name = tensor("obj_401_cast_fp16")]; + tensor var_6337 = const()[name = tensor("op_6337"), val = tensor([1, 1])]; + tensor var_6339 = const()[name = tensor("op_6339"), val = tensor([1, 1])]; + tensor query_115_pad_type_0 = const()[name = tensor("query_115_pad_type_0"), val = tensor("custom")]; + tensor query_115_pad_0 = const()[name = tensor("query_115_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_28_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1616323072)))]; + tensor layers_28_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_28_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1619599936)))]; + tensor query_115_cast_fp16 = conv(bias = layers_28_encoder_attn_q_proj_bias_to_fp16, dilations = var_6339, groups = var_6215, pad = query_115_pad_0, pad_type = query_115_pad_type_0, strides = var_6337, weight = layers_28_encoder_attn_q_proj_weight_to_fp16, x = obj_401_cast_fp16)[name = tensor("query_115_cast_fp16")]; + tensor var_6343 = const()[name = tensor("op_6343"), val = tensor([1, 1])]; + tensor var_6345 = const()[name = tensor("op_6345"), val = tensor([1, 1])]; + tensor key_115_pad_type_0 = const()[name = tensor("key_115_pad_type_0"), val = tensor("custom")]; + tensor key_115_pad_0 = const()[name = tensor("key_115_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_28_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1619602560)))]; + tensor key_115_cast_fp16 = conv(dilations = var_6345, groups = var_6215, pad = key_115_pad_0, pad_type = key_115_pad_type_0, strides = var_6343, weight = layers_28_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_115_cast_fp16")]; + tensor var_6350 = const()[name = tensor("op_6350"), val = tensor([1, 1])]; + tensor var_6352 = const()[name = tensor("op_6352"), val = tensor([1, 1])]; + tensor value_115_pad_type_0 = const()[name = tensor("value_115_pad_type_0"), val = tensor("custom")]; + tensor value_115_pad_0 = const()[name = tensor("value_115_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_28_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1622879424)))]; + tensor layers_28_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_28_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1626156288)))]; + tensor value_115_cast_fp16 = conv(bias = layers_28_encoder_attn_v_proj_bias_to_fp16, dilations = var_6352, groups = var_6215, pad = value_115_pad_0, pad_type = value_115_pad_type_0, strides = var_6350, weight = layers_28_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_115_cast_fp16")]; + tensor var_6356 = const()[name = tensor("op_6356"), val = tensor([1, 20, 64, -1])]; + tensor var_6357_cast_fp16 = reshape(shape = var_6356, x = query_115_cast_fp16)[name = tensor("op_6357_cast_fp16")]; + tensor var_6358_to_fp16 = const()[name = tensor("op_6358_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6359_cast_fp16 = mul(x = var_6357_cast_fp16, y = var_6358_to_fp16)[name = tensor("op_6359_cast_fp16")]; + tensor var_6360 = const()[name = tensor("op_6360"), val = tensor([1, 20, 64, -1])]; + tensor var_6361_cast_fp16 = reshape(shape = var_6360, x = key_115_cast_fp16)[name = tensor("op_6361_cast_fp16")]; + tensor mh_w_173_transpose_x_0 = const()[name = tensor("mh_w_173_transpose_x_0"), val = tensor(true)]; + tensor mh_w_173_transpose_y_0 = const()[name = tensor("mh_w_173_transpose_y_0"), val = tensor(false)]; + tensor mh_w_173_cast_fp16 = matmul(transpose_x = mh_w_173_transpose_x_0, transpose_y = mh_w_173_transpose_y_0, x = var_6359_cast_fp16, y = var_6361_cast_fp16)[name = tensor("mh_w_173_cast_fp16")]; + tensor obj_405_cast_fp16 = softmax(axis = var_6208, x = mh_w_173_cast_fp16)[name = tensor("obj_405_cast_fp16")]; + tensor var_6365 = const()[name = tensor("op_6365"), val = tensor([1, 20, 64, -1])]; + tensor var_6366_cast_fp16 = reshape(shape = var_6365, x = value_115_cast_fp16)[name = tensor("op_6366_cast_fp16")]; + tensor attn_115_transpose_x_0 = const()[name = tensor("attn_115_transpose_x_0"), val = tensor(false)]; + tensor attn_115_transpose_y_0 = const()[name = tensor("attn_115_transpose_y_0"), val = tensor(true)]; + tensor attn_115_cast_fp16 = matmul(transpose_x = attn_115_transpose_x_0, transpose_y = attn_115_transpose_y_0, x = var_6366_cast_fp16, y = obj_405_cast_fp16)[name = tensor("attn_115_cast_fp16")]; + tensor var_6369 = const()[name = tensor("op_6369"), val = tensor([1, 1280, 1, -1])]; + tensor input_283_cast_fp16 = reshape(shape = var_6369, x = attn_115_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor var_6373 = const()[name = tensor("op_6373"), val = tensor([1, 1])]; + tensor var_6375 = const()[name = tensor("op_6375"), val = tensor([1, 1])]; + tensor obj_403_pad_type_0 = const()[name = tensor("obj_403_pad_type_0"), val = tensor("custom")]; + tensor obj_403_pad_0 = const()[name = tensor("obj_403_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_28_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1626158912)))]; + tensor layers_28_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_28_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1629435776)))]; + tensor obj_403_cast_fp16 = conv(bias = layers_28_encoder_attn_o_proj_bias_to_fp16, dilations = var_6375, groups = var_6215, pad = obj_403_pad_0, pad_type = obj_403_pad_type_0, strides = var_6373, weight = layers_28_encoder_attn_o_proj_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("obj_403_cast_fp16")]; + tensor inputs_173_cast_fp16 = add(x = inputs_171_cast_fp16, y = obj_403_cast_fp16)[name = tensor("inputs_173_cast_fp16")]; + tensor var_6381 = const()[name = tensor("op_6381"), val = tensor([1])]; + tensor channels_mean_173_cast_fp16 = reduce_mean(axes = var_6381, keep_dims = var_6216, x = inputs_173_cast_fp16)[name = tensor("channels_mean_173_cast_fp16")]; + tensor zero_mean_173_cast_fp16 = sub(x = inputs_173_cast_fp16, y = channels_mean_173_cast_fp16)[name = tensor("zero_mean_173_cast_fp16")]; + tensor zero_mean_sq_173_cast_fp16 = mul(x = zero_mean_173_cast_fp16, y = zero_mean_173_cast_fp16)[name = tensor("zero_mean_sq_173_cast_fp16")]; + tensor var_6385 = const()[name = tensor("op_6385"), val = tensor([1])]; + tensor var_6386_cast_fp16 = reduce_mean(axes = var_6385, keep_dims = var_6216, x = zero_mean_sq_173_cast_fp16)[name = tensor("op_6386_cast_fp16")]; + tensor var_6387_to_fp16 = const()[name = tensor("op_6387_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6388_cast_fp16 = add(x = var_6386_cast_fp16, y = var_6387_to_fp16)[name = tensor("op_6388_cast_fp16")]; + tensor denom_173_epsilon_0_to_fp16 = const()[name = tensor("denom_173_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_173_cast_fp16 = rsqrt(epsilon = denom_173_epsilon_0_to_fp16, x = var_6388_cast_fp16)[name = tensor("denom_173_cast_fp16")]; + tensor out_173_cast_fp16 = mul(x = zero_mean_173_cast_fp16, y = denom_173_cast_fp16)[name = tensor("out_173_cast_fp16")]; + tensor input_285_gamma_0_to_fp16 = const()[name = tensor("input_285_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1629438400)))]; + tensor input_285_beta_0_to_fp16 = const()[name = tensor("input_285_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1629441024)))]; + tensor input_285_epsilon_0_to_fp16 = const()[name = tensor("input_285_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_285_cast_fp16 = batch_norm(beta = input_285_beta_0_to_fp16, epsilon = input_285_epsilon_0_to_fp16, gamma = input_285_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_173_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor var_6399 = const()[name = tensor("op_6399"), val = tensor([1, 1])]; + tensor var_6401 = const()[name = tensor("op_6401"), val = tensor([1, 1])]; + tensor input_287_pad_type_0 = const()[name = tensor("input_287_pad_type_0"), val = tensor("custom")]; + tensor input_287_pad_0 = const()[name = tensor("input_287_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_fc1_weight_to_fp16 = const()[name = tensor("layers_28_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1629443648)))]; + tensor layers_28_fc1_bias_to_fp16 = const()[name = tensor("layers_28_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1642550912)))]; + tensor input_287_cast_fp16 = conv(bias = layers_28_fc1_bias_to_fp16, dilations = var_6401, groups = var_6215, pad = input_287_pad_0, pad_type = input_287_pad_type_0, strides = var_6399, weight = layers_28_fc1_weight_to_fp16, x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor input_289_mode_0 = const()[name = tensor("input_289_mode_0"), val = tensor("EXACT")]; + tensor input_289_cast_fp16 = gelu(mode = input_289_mode_0, x = input_287_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor var_6407 = const()[name = tensor("op_6407"), val = tensor([1, 1])]; + tensor var_6409 = const()[name = tensor("op_6409"), val = tensor([1, 1])]; + tensor hidden_states_59_pad_type_0 = const()[name = tensor("hidden_states_59_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_59_pad_0 = const()[name = tensor("hidden_states_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_fc2_weight_to_fp16 = const()[name = tensor("layers_28_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1642561216)))]; + tensor layers_28_fc2_bias_to_fp16 = const()[name = tensor("layers_28_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1655668480)))]; + tensor hidden_states_59_cast_fp16 = conv(bias = layers_28_fc2_bias_to_fp16, dilations = var_6409, groups = var_6215, pad = hidden_states_59_pad_0, pad_type = hidden_states_59_pad_type_0, strides = var_6407, weight = layers_28_fc2_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("hidden_states_59_cast_fp16")]; + tensor inputs_175_cast_fp16 = add(x = inputs_173_cast_fp16, y = hidden_states_59_cast_fp16)[name = tensor("inputs_175_cast_fp16")]; + tensor var_6422 = const()[name = tensor("op_6422"), val = tensor(3)]; + tensor var_6429 = const()[name = tensor("op_6429"), val = tensor(1)]; + tensor var_6430 = const()[name = tensor("op_6430"), val = tensor(true)]; + tensor var_6442 = const()[name = tensor("op_6442"), val = tensor([1])]; + tensor channels_mean_175_cast_fp16 = reduce_mean(axes = var_6442, keep_dims = var_6430, x = inputs_175_cast_fp16)[name = tensor("channels_mean_175_cast_fp16")]; + tensor zero_mean_175_cast_fp16 = sub(x = inputs_175_cast_fp16, y = channels_mean_175_cast_fp16)[name = tensor("zero_mean_175_cast_fp16")]; + tensor zero_mean_sq_175_cast_fp16 = mul(x = zero_mean_175_cast_fp16, y = zero_mean_175_cast_fp16)[name = tensor("zero_mean_sq_175_cast_fp16")]; + tensor var_6446 = const()[name = tensor("op_6446"), val = tensor([1])]; + tensor var_6447_cast_fp16 = reduce_mean(axes = var_6446, keep_dims = var_6430, x = zero_mean_sq_175_cast_fp16)[name = tensor("op_6447_cast_fp16")]; + tensor