diff --git "a/openai_whisper-large-v2_turbo_1022MB/TextDecoder.mlmodelc/model.mil" "b/openai_whisper-large-v2_turbo_1022MB/TextDecoder.mlmodelc/model.mil" deleted file mode 100644--- "a/openai_whisper-large-v2_turbo_1022MB/TextDecoder.mlmodelc/model.mil" +++ /dev/null @@ -1,9414 +0,0 @@ -program(1.0) -[buildInfo = dict, tensor>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}})] -{ - 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 var_80_validate_indices_0 = const()[name = tensor("op_80_validate_indices_0"), val = tensor(false)]; - 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, validate_indices = var_80_validate_indices_0, 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 var_84_validate_indices_0 = const()[name = tensor("op_84_validate_indices_0"), val = tensor(false)]; - 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(132774528)))]; - tensor cache_length_to_int16_dtype_0 = const()[name = tensor("cache_length_to_int16_dtype_0"), val = tensor("int16")]; - tensor cast_0 = cast(dtype = cache_length_to_int16_dtype_0, x = cache_length)[name = tensor("cast_0")]; - tensor var_84_cast_fp16_cast_int16 = gather(axis = var_84_axis_0, batch_dims = var_84_batch_dims_0, indices = cast_0, validate_indices = var_84_validate_indices_0, x = embed_positions_weight_to_fp16)[name = tensor("op_84_cast_fp16_cast_int16")]; - tensor hidden_states_1_cast_fp16 = add(x = var_80_cast_fp16, y = var_84_cast_fp16_cast_int16)[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 = const()[name = tensor("denom_1_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_1_cast_fp16 = rsqrt(epsilon = denom_1_epsilon_0, 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(133921472)))]; - 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(133924096)))]; - 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(133926720)))]; - 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(133929344)))]; - 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_221 = const()[name = tensor("op_221"), val = tensor([1, 1])]; - tensor var_223 = const()[name = tensor("op_223"), val = tensor([1, 1])]; - tensor pretrained_out_1_pad_type_0 = const()[name = tensor("pretrained_out_1_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_1_pad_0 = const()[name = tensor("pretrained_out_1_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133931968))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134751232))), name = tensor("layers_0_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_0_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134751360)))]; - tensor pretrained_out_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_223, groups = var_183, pad = pretrained_out_1_pad_0, pad_type = pretrained_out_1_pad_type_0, strides = var_221, weight = layers_0_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor("pretrained_out_1_cast_fp16")]; - tensor var_227 = const()[name = tensor("op_227"), val = tensor([1, 1])]; - tensor var_229 = const()[name = tensor("op_229"), val = tensor([1, 1])]; - tensor input_1_pad_type_0 = const()[name = tensor("input_1_pad_type_0"), val = tensor("custom")]; - tensor input_1_pad_0 = const()[name = tensor("input_1_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134753984)))]; - tensor input_1_cast_fp16 = conv(dilations = var_229, groups = var_183, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = var_227, weight = layers_0_self_attn_q_proj_loraA_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("input_1_cast_fp16")]; - tensor var_233 = const()[name = tensor("op_233"), val = tensor([1, 1])]; - tensor var_235 = const()[name = tensor("op_235"), val = tensor([1, 1])]; - tensor lora_out_1_pad_type_0 = const()[name = tensor("lora_out_1_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1_pad_0 = const()[name = tensor("lora_out_1_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_3_weight_0_to_fp16 = const()[name = tensor("lora_out_3_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134795008)))]; - tensor lora_out_3_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_235, groups = var_183, pad = lora_out_1_pad_0, pad_type = lora_out_1_pad_type_0, strides = var_233, weight = lora_out_3_weight_0_to_fp16, x = input_1_cast_fp16)[name = tensor("lora_out_3_cast_fp16")]; - tensor query_1_cast_fp16 = add(x = pretrained_out_1_cast_fp16, y = lora_out_3_cast_fp16)[name = tensor("query_1_cast_fp16")]; - tensor var_245 = const()[name = tensor("op_245"), val = tensor([1, 1])]; - tensor var_247 = const()[name = tensor("op_247"), val = tensor([1, 1])]; - tensor pretrained_out_3_pad_type_0 = const()[name = tensor("pretrained_out_3_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_3_pad_0 = const()[name = tensor("pretrained_out_3_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134836032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135655296))), name = tensor("layers_0_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_3_cast_fp16 = conv(dilations = var_247, groups = var_183, pad = pretrained_out_3_pad_0, pad_type = pretrained_out_3_pad_type_0, strides = var_245, weight = layers_0_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor("pretrained_out_3_cast_fp16")]; - tensor var_251 = const()[name = tensor("op_251"), val = tensor([1, 1])]; - tensor var_253 = const()[name = tensor("op_253"), val = tensor([1, 1])]; - tensor input_3_pad_type_0 = const()[name = tensor("input_3_pad_type_0"), val = tensor("custom")]; - tensor input_3_pad_0 = const()[name = tensor("input_3_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135655424)))]; - tensor input_3_cast_fp16 = conv(dilations = var_253, groups = var_183, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = var_251, weight = layers_0_self_attn_k_proj_loraA_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("input_3_cast_fp16")]; - tensor var_257 = const()[name = tensor("op_257"), val = tensor([1, 1])]; - tensor var_259 = const()[name = tensor("op_259"), val = tensor([1, 1])]; - tensor lora_out_5_pad_type_0 = const()[name = tensor("lora_out_5_pad_type_0"), val = tensor("custom")]; - tensor lora_out_5_pad_0 = const()[name = tensor("lora_out_5_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_7_weight_0_to_fp16 = const()[name = tensor("lora_out_7_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135696448)))]; - tensor lora_out_7_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_259, groups = var_183, pad = lora_out_5_pad_0, pad_type = lora_out_5_pad_type_0, strides = var_257, weight = lora_out_7_weight_0_to_fp16, x = input_3_cast_fp16)[name = tensor("lora_out_7_cast_fp16")]; - tensor current_key_1_cast_fp16 = add(x = pretrained_out_3_cast_fp16, y = lora_out_7_cast_fp16)[name = tensor("current_key_1_cast_fp16")]; - tensor var_270 = const()[name = tensor("op_270"), val = tensor([1, 1])]; - tensor var_272 = const()[name = tensor("op_272"), val = tensor([1, 1])]; - tensor pretrained_out_5_pad_type_0 = const()[name = tensor("pretrained_out_5_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_5_pad_0 = const()[name = tensor("pretrained_out_5_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135737472))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136556736))), name = tensor("layers_0_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_0_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136556864)))]; - tensor pretrained_out_5_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_272, groups = var_183, pad = pretrained_out_5_pad_0, pad_type = pretrained_out_5_pad_type_0, strides = var_270, weight = layers_0_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor("pretrained_out_5_cast_fp16")]; - tensor var_276 = const()[name = tensor("op_276"), val = tensor([1, 1])]; - tensor var_278 = const()[name = tensor("op_278"), val = tensor([1, 1])]; - tensor input_5_pad_type_0 = const()[name = tensor("input_5_pad_type_0"), val = tensor("custom")]; - tensor input_5_pad_0 = const()[name = tensor("input_5_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136559488)))]; - tensor input_5_cast_fp16 = conv(dilations = var_278, groups = var_183, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = var_276, weight = layers_0_self_attn_v_proj_loraA_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("input_5_cast_fp16")]; - tensor var_282 = const()[name = tensor("op_282"), val = tensor([1, 1])]; - tensor var_284 = const()[name = tensor("op_284"), val = tensor([1, 1])]; - tensor lora_out_9_pad_type_0 = const()[name = tensor("lora_out_9_pad_type_0"), val = tensor("custom")]; - tensor lora_out_9_pad_0 = const()[name = tensor("lora_out_9_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_11_weight_0_to_fp16 = const()[name = tensor("lora_out_11_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136600512)))]; - tensor lora_out_11_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_284, groups = var_183, pad = lora_out_9_pad_0, pad_type = lora_out_9_pad_type_0, strides = var_282, weight = lora_out_11_weight_0_to_fp16, x = input_5_cast_fp16)[name = tensor("lora_out_11_cast_fp16")]; - tensor current_value_1_cast_fp16 = add(x = pretrained_out_5_cast_fp16, y = lora_out_11_cast_fp16)[name = tensor("current_value_1_cast_fp16")]; - tensor var_291_axes_0 = const()[name = tensor("op_291_axes_0"), val = tensor([1])]; - tensor var_291_cast_fp16 = expand_dims(axes = var_291_axes_0, x = kv_cache_update_mask)[name = tensor("op_291_cast_fp16")]; - tensor var_292_axes_0 = const()[name = tensor("op_292_axes_0"), val = tensor([2])]; - tensor var_292_cast_fp16 = expand_dims(axes = var_292_axes_0, x = var_291_cast_fp16)[name = tensor("op_292_cast_fp16")]; - tensor var_294_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_294_cast_fp16")]; - tensor var_177_to_fp16 = const()[name = tensor("op_177_to_fp16"), val = tensor(0x1p+0)]; - tensor var_295_cast_fp16 = sub(x = var_177_to_fp16, y = var_292_cast_fp16)[name = tensor("op_295_cast_fp16")]; - tensor var_296_cast_fp16 = mul(x = var_103_cast_fp16_0, y = var_295_cast_fp16)[name = tensor("op_296_cast_fp16")]; - tensor key_1_cast_fp16 = add(x = var_294_cast_fp16, y = var_296_cast_fp16)[name = tensor("key_1_cast_fp16")]; - tensor var_298_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_298_cast_fp16")]; - tensor var_300_cast_fp16 = mul(x = var_138_cast_fp16_0, y = var_295_cast_fp16)[name = tensor("op_300_cast_fp16")]; - tensor value_1_cast_fp16 = add(x = var_298_cast_fp16, y = var_300_cast_fp16)[name = tensor("value_1_cast_fp16")]; - tensor var_303 = const()[name = tensor("op_303"), val = tensor([1, 20, 64, -1])]; - tensor var_304_cast_fp16 = reshape(shape = var_303, x = query_1_cast_fp16)[name = tensor("op_304_cast_fp16")]; - tensor var_305_to_fp16 = const()[name = tensor("op_305_to_fp16"), val = tensor(0x1p-3)]; - tensor var_306_cast_fp16 = mul(x = var_304_cast_fp16, y = var_305_to_fp16)[name = tensor("op_306_cast_fp16")]; - tensor var_307 = const()[name = tensor("op_307"), val = tensor([1, 20, 64, -1])]; - tensor var_308_cast_fp16 = reshape(shape = var_307, x = key_1_cast_fp16)[name = tensor("op_308_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_306_cast_fp16, y = var_308_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; - tensor var_312_axes_0 = const()[name = tensor("op_312_axes_0"), val = tensor([1])]; - tensor var_312_cast_fp16 = expand_dims(axes = var_312_axes_0, x = decoder_key_padding_mask)[name = tensor("op_312_cast_fp16")]; - tensor var_313_axes_0 = const()[name = tensor("op_313_axes_0"), val = tensor([2])]; - tensor var_313_cast_fp16 = expand_dims(axes = var_313_axes_0, x = var_312_cast_fp16)[name = tensor("op_313_cast_fp16")]; - tensor mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; - tensor var_316_cast_fp16 = softmax(axis = var_176, x = mh_w_3_cast_fp16)[name = tensor("op_316_cast_fp16")]; - tensor var_317 = const()[name = tensor("op_317"), val = tensor([1, 20, 64, -1])]; - tensor var_318_cast_fp16 = reshape(shape = var_317, x = value_1_cast_fp16)[name = tensor("op_318_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_318_cast_fp16, y = var_316_cast_fp16)[name = tensor("attn_1_cast_fp16")]; - tensor var_321 = const()[name = tensor("op_321"), val = tensor([1, 1280, 1, -1])]; - tensor input_7_cast_fp16 = reshape(shape = var_321, x = attn_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; - tensor var_328 = const()[name = tensor("op_328"), val = tensor([1, 1])]; - tensor var_330 = const()[name = tensor("op_330"), val = tensor([1, 1])]; - tensor pretrained_out_7_pad_type_0 = const()[name = tensor("pretrained_out_7_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_7_pad_0 = const()[name = tensor("pretrained_out_7_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136641536))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137460800))), name = tensor("layers_0_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_0_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137460928)))]; - tensor pretrained_out_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_330, groups = var_183, pad = pretrained_out_7_pad_0, pad_type = pretrained_out_7_pad_type_0, strides = var_328, weight = layers_0_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_7_cast_fp16)[name = tensor("pretrained_out_7_cast_fp16")]; - tensor var_334 = const()[name = tensor("op_334"), val = tensor([1, 1])]; - tensor var_336 = const()[name = tensor("op_336"), val = tensor([1, 1])]; - tensor input_9_pad_type_0 = const()[name = tensor("input_9_pad_type_0"), val = tensor("custom")]; - tensor input_9_pad_0 = const()[name = tensor("input_9_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137463552)))]; - tensor input_9_cast_fp16 = conv(dilations = var_336, groups = var_183, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = var_334, weight = layers_0_self_attn_o_proj_loraA_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; - tensor var_340 = const()[name = tensor("op_340"), val = tensor([1, 1])]; - tensor var_342 = const()[name = tensor("op_342"), val = tensor([1, 1])]; - tensor lora_out_13_pad_type_0 = const()[name = tensor("lora_out_13_pad_type_0"), val = tensor("custom")]; - tensor lora_out_13_pad_0 = const()[name = tensor("lora_out_13_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_15_weight_0_to_fp16 = const()[name = tensor("lora_out_15_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137504576)))]; - tensor lora_out_15_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_342, groups = var_183, pad = lora_out_13_pad_0, pad_type = lora_out_13_pad_type_0, strides = var_340, weight = lora_out_15_weight_0_to_fp16, x = input_9_cast_fp16)[name = tensor("lora_out_15_cast_fp16")]; - tensor obj_7_cast_fp16 = add(x = pretrained_out_7_cast_fp16, y = lora_out_15_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_355 = const()[name = tensor("op_355"), val = tensor([1])]; - tensor channels_mean_3_cast_fp16 = reduce_mean(axes = var_355, 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_359 = const()[name = tensor("op_359"), val = tensor([1])]; - tensor var_360_cast_fp16 = reduce_mean(axes = var_359, keep_dims = var_184, x = zero_mean_sq_3_cast_fp16)[name = tensor("op_360_cast_fp16")]; - tensor var_361_to_fp16 = const()[name = tensor("op_361_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_362_cast_fp16 = add(x = var_360_cast_fp16, y = var_361_to_fp16)[name = tensor("op_362_cast_fp16")]; - tensor denom_3_epsilon_0 = const()[name = tensor("denom_3_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_3_cast_fp16 = rsqrt(epsilon = denom_3_epsilon_0, x = var_362_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(137545600)))]; - 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(137548224)))]; - 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_380 = const()[name = tensor("op_380"), val = tensor([1, 1])]; - tensor var_382 = const()[name = tensor("op_382"), val = tensor([1, 1])]; - tensor pretrained_out_9_pad_type_0 = const()[name = tensor("pretrained_out_9_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_9_pad_0 = const()[name = tensor("pretrained_out_9_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137550848))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138370112))), name = tensor("layers_0_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_0_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138370240)))]; - tensor pretrained_out_9_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_382, groups = var_183, pad = pretrained_out_9_pad_0, pad_type = pretrained_out_9_pad_type_0, strides = var_380, weight = layers_0_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = tensor("pretrained_out_9_cast_fp16")]; - tensor var_386 = const()[name = tensor("op_386"), val = tensor([1, 1])]; - tensor var_388 = const()[name = tensor("op_388"), val = tensor([1, 1])]; - tensor input_11_pad_type_0 = const()[name = tensor("input_11_pad_type_0"), val = tensor("custom")]; - tensor input_11_pad_0 = const()[name = tensor("input_11_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138372864)))]; - tensor input_11_cast_fp16 = conv(dilations = var_388, groups = var_183, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = var_386, weight = layers_0_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; - tensor var_392 = const()[name = tensor("op_392"), val = tensor([1, 1])]; - tensor var_394 = const()[name = tensor("op_394"), val = tensor([1, 1])]; - tensor lora_out_17_pad_type_0 = const()[name = tensor("lora_out_17_pad_type_0"), val = tensor("custom")]; - tensor lora_out_17_pad_0 = const()[name = tensor("lora_out_17_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_19_weight_0_to_fp16 = const()[name = tensor("lora_out_19_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138413888)))]; - tensor lora_out_19_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_394, groups = var_183, pad = lora_out_17_pad_0, pad_type = lora_out_17_pad_type_0, strides = var_392, weight = lora_out_19_weight_0_to_fp16, x = input_11_cast_fp16)[name = tensor("lora_out_19_cast_fp16")]; - tensor query_3_cast_fp16 = add(x = pretrained_out_9_cast_fp16, y = lora_out_19_cast_fp16)[name = tensor("query_3_cast_fp16")]; - tensor var_404 = const()[name = tensor("op_404"), val = tensor([1, 1])]; - tensor var_406 = const()[name = tensor("op_406"), val = tensor([1, 1])]; - tensor pretrained_out_11_pad_type_0 = const()[name = tensor("pretrained_out_11_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_11_pad_0 = const()[name = tensor("pretrained_out_11_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138454912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139274176))), name = tensor("layers_0_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_11_cast_fp16 = conv(dilations = var_406, groups = var_183, pad = pretrained_out_11_pad_0, pad_type = pretrained_out_11_pad_type_0, strides = var_404, weight = layers_0_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_11_cast_fp16")]; - tensor var_410 = const()[name = tensor("op_410"), val = tensor([1, 1])]; - tensor var_412 = const()[name = tensor("op_412"), val = tensor([1, 1])]; - tensor input_13_pad_type_0 = const()[name = tensor("input_13_pad_type_0"), val = tensor("custom")]; - tensor input_13_pad_0 = const()[name = tensor("input_13_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139274304)))]; - tensor input_13_cast_fp16 = conv(dilations = var_412, groups = var_183, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = var_410, weight = layers_0_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_13_cast_fp16")]; - tensor var_416 = const()[name = tensor("op_416"), val = tensor([1, 1])]; - tensor var_418 = const()[name = tensor("op_418"), val = tensor([1, 1])]; - tensor lora_out_21_pad_type_0 = const()[name = tensor("lora_out_21_pad_type_0"), val = tensor("custom")]; - tensor lora_out_21_pad_0 = const()[name = tensor("lora_out_21_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_23_weight_0_to_fp16 = const()[name = tensor("lora_out_23_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139315328)))]; - tensor lora_out_23_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_418, groups = var_183, pad = lora_out_21_pad_0, pad_type = lora_out_21_pad_type_0, strides = var_416, weight = lora_out_23_weight_0_to_fp16, x = input_13_cast_fp16)[name = tensor("lora_out_23_cast_fp16")]; - tensor key_3_cast_fp16 = add(x = pretrained_out_11_cast_fp16, y = lora_out_23_cast_fp16)[name = tensor("key_3_cast_fp16")]; - tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 1])]; - tensor var_431 = const()[name = tensor("op_431"), val = tensor([1, 1])]; - tensor pretrained_out_13_pad_type_0 = const()[name = tensor("pretrained_out_13_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_13_pad_0 = const()[name = tensor("pretrained_out_13_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139356352))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140175616))), name = tensor("layers_0_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_0_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140175744)))]; - tensor pretrained_out_13_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_431, groups = var_183, pad = pretrained_out_13_pad_0, pad_type = pretrained_out_13_pad_type_0, strides = var_429, weight = layers_0_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_13_cast_fp16")]; - tensor var_435 = const()[name = tensor("op_435"), val = tensor([1, 1])]; - tensor var_437 = const()[name = tensor("op_437"), val = tensor([1, 1])]; - tensor input_15_pad_type_0 = const()[name = tensor("input_15_pad_type_0"), val = tensor("custom")]; - tensor input_15_pad_0 = const()[name = tensor("input_15_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140178368)))]; - tensor input_15_cast_fp16 = conv(dilations = var_437, groups = var_183, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = var_435, weight = layers_0_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_15_cast_fp16")]; - tensor var_441 = const()[name = tensor("op_441"), val = tensor([1, 1])]; - tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1])]; - tensor lora_out_25_pad_type_0 = const()[name = tensor("lora_out_25_pad_type_0"), val = tensor("custom")]; - tensor lora_out_25_pad_0 = const()[name = tensor("lora_out_25_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_27_weight_0_to_fp16 = const()[name = tensor("lora_out_27_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140219392)))]; - tensor lora_out_27_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_443, groups = var_183, pad = lora_out_25_pad_0, pad_type = lora_out_25_pad_type_0, strides = var_441, weight = lora_out_27_weight_0_to_fp16, x = input_15_cast_fp16)[name = tensor("lora_out_27_cast_fp16")]; - tensor value_3_cast_fp16 = add(x = pretrained_out_13_cast_fp16, y = lora_out_27_cast_fp16)[name = tensor("value_3_cast_fp16")]; - tensor var_450 = const()[name = tensor("op_450"), val = tensor([1, 20, 64, -1])]; - tensor var_451_cast_fp16 = reshape(shape = var_450, x = query_3_cast_fp16)[name = tensor("op_451_cast_fp16")]; - tensor var_452_to_fp16 = const()[name = tensor("op_452_to_fp16"), val = tensor(0x1p-3)]; - tensor var_453_cast_fp16 = mul(x = var_451_cast_fp16, y = var_452_to_fp16)[name = tensor("op_453_cast_fp16")]; - tensor var_454 = const()[name = tensor("op_454"), val = tensor([1, 20, 64, -1])]; - tensor var_455_cast_fp16 = reshape(shape = var_454, x = key_3_cast_fp16)[name = tensor("op_455_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_453_cast_fp16, y = var_455_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; - tensor var_458_cast_fp16 = softmax(axis = var_176, x = mh_w_5_cast_fp16)[name = tensor("op_458_cast_fp16")]; - tensor var_459 = const()[name = tensor("op_459"), val = tensor([1, 20, 64, -1])]; - tensor var_460_cast_fp16 = reshape(shape = var_459, x = value_3_cast_fp16)[name = tensor("op_460_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_460_cast_fp16, y = var_458_cast_fp16)[name = tensor("attn_3_cast_fp16")]; - tensor var_463 = const()[name = tensor("op_463"), val = tensor([1, 1280, 1, -1])]; - tensor input_17_cast_fp16 = reshape(shape = var_463, x = attn_3_cast_fp16)[name = tensor("input_17_cast_fp16")]; - tensor var_470 = const()[name = tensor("op_470"), val = tensor([1, 1])]; - tensor var_472 = const()[name = tensor("op_472"), val = tensor([1, 1])]; - tensor pretrained_out_15_pad_type_0 = const()[name = tensor("pretrained_out_15_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_15_pad_0 = const()[name = tensor("pretrained_out_15_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140260416))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141079680))), name = tensor("layers_0_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_0_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141079808)))]; - tensor pretrained_out_15_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_472, groups = var_183, pad = pretrained_out_15_pad_0, pad_type = pretrained_out_15_pad_type_0, strides = var_470, weight = layers_0_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_17_cast_fp16)[name = tensor("pretrained_out_15_cast_fp16")]; - tensor var_476 = const()[name = tensor("op_476"), val = tensor([1, 1])]; - tensor var_478 = const()[name = tensor("op_478"), val = tensor([1, 1])]; - tensor input_19_pad_type_0 = const()[name = tensor("input_19_pad_type_0"), val = tensor("custom")]; - tensor input_19_pad_0 = const()[name = tensor("input_19_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141082432)))]; - tensor input_19_cast_fp16 = conv(dilations = var_478, groups = var_183, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = var_476, weight = layers_0_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; - tensor var_482 = const()[name = tensor("op_482"), val = tensor([1, 1])]; - tensor var_484 = const()[name = tensor("op_484"), val = tensor([1, 1])]; - tensor lora_out_29_pad_type_0 = const()[name = tensor("lora_out_29_pad_type_0"), val = tensor("custom")]; - tensor lora_out_29_pad_0 = const()[name = tensor("lora_out_29_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_31_weight_0_to_fp16 = const()[name = tensor("lora_out_31_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141123456)))]; - tensor lora_out_31_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_484, groups = var_183, pad = lora_out_29_pad_0, pad_type = lora_out_29_pad_type_0, strides = var_482, weight = lora_out_31_weight_0_to_fp16, x = input_19_cast_fp16)[name = tensor("lora_out_31_cast_fp16")]; - tensor obj_11_cast_fp16 = add(x = pretrained_out_15_cast_fp16, y = lora_out_31_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_493 = const()[name = tensor("op_493"), val = tensor([1])]; - tensor channels_mean_5_cast_fp16 = reduce_mean(axes = var_493, 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_497 = const()[name = tensor("op_497"), val = tensor([1])]; - tensor var_498_cast_fp16 = reduce_mean(axes = var_497, keep_dims = var_184, x = zero_mean_sq_5_cast_fp16)[name = tensor("op_498_cast_fp16")]; - tensor var_499_to_fp16 = const()[name = tensor("op_499_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_500_cast_fp16 = add(x = var_498_cast_fp16, y = var_499_to_fp16)[name = tensor("op_500_cast_fp16")]; - tensor denom_5_epsilon_0 = const()[name = tensor("denom_5_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_5_cast_fp16 = rsqrt(epsilon = denom_5_epsilon_0, x = var_500_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_21_gamma_0_to_fp16 = const()[name = tensor("input_21_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141164480)))]; - tensor input_21_beta_0_to_fp16 = const()[name = tensor("input_21_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141167104)))]; - tensor input_21_epsilon_0_to_fp16 = const()[name = tensor("input_21_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_21_cast_fp16 = batch_norm(beta = input_21_beta_0_to_fp16, epsilon = input_21_epsilon_0_to_fp16, gamma = input_21_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_21_cast_fp16")]; - tensor var_514 = const()[name = tensor("op_514"), val = tensor([1, 1])]; - tensor var_516 = const()[name = tensor("op_516"), val = tensor([1, 1])]; - tensor pretrained_out_17_pad_type_0 = const()[name = tensor("pretrained_out_17_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_17_pad_0 = const()[name = tensor("pretrained_out_17_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141169728))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144446592))), name = tensor("layers_0_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_0_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144446720)))]; - tensor