var_6448_to_fp16 = const()[name = tensor("op_6448_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6449_cast_fp16 = add(x = var_6447_cast_fp16, y = var_6448_to_fp16)[name = tensor("op_6449_cast_fp16")]; + tensor denom_175_epsilon_0_to_fp16 = const()[name = tensor("denom_175_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_175_cast_fp16 = rsqrt(epsilon = denom_175_epsilon_0_to_fp16, x = var_6449_cast_fp16)[name = tensor("denom_175_cast_fp16")]; + tensor out_175_cast_fp16 = mul(x = zero_mean_175_cast_fp16, y = denom_175_cast_fp16)[name = tensor("out_175_cast_fp16")]; + tensor obj_407_gamma_0_to_fp16 = const()[name = tensor("obj_407_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1655671104)))]; + tensor obj_407_beta_0_to_fp16 = const()[name = tensor("obj_407_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1655673728)))]; + tensor obj_407_epsilon_0_to_fp16 = const()[name = tensor("obj_407_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_407_cast_fp16 = batch_norm(beta = obj_407_beta_0_to_fp16, epsilon = obj_407_epsilon_0_to_fp16, gamma = obj_407_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_175_cast_fp16)[name = tensor("obj_407_cast_fp16")]; + tensor var_6464 = const()[name = tensor("op_6464"), val = tensor([1, 1])]; + tensor var_6466 = const()[name = tensor("op_6466"), val = tensor([1, 1])]; + tensor query_117_pad_type_0 = const()[name = tensor("query_117_pad_type_0"), val = tensor("custom")]; + tensor query_117_pad_0 = const()[name = tensor("query_117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1655676352)))]; + tensor layers_29_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1658953216)))]; + tensor query_117_cast_fp16 = conv(bias = layers_29_self_attn_q_proj_bias_to_fp16, dilations = var_6466, groups = var_6429, pad = query_117_pad_0, pad_type = query_117_pad_type_0, strides = var_6464, weight = layers_29_self_attn_q_proj_weight_to_fp16, x = obj_407_cast_fp16)[name = tensor("query_117_cast_fp16")]; + tensor var_6470 = const()[name = tensor("op_6470"), val = tensor([1, 1])]; + tensor var_6472 = const()[name = tensor("op_6472"), val = tensor([1, 1])]; + tensor current_key_59_pad_type_0 = const()[name = tensor("current_key_59_pad_type_0"), val = tensor("custom")]; + tensor current_key_59_pad_0 = const()[name = tensor("current_key_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1658955840)))]; + tensor current_key_59_cast_fp16 = conv(dilations = var_6472, groups = var_6429, pad = current_key_59_pad_0, pad_type = current_key_59_pad_type_0, strides = var_6470, weight = layers_29_self_attn_k_proj_weight_to_fp16, x = obj_407_cast_fp16)[name = tensor("current_key_59_cast_fp16")]; + tensor var_6477 = const()[name = tensor("op_6477"), val = tensor([1, 1])]; + tensor var_6479 = const()[name = tensor("op_6479"), val = tensor([1, 1])]; + tensor current_value_59_pad_type_0 = const()[name = tensor("current_value_59_pad_type_0"), val = tensor("custom")]; + tensor current_value_59_pad_0 = const()[name = tensor("current_value_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1662232704)))]; + tensor layers_29_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1665509568)))]; + tensor current_value_59_cast_fp16 = conv(bias = layers_29_self_attn_v_proj_bias_to_fp16, dilations = var_6479, groups = var_6429, pad = current_value_59_pad_0, pad_type = current_value_59_pad_type_0, strides = var_6477, weight = layers_29_self_attn_v_proj_weight_to_fp16, x = obj_407_cast_fp16)[name = tensor("current_value_59_cast_fp16")]; + tensor var_6486_cast_fp16 = mul(x = current_key_59_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6486_cast_fp16")]; + tensor var_6488_cast_fp16 = mul(x = var_103_cast_fp16_29, y = var_241_cast_fp16)[name = tensor("op_6488_cast_fp16")]; + tensor key_117_cast_fp16 = add(x = var_6486_cast_fp16, y = var_6488_cast_fp16)[name = tensor("key_117_cast_fp16")]; + tensor var_6490_cast_fp16 = mul(x = current_value_59_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6490_cast_fp16")]; + tensor var_6492_cast_fp16 = mul(x = var_138_cast_fp16_29, y = var_241_cast_fp16)[name = tensor("op_6492_cast_fp16")]; + tensor value_117_cast_fp16 = add(x = var_6490_cast_fp16, y = var_6492_cast_fp16)[name = tensor("value_117_cast_fp16")]; + tensor var_6495 = const()[name = tensor("op_6495"), val = tensor([1, 20, 64, -1])]; + tensor var_6496_cast_fp16 = reshape(shape = var_6495, x = query_117_cast_fp16)[name = tensor("op_6496_cast_fp16")]; + tensor var_6497_to_fp16 = const()[name = tensor("op_6497_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6498_cast_fp16 = mul(x = var_6496_cast_fp16, y = var_6497_to_fp16)[name = tensor("op_6498_cast_fp16")]; + tensor var_6499 = const()[name = tensor("op_6499"), val = tensor([1, 20, 64, -1])]; + tensor var_6500_cast_fp16 = reshape(shape = var_6499, x = key_117_cast_fp16)[name = tensor("op_6500_cast_fp16")]; + tensor mh_w_175_transpose_x_0 = const()[name = tensor("mh_w_175_transpose_x_0"), val = tensor(true)]; + tensor mh_w_175_transpose_y_0 = const()[name = tensor("mh_w_175_transpose_y_0"), val = tensor(false)]; + tensor mh_w_175_cast_fp16 = matmul(transpose_x = mh_w_175_transpose_x_0, transpose_y = mh_w_175_transpose_y_0, x = var_6498_cast_fp16, y = var_6500_cast_fp16)[name = tensor("mh_w_175_cast_fp16")]; + tensor mh_w_177_cast_fp16 = add(x = mh_w_175_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_177_cast_fp16")]; + tensor var_6508_cast_fp16 = softmax(axis = var_6422, x = mh_w_177_cast_fp16)[name = tensor("op_6508_cast_fp16")]; + tensor var_6509 = const()[name = tensor("op_6509"), val = tensor([1, 20, 64, -1])]; + tensor var_6510_cast_fp16 = reshape(shape = var_6509, x = value_117_cast_fp16)[name = tensor("op_6510_cast_fp16")]; + tensor attn_117_transpose_x_0 = const()[name = tensor("attn_117_transpose_x_0"), val = tensor(false)]; + tensor attn_117_transpose_y_0 = const()[name = tensor("attn_117_transpose_y_0"), val = tensor(true)]; + tensor attn_117_cast_fp16 = matmul(transpose_x = attn_117_transpose_x_0, transpose_y = attn_117_transpose_y_0, x = var_6510_cast_fp16, y = var_6508_cast_fp16)[name = tensor("attn_117_cast_fp16")]; + tensor var_6513 = const()[name = tensor("op_6513"), val = tensor([1, 1280, 1, -1])]; + tensor input_291_cast_fp16 = reshape(shape = var_6513, x = attn_117_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor var_6517 = const()[name = tensor("op_6517"), val = tensor([1, 1])]; + tensor var_6519 = const()[name = tensor("op_6519"), val = tensor([1, 1])]; + tensor obj_413_pad_type_0 = const()[name = tensor("obj_413_pad_type_0"), val = tensor("custom")]; + tensor obj_413_pad_0 = const()[name = tensor("obj_413_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1665512192)))]; + tensor layers_29_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1668789056)))]; + tensor obj_413_cast_fp16 = conv(bias = layers_29_self_attn_o_proj_bias_to_fp16, dilations = var_6519, groups = var_6429, pad = obj_413_pad_0, pad_type = obj_413_pad_type_0, strides = var_6517, weight = layers_29_self_attn_o_proj_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("obj_413_cast_fp16")]; + tensor inputs_177_cast_fp16 = add(x = inputs_175_cast_fp16, y = obj_413_cast_fp16)[name = tensor("inputs_177_cast_fp16")]; + tensor var_6529 = const()[name = tensor("op_6529"), val = tensor([1])]; + tensor channels_mean_177_cast_fp16 = reduce_mean(axes = var_6529, keep_dims = var_6430, x = inputs_177_cast_fp16)[name = tensor("channels_mean_177_cast_fp16")]; + tensor zero_mean_177_cast_fp16 = sub(x = inputs_177_cast_fp16, y = channels_mean_177_cast_fp16)[name = tensor("zero_mean_177_cast_fp16")]; + tensor zero_mean_sq_177_cast_fp16 = mul(x = zero_mean_177_cast_fp16, y = zero_mean_177_cast_fp16)[name = tensor("zero_mean_sq_177_cast_fp16")]; + tensor var_6533 = const()[name = tensor("op_6533"), val = tensor([1])]; + tensor var_6534_cast_fp16 = reduce_mean(axes = var_6533, keep_dims = var_6430, x = zero_mean_sq_177_cast_fp16)[name = tensor("op_6534_cast_fp16")]; + tensor var_6535_to_fp16 = const()[name = tensor("op_6535_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6536_cast_fp16 = add(x = var_6534_cast_fp16, y = var_6535_to_fp16)[name = tensor("op_6536_cast_fp16")]; + tensor denom_177_epsilon_0_to_fp16 = const()[name = tensor("denom_177_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_177_cast_fp16 = rsqrt(epsilon = denom_177_epsilon_0_to_fp16, x = var_6536_cast_fp16)[name = tensor("denom_177_cast_fp16")]; + tensor out_177_cast_fp16 = mul(x = zero_mean_177_cast_fp16, y = denom_177_cast_fp16)[name = tensor("out_177_cast_fp16")]; + tensor obj_415_gamma_0_to_fp16 = const()[name = tensor("obj_415_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1668791680)))]; + tensor obj_415_beta_0_to_fp16 = const()[name = tensor("obj_415_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1668794304)))]; + tensor obj_415_epsilon_0_to_fp16 = const()[name = tensor("obj_415_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_415_cast_fp16 = batch_norm(beta = obj_415_beta_0_to_fp16, epsilon = obj_415_epsilon_0_to_fp16, gamma = obj_415_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_177_cast_fp16)[name = tensor("obj_415_cast_fp16")]; + tensor var_6551 = const()[name = tensor("op_6551"), val = tensor([1, 1])]; + tensor var_6553 = const()[name = tensor("op_6553"), val = tensor([1, 1])]; + tensor query_119_pad_type_0 = const()[name = tensor("query_119_pad_type_0"), val = tensor("custom")]; + tensor query_119_pad_0 = const()[name = tensor("query_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_29_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1668796928)))]; + tensor layers_29_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_29_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1672073792)))]; + tensor query_119_cast_fp16 = conv(bias = layers_29_encoder_attn_q_proj_bias_to_fp16, dilations = var_6553, groups = var_6429, pad = query_119_pad_0, pad_type = query_119_pad_type_0, strides = var_6551, weight = layers_29_encoder_attn_q_proj_weight_to_fp16, x = obj_415_cast_fp16)[name = tensor("query_119_cast_fp16")]; + tensor var_6557 = const()[name = tensor("op_6557"), val = tensor([1, 1])]; + tensor var_6559 = const()[name = tensor("op_6559"), val = tensor([1, 1])]; + tensor key_119_pad_type_0 = const()[name = tensor("key_119_pad_type_0"), val = tensor("custom")]; + tensor key_119_pad_0 = const()[name = tensor("key_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_29_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1672076416)))]; + tensor key_119_cast_fp16 = conv(dilations = var_6559, groups = var_6429, pad = key_119_pad_0, pad_type = key_119_pad_type_0, strides = var_6557, weight = layers_29_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_119_cast_fp16")]; + tensor var_6564 = const()[name = tensor("op_6564"), val = tensor([1, 1])]; + tensor var_6566 = const()[name = tensor("op_6566"), val = tensor([1, 1])]; + tensor value_119_pad_type_0 = const()[name = tensor("value_119_pad_type_0"), val = tensor("custom")]; + tensor value_119_pad_0 = const()[name = tensor("value_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_29_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1675353280)))]; + tensor layers_29_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_29_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1678630144)))]; + tensor value_119_cast_fp16 = conv(bias = layers_29_encoder_attn_v_proj_bias_to_fp16, dilations = var_6566, groups = var_6429, pad = value_119_pad_0, pad_type = value_119_pad_type_0, strides = var_6564, weight = layers_29_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_119_cast_fp16")]; + tensor var_6570 = const()[name = tensor("op_6570"), val = tensor([1, 20, 64, -1])]; + tensor var_6571_cast_fp16 = reshape(shape = var_6570, x = query_119_cast_fp16)[name = tensor("op_6571_cast_fp16")]; + tensor var_6572_to_fp16 = const()[name = tensor("op_6572_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6573_cast_fp16 = mul(x = var_6571_cast_fp16, y = var_6572_to_fp16)[name = tensor("op_6573_cast_fp16")]; + tensor var_6574 = const()[name = tensor("op_6574"), val = tensor([1, 20, 64, -1])]; + tensor var_6575_cast_fp16 = reshape(shape = var_6574, x = key_119_cast_fp16)[name = tensor("op_6575_cast_fp16")]; + tensor mh_w_179_transpose_x_0 = const()[name = tensor("mh_w_179_transpose_x_0"), val = tensor(true)]; + tensor mh_w_179_transpose_y_0 = const()[name = tensor("mh_w_179_transpose_y_0"), val = tensor(false)]; + tensor mh_w_179_cast_fp16 = matmul(transpose_x = mh_w_179_transpose_x_0, transpose_y = mh_w_179_transpose_y_0, x = var_6573_cast_fp16, y = var_6575_cast_fp16)[name = tensor("mh_w_179_cast_fp16")]; + tensor obj_419_cast_fp16 = softmax(axis = var_6422, x = mh_w_179_cast_fp16)[name = tensor("obj_419_cast_fp16")]; + tensor var_6579 = const()[name = tensor("op_6579"), val = tensor([1, 20, 64, -1])]; + tensor var_6580_cast_fp16 = reshape(shape = var_6579, x = value_119_cast_fp16)[name = tensor("op_6580_cast_fp16")]; + tensor attn_119_transpose_x_0 = const()[name = tensor("attn_119_transpose_x_0"), val = tensor(false)]; + tensor attn_119_transpose_y_0 = const()[name = tensor("attn_119_transpose_y_0"), val = tensor(true)]; + tensor attn_119_cast_fp16 = matmul(transpose_x = attn_119_transpose_x_0, transpose_y = attn_119_transpose_y_0, x = var_6580_cast_fp16, y = obj_419_cast_fp16)[name = tensor("attn_119_cast_fp16")]; + tensor var_6583 = const()[name = tensor("op_6583"), val = tensor([1, 1280, 1, -1])]; + tensor input_293_cast_fp16 = reshape(shape = var_6583, x = attn_119_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor var_6587 = const()[name = tensor("op_6587"), val = tensor([1, 1])]; + tensor var_6589 = const()[name = tensor("op_6589"), val = tensor([1, 1])]; + tensor obj_417_pad_type_0 = const()[name = tensor("obj_417_pad_type_0"), val = tensor("custom")]; + tensor obj_417_pad_0 = const()[name = tensor("obj_417_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_29_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1678632768)))]; + tensor layers_29_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_29_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1681909632)))]; + tensor obj_417_cast_fp16 = conv(bias = layers_29_encoder_attn_o_proj_bias_to_fp16, dilations = var_6589, groups = var_6429, pad = obj_417_pad_0, pad_type = obj_417_pad_type_0, strides = var_6587, weight = layers_29_encoder_attn_o_proj_weight_to_fp16, x = input_293_cast_fp16)[name = tensor("obj_417_cast_fp16")]; + tensor inputs_179_cast_fp16 = add(x = inputs_177_cast_fp16, y = obj_417_cast_fp16)[name = tensor("inputs_179_cast_fp16")]; + tensor var_6595 = const()[name = tensor("op_6595"), val = tensor([1])]; + tensor channels_mean_179_cast_fp16 = reduce_mean(axes = var_6595, keep_dims = var_6430, x = inputs_179_cast_fp16)[name = tensor("channels_mean_179_cast_fp16")]; + tensor zero_mean_179_cast_fp16 = sub(x = inputs_179_cast_fp16, y = channels_mean_179_cast_fp16)[name = tensor("zero_mean_179_cast_fp16")]; + tensor zero_mean_sq_179_cast_fp16 = mul(x = zero_mean_179_cast_fp16, y = zero_mean_179_cast_fp16)[name = tensor("zero_mean_sq_179_cast_fp16")]; + tensor var_6599 = const()[name = tensor("op_6599"), val = tensor([1])]; + tensor var_6600_cast_fp16 = reduce_mean(axes = var_6599, keep_dims = var_6430, x = zero_mean_sq_179_cast_fp16)[name = tensor("op_6600_cast_fp16")]; + tensor var_6601_to_fp16 = const()[name = tensor("op_6601_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6602_cast_fp16 = add(x = var_6600_cast_fp16, y = var_6601_to_fp16)[name = tensor("op_6602_cast_fp16")]; + tensor denom_179_epsilon_0_to_fp16 = const()[name = tensor("denom_179_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_179_cast_fp16 = rsqrt(epsilon = denom_179_epsilon_0_to_fp16, x = var_6602_cast_fp16)[name = tensor("denom_179_cast_fp16")]; + tensor out_179_cast_fp16 = mul(x = zero_mean_179_cast_fp16, y = denom_179_cast_fp16)[name = tensor("out_179_cast_fp16")]; + tensor input_295_gamma_0_to_fp16 = const()[name = tensor("input_295_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1681912256)))]; + tensor input_295_beta_0_to_fp16 = const()[name = tensor("input_295_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1681914880)))]; + tensor input_295_epsilon_0_to_fp16 = const()[name = tensor("input_295_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_295_cast_fp16 = batch_norm(beta = input_295_beta_0_to_fp16, epsilon = input_295_epsilon_0_to_fp16, gamma = input_295_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_179_cast_fp16)[name = tensor("input_295_cast_fp16")]; + tensor var_6613 = const()[name = tensor("op_6613"), val = tensor([1, 1])]; + tensor var_6615 = const()[name = tensor("op_6615"), val = tensor([1, 1])]; + tensor input_297_pad_type_0 = const()[name = tensor("input_297_pad_type_0"), val = tensor("custom")]; + tensor input_297_pad_0 = const()[name = tensor("input_297_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_fc1_weight_to_fp16 = const()[name = tensor("layers_29_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1681917504)))]; + tensor layers_29_fc1_bias_to_fp16 = const()[name = tensor("layers_29_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1695024768)))]; + tensor input_297_cast_fp16 = conv(bias = layers_29_fc1_bias_to_fp16, dilations = var_6615, groups = var_6429, pad = input_297_pad_0, pad_type = input_297_pad_type_0, strides = var_6613, weight = layers_29_fc1_weight_to_fp16, x = input_295_cast_fp16)[name = tensor("input_297_cast_fp16")]; + tensor input_299_mode_0 = const()[name = tensor("input_299_mode_0"), val = tensor("EXACT")]; + tensor input_299_cast_fp16 = gelu(mode = input_299_mode_0, x = input_297_cast_fp16)[name = tensor("input_299_cast_fp16")]; + tensor var_6621 = const()[name = tensor("op_6621"), val = tensor([1, 1])]; + tensor var_6623 = const()[name = tensor("op_6623"), val = tensor([1, 1])]; + tensor hidden_states_61_pad_type_0 = const()[name = tensor("hidden_states_61_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_61_pad_0 = const()[name = tensor("hidden_states_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_fc2_weight_to_fp16 = const()[name = tensor("layers_29_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1695035072)))]; + tensor layers_29_fc2_bias_to_fp16 = const()[name = tensor("layers_29_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1708142336)))]; + tensor hidden_states_61_cast_fp16 = conv(bias = layers_29_fc2_bias_to_fp16, dilations = var_6623, groups = var_6429, pad = hidden_states_61_pad_0, pad_type = hidden_states_61_pad_type_0, strides = var_6621, weight = layers_29_fc2_weight_to_fp16, x = input_299_cast_fp16)[name = tensor("hidden_states_61_cast_fp16")]; + tensor inputs_181_cast_fp16 = add(x = inputs_179_cast_fp16, y = hidden_states_61_cast_fp16)[name = tensor("inputs_181_cast_fp16")]; + tensor var_6636 = const()[name = tensor("op_6636"), val = tensor(3)]; + tensor var_6643 = const()[name = tensor("op_6643"), val = tensor(1)]; + tensor var_6644 = const()[name = tensor("op_6644"), val = tensor(true)]; + tensor var_6656 = const()[name = tensor("op_6656"), val = tensor([1])]; + tensor channels_mean_181_cast_fp16 = reduce_mean(axes = var_6656, keep_dims = var_6644, x = inputs_181_cast_fp16)[name = tensor("channels_mean_181_cast_fp16")]; + tensor zero_mean_181_cast_fp16 = sub(x = inputs_181_cast_fp16, y = channels_mean_181_cast_fp16)[name = tensor("zero_mean_181_cast_fp16")]; + tensor zero_mean_sq_181_cast_fp16 = mul(x = zero_mean_181_cast_fp16, y = zero_mean_181_cast_fp16)[name = tensor("zero_mean_sq_181_cast_fp16")]; + tensor var_6660 = const()[name = tensor("op_6660"), val = tensor([1])]; + tensor var_6661_cast_fp16 = reduce_mean(axes = var_6660, keep_dims = var_6644, x = zero_mean_sq_181_cast_fp16)[name = tensor("op_6661_cast_fp16")]; + tensor var_6662_to_fp16 = const()[name = tensor("op_6662_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6663_cast_fp16 = add(x = var_6661_cast_fp16, y = var_6662_to_fp16)[name = tensor("op_6663_cast_fp16")]; + tensor denom_181_epsilon_0_to_fp16 = const()[name = tensor("denom_181_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_181_cast_fp16 = rsqrt(epsilon = denom_181_epsilon_0_to_fp16, x = var_6663_cast_fp16)[name = tensor("denom_181_cast_fp16")]; + tensor out_181_cast_fp16 = mul(x = zero_mean_181_cast_fp16, y = denom_181_cast_fp16)[name = tensor("out_181_cast_fp16")]; + tensor obj_421_gamma_0_to_fp16 = const()[name = tensor("obj_421_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1708144960)))]; + tensor obj_421_beta_0_to_fp16 = const()[name = tensor("obj_421_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1708147584)))]; + tensor obj_421_epsilon_0_to_fp16 = const()[name = tensor("obj_421_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_421_cast_fp16 = batch_norm(beta = obj_421_beta_0_to_fp16, epsilon = obj_421_epsilon_0_to_fp16, gamma = obj_421_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_181_cast_fp16)[name = tensor("obj_421_cast_fp16")]; + tensor var_6678 = const()[name = tensor("op_6678"), val = tensor([1, 1])]; + tensor var_6680 = const()[name = tensor("op_6680"), val = tensor([1, 1])]; + tensor query_121_pad_type_0 = const()[name = tensor("query_121_pad_type_0"), val = tensor("custom")]; + tensor