pretrained_out_17_cast_fp16 = conv(bias = layers_0_fc1_pretrained_bias_to_fp16, dilations = var_516, groups = var_183, pad = pretrained_out_17_pad_0, pad_type = pretrained_out_17_pad_type_0, strides = var_514, weight = layers_0_fc1_pretrained_weight_to_fp16_palettized, x = input_21_cast_fp16)[name = tensor("pretrained_out_17_cast_fp16")]; - tensor var_520 = const()[name = tensor("op_520"), val = tensor([1, 1])]; - tensor var_522 = const()[name = tensor("op_522"), val = tensor([1, 1])]; - tensor input_23_pad_type_0 = const()[name = tensor("input_23_pad_type_0"), val = tensor("custom")]; - tensor input_23_pad_0 = const()[name = tensor("input_23_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_0_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144457024)))]; - tensor input_23_cast_fp16 = conv(dilations = var_522, groups = var_183, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = var_520, weight = layers_0_fc1_loraA_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; - tensor var_526 = const()[name = tensor("op_526"), val = tensor([1, 1])]; - tensor var_528 = const()[name = tensor("op_528"), val = tensor([1, 1])]; - tensor lora_out_33_pad_type_0 = const()[name = tensor("lora_out_33_pad_type_0"), val = tensor("custom")]; - tensor lora_out_33_pad_0 = const()[name = tensor("lora_out_33_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_35_weight_0_to_fp16 = const()[name = tensor("lora_out_35_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144498048)))]; - tensor lora_out_35_bias_0_to_fp16 = const()[name = tensor("lora_out_35_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144661952)))]; - tensor lora_out_35_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_528, groups = var_183, pad = lora_out_33_pad_0, pad_type = lora_out_33_pad_type_0, strides = var_526, weight = lora_out_35_weight_0_to_fp16, x = input_23_cast_fp16)[name = tensor("lora_out_35_cast_fp16")]; - tensor input_25_cast_fp16 = add(x = pretrained_out_17_cast_fp16, y = lora_out_35_cast_fp16)[name = tensor("input_25_cast_fp16")]; - tensor input_27_mode_0 = const()[name = tensor("input_27_mode_0"), val = tensor("EXACT")]; - tensor input_27_cast_fp16 = gelu(mode = input_27_mode_0, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; - tensor var_540 = const()[name = tensor("op_540"), val = tensor([1, 1])]; - tensor var_542 = const()[name = tensor("op_542"), val = tensor([1, 1])]; - tensor pretrained_out_19_pad_type_0 = const()[name = tensor("pretrained_out_19_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_19_pad_0 = const()[name = tensor("pretrained_out_19_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144672256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147949120))), name = tensor("layers_0_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_0_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147949248)))]; - tensor pretrained_out_19_cast_fp16 = conv(bias = layers_0_fc2_pretrained_bias_to_fp16, dilations = var_542, groups = var_183, pad = pretrained_out_19_pad_0, pad_type = pretrained_out_19_pad_type_0, strides = var_540, weight = layers_0_fc2_pretrained_weight_to_fp16_palettized, x = input_27_cast_fp16)[name = tensor("pretrained_out_19_cast_fp16")]; - tensor var_546 = const()[name = tensor("op_546"), val = tensor([1, 1])]; - tensor var_548 = const()[name = tensor("op_548"), val = tensor([1, 1])]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("custom")]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_0_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_0_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147951872)))]; - tensor input_29_cast_fp16 = conv(dilations = var_548, groups = var_183, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = var_546, weight = layers_0_fc2_loraA_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; - tensor var_552 = const()[name = tensor("op_552"), val = tensor([1, 1])]; - tensor var_554 = const()[name = tensor("op_554"), val = tensor([1, 1])]; - tensor lora_out_37_pad_type_0 = const()[name = tensor("lora_out_37_pad_type_0"), val = tensor("custom")]; - tensor lora_out_37_pad_0 = const()[name = tensor("lora_out_37_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_39_weight_0_to_fp16 = const()[name = tensor("lora_out_39_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148115776)))]; - tensor lora_out_39_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_554, groups = var_183, pad = lora_out_37_pad_0, pad_type = lora_out_37_pad_type_0, strides = var_552, weight = lora_out_39_weight_0_to_fp16, x = input_29_cast_fp16)[name = tensor("lora_out_39_cast_fp16")]; - tensor hidden_states_3_cast_fp16 = add(x = pretrained_out_19_cast_fp16, y = lora_out_39_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_570 = const()[name = tensor("op_570"), val = tensor(3)]; - tensor var_577 = const()[name = tensor("op_577"), val = tensor(1)]; - tensor var_578 = const()[name = tensor("op_578"), val = tensor(true)]; - tensor var_590 = const()[name = tensor("op_590"), val = tensor([1])]; - tensor channels_mean_7_cast_fp16 = reduce_mean(axes = var_590, keep_dims = var_578, 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_594 = const()[name = tensor("op_594"), val = tensor([1])]; - tensor var_595_cast_fp16 = reduce_mean(axes = var_594, keep_dims = var_578, x = zero_mean_sq_7_cast_fp16)[name = tensor("op_595_cast_fp16")]; - tensor var_596_to_fp16 = const()[name = tensor("op_596_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_597_cast_fp16 = add(x = var_595_cast_fp16, y = var_596_to_fp16)[name = tensor("op_597_cast_fp16")]; - tensor denom_7_epsilon_0 = const()[name = tensor("denom_7_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_7_cast_fp16 = rsqrt(epsilon = denom_7_epsilon_0, x = var_597_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_13_gamma_0_to_fp16 = const()[name = tensor("obj_13_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148156800)))]; - tensor obj_13_beta_0_to_fp16 = const()[name = tensor("obj_13_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148159424)))]; - tensor obj_13_epsilon_0_to_fp16 = const()[name = tensor("obj_13_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_13_cast_fp16 = batch_norm(beta = obj_13_beta_0_to_fp16, epsilon = obj_13_epsilon_0_to_fp16, gamma = obj_13_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_13_cast_fp16")]; - tensor var_615 = const()[name = tensor("op_615"), val = tensor([1, 1])]; - tensor var_617 = const()[name = tensor("op_617"), val = tensor([1, 1])]; - tensor pretrained_out_21_pad_type_0 = const()[name = tensor("pretrained_out_21_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_21_pad_0 = const()[name = tensor("pretrained_out_21_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148162048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148981312))), name = tensor("layers_1_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_1_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148981440)))]; - tensor pretrained_out_21_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_617, groups = var_577, pad = pretrained_out_21_pad_0, pad_type = pretrained_out_21_pad_type_0, strides = var_615, weight = layers_1_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_13_cast_fp16)[name = tensor("pretrained_out_21_cast_fp16")]; - tensor var_621 = const()[name = tensor("op_621"), val = tensor([1, 1])]; - tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 1])]; - tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("custom")]; - tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148984064)))]; - tensor input_31_cast_fp16 = conv(dilations = var_623, groups = var_577, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = var_621, weight = layers_1_self_attn_q_proj_loraA_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor("input_31_cast_fp16")]; - tensor var_627 = const()[name = tensor("op_627"), val = tensor([1, 1])]; - tensor var_629 = const()[name = tensor("op_629"), val = tensor([1, 1])]; - tensor lora_out_41_pad_type_0 = const()[name = tensor("lora_out_41_pad_type_0"), val = tensor("custom")]; - tensor lora_out_41_pad_0 = const()[name = tensor("lora_out_41_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_43_weight_0_to_fp16 = const()[name = tensor("lora_out_43_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149025088)))]; - tensor lora_out_43_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_629, groups = var_577, pad = lora_out_41_pad_0, pad_type = lora_out_41_pad_type_0, strides = var_627, weight = lora_out_43_weight_0_to_fp16, x = input_31_cast_fp16)[name = tensor("lora_out_43_cast_fp16")]; - tensor query_5_cast_fp16 = add(x = pretrained_out_21_cast_fp16, y = lora_out_43_cast_fp16)[name = tensor("query_5_cast_fp16")]; - tensor var_639 = const()[name = tensor("op_639"), val = tensor([1, 1])]; - tensor var_641 = const()[name = tensor("op_641"), val = tensor([1, 1])]; - tensor pretrained_out_23_pad_type_0 = const()[name = tensor("pretrained_out_23_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_23_pad_0 = const()[name = tensor("pretrained_out_23_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149066112))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149885376))), name = tensor("layers_1_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_23_cast_fp16 = conv(dilations = var_641, groups = var_577, pad = pretrained_out_23_pad_0, pad_type = pretrained_out_23_pad_type_0, strides = var_639, weight = layers_1_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_13_cast_fp16)[name = tensor("pretrained_out_23_cast_fp16")]; - tensor var_645 = const()[name = tensor("op_645"), val = tensor([1, 1])]; - tensor var_647 = const()[name = tensor("op_647"), val = tensor([1, 1])]; - tensor input_33_pad_type_0 = const()[name = tensor("input_33_pad_type_0"), val = tensor("custom")]; - tensor input_33_pad_0 = const()[name = tensor("input_33_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149885504)))]; - tensor input_33_cast_fp16 = conv(dilations = var_647, groups = var_577, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = var_645, weight = layers_1_self_attn_k_proj_loraA_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor("input_33_cast_fp16")]; - tensor var_651 = const()[name = tensor("op_651"), val = tensor([1, 1])]; - tensor var_653 = const()[name = tensor("op_653"), val = tensor([1, 1])]; - tensor lora_out_45_pad_type_0 = const()[name = tensor("lora_out_45_pad_type_0"), val = tensor("custom")]; - tensor lora_out_45_pad_0 = const()[name = tensor("lora_out_45_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_47_weight_0_to_fp16 = const()[name = tensor("lora_out_47_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149926528)))]; - tensor lora_out_47_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_653, groups = var_577, pad = lora_out_45_pad_0, pad_type = lora_out_45_pad_type_0, strides = var_651, weight = lora_out_47_weight_0_to_fp16, x = input_33_cast_fp16)[name = tensor("lora_out_47_cast_fp16")]; - tensor current_key_3_cast_fp16 = add(x = pretrained_out_23_cast_fp16, y = lora_out_47_cast_fp16)[name = tensor("current_key_3_cast_fp16")]; - tensor var_664 = const()[name = tensor("op_664"), val = tensor([1, 1])]; - tensor var_666 = const()[name = tensor("op_666"), val = tensor([1, 1])]; - tensor pretrained_out_25_pad_type_0 = const()[name = tensor("pretrained_out_25_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_25_pad_0 = const()[name = tensor("pretrained_out_25_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149967552))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150786816))), name = tensor("layers_1_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_1_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150786944)))]; - tensor pretrained_out_25_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_666, groups = var_577, pad = pretrained_out_25_pad_0, pad_type = pretrained_out_25_pad_type_0, strides = var_664, weight = layers_1_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_13_cast_fp16)[name = tensor("pretrained_out_25_cast_fp16")]; - tensor var_670 = const()[name = tensor("op_670"), val = tensor([1, 1])]; - tensor var_672 = const()[name = tensor("op_672"), val = tensor([1, 1])]; - tensor input_35_pad_type_0 = const()[name = tensor("input_35_pad_type_0"), val = tensor("custom")]; - tensor input_35_pad_0 = const()[name = tensor("input_35_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150789568)))]; - tensor input_35_cast_fp16 = conv(dilations = var_672, groups = var_577, pad = input_35_pad_0, pad_type = input_35_pad_type_0, strides = var_670, weight = layers_1_self_attn_v_proj_loraA_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor("input_35_cast_fp16")]; - tensor var_676 = const()[name = tensor("op_676"), val = tensor([1, 1])]; - tensor var_678 = const()[name = tensor("op_678"), val = tensor([1, 1])]; - tensor lora_out_49_pad_type_0 = const()[name = tensor("lora_out_49_pad_type_0"), val = tensor("custom")]; - tensor lora_out_49_pad_0 = const()[name = tensor("lora_out_49_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_51_weight_0_to_fp16 = const()[name = tensor("lora_out_51_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150830592)))]; - tensor lora_out_51_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_678, groups = var_577, pad = lora_out_49_pad_0, pad_type = lora_out_49_pad_type_0, strides = var_676, weight = lora_out_51_weight_0_to_fp16, x = input_35_cast_fp16)[name = tensor("lora_out_51_cast_fp16")]; - tensor current_value_3_cast_fp16 = add(x = pretrained_out_25_cast_fp16, y = lora_out_51_cast_fp16)[name = tensor("current_value_3_cast_fp16")]; - tensor var_688_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_688_cast_fp16")]; - tensor var_690_cast_fp16 = mul(x = var_103_cast_fp16_1, y = var_295_cast_fp16)[name = tensor("op_690_cast_fp16")]; - tensor key_5_cast_fp16 = add(x = var_688_cast_fp16, y = var_690_cast_fp16)[name = tensor("key_5_cast_fp16")]; - tensor var_692_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_692_cast_fp16")]; - tensor var_694_cast_fp16 = mul(x = var_138_cast_fp16_1, y = var_295_cast_fp16)[name = tensor("op_694_cast_fp16")]; - tensor value_5_cast_fp16 = add(x = var_692_cast_fp16, y = var_694_cast_fp16)[name = tensor("value_5_cast_fp16")]; - tensor var_697 = const()[name = tensor("op_697"), val = tensor([1, 20, 64, -1])]; - tensor var_698_cast_fp16 = reshape(shape = var_697, x = query_5_cast_fp16)[name = tensor("op_698_cast_fp16")]; - tensor var_699_to_fp16 = const()[name = tensor("op_699_to_fp16"), val = tensor(0x1p-3)]; - tensor var_700_cast_fp16 = mul(x = var_698_cast_fp16, y = var_699_to_fp16)[name = tensor("op_700_cast_fp16")]; - tensor var_701 = const()[name = tensor("op_701"), val = tensor([1, 20, 64, -1])]; - tensor var_702_cast_fp16 = reshape(shape = var_701, x = key_5_cast_fp16)[name = tensor("op_702_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_700_cast_fp16, y = var_702_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; - tensor mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; - tensor var_710_cast_fp16 = softmax(axis = var_570, x = mh_w_9_cast_fp16)[name = tensor("op_710_cast_fp16")]; - tensor var_711 = const()[name = tensor("op_711"), val = tensor([1, 20, 64, -1])]; - tensor var_712_cast_fp16 = reshape(shape = var_711, x = value_5_cast_fp16)[name = tensor("op_712_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_712_cast_fp16, y = var_710_cast_fp16)[name = tensor("attn_5_cast_fp16")]; - tensor var_715 = const()[name = tensor("op_715"), val = tensor([1, 1280, 1, -1])]; - tensor input_37_cast_fp16 = reshape(shape = var_715, x = attn_5_cast_fp16)[name = tensor("input_37_cast_fp16")]; - tensor var_722 = const()[name = tensor("op_722"), val = tensor([1, 1])]; - tensor var_724 = const()[name = tensor("op_724"), val = tensor([1, 1])]; - tensor pretrained_out_27_pad_type_0 = const()[name = tensor("pretrained_out_27_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_27_pad_0 = const()[name = tensor("pretrained_out_27_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150871616))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151690880))), name = tensor("layers_1_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_1_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151691008)))]; - tensor pretrained_out_27_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_724, groups = var_577, pad = pretrained_out_27_pad_0, pad_type = pretrained_out_27_pad_type_0, strides = var_722, weight = layers_1_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_37_cast_fp16)[name = tensor("pretrained_out_27_cast_fp16")]; - tensor var_728 = const()[name = tensor("op_728"), val = tensor([1, 1])]; - tensor var_730 = const()[name = tensor("op_730"), val = tensor([1, 1])]; - tensor input_39_pad_type_0 = const()[name = tensor("input_39_pad_type_0"), val = tensor("custom")]; - tensor input_39_pad_0 = const()[name = tensor("input_39_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151693632)))]; - tensor input_39_cast_fp16 = conv(dilations = var_730, groups = var_577, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = var_728, weight = layers_1_self_attn_o_proj_loraA_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; - tensor var_734 = const()[name = tensor("op_734"), val = tensor([1, 1])]; - tensor var_736 = const()[name = tensor("op_736"), val = tensor([1, 1])]; - tensor lora_out_53_pad_type_0 = const()[name = tensor("lora_out_53_pad_type_0"), val = tensor("custom")]; - tensor lora_out_53_pad_0 = const()[name = tensor("lora_out_53_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_55_weight_0_to_fp16 = const()[name = tensor("lora_out_55_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151734656)))]; - tensor lora_out_55_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_736, groups = var_577, pad = lora_out_53_pad_0, pad_type = lora_out_53_pad_type_0, strides = var_734, weight = lora_out_55_weight_0_to_fp16, x = input_39_cast_fp16)[name = tensor("lora_out_55_cast_fp16")]; - tensor obj_19_cast_fp16 = add(x = pretrained_out_27_cast_fp16, y = lora_out_55_cast_fp16)[name = tensor("obj_19_cast_fp16")]; - tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_19_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; - tensor var_749 = const()[name = tensor("op_749"), val = tensor([1])]; - tensor channels_mean_9_cast_fp16 = reduce_mean(axes = var_749, keep_dims = var_578, 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_753 = const()[name = tensor("op_753"), val = tensor([1])]; - tensor var_754_cast_fp16 = reduce_mean(axes = var_753, keep_dims = var_578, x = zero_mean_sq_9_cast_fp16)[name = tensor("op_754_cast_fp16")]; - tensor var_755_to_fp16 = const()[name = tensor("op_755_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_756_cast_fp16 = add(x = var_754_cast_fp16, y = var_755_to_fp16)[name = tensor("op_756_cast_fp16")]; - tensor denom_9_epsilon_0 = const()[name = tensor("denom_9_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_9_cast_fp16 = rsqrt(epsilon = denom_9_epsilon_0, x = var_756_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_21_gamma_0_to_fp16 = const()[name = tensor("obj_21_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151775680)))]; - tensor obj_21_beta_0_to_fp16 = const()[name = tensor("obj_21_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151778304)))]; - tensor obj_21_epsilon_0_to_fp16 = const()[name = tensor("obj_21_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_21_cast_fp16 = batch_norm(beta = obj_21_beta_0_to_fp16, epsilon = obj_21_epsilon_0_to_fp16, gamma = obj_21_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_21_cast_fp16")]; - tensor var_774 = const()[name = tensor("op_774"), val = tensor([1, 1])]; - tensor var_776 = const()[name = tensor("op_776"), val = tensor([1, 1])]; - tensor pretrained_out_29_pad_type_0 = const()[name = tensor("pretrained_out_29_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_29_pad_0 = const()[name = tensor("pretrained_out_29_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151780928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152600192))), name = tensor("layers_1_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_1_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152600320)))]; - tensor pretrained_out_29_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_776, groups = var_577, pad = pretrained_out_29_pad_0, pad_type = pretrained_out_29_pad_type_0, strides = var_774, weight = layers_1_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_21_cast_fp16)[name = tensor("pretrained_out_29_cast_fp16")]; - tensor var_780 = const()[name = tensor("op_780"), val = tensor([1, 1])]; - tensor var_782 = const()[name = tensor("op_782"), val = tensor([1, 1])]; - tensor input_41_pad_type_0 = const()[name = tensor("input_41_pad_type_0"), val = tensor("custom")]; - tensor input_41_pad_0 = const()[name = tensor("input_41_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152602944)))]; - tensor input_41_cast_fp16 = conv(dilations = var_782, groups = var_577, pad = input_41_pad_0, pad_type = input_41_pad_type_0, strides = var_780, weight = layers_1_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_21_cast_fp16)[name = tensor("input_41_cast_fp16")]; - tensor var_786 = const()[name = tensor("op_786"), val = tensor([1, 1])]; - tensor var_788 = const()[name = tensor("op_788"), val = tensor([1, 1])]; - tensor lora_out_57_pad_type_0 = const()[name = tensor("lora_out_57_pad_type_0"), val = tensor("custom")]; - tensor lora_out_57_pad_0 = const()[name = tensor("lora_out_57_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_59_weight_0_to_fp16 = const()[name = tensor("lora_out_59_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152643968)))]; - tensor lora_out_59_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_788, groups = var_577, pad = lora_out_57_pad_0, pad_type = lora_out_57_pad_type_0, strides = var_786, weight = lora_out_59_weight_0_to_fp16, x = input_41_cast_fp16)[name = tensor("lora_out_59_cast_fp16")]; - tensor query_7_cast_fp16 = add(x = pretrained_out_29_cast_fp16, y = lora_out_59_cast_fp16)[name = tensor("query_7_cast_fp16")]; - tensor var_798 = const()[name = tensor("op_798"), val = tensor([1, 1])]; - tensor var_800 = const()[name = tensor("op_800"), val = tensor([1, 1])]; - tensor pretrained_out_31_pad_type_0 = const()[name = tensor("pretrained_out_31_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_31_pad_0 = const()[name = tensor("pretrained_out_31_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152684992))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153504256))), name = tensor("layers_1_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_31_cast_fp16 = conv(dilations = var_800, groups = var_577, pad = pretrained_out_31_pad_0, pad_type = pretrained_out_31_pad_type_0, strides = var_798, weight = layers_1_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_31_cast_fp16")]; - tensor var_804 = const()[name = tensor("op_804"), val = tensor([1, 1])]; - tensor var_806 = const()[name = tensor("op_806"), val = tensor([1, 1])]; - tensor input_43_pad_type_0 = const()[name = tensor("input_43_pad_type_0"), val = tensor("custom")]; - tensor input_43_pad_0 = const()[name = tensor("input_43_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153504384)))]; - tensor input_43_cast_fp16 = conv(dilations = var_806, groups = var_577, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = var_804, weight = layers_1_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_43_cast_fp16")]; - tensor var_810 = const()[name = tensor("op_810"), val = tensor([1, 1])]; - tensor var_812 = const()[name = tensor("op_812"), val = tensor([1, 1])]; - tensor lora_out_61_pad_type_0 = const()[name = tensor("lora_out_61_pad_type_0"), val = tensor("custom")]; - tensor lora_out_61_pad_0 = const()[name = tensor("lora_out_61_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_63_weight_0_to_fp16 = const()[name = tensor("lora_out_63_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153545408)))]; - tensor lora_out_63_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_812, groups = var_577, pad = lora_out_61_pad_0, pad_type = lora_out_61_pad_type_0, strides = var_810, weight = lora_out_63_weight_0_to_fp16, x = input_43_cast_fp16)[name = tensor("lora_out_63_cast_fp16")]; - tensor key_7_cast_fp16 = add(x = pretrained_out_31_cast_fp16, y = lora_out_63_cast_fp16)[name = tensor("key_7_cast_fp16")]; - tensor var_823 = const()[name = tensor("op_823"), val = tensor([1, 1])]; - tensor var_825 = const()[name = tensor("op_825"), val = tensor([1, 1])]; - tensor pretrained_out_33_pad_type_0 = const()[name = tensor("pretrained_out_33_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_33_pad_0 = const()[name = tensor("pretrained_out_33_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153586432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154405696))), name = tensor("layers_1_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_1_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154405824)))]; - tensor pretrained_out_33_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_825, groups = var_577, pad = pretrained_out_33_pad_0, pad_type = pretrained_out_33_pad_type_0, strides = var_823, weight = layers_1_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_33_cast_fp16")]; - tensor var_829 = const()[name = tensor("op_829"), val = tensor([1, 1])]; - tensor var_831 = const()[name = tensor("op_831"), val = tensor([1, 1])]; - tensor input_45_pad_type_0 = const()[name = tensor("input_45_pad_type_0"), val = tensor("custom")]; - tensor input_45_pad_0 = const()[name = tensor("input_45_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154408448)))]; - tensor input_45_cast_fp16 = conv(dilations = var_831, groups = var_577, pad = input_45_pad_0, pad_type = input_45_pad_type_0, strides = var_829, weight = layers_1_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_45_cast_fp16")]; - tensor var_835 = const()[name = tensor("op_835"), val = tensor([1, 1])]; - tensor var_837 = const()[name = tensor("op_837"), val = tensor([1, 1])]; - tensor lora_out_65_pad_type_0 = const()[name = tensor("lora_out_65_pad_type_0"), val = tensor("custom")]; - tensor lora_out_65_pad_0 = const()[name = tensor("lora_out_65_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_67_weight_0_to_fp16 = const()[name = tensor("lora_out_67_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154449472)))]; - tensor lora_out_67_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_837, groups = var_577, pad = lora_out_65_pad_0, pad_type = lora_out_65_pad_type_0, strides = var_835, weight = lora_out_67_weight_0_to_fp16, x = input_45_cast_fp16)[name = tensor("lora_out_67_cast_fp16")]; - tensor value_7_cast_fp16 = add(x = pretrained_out_33_cast_fp16, y = lora_out_67_cast_fp16)[name = tensor("value_7_cast_fp16")]; - tensor var_844 = const()[name = tensor("op_844"), val = tensor([1, 20, 64, -1])]; - tensor var_845_cast_fp16 = reshape(shape = var_844, x = query_7_cast_fp16)[name = tensor("op_845_cast_fp16")]; - tensor var_846_to_fp16 = const()[name = tensor("op_846_to_fp16"), val = tensor(0x1p-3)]; - tensor var_847_cast_fp16 = mul(x = var_845_cast_fp16, y = var_846_to_fp16)[name = tensor("op_847_cast_fp16")]; - tensor var_848 = const()[name = tensor("op_848"), val = tensor([1, 20, 64, -1])]; - tensor var_849_cast_fp16 = reshape(shape = var_848, x = key_7_cast_fp16)[name = tensor("op_849_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_847_cast_fp16, y = var_849_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; - tensor var_852_cast_fp16 = softmax(axis = var_570, x = mh_w_11_cast_fp16)[name = tensor("op_852_cast_fp16")]; - tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 20, 64, -1])]; - tensor var_854_cast_fp16 = reshape(shape = var_853, x = value_7_cast_fp16)[name = tensor("op_854_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_854_cast_fp16, y = var_852_cast_fp16)[name = tensor("attn_7_cast_fp16")]; - tensor var_857 = const()[name = tensor("op_857"), val = tensor([1, 1280, 1, -1])]; - tensor input_47_cast_fp16 = reshape(shape = var_857, x = attn_7_cast_fp16)[name = tensor("input_47_cast_fp16")]; - tensor var_864 = const()[name = tensor("op_864"), val = tensor([1, 1])]; - tensor var_866 = const()[name = tensor("op_866"), val = tensor([1, 1])]; - tensor pretrained_out_35_pad_type_0 = const()[name = tensor("pretrained_out_35_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_35_pad_0 = const()[name = tensor("pretrained_out_35_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154490496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155309760))), name = tensor("layers_1_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_1_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155309888)))]; - tensor pretrained_out_35_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_866, groups = var_577, pad = pretrained_out_35_pad_0, pad_type = pretrained_out_35_pad_type_0, strides = var_864, weight = layers_1_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_47_cast_fp16)[name = tensor("pretrained_out_35_cast_fp16")]; - tensor var_870 = const()[name = tensor("op_870"), val = tensor([1, 1])]; - tensor var_872 = const()[name = tensor("op_872"), val = tensor([1, 1])]; - tensor input_49_pad_type_0 = const()[name = tensor("input_49_pad_type_0"), val = tensor("custom")]; - tensor input_49_pad_0 = const()[name = tensor("input_49_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155312512)))]; - tensor input_49_cast_fp16 = conv(dilations = var_872, groups = var_577, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = var_870, weight = layers_1_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; - tensor var_876 = const()[name = tensor("op_876"), val = tensor([1, 1])]; - tensor var_878 = const()[name = tensor("op_878"), val = tensor([1, 1])]; - tensor lora_out_69_pad_type_0 = const()[name = tensor("lora_out_69_pad_type_0"), val = tensor("custom")]; - tensor lora_out_69_pad_0 = const()[name = tensor("lora_out_69_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_71_weight_0_to_fp16 = const()[name = tensor("lora_out_71_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155353536)))]; - tensor lora_out_71_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_878, groups = var_577, pad = lora_out_69_pad_0, pad_type = lora_out_69_pad_type_0, strides = var_876, weight = lora_out_71_weight_0_to_fp16, x = input_49_cast_fp16)[name = tensor("lora_out_71_cast_fp16")]; - tensor obj_23_cast_fp16 = add(x = pretrained_out_35_cast_fp16, y = lora_out_71_cast_fp16)[name = tensor("obj_23_cast_fp16")]; - tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_23_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; - tensor var_887 = const()[name = tensor("op_887"), val = tensor([1])]; - tensor channels_mean_11_cast_fp16 = reduce_mean(axes = var_887, keep_dims = var_578, 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_891 = const()[name = tensor("op_891"), val = tensor([1])]; - tensor var_892_cast_fp16 = reduce_mean(axes = var_891, keep_dims = var_578, x = zero_mean_sq_11_cast_fp16)[name = tensor("op_892_cast_fp16")]; - tensor var_893_to_fp16 = const()[name = tensor("op_893_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_894_cast_fp16 = add(x = var_892_cast_fp16, y = var_893_to_fp16)[name = tensor("op_894_cast_fp16")]; - tensor denom_11_epsilon_0 = const()[name = tensor("denom_11_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_11_cast_fp16 = rsqrt(epsilon = denom_11_epsilon_0, x = var_894_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_51_gamma_0_to_fp16 = const()[name = tensor("input_51_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155394560)))]; - tensor input_51_beta_0_to_fp16 = const()[name = tensor("input_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155397184)))]; - tensor input_51_epsilon_0_to_fp16 = const()[name = tensor("input_51_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_51_cast_fp16 = batch_norm(beta = input_51_beta_0_to_fp16, epsilon = input_51_epsilon_0_to_fp16, gamma = input_51_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_51_cast_fp16")]; - tensor var_908 = const()[name = tensor("op_908"), val = tensor([1, 1])]; - tensor var_910 = const()[name = tensor("op_910"), val = tensor([1, 1])]; - tensor pretrained_out_37_pad_type_0 = const()[name = tensor("pretrained_out_37_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_37_pad_0 = const()[name = tensor("pretrained_out_37_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155399808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158676672))), name = tensor("layers_1_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_1_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158676800)))]; - tensor pretrained_out_37_cast_fp16 = conv(bias = layers_1_fc1_pretrained_bias_to_fp16, dilations = var_910, groups = var_577, pad = pretrained_out_37_pad_0, pad_type = pretrained_out_37_pad_type_0, strides = var_908, weight = layers_1_fc1_pretrained_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = tensor("pretrained_out_37_cast_fp16")]; - tensor var_914 = const()[name = tensor("op_914"), val = tensor([1, 1])]; - tensor var_916 = const()[name = tensor("op_916"), val = tensor([1, 1])]; - tensor input_53_pad_type_0 = const()[name = tensor("input_53_pad_type_0"), val = tensor("custom")]; - tensor input_53_pad_0 = const()[name = tensor("input_53_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_1_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158687104)))]; - tensor input_53_cast_fp16 = conv(dilations = var_916, groups = var_577, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = var_914, weight = layers_1_fc1_loraA_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; - tensor var_920 = const()[name = tensor("op_920"), val = tensor([1, 1])]; - tensor var_922 = const()[name = tensor("op_922"), val = tensor([1, 1])]; - tensor lora_out_73_pad_type_0 = const()[name = tensor("lora_out_73_pad_type_0"), val = tensor("custom")]; - tensor lora_out_73_pad_0 = const()[name = tensor("lora_out_73_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_75_weight_0_to_fp16 = const()[name = tensor("lora_out_75_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158728128)))]; - tensor lora_out_75_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_922, groups = var_577, pad = lora_out_73_pad_0, pad_type = lora_out_73_pad_type_0, strides = var_920, weight = lora_out_75_weight_0_to_fp16, x = input_53_cast_fp16)[name = tensor("lora_out_75_cast_fp16")]; - tensor input_55_cast_fp16 = add(x = pretrained_out_37_cast_fp16, y = lora_out_75_cast_fp16)[name = tensor("input_55_cast_fp16")]; - tensor input_57_mode_0 = const()[name = tensor("input_57_mode_0"), val = tensor("EXACT")]; - tensor input_57_cast_fp16 = gelu(mode = input_57_mode_0, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; - tensor var_934 = const()[name = tensor("op_934"), val = tensor([1, 1])]; - tensor var_936 = const()[name = tensor("op_936"), val = tensor([1, 1])]; - tensor pretrained_out_39_pad_type_0 = const()[name = tensor("pretrained_out_39_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_39_pad_0 = const()[name = tensor("pretrained_out_39_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158892032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162168896))), name = tensor("layers_1_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_1_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162169024)))]; - tensor pretrained_out_39_cast_fp16 = conv(bias = layers_1_fc2_pretrained_bias_to_fp16, dilations = var_936, groups = var_577, pad = pretrained_out_39_pad_0, pad_type = pretrained_out_39_pad_type_0, strides = var_934, weight = layers_1_fc2_pretrained_weight_to_fp16_palettized, x = input_57_cast_fp16)[name = tensor("pretrained_out_39_cast_fp16")]; - tensor var_940 = const()[name = tensor("op_940"), val = tensor([1, 1])]; - tensor var_942 = const()[name = tensor("op_942"), val = tensor([1, 1])]; - tensor input_59_pad_type_0 = const()[name = tensor("input_59_pad_type_0"), val = tensor("custom")]; - tensor input_59_pad_0 = const()[name = tensor("input_59_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_1_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_1_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162171648)))]; - tensor input_59_cast_fp16 = conv(dilations = var_942, groups = var_577, pad = input_59_pad_0, pad_type = input_59_pad_type_0, strides = var_940, weight = layers_1_fc2_loraA_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; - tensor var_946 = const()[name = tensor("op_946"), val = tensor([1, 1])]; - tensor var_948 = const()[name = tensor("op_948"), val = tensor([1, 1])]; - tensor lora_out_77_pad_type_0 = const()[name = tensor("lora_out_77_pad_type_0"), val = tensor("custom")]; - tensor lora_out_77_pad_0 = const()[name = tensor("lora_out_77_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_79_weight_0_to_fp16 = const()[name = tensor("lora_out_79_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162335552)))]; - tensor lora_out_79_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_948, groups = var_577, pad = lora_out_77_pad_0, pad_type = lora_out_77_pad_type_0, strides = var_946, weight = lora_out_79_weight_0_to_fp16, x = input_59_cast_fp16)[name = tensor("lora_out_79_cast_fp16")]; - tensor hidden_states_5_cast_fp16 = add(x = pretrained_out_39_cast_fp16, y = lora_out_79_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_964 = const()[name = tensor("op_964"), val = tensor(3)]; - tensor var_971 = const()[name = tensor("op_971"), val = tensor(1)]; - tensor var_972 = const()[name = tensor("op_972"), val = tensor(true)]; - tensor var_984 = const()[name = tensor("op_984"), val = tensor([1])]; - tensor channels_mean_13_cast_fp16 = reduce_mean(axes = var_984, keep_dims = var_972, 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_988 = const()[name = tensor("op_988"), val = tensor([1])]; - tensor