query_121_pad_0 = const()[name = tensor("query_121_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1708150208)))]; + tensor layers_30_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1711427072)))]; + tensor query_121_cast_fp16 = conv(bias = layers_30_self_attn_q_proj_bias_to_fp16, dilations = var_6680, groups = var_6643, pad = query_121_pad_0, pad_type = query_121_pad_type_0, strides = var_6678, weight = layers_30_self_attn_q_proj_weight_to_fp16, x = obj_421_cast_fp16)[name = tensor("query_121_cast_fp16")]; + tensor var_6684 = const()[name = tensor("op_6684"), val = tensor([1, 1])]; + tensor var_6686 = const()[name = tensor("op_6686"), val = tensor([1, 1])]; + tensor current_key_61_pad_type_0 = const()[name = tensor("current_key_61_pad_type_0"), val = tensor("custom")]; + tensor current_key_61_pad_0 = const()[name = tensor("current_key_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1711429696)))]; + tensor current_key_61_cast_fp16 = conv(dilations = var_6686, groups = var_6643, pad = current_key_61_pad_0, pad_type = current_key_61_pad_type_0, strides = var_6684, weight = layers_30_self_attn_k_proj_weight_to_fp16, x = obj_421_cast_fp16)[name = tensor("current_key_61_cast_fp16")]; + tensor var_6691 = const()[name = tensor("op_6691"), val = tensor([1, 1])]; + tensor var_6693 = const()[name = tensor("op_6693"), val = tensor([1, 1])]; + tensor current_value_61_pad_type_0 = const()[name = tensor("current_value_61_pad_type_0"), val = tensor("custom")]; + tensor current_value_61_pad_0 = const()[name = tensor("current_value_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1714706560)))]; + tensor layers_30_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1717983424)))]; + tensor current_value_61_cast_fp16 = conv(bias = layers_30_self_attn_v_proj_bias_to_fp16, dilations = var_6693, groups = var_6643, pad = current_value_61_pad_0, pad_type = current_value_61_pad_type_0, strides = var_6691, weight = layers_30_self_attn_v_proj_weight_to_fp16, x = obj_421_cast_fp16)[name = tensor("current_value_61_cast_fp16")]; + tensor var_6700_cast_fp16 = mul(x = current_key_61_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6700_cast_fp16")]; + tensor var_6702_cast_fp16 = mul(x = var_103_cast_fp16_30, y = var_241_cast_fp16)[name = tensor("op_6702_cast_fp16")]; + tensor key_121_cast_fp16 = add(x = var_6700_cast_fp16, y = var_6702_cast_fp16)[name = tensor("key_121_cast_fp16")]; + tensor var_6704_cast_fp16 = mul(x = current_value_61_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6704_cast_fp16")]; + tensor var_6706_cast_fp16 = mul(x = var_138_cast_fp16_30, y = var_241_cast_fp16)[name = tensor("op_6706_cast_fp16")]; + tensor value_121_cast_fp16 = add(x = var_6704_cast_fp16, y = var_6706_cast_fp16)[name = tensor("value_121_cast_fp16")]; + tensor var_6709 = const()[name = tensor("op_6709"), val = tensor([1, 20, 64, -1])]; + tensor var_6710_cast_fp16 = reshape(shape = var_6709, x = query_121_cast_fp16)[name = tensor("op_6710_cast_fp16")]; + tensor var_6711_to_fp16 = const()[name = tensor("op_6711_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6712_cast_fp16 = mul(x = var_6710_cast_fp16, y = var_6711_to_fp16)[name = tensor("op_6712_cast_fp16")]; + tensor var_6713 = const()[name = tensor("op_6713"), val = tensor([1, 20, 64, -1])]; + tensor var_6714_cast_fp16 = reshape(shape = var_6713, x = key_121_cast_fp16)[name = tensor("op_6714_cast_fp16")]; + tensor mh_w_181_transpose_x_0 = const()[name = tensor("mh_w_181_transpose_x_0"), val = tensor(true)]; + tensor mh_w_181_transpose_y_0 = const()[name = tensor("mh_w_181_transpose_y_0"), val = tensor(false)]; + tensor mh_w_181_cast_fp16 = matmul(transpose_x = mh_w_181_transpose_x_0, transpose_y = mh_w_181_transpose_y_0, x = var_6712_cast_fp16, y = var_6714_cast_fp16)[name = tensor("mh_w_181_cast_fp16")]; + tensor mh_w_183_cast_fp16 = add(x = mh_w_181_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_183_cast_fp16")]; + tensor var_6722_cast_fp16 = softmax(axis = var_6636, x = mh_w_183_cast_fp16)[name = tensor("op_6722_cast_fp16")]; + tensor var_6723 = const()[name = tensor("op_6723"), val = tensor([1, 20, 64, -1])]; + tensor var_6724_cast_fp16 = reshape(shape = var_6723, x = value_121_cast_fp16)[name = tensor("op_6724_cast_fp16")]; + tensor attn_121_transpose_x_0 = const()[name = tensor("attn_121_transpose_x_0"), val = tensor(false)]; + tensor attn_121_transpose_y_0 = const()[name = tensor("attn_121_transpose_y_0"), val = tensor(true)]; + tensor attn_121_cast_fp16 = matmul(transpose_x = attn_121_transpose_x_0, transpose_y = attn_121_transpose_y_0, x = var_6724_cast_fp16, y = var_6722_cast_fp16)[name = tensor("attn_121_cast_fp16")]; + tensor var_6727 = const()[name = tensor("op_6727"), val = tensor([1, 1280, 1, -1])]; + tensor input_301_cast_fp16 = reshape(shape = var_6727, x = attn_121_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor var_6731 = const()[name = tensor("op_6731"), val = tensor([1, 1])]; + tensor var_6733 = const()[name = tensor("op_6733"), val = tensor([1, 1])]; + tensor obj_427_pad_type_0 = const()[name = tensor("obj_427_pad_type_0"), val = tensor("custom")]; + tensor obj_427_pad_0 = const()[name = tensor("obj_427_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1717986048)))]; + tensor layers_30_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1721262912)))]; + tensor obj_427_cast_fp16 = conv(bias = layers_30_self_attn_o_proj_bias_to_fp16, dilations = var_6733, groups = var_6643, pad = obj_427_pad_0, pad_type = obj_427_pad_type_0, strides = var_6731, weight = layers_30_self_attn_o_proj_weight_to_fp16, x = input_301_cast_fp16)[name = tensor("obj_427_cast_fp16")]; + tensor inputs_183_cast_fp16 = add(x = inputs_181_cast_fp16, y = obj_427_cast_fp16)[name = tensor("inputs_183_cast_fp16")]; + tensor var_6743 = const()[name = tensor("op_6743"), val = tensor([1])]; + tensor channels_mean_183_cast_fp16 = reduce_mean(axes = var_6743, keep_dims = var_6644, x = inputs_183_cast_fp16)[name = tensor("channels_mean_183_cast_fp16")]; + tensor zero_mean_183_cast_fp16 = sub(x = inputs_183_cast_fp16, y = channels_mean_183_cast_fp16)[name = tensor("zero_mean_183_cast_fp16")]; + tensor zero_mean_sq_183_cast_fp16 = mul(x = zero_mean_183_cast_fp16, y = zero_mean_183_cast_fp16)[name = tensor("zero_mean_sq_183_cast_fp16")]; + tensor var_6747 = const()[name = tensor("op_6747"), val = tensor([1])]; + tensor var_6748_cast_fp16 = reduce_mean(axes = var_6747, keep_dims = var_6644, x = zero_mean_sq_183_cast_fp16)[name = tensor("op_6748_cast_fp16")]; + tensor var_6749_to_fp16 = const()[name = tensor("op_6749_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6750_cast_fp16 = add(x = var_6748_cast_fp16, y = var_6749_to_fp16)[name = tensor("op_6750_cast_fp16")]; + tensor denom_183_epsilon_0_to_fp16 = const()[name = tensor("denom_183_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_183_cast_fp16 = rsqrt(epsilon = denom_183_epsilon_0_to_fp16, x = var_6750_cast_fp16)[name = tensor("denom_183_cast_fp16")]; + tensor out_183_cast_fp16 = mul(x = zero_mean_183_cast_fp16, y = denom_183_cast_fp16)[name = tensor("out_183_cast_fp16")]; + tensor obj_429_gamma_0_to_fp16 = const()[name = tensor("obj_429_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1721265536)))]; + tensor obj_429_beta_0_to_fp16 = const()[name = tensor("obj_429_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1721268160)))]; + tensor obj_429_epsilon_0_to_fp16 = const()[name = tensor("obj_429_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_429_cast_fp16 = batch_norm(beta = obj_429_beta_0_to_fp16, epsilon = obj_429_epsilon_0_to_fp16, gamma = obj_429_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_183_cast_fp16)[name = tensor("obj_429_cast_fp16")]; + tensor var_6765 = const()[name = tensor("op_6765"), val = tensor([1, 1])]; + tensor var_6767 = const()[name = tensor("op_6767"), val = tensor([1, 1])]; + tensor query_123_pad_type_0 = const()[name = tensor("query_123_pad_type_0"), val = tensor("custom")]; + tensor query_123_pad_0 = const()[name = tensor("query_123_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_30_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1721270784)))]; + tensor layers_30_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_30_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1724547648)))]; + tensor query_123_cast_fp16 = conv(bias = layers_30_encoder_attn_q_proj_bias_to_fp16, dilations = var_6767, groups = var_6643, pad = query_123_pad_0, pad_type = query_123_pad_type_0, strides = var_6765, weight = layers_30_encoder_attn_q_proj_weight_to_fp16, x = obj_429_cast_fp16)[name = tensor("query_123_cast_fp16")]; + tensor var_6771 = const()[name = tensor("op_6771"), val = tensor([1, 1])]; + tensor var_6773 = const()[name = tensor("op_6773"), val = tensor([1, 1])]; + tensor key_123_pad_type_0 = const()[name = tensor("key_123_pad_type_0"), val = tensor("custom")]; + tensor key_123_pad_0 = const()[name = tensor("key_123_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_30_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1724550272)))]; + tensor key_123_cast_fp16 = conv(dilations = var_6773, groups = var_6643, pad = key_123_pad_0, pad_type = key_123_pad_type_0, strides = var_6771, weight = layers_30_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_123_cast_fp16")]; + tensor var_6778 = const()[name = tensor("op_6778"), val = tensor([1, 1])]; + tensor var_6780 = const()[name = tensor("op_6780"), val = tensor([1, 1])]; + tensor value_123_pad_type_0 = const()[name = tensor("value_123_pad_type_0"), val = tensor("custom")]; + tensor value_123_pad_0 = const()[name = tensor("value_123_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_30_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1727827136)))]; + tensor layers_30_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_30_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1731104000)))]; + tensor value_123_cast_fp16 = conv(bias = layers_30_encoder_attn_v_proj_bias_to_fp16, dilations = var_6780, groups = var_6643, pad = value_123_pad_0, pad_type = value_123_pad_type_0, strides = var_6778, weight = layers_30_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_123_cast_fp16")]; + tensor var_6784 = const()[name = tensor("op_6784"), val = tensor([1, 20, 64, -1])]; + tensor var_6785_cast_fp16 = reshape(shape = var_6784, x = query_123_cast_fp16)[name = tensor("op_6785_cast_fp16")]; + tensor var_6786_to_fp16 = const()[name = tensor("op_6786_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6787_cast_fp16 = mul(x = var_6785_cast_fp16, y = var_6786_to_fp16)[name = tensor("op_6787_cast_fp16")]; + tensor var_6788 = const()[name = tensor("op_6788"), val = tensor([1, 20, 64, -1])]; + tensor var_6789_cast_fp16 = reshape(shape = var_6788, x = key_123_cast_fp16)[name = tensor("op_6789_cast_fp16")]; + tensor mh_w_185_transpose_x_0 = const()[name = tensor("mh_w_185_transpose_x_0"), val = tensor(true)]; + tensor mh_w_185_transpose_y_0 = const()[name = tensor("mh_w_185_transpose_y_0"), val = tensor(false)]; + tensor mh_w_185_cast_fp16 = matmul(transpose_x = mh_w_185_transpose_x_0, transpose_y = mh_w_185_transpose_y_0, x = var_6787_cast_fp16, y = var_6789_cast_fp16)[name = tensor("mh_w_185_cast_fp16")]; + tensor obj_433_cast_fp16 = softmax(axis = var_6636, x = mh_w_185_cast_fp16)[name = tensor("obj_433_cast_fp16")]; + tensor var_6793 = const()[name = tensor("op_6793"), val = tensor([1, 20, 64, -1])]; + tensor var_6794_cast_fp16 = reshape(shape = var_6793, x = value_123_cast_fp16)[name = tensor("op_6794_cast_fp16")]; + tensor attn_123_transpose_x_0 = const()[name = tensor("attn_123_transpose_x_0"), val = tensor(false)]; + tensor attn_123_transpose_y_0 = const()[name = tensor("attn_123_transpose_y_0"), val = tensor(true)]; + tensor attn_123_cast_fp16 = matmul(transpose_x = attn_123_transpose_x_0, transpose_y = attn_123_transpose_y_0, x = var_6794_cast_fp16, y = obj_433_cast_fp16)[name = tensor("attn_123_cast_fp16")]; + tensor var_6797 = const()[name = tensor("op_6797"), val = tensor([1, 1280, 1, -1])]; + tensor input_303_cast_fp16 = reshape(shape = var_6797, x = attn_123_cast_fp16)[name = tensor("input_303_cast_fp16")]; + tensor var_6801 = const()[name = tensor("op_6801"), val = tensor([1, 1])]; + tensor var_6803 = const()[name = tensor("op_6803"), val = tensor([1, 1])]; + tensor obj_431_pad_type_0 = const()[name = tensor("obj_431_pad_type_0"), val = tensor("custom")]; + tensor obj_431_pad_0 = const()[name = tensor("obj_431_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_30_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1731106624)))]; + tensor layers_30_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_30_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1734383488)))]; + tensor obj_431_cast_fp16 = conv(bias = layers_30_encoder_attn_o_proj_bias_to_fp16, dilations = var_6803, groups = var_6643, pad = obj_431_pad_0, pad_type = obj_431_pad_type_0, strides = var_6801, weight = layers_30_encoder_attn_o_proj_weight_to_fp16, x = input_303_cast_fp16)[name = tensor("obj_431_cast_fp16")]; + tensor inputs_185_cast_fp16 = add(x = inputs_183_cast_fp16, y = obj_431_cast_fp16)[name = tensor("inputs_185_cast_fp16")]; + tensor var_6809 = const()[name = tensor("op_6809"), val = tensor([1])]; + tensor channels_mean_185_cast_fp16 = reduce_mean(axes = var_6809, keep_dims = var_6644, x = inputs_185_cast_fp16)[name = tensor("channels_mean_185_cast_fp16")]; + tensor zero_mean_185_cast_fp16 = sub(x = inputs_185_cast_fp16, y = channels_mean_185_cast_fp16)[name = tensor("zero_mean_185_cast_fp16")]; + tensor zero_mean_sq_185_cast_fp16 = mul(x = zero_mean_185_cast_fp16, y = zero_mean_185_cast_fp16)[name = tensor("zero_mean_sq_185_cast_fp16")]; + tensor var_6813 = const()[name = tensor("op_6813"), val = tensor([1])]; + tensor var_6814_cast_fp16 = reduce_mean(axes = var_6813, keep_dims = var_6644, x = zero_mean_sq_185_cast_fp16)[name = tensor("op_6814_cast_fp16")]; + tensor var_6815_to_fp16 = const()[name = tensor("op_6815_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6816_cast_fp16 = add(x = var_6814_cast_fp16, y = var_6815_to_fp16)[name = tensor("op_6816_cast_fp16")]; + tensor denom_185_epsilon_0_to_fp16 = const()[name = tensor("denom_185_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_185_cast_fp16 = rsqrt(epsilon = denom_185_epsilon_0_to_fp16, x = var_6816_cast_fp16)[name = tensor("denom_185_cast_fp16")]; + tensor out_185_cast_fp16 = mul(x = zero_mean_185_cast_fp16, y = denom_185_cast_fp16)[name = tensor("out_185_cast_fp16")]; + tensor input_305_gamma_0_to_fp16 = const()[name = tensor("input_305_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1734386112)))]; + tensor input_305_beta_0_to_fp16 = const()[name = tensor("input_305_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1734388736)))]; + tensor input_305_epsilon_0_to_fp16 = const()[name = tensor("input_305_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_305_cast_fp16 = batch_norm(beta = input_305_beta_0_to_fp16, epsilon = input_305_epsilon_0_to_fp16, gamma = input_305_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_185_cast_fp16)[name = tensor("input_305_cast_fp16")]; + tensor var_6827 = const()[name = tensor("op_6827"), val = tensor([1, 1])]; + tensor var_6829 = const()[name = tensor("op_6829"), val = tensor([1, 1])]; + tensor input_307_pad_type_0 = const()[name = tensor("input_307_pad_type_0"), val = tensor("custom")]; + tensor input_307_pad_0 = const()[name = tensor("input_307_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_fc1_weight_to_fp16 = const()[name = tensor("layers_30_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1734391360)))]; + tensor layers_30_fc1_bias_to_fp16 = const()[name = tensor("layers_30_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1747498624)))]; + tensor input_307_cast_fp16 = conv(bias = layers_30_fc1_bias_to_fp16, dilations = var_6829, groups = var_6643, pad = input_307_pad_0, pad_type = input_307_pad_type_0, strides = var_6827, weight = layers_30_fc1_weight_to_fp16, x = input_305_cast_fp16)[name = tensor("input_307_cast_fp16")]; + tensor input_309_mode_0 = const()[name = tensor("input_309_mode_0"), val = tensor("EXACT")]; + tensor input_309_cast_fp16 = gelu(mode = input_309_mode_0, x = input_307_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor var_6835 = const()[name = tensor("op_6835"), val = tensor([1, 1])]; + tensor var_6837 = const()[name = tensor("op_6837"), val = tensor([1, 1])]; + tensor hidden_states_63_pad_type_0 = const()[name = tensor("hidden_states_63_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_63_pad_0 = const()[name = tensor("hidden_states_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_fc2_weight_to_fp16 = const()[name = tensor("layers_30_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1747508928)))]; + tensor layers_30_fc2_bias_to_fp16 = const()[name = tensor("layers_30_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1760616192)))]; + tensor hidden_states_63_cast_fp16 = conv(bias = layers_30_fc2_bias_to_fp16, dilations = var_6837, groups = var_6643, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = var_6835, weight = layers_30_fc2_weight_to_fp16, x = input_309_cast_fp16)[name = tensor("hidden_states_63_cast_fp16")]; + tensor inputs_187_cast_fp16 = add(x = inputs_185_cast_fp16, y = hidden_states_63_cast_fp16)[name = tensor("inputs_187_cast_fp16")]; + tensor var_6850 = const()[name = tensor("op_6850"), val = tensor(3)]; + tensor var_6857 = const()[name = tensor("op_6857"), val = tensor(1)]; + tensor var_6858 = const()[name = tensor("op_6858"), val = tensor(true)]; + tensor var_6870 = const()[name = tensor("op_6870"), val = tensor([1])]; + tensor channels_mean_187_cast_fp16 = reduce_mean(axes = var_6870, keep_dims = var_6858, x = inputs_187_cast_fp16)[name = tensor("channels_mean_187_cast_fp16")]; + tensor zero_mean_187_cast_fp16 = sub(x = inputs_187_cast_fp16, y = channels_mean_187_cast_fp16)[name = tensor("zero_mean_187_cast_fp16")]; + tensor zero_mean_sq_187_cast_fp16 = mul(x = zero_mean_187_cast_fp16, y = zero_mean_187_cast_fp16)[name = tensor("zero_mean_sq_187_cast_fp16")]; + tensor var_6874 = const()[name = tensor("op_6874"), val = tensor([1])]; + tensor var_6875_cast_fp16 = reduce_mean(axes = var_6874, keep_dims = var_6858, x = zero_mean_sq_187_cast_fp16)[name = tensor("op_6875_cast_fp16")]; + tensor var_6876_to_fp16 = const()[name = tensor("op_6876_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6877_cast_fp16 = add(x = var_6875_cast_fp16, y = var_6876_to_fp16)[name = tensor("op_6877_cast_fp16")]; + tensor denom_187_epsilon_0_to_fp16 = const()[name = tensor("denom_187_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_187_cast_fp16 = rsqrt(epsilon = denom_187_epsilon_0_to_fp16, x = var_6877_cast_fp16)[name = tensor("denom_187_cast_fp16")]; + tensor out_187_cast_fp16 = mul(x = zero_mean_187_cast_fp16, y = denom_187_cast_fp16)[name = tensor("out_187_cast_fp16")]; + tensor obj_435_gamma_0_to_fp16 = const()[name = tensor("obj_435_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1760618816)))]; + tensor obj_435_beta_0_to_fp16 = const()[name = tensor("obj_435_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1760621440)))]; + tensor obj_435_epsilon_0_to_fp16 = const()[name = tensor("obj_435_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_435_cast_fp16 = batch_norm(beta = obj_435_beta_0_to_fp16, epsilon = obj_435_epsilon_0_to_fp16, gamma = obj_435_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_187_cast_fp16)[name = tensor("obj_435_cast_fp16")]; + tensor var_6892 = const()[name = tensor("op_6892"), val = tensor([1, 1])]; + tensor var_6894 = const()[name = tensor("op_6894"), val = tensor([1, 1])]; + tensor query_125_pad_type_0 = const()[name = tensor("query_125_pad_type_0"), val = tensor("custom")]; + tensor query_125_pad_0 = const()[name = tensor("query_125_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1760624064)))]; + tensor layers_31_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1763900928)))]; + tensor query_125_cast_fp16 = conv(bias = layers_31_self_attn_q_proj_bias_to_fp16, dilations = var_6894, groups = var_6857, pad = query_125_pad_0, pad_type = query_125_pad_type_0, strides = var_6892, weight = layers_31_self_attn_q_proj_weight_to_fp16, x = obj_435_cast_fp16)[name = tensor("query_125_cast_fp16")]; + tensor var_6898 = const()[name = tensor("op_6898"), val = tensor([1, 1])]; + tensor var_6900 = const()[name = tensor("op_6900"), val = tensor([1, 1])]; + tensor current_key_pad_type_0 = const()[name = tensor("current_key_pad_type_0"), val = tensor("custom")]; + tensor current_key_pad_0 = const()[name = tensor("current_key_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1763903552)))]; + tensor current_key_cast_fp16 = conv(dilations = var_6900, groups = var_6857, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = var_6898, weight = layers_31_self_attn_k_proj_weight_to_fp16, x = obj_435_cast_fp16)[name = tensor("current_key_cast_fp16")]; + tensor var_6905 = const()[name = tensor("op_6905"), val = tensor([1, 1])]; + tensor var_6907 = const()[name = tensor("op_6907"), val = tensor([1, 1])]; + tensor current_value_pad_type_0 = const()[name = tensor("current_value_pad_type_0"), val = tensor("custom")]; + tensor current_value_pad_0 = const()[name = tensor("current_value_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1767180416)))]; + tensor layers_31_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1770457280)))]; + tensor current_value_cast_fp16 = conv(bias = layers_31_self_attn_v_proj_bias_to_fp16, dilations = var_6907, groups = var_6857, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = var_6905, weight = layers_31_self_attn_v_proj_weight_to_fp16, x = obj_435_cast_fp16)[name = tensor("current_value_cast_fp16")]; + tensor var_6914_cast_fp16 = mul(x = current_key_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6914_cast_fp16")]; + tensor var_6916_cast_fp16 = mul(x = var_103_cast_fp16_31, y = var_241_cast_fp16)[name = tensor("op_6916_cast_fp16")]; + tensor key_125_cast_fp16 = add(x = var_6914_cast_fp16, y = var_6916_cast_fp16)[name = tensor("key_125_cast_fp16")]; + tensor var_6918_cast_fp16 = mul(x = current_value_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6918_cast_fp16")]; + tensor var_6920_cast_fp16 = mul(x = var_138_cast_fp16_31, y = var_241_cast_fp16)[name = tensor("op_6920_cast_fp16")]; + tensor value_125_cast_fp16 = add(x = var_6918_cast_fp16, y = var_6920_cast_fp16)[name = tensor("value_125_cast_fp16")]; + tensor var_6923 = const()[name = tensor("op_6923"), val = tensor([1, 20, 64, -1])]; + tensor var_6924_cast_fp16 = reshape(shape = var_6923, x = query_125_cast_fp16)[name = tensor("op_6924_cast_fp16")]; + tensor var_6925_to_fp16 = const()[name = tensor("op_6925_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6926_cast_fp16 = mul(x = var_6924_cast_fp16, y = var_6925_to_fp16)[name = tensor("op_6926_cast_fp16")]; + tensor var_6927 = const()[name = tensor("op_6927"), val = tensor([1, 20, 64, -1])]; + tensor var_6928_cast_fp16 = reshape(shape = var_6927, x = key_125_cast_fp16)[name = tensor("op_6928_cast_fp16")]; + tensor mh_w_187_transpose_x_0 = const()[name = tensor("mh_w_187_transpose_x_0"), val = tensor(true)]; + tensor mh_w_187_transpose_y_0 = const()[name = tensor("mh_w_187_transpose_y_0"), val = tensor(false)]; + tensor mh_w_187_cast_fp16 = matmul(transpose_x = mh_w_187_transpose_x_0, transpose_y = mh_w_187_transpose_y_0, x = var_6926_cast_fp16, y = var_6928_cast_fp16)[name = tensor("mh_w_187_cast_fp16")]; + tensor mh_w_189_cast_fp16 = add(x = mh_w_187_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_189_cast_fp16")]; + tensor var_6936_cast_fp16 = softmax(axis = var_6850, x = mh_w_189_cast_fp16)[name = tensor("op_6936_cast_fp16")]; + tensor var_6937 = const()[name = tensor("op_6937"), val = tensor([1, 20, 64, -1])]; + tensor var_6938_cast_fp16 = reshape(shape = var_6937, x = value_125_cast_fp16)[name = tensor("op_6938_cast_fp16")]; + tensor attn_125_transpose_x_0 = const()[name = tensor("attn_125_transpose_x_0"), val = tensor(false)]; + tensor attn_125_transpose_y_0 = const()[name = tensor("attn_125_transpose_y_0"), val = tensor(true)]; + tensor attn_125_cast_fp16 = matmul(transpose_x = attn_125_transpose_x_0, transpose_y = attn_125_transpose_y_0, x = var_6938_cast_fp16, y = var_6936_cast_fp16)[name = tensor("attn_125_cast_fp16")]; + tensor var_6941 = const()[name = tensor("op_6941"), val = tensor([1, 1280, 1, -1])]; + tensor input_311_cast_fp16 = reshape(shape = var_6941, x = attn_125_cast_fp16)[name = tensor("input_311_cast_fp16")]; + tensor var_6945 = const()[name = tensor("op_6945"), val = tensor([1, 1])]; + tensor var_6947 = const()[name = tensor("op_6947"), val = tensor([1, 1])]; + tensor obj_441_pad_type_0 = const()[name = tensor("obj_441_pad_type_0"), val = tensor("custom")]; + tensor obj_441_pad_0 = const()[name = tensor("obj_441_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1770459904)))]; + tensor layers_31_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1773736768)))]; + tensor obj_441_cast_fp16 = conv(bias = layers_31_self_attn_o_proj_bias_to_fp16, dilations = var_6947, groups = var_6857, pad = obj_441_pad_0, pad_type = obj_441_pad_type_0, strides = var_6945, weight = layers_31_self_attn_o_proj_weight_to_fp16, x = input_311_cast_fp16)[name = tensor("obj_441_cast_fp16")]; + tensor inputs_189_cast_fp16 = add(x = inputs_187_cast_fp16, y = obj_441_cast_fp16)[name = tensor("inputs_189_cast_fp16")]; + tensor var_6957 = const()[name = tensor("op_6957"), val = tensor([1])]; + tensor channels_mean_189_cast_fp16 = reduce_mean(axes = var_6957, keep_dims = var_6858, x = inputs_189_cast_fp16)[name = tensor("channels_mean_189_cast_fp16")]; + tensor zero_mean_189_cast_fp16 = sub(x = inputs_189_cast_fp16, y = channels_mean_189_cast_fp16)[name = tensor("zero_mean_189_cast_fp16")]; + tensor zero_mean_sq_189_cast_fp16 = mul(x = zero_mean_189_cast_fp16, y = zero_mean_189_cast_fp16)[name = tensor("zero_mean_sq_189_cast_fp16")]; + tensor var_6961 = const()[name = tensor("op_6961"), val = tensor([1])]; + tensor var_6962_cast_fp16 = reduce_mean(axes = var_6961, keep_dims = var_6858, x = zero_mean_sq_189_cast_fp16)[name = tensor("op_6962_cast_fp16")]; + tensor var_6963_to_fp16 = const()[name = tensor("op_6963_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6964_cast_fp16 = add(x = var_6962_cast_fp16, y = var_6963_to_fp16)[name = tensor("op_6964_cast_fp16")]; + tensor denom_189_epsilon_0_to_fp16 = const()[name = tensor("denom_189_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_189_cast_fp16 = rsqrt(epsilon = denom_189_epsilon_0_to_fp16, x = var_6964_cast_fp16)[name = tensor("denom_189_cast_fp16")]; + tensor out_189_cast_fp16 = mul(x = zero_mean_189_cast_fp16, y = denom_189_cast_fp16)[name = tensor("out_189_cast_fp16")]; + tensor obj_443_gamma_0_to_fp16 = const()[name = tensor("obj_443_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1773739392)))]; + tensor obj_443_beta_0_to_fp16 = const()[name = tensor("obj_443_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1773742016)))]; + tensor obj_443_epsilon_0_to_fp16 = const()[name = tensor("obj_443_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_443_cast_fp16 = batch_norm(beta = obj_443_beta_0_to_fp16, epsilon = obj_443_epsilon_0_to_fp16, gamma = obj_443_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_189_cast_fp16)[name = tensor("obj_443_cast_fp16")]; + tensor var_6979 = const()[name = tensor("op_6979"), val = tensor([1, 1])]; + tensor var_6981 = const()[name = tensor("op_6981"), val = tensor([1, 1])]; + tensor query_pad_type_0 = const()[name = tensor("query_pad_type_0"), val = tensor("custom")]; + tensor query_pad_0 = const()[name = tensor("query_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_31_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1773744640)))]; + tensor layers_31_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_31_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1777021504)))]; + tensor query_cast_fp16 = conv(bias = layers_31_encoder_attn_q_proj_bias_to_fp16, dilations = var_6981, groups = var_6857, pad = query_pad_0, pad_type = query_pad_type_0, strides = var_6979, weight = layers_31_encoder_attn_q_proj_weight_to_fp16, x = obj_443_cast_fp16)[name = tensor("query_cast_fp16")]; + tensor var_6985 = const()[name = tensor("op_6985"), val = tensor([1, 1])]; + tensor var_6987 = const()[name = tensor("op_6987"), val = tensor([1, 1])]; + tensor key_pad_type_0 = const()[name = tensor("key_pad_type_0"), val = tensor("custom")]; + tensor key_pad_0 = const()[name = tensor("key_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_31_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1777024128)))]; + tensor key_cast_fp16 = conv(dilations = var_6987, groups = var_6857, pad = key_pad_0, pad_type = key_pad_type_0, strides = var_6985, weight = layers_31_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_cast_fp16")]; + tensor var_6992 = const()[name = tensor("op_6992"), val = tensor([1, 1])]; + tensor var_6994 = const()[name = tensor("op_6994"), val = tensor([1, 1])]; + tensor value_pad_type_0 = const()[name = tensor("value_pad_type_0"), val = tensor("custom")]; + tensor value_pad_0 = const()[name = tensor("value_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_31_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1780300992)))]; + tensor layers_31_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_31_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1783577856)))]; + tensor value_cast_fp16 = conv(bias = layers_31_encoder_attn_v_proj_bias_to_fp16, dilations = var_6994, groups = var_6857, pad = value_pad_0, pad_type = value_pad_type_0, strides = var_6992, weight = layers_31_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_cast_fp16")]; + tensor var_6998 = const()[name = tensor("op_6998"), val = tensor([1, 20, 64, -1])]; + tensor var_6999_cast_fp16 = reshape(shape = var_6998, x = query_cast_fp16)[name = tensor("op_6999_cast_fp16")]; + tensor var_7000_to_fp16 = const()[name = tensor("op_7000_to_fp16"), val = tensor(0x1p-3)]; + tensor var_7001_cast_fp16 = mul(x = var_6999_cast_fp16, y = var_7000_to_fp16)[name = tensor("op_7001_cast_fp16")]; + tensor var_7002 = const()[name = tensor("op_7002"), val = tensor([1, 20, 64, -1])]; + tensor var_7003_cast_fp16 = reshape(shape = var_7002, x = key_cast_fp16)[name = tensor("op_7003_cast_fp16")]; + tensor mh_w_transpose_x_0 = const()[name = tensor("mh_w_transpose_x_0"), val = tensor(true)]; + tensor mh_w_transpose_y_0 = const()[name = tensor("mh_w_transpose_y_0"), val = tensor(false)]; + tensor mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_7001_cast_fp16, y = var_7003_cast_fp16)[name = tensor("mh_w_cast_fp16")]; + tensor obj_447_cast_fp16 = softmax(axis = var_6850, x = mh_w_cast_fp16)[name = tensor("obj_447_cast_fp16")]; + tensor var_7007 = const()[name = tensor("op_7007"), val = tensor([1, 20, 64, -1])]; + tensor var_7008_cast_fp16 = reshape(shape = var_7007, x = value_cast_fp16)[name = tensor("op_7008_cast_fp16")]; + tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; + tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(true)]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_7008_cast_fp16, y = obj_447_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_7011 = const()[name = tensor("op_7011"), val = tensor([1, 1280, 1, -1])]; + tensor input_313_cast_fp16 = reshape(shape = var_7011, x = attn_cast_fp16)[name = tensor("input_313_cast_fp16")]; + tensor var_7015 = const()[name = tensor("op_7015"), val = tensor([1, 1])]; + tensor var_7017 = const()[name = tensor("op_7017"), val = tensor([1, 1])]; + tensor obj_445_pad_type_0 = const()[name = tensor("obj_445_pad_type_0"), val = tensor("custom")]; + tensor obj_445_pad_0 = const()[name = tensor("obj_445_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_31_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1783580480)))]; + tensor layers_31_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_31_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1786857344)))]; + tensor obj_445_cast_fp16 = conv(bias = layers_31_encoder_attn_o_proj_bias_to_fp16, dilations = var_7017, groups = var_6857, pad = obj_445_pad_0, pad_type = obj_445_pad_type_0, strides = var_7015, weight = layers_31_encoder_attn_o_proj_weight_to_fp16, x = input_313_cast_fp16)[name = tensor("obj_445_cast_fp16")]; + tensor inputs_191_cast_fp16 = add(x = inputs_189_cast_fp16, y = obj_445_cast_fp16)[name = tensor("inputs_191_cast_fp16")]; + tensor var_7023 = const()[name = tensor("op_7023"), val = tensor([1])]; + tensor channels_mean_191_cast_fp16 = reduce_mean(axes = var_7023, keep_dims = var_6858, x = inputs_191_cast_fp16)[name = tensor("channels_mean_191_cast_fp16")]; + tensor zero_mean_191_cast_fp16 = sub(x = inputs_191_cast_fp16, y = channels_mean_191_cast_fp16)[name = tensor("zero_mean_191_cast_fp16")]; + tensor zero_mean_sq_191_cast_fp16 = mul(x = zero_mean_191_cast_fp16, y = zero_mean_191_cast_fp16)[name = tensor("zero_mean_sq_191_cast_fp16")]; + tensor var_7027 = const()[name = tensor("op_7027"), val = tensor([1])]; + tensor var_7028_cast_fp16 = reduce_mean(axes = var_7027, keep_dims = var_6858, x = zero_mean_sq_191_cast_fp16)[name = tensor("op_7028_cast_fp16")]; + tensor var_7029_to_fp16 = const()[name = tensor("op_7029_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_7030_cast_fp16 = add(x = var_7028_cast_fp16, y = var_7029_to_fp16)[name = tensor("op_7030_cast_fp16")]; + tensor denom_191_epsilon_0_to_fp16 = const()[name = tensor("denom_191_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_191_cast_fp16 = rsqrt(epsilon = denom_191_epsilon_0_to_fp16, x = var_7030_cast_fp16)[name = tensor("denom_191_cast_fp16")]; + tensor out_191_cast_fp16 = mul(x = zero_mean_191_cast_fp16, y = denom_191_cast_fp16)[name = tensor("out_191_cast_fp16")]; + tensor input_315_gamma_0_to_fp16 = const()[name = tensor("input_315_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1786859968)))]; + tensor input_315_beta_0_to_fp16 = const()[name = tensor("input_315_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1786862592)))]; + tensor input_315_epsilon_0_to_fp16 = const()[name = tensor("input_315_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_315_cast_fp16 = batch_norm(beta = input_315_beta_0_to_fp16, epsilon = input_315_epsilon_0_to_fp16, gamma = input_315_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_191_cast_fp16)[name = tensor("input_315_cast_fp16")]; + tensor var_7041 = const()[name = tensor("op_7041"), val = tensor([1, 1])]; + tensor var_7043 = const()[name = tensor("op_7043"), val = tensor([1, 1])]; + tensor input_317_pad_type_0 = const()[name = tensor("input_317_pad_type_0"), val = tensor("custom")]; + tensor input_317_pad_0 = const()[name = tensor("input_317_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_fc1_weight_to_fp16 = const()[name = tensor("layers_31_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1786865216)))]; + tensor layers_31_fc1_bias_to_fp16 = const()[name = tensor("layers_31_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1799972480)))]; + tensor input_317_cast_fp16 = conv(bias = layers_31_fc1_bias_to_fp16, dilations = var_7043, groups = var_6857, pad = input_317_pad_0, pad_type = input_317_pad_type_0, strides = var_7041, weight = layers_31_fc1_weight_to_fp16, x = input_315_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_317_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_7049 = const()[name = tensor("op_7049"), val = tensor([1, 1])]; + tensor var_7051 = const()[name = tensor("op_7051"), val = tensor([1, 1])]; + tensor hidden_states_65_pad_type_0 = const()[name = tensor("hidden_states_65_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_65_pad_0 = const()[name = tensor("hidden_states_65_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_fc2_weight_to_fp16 = const()[name = tensor("layers_31_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1799982784)))]; + tensor layers_31_fc2_bias_to_fp16 = const()[name = tensor("layers_31_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1813090048)))]; + tensor hidden_states_65_cast_fp16 = conv(bias = layers_31_fc2_bias_to_fp16, dilations = var_7051, groups = var_6857, pad = hidden_states_65_pad_0, pad_type = hidden_states_65_pad_type_0, strides = var_7049, weight = layers_31_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor("hidden_states_65_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_191_cast_fp16, y = hidden_states_65_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor var_7061 = const()[name = tensor("op_7061"), val = tensor(true)]; + tensor var_7065 = const()[name = tensor("op_7065"), val = tensor([1])]; + tensor channels_mean_cast_fp16 = reduce_mean(axes = var_7065, keep_dims = var_7061, x = inputs_cast_fp16)[name = tensor("channels_mean_cast_fp16")]; + tensor zero_mean_cast_fp16 = sub(x = inputs_cast_fp16, y = channels_mean_cast_fp16)[name = tensor("zero_mean_cast_fp16")]; + tensor zero_mean_sq_cast_fp16 = mul(x = zero_mean_cast_fp16, y = zero_mean_cast_fp16)[name = tensor("zero_mean_sq_cast_fp16")]; + tensor var_7069 = const()[name = tensor("op_7069"), val = tensor([1])]; + tensor var_7070_cast_fp16 = reduce_mean(axes = var_7069, keep_dims = var_7061, x = zero_mean_sq_cast_fp16)[name = tensor("op_7070_cast_fp16")]; + tensor var_7071_to_fp16 = const()[name = tensor("op_7071_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_7072_cast_fp16 = add(x = var_7070_cast_fp16, y = var_7071_to_fp16)[name = tensor("op_7072_cast_fp16")]; + tensor denom_epsilon_0_to_fp16 = const()[name = tensor("denom_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_cast_fp16 = rsqrt(epsilon = denom_epsilon_0_to_fp16, x = var_7072_cast_fp16)[name = tensor("denom_cast_fp16")]; + tensor out_cast_fp16 = mul(x = zero_mean_cast_fp16, y = denom_cast_fp16)[name = tensor("out_cast_fp16")]; + tensor hidden_states_gamma_0_to_fp16 = const()[name = tensor("hidden_states_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1813092672)))]; + tensor hidden_states_beta_0_to_fp16 = const()[name = tensor("hidden_states_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1813095296)))]; + tensor hidden_states_epsilon_0_to_fp16 = const()[name = tensor("hidden_states_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor("hidden_states_cast_fp16")]; + tensor var_7082_axes_0 = const()[name = tensor("op_7082_axes_0"), val = tensor([2])]; + tensor var_7082_cast_fp16 = squeeze(axes = var_7082_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_7082_cast_fp16")]; + tensor var_7085_perm_0 = const()[name = tensor("op_7085_perm_0"), val = tensor([0, 2, 1])]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1813097920)))]; + tensor transpose_0 = transpose(perm = var_7085_perm_0, x = var_7082_cast_fp16)[name = tensor("transpose_0")]; + tensor logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = transpose_0)[name = tensor("linear_0_cast_fp16")]; + tensor var_7089 = const()[name = tensor("op_7089"), val = tensor(1)]; + tensor obj_451_interleave_0 = const()[name = tensor("obj_451_interleave_0"), val = tensor(false)]; + tensor key_cache_updates = concat(axis = var_7089, interleave = obj_451_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_7_cast_fp16, current_key_9_cast_fp16, current_key_11_cast_fp16, current_key_13_cast_fp16, current_key_15_cast_fp16, current_key_17_cast_fp16, current_key_19_cast_fp16, current_key_21_cast_fp16, current_key_23_cast_fp16, current_key_25_cast_fp16, current_key_27_cast_fp16, current_key_29_cast_fp16, current_key_31_cast_fp16, current_key_33_cast_fp16, current_key_35_cast_fp16, current_key_37_cast_fp16, current_key_39_cast_fp16, current_key_41_cast_fp16, current_key_43_cast_fp16, current_key_45_cast_fp16, current_key_47_cast_fp16, current_key_49_cast_fp16, current_key_51_cast_fp16, current_key_53_cast_fp16, current_key_55_cast_fp16, current_key_57_cast_fp16, current_key_59_cast_fp16, current_key_61_cast_fp16, current_key_cast_fp16))[name = tensor("obj_451_cast_fp16")]; + tensor var_7092 = const()[name = tensor("op_7092"), val = tensor(1)]; + tensor obj_453_interleave_0 = const()[name = tensor("obj_453_interleave_0"), val = tensor(false)]; + tensor value_cache_updates = concat(axis = var_7092, interleave = obj_453_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_7_cast_fp16, current_value_9_cast_fp16, current_value_11_cast_fp16, current_value_13_cast_fp16, current_value_15_cast_fp16, current_value_17_cast_fp16, current_value_19_cast_fp16, current_value_21_cast_fp16, current_value_23_cast_fp16, current_value_25_cast_fp16, current_value_27_cast_fp16, current_value_29_cast_fp16, current_value_31_cast_fp16, current_value_33_cast_fp16, current_value_35_cast_fp16, current_value_37_cast_fp16, current_value_39_cast_fp16, current_value_41_cast_fp16, current_value_43_cast_fp16, current_value_45_cast_fp16, current_value_47_cast_fp16, current_value_49_cast_fp16, current_value_51_cast_fp16, current_value_53_cast_fp16, current_value_55_cast_fp16, current_value_57_cast_fp16, current_value_59_cast_fp16, current_value_61_cast_fp16, current_value_cast_fp16))[name = tensor("obj_453_cast_fp16")]; + tensor var_7103_begin_0 = const()[name = tensor("op_7103_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7103_end_0 = const()[name = tensor("op_7103_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7103_end_mask_0 = const()[name = tensor("op_7103_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7103_cast_fp16 = slice_by_index(begin = var_7103_begin_0, end = var_7103_end_0, end_mask = var_7103_end_mask_0, x = obj_111_cast_fp16)[name = tensor("op_7103_cast_fp16")]; + tensor var_7106_begin_0 = const()[name = tensor("op_7106_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7106_end_0 = const()[name = tensor("op_7106_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7106_end_mask_0 = const()[name = tensor("op_7106_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7106_squeeze_mask_0 = const()[name = tensor("op_7106_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7106_cast_fp16 = slice_by_index(begin = var_7106_begin_0, end = var_7106_end_0, end_mask = var_7106_end_mask_0, squeeze_mask = var_7106_squeeze_mask_0, x = var_7103_cast_fp16)[name = tensor("op_7106_cast_fp16")]; + tensor var_7121_begin_0 = const()[name = tensor("op_7121_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_7121_end_0 = const()[name = tensor("op_7121_end_0"), val = tensor([1, 18, 1, 1500])]; + tensor var_7121_end_mask_0 = const()[name = tensor("op_7121_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7121_cast_fp16 = slice_by_index(begin = var_7121_begin_0, end = var_7121_end_0, end_mask = var_7121_end_mask_0, x = obj_153_cast_fp16)[name = tensor("op_7121_cast_fp16")]; + tensor var_7124_begin_0 = const()[name = tensor("op_7124_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7124_end_0 = const()[name = tensor("op_7124_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7124_end_mask_0 = const()[name = tensor("op_7124_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7124_squeeze_mask_0 = const()[name = tensor("op_7124_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7124_cast_fp16 = slice_by_index(begin = var_7124_begin_0, end = var_7124_end_0, end_mask = var_7124_end_mask_0, squeeze_mask = var_7124_squeeze_mask_0, x = var_7121_cast_fp16)[name = tensor("op_7124_cast_fp16")]; + tensor var_7139_begin_0 = const()[name = tensor("op_7139_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_7139_end_0 = const()[name = tensor("op_7139_end_0"), val = tensor([1, 19, 1, 1500])]; + tensor var_7139_end_mask_0 = const()[name = tensor("op_7139_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7139_cast_fp16 = slice_by_index(begin = var_7139_begin_0, end = var_7139_end_0, end_mask = var_7139_end_mask_0, x = obj_181_cast_fp16)[name = tensor("op_7139_cast_fp16")]; + tensor var_7142_begin_0 = const()[name = tensor("op_7142_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7142_end_0 = const()[name = tensor("op_7142_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7142_end_mask_0 = const()[name = tensor("op_7142_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7142_squeeze_mask_0 = const()[name = tensor("op_7142_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7142_cast_fp16 = slice_by_index(begin = var_7142_begin_0, end = var_7142_end_0, end_mask = var_7142_end_mask_0, squeeze_mask = var_7142_squeeze_mask_0, x = var_7139_cast_fp16)[name = tensor("op_7142_cast_fp16")]; + tensor var_7157_begin_0 = const()[name = tensor("op_7157_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_7157_end_0 = const()[name = tensor("op_7157_end_0"), val = tensor([1, 13, 1, 1500])]; + tensor var_7157_end_mask_0 = const()[name = tensor("op_7157_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7157_cast_fp16 = slice_by_index(begin = var_7157_begin_0, end = var_7157_end_0, end_mask = var_7157_end_mask_0, x = obj_195_cast_fp16)[name = tensor("op_7157_cast_fp16")]; + tensor var_7160_begin_0 = const()[name = tensor("op_7160_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7160_end_0 = const()[name = tensor("op_7160_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7160_end_mask_0 = const()[name = tensor("op_7160_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7160_squeeze_mask_0 = const()[name = tensor("op_7160_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7160_cast_fp16 = slice_by_index(begin = var_7160_begin_0, end = var_7160_end_0, end_mask = var_7160_end_mask_0, squeeze_mask = var_7160_squeeze_mask_0, x = var_7157_cast_fp16)[name = tensor("op_7160_cast_fp16")]; + tensor var_7175_begin_0 = const()[name = tensor("op_7175_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_7175_end_0 = const()[name = tensor("op_7175_end_0"), val = tensor([1, 2, 1, 1500])]; + tensor var_7175_end_mask_0 = const()[name = tensor("op_7175_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7175_cast_fp16 = slice_by_index(begin = var_7175_begin_0, end = var_7175_end_0, end_mask = var_7175_end_mask_0, x = obj_237_cast_fp16)[name = tensor("op_7175_cast_fp16")]; + tensor var_7178_begin_0 = const()[name = tensor("op_7178_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7178_end_0 = const()[name = tensor("op_7178_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7178_end_mask_0 = const()[name = tensor("op_7178_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7178_squeeze_mask_0 = const()[name = tensor("op_7178_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7178_cast_fp16 = slice_by_index(begin = var_7178_begin_0, end = var_7178_end_0, end_mask = var_7178_end_mask_0, squeeze_mask = var_7178_squeeze_mask_0, x = var_7175_cast_fp16)[name = tensor("op_7178_cast_fp16")]; + tensor var_7193_begin_0 = const()[name = tensor("op_7193_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_7193_end_0 = const()[name = tensor("op_7193_end_0"), val = tensor([1, 15, 1, 1500])]; + tensor var_7193_end_mask_0 = const()[name = tensor("op_7193_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7193_cast_fp16 = slice_by_index(begin = var_7193_begin_0, end = var_7193_end_0, end_mask = var_7193_end_mask_0, x = obj_251_cast_fp16)[name = tensor("op_7193_cast_fp16")]; + tensor var_7196_begin_0 = const()[name = tensor("op_7196_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7196_end_0 = const()[name = tensor("op_7196_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7196_end_mask_0 = const()[name = tensor("op_7196_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7196_squeeze_mask_0 = const()[name = tensor("op_7196_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7196_cast_fp16 = slice_by_index(begin = var_7196_begin_0, end = var_7196_end_0, end_mask = var_7196_end_mask_0, squeeze_mask = var_7196_squeeze_mask_0, x = var_7193_cast_fp16)[name = tensor("op_7196_cast_fp16")]; + tensor var_7211_begin_0 = const()[name = tensor("op_7211_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_7211_end_0 = const()[name = tensor("op_7211_end_0"), val = tensor([1, 12, 1, 1500])]; + tensor var_7211_end_mask_0 = const()[name = tensor("op_7211_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7211_cast_fp16 = slice_by_index(begin = var_7211_begin_0, end = var_7211_end_0, end_mask = var_7211_end_mask_0, x = obj_279_cast_fp16)[name = tensor("op_7211_cast_fp16")]; + tensor var_7214_begin_0 = const()[name = tensor("op_7214_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7214_end_0 = const()[name = tensor("op_7214_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7214_end_mask_0 = const()[name = tensor("op_7214_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7214_squeeze_mask_0 = const()[name = tensor("op_7214_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7214_cast_fp16 = slice_by_index(begin = var_7214_begin_0, end = var_7214_end_0, end_mask = var_7214_end_mask_0, squeeze_mask = var_7214_squeeze_mask_0, x = var_7211_cast_fp16)[name = tensor("op_7214_cast_fp16")]; + tensor var_7229_begin_0 = const()[name = tensor("op_7229_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_7229_end_0 = const()[name = tensor("op_7229_end_0"), val = tensor([1, 5, 1, 1500])]; + tensor var_7229_end_mask_0 = const()[name = tensor("op_7229_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7229_cast_fp16 = slice_by_index(begin = var_7229_begin_0, end = var_7229_end_0, end_mask = var_7229_end_mask_0, x = obj_307_cast_fp16)[name = tensor("op_7229_cast_fp16")]; + tensor var_7232_begin_0 = const()[name = tensor("op_7232_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7232_end_0 = const()[name = tensor("op_7232_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7232_end_mask_0 = const()[name = tensor("op_7232_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7232_squeeze_mask_0 = const()[name = tensor("op_7232_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7232_cast_fp16 = slice_by_index(begin = var_7232_begin_0, end = var_7232_end_0, end_mask = var_7232_end_mask_0, squeeze_mask = var_7232_squeeze_mask_0, x = var_7229_cast_fp16)[name = tensor("op_7232_cast_fp16")]; + tensor var_7247_begin_0 = const()[name = tensor("op_7247_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_7247_end_0 = const()[name = tensor("op_7247_end_0"), val = tensor([1, 2, 1, 1500])]; + tensor var_7247_end_mask_0 = const()[name = tensor("op_7247_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7247_cast_fp16 = slice_by_index(begin = var_7247_begin_0, end = var_7247_end_0, end_mask = var_7247_end_mask_0, x = obj_349_cast_fp16)[name = tensor("op_7247_cast_fp16")]; + tensor var_7250_begin_0 = const()[name = tensor("op_7250_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7250_end_0 = const()[name = tensor("op_7250_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7250_end_mask_0 = const()[name = tensor("op_7250_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7250_squeeze_mask_0 = const()[name = tensor("op_7250_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7250_cast_fp16 = slice_by_index(begin = var_7250_begin_0, end = var_7250_end_0, end_mask = var_7250_end_mask_0, squeeze_mask = var_7250_squeeze_mask_0, x = var_7247_cast_fp16)[name = tensor("op_7250_cast_fp16")]; + tensor var_7265_begin_0 = const()[name = tensor("op_7265_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_7265_end_0 = const()[name = tensor("op_7265_end_0"), val = tensor([1, 7, 1, 1500])]; + tensor var_7265_end_mask_0 = const()[name = tensor("op_7265_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7265_cast_fp16 = slice_by_index(begin = var_7265_begin_0, end = var_7265_end_0, end_mask = var_7265_end_mask_0, x = obj_363_cast_fp16)[name = tensor("op_7265_cast_fp16")]; + tensor var_7268_begin_0 = const()[name = tensor("op_7268_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7268_end_0 = const()[name = tensor("op_7268_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7268_end_mask_0 = const()[name = tensor("op_7268_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7268_squeeze_mask_0 = const()[name = tensor("op_7268_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7268_cast_fp16 = slice_by_index(begin = var_7268_begin_0, end = var_7268_end_0, end_mask = var_7268_end_mask_0, squeeze_mask = var_7268_squeeze_mask_0, x = var_7265_cast_fp16)[name = tensor("op_7268_cast_fp16")]; + tensor var_7275 = const()[name = tensor("op_7275"), val = tensor(1)]; + tensor var_7276_interleave_0 = const()[name = tensor("op_7276_interleave_0"), val = tensor(false)]; + tensor var_7276_cast_fp16 = concat(axis = var_7275, interleave = var_7276_interleave_0, values = (var_7106_cast_fp16, var_7124_cast_fp16, var_7142_cast_fp16, var_7160_cast_fp16, var_7178_cast_fp16, var_7196_cast_fp16, var_7214_cast_fp16, var_7232_cast_fp16, var_7250_cast_fp16, var_7268_cast_fp16))[name = tensor("op_7276_cast_fp16")]; + tensor var_7278 = const()[name = tensor("op_7278"), val = tensor([1])]; + tensor var_7279 = const()[name = tensor("op_7279"), val = tensor(false)]; + tensor alignment_heads_weights = reduce_mean(axes = var_7278, keep_dims = var_7279, x = var_7276_cast_fp16)[name = tensor("obj_cast_fp16")]; + } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); +} \ No newline at end of file