var_989_cast_fp16 = reduce_mean(axes = var_988, keep_dims = var_972, x = zero_mean_sq_13_cast_fp16)[name = tensor("op_989_cast_fp16")]; - tensor var_990_to_fp16 = const()[name = tensor("op_990_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_991_cast_fp16 = add(x = var_989_cast_fp16, y = var_990_to_fp16)[name = tensor("op_991_cast_fp16")]; - tensor denom_13_epsilon_0 = const()[name = tensor("denom_13_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_13_cast_fp16 = rsqrt(epsilon = denom_13_epsilon_0, x = var_991_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_25_gamma_0_to_fp16 = const()[name = tensor("obj_25_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162376576)))]; - tensor obj_25_beta_0_to_fp16 = const()[name = tensor("obj_25_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162379200)))]; - tensor obj_25_epsilon_0_to_fp16 = const()[name = tensor("obj_25_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_25_cast_fp16 = batch_norm(beta = obj_25_beta_0_to_fp16, epsilon = obj_25_epsilon_0_to_fp16, gamma = obj_25_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_25_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 pretrained_out_41_pad_type_0 = const()[name = tensor("pretrained_out_41_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_41_pad_0 = const()[name = tensor("pretrained_out_41_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162381824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163201088))), name = tensor("layers_2_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_2_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163201216)))]; - tensor pretrained_out_41_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_1011, groups = var_971, pad = pretrained_out_41_pad_0, pad_type = pretrained_out_41_pad_type_0, strides = var_1009, weight = layers_2_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_25_cast_fp16)[name = tensor("pretrained_out_41_cast_fp16")]; - tensor var_1015 = const()[name = tensor("op_1015"), val = tensor([1, 1])]; - tensor var_1017 = const()[name = tensor("op_1017"), val = tensor([1, 1])]; - tensor input_61_pad_type_0 = const()[name = tensor("input_61_pad_type_0"), val = tensor("custom")]; - tensor input_61_pad_0 = const()[name = tensor("input_61_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163203840)))]; - tensor input_61_cast_fp16 = conv(dilations = var_1017, groups = var_971, pad = input_61_pad_0, pad_type = input_61_pad_type_0, strides = var_1015, weight = layers_2_self_attn_q_proj_loraA_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor("input_61_cast_fp16")]; - tensor var_1021 = const()[name = tensor("op_1021"), val = tensor([1, 1])]; - tensor var_1023 = const()[name = tensor("op_1023"), val = tensor([1, 1])]; - tensor lora_out_81_pad_type_0 = const()[name = tensor("lora_out_81_pad_type_0"), val = tensor("custom")]; - tensor lora_out_81_pad_0 = const()[name = tensor("lora_out_81_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_83_weight_0_to_fp16 = const()[name = tensor("lora_out_83_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163244864)))]; - tensor lora_out_83_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1023, groups = var_971, pad = lora_out_81_pad_0, pad_type = lora_out_81_pad_type_0, strides = var_1021, weight = lora_out_83_weight_0_to_fp16, x = input_61_cast_fp16)[name = tensor("lora_out_83_cast_fp16")]; - tensor query_9_cast_fp16 = add(x = pretrained_out_41_cast_fp16, y = lora_out_83_cast_fp16)[name = tensor("query_9_cast_fp16")]; - tensor var_1033 = const()[name = tensor("op_1033"), val = tensor([1, 1])]; - tensor var_1035 = const()[name = tensor("op_1035"), val = tensor([1, 1])]; - tensor pretrained_out_43_pad_type_0 = const()[name = tensor("pretrained_out_43_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_43_pad_0 = const()[name = tensor("pretrained_out_43_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163285888))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164105152))), name = tensor("layers_2_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_43_cast_fp16 = conv(dilations = var_1035, groups = var_971, pad = pretrained_out_43_pad_0, pad_type = pretrained_out_43_pad_type_0, strides = var_1033, weight = layers_2_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_25_cast_fp16)[name = tensor("pretrained_out_43_cast_fp16")]; - tensor var_1039 = const()[name = tensor("op_1039"), val = tensor([1, 1])]; - tensor var_1041 = const()[name = tensor("op_1041"), val = tensor([1, 1])]; - tensor input_63_pad_type_0 = const()[name = tensor("input_63_pad_type_0"), val = tensor("custom")]; - tensor input_63_pad_0 = const()[name = tensor("input_63_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164105280)))]; - tensor input_63_cast_fp16 = conv(dilations = var_1041, groups = var_971, pad = input_63_pad_0, pad_type = input_63_pad_type_0, strides = var_1039, weight = layers_2_self_attn_k_proj_loraA_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor("input_63_cast_fp16")]; - tensor var_1045 = const()[name = tensor("op_1045"), val = tensor([1, 1])]; - tensor var_1047 = const()[name = tensor("op_1047"), val = tensor([1, 1])]; - tensor lora_out_85_pad_type_0 = const()[name = tensor("lora_out_85_pad_type_0"), val = tensor("custom")]; - tensor lora_out_85_pad_0 = const()[name = tensor("lora_out_85_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_87_weight_0_to_fp16 = const()[name = tensor("lora_out_87_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164146304)))]; - tensor lora_out_87_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1047, groups = var_971, pad = lora_out_85_pad_0, pad_type = lora_out_85_pad_type_0, strides = var_1045, weight = lora_out_87_weight_0_to_fp16, x = input_63_cast_fp16)[name = tensor("lora_out_87_cast_fp16")]; - tensor current_key_5_cast_fp16 = add(x = pretrained_out_43_cast_fp16, y = lora_out_87_cast_fp16)[name = tensor("current_key_5_cast_fp16")]; - tensor var_1058 = const()[name = tensor("op_1058"), val = tensor([1, 1])]; - tensor var_1060 = const()[name = tensor("op_1060"), val = tensor([1, 1])]; - tensor pretrained_out_45_pad_type_0 = const()[name = tensor("pretrained_out_45_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_45_pad_0 = const()[name = tensor("pretrained_out_45_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164187328))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165006592))), name = tensor("layers_2_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_2_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165006720)))]; - tensor pretrained_out_45_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_1060, groups = var_971, pad = pretrained_out_45_pad_0, pad_type = pretrained_out_45_pad_type_0, strides = var_1058, weight = layers_2_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_25_cast_fp16)[name = tensor("pretrained_out_45_cast_fp16")]; - tensor var_1064 = const()[name = tensor("op_1064"), val = tensor([1, 1])]; - tensor var_1066 = const()[name = tensor("op_1066"), val = tensor([1, 1])]; - tensor input_65_pad_type_0 = const()[name = tensor("input_65_pad_type_0"), val = tensor("custom")]; - tensor input_65_pad_0 = const()[name = tensor("input_65_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165009344)))]; - tensor input_65_cast_fp16 = conv(dilations = var_1066, groups = var_971, pad = input_65_pad_0, pad_type = input_65_pad_type_0, strides = var_1064, weight = layers_2_self_attn_v_proj_loraA_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor("input_65_cast_fp16")]; - tensor var_1070 = const()[name = tensor("op_1070"), val = tensor([1, 1])]; - tensor var_1072 = const()[name = tensor("op_1072"), val = tensor([1, 1])]; - tensor lora_out_89_pad_type_0 = const()[name = tensor("lora_out_89_pad_type_0"), val = tensor("custom")]; - tensor lora_out_89_pad_0 = const()[name = tensor("lora_out_89_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_91_weight_0_to_fp16 = const()[name = tensor("lora_out_91_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165050368)))]; - tensor lora_out_91_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1072, groups = var_971, pad = lora_out_89_pad_0, pad_type = lora_out_89_pad_type_0, strides = var_1070, weight = lora_out_91_weight_0_to_fp16, x = input_65_cast_fp16)[name = tensor("lora_out_91_cast_fp16")]; - tensor current_value_5_cast_fp16 = add(x = pretrained_out_45_cast_fp16, y = lora_out_91_cast_fp16)[name = tensor("current_value_5_cast_fp16")]; - tensor var_1082_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_1082_cast_fp16")]; - tensor var_1084_cast_fp16 = mul(x = var_103_cast_fp16_2, y = var_295_cast_fp16)[name = tensor("op_1084_cast_fp16")]; - tensor key_9_cast_fp16 = add(x = var_1082_cast_fp16, y = var_1084_cast_fp16)[name = tensor("key_9_cast_fp16")]; - tensor var_1086_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_1086_cast_fp16")]; - tensor var_1088_cast_fp16 = mul(x = var_138_cast_fp16_2, y = var_295_cast_fp16)[name = tensor("op_1088_cast_fp16")]; - tensor value_9_cast_fp16 = add(x = var_1086_cast_fp16, y = var_1088_cast_fp16)[name = tensor("value_9_cast_fp16")]; - tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([1, 20, 64, -1])]; - tensor var_1092_cast_fp16 = reshape(shape = var_1091, x = query_9_cast_fp16)[name = tensor("op_1092_cast_fp16")]; - tensor var_1093_to_fp16 = const()[name = tensor("op_1093_to_fp16"), val = tensor(0x1p-3)]; - tensor var_1094_cast_fp16 = mul(x = var_1092_cast_fp16, y = var_1093_to_fp16)[name = tensor("op_1094_cast_fp16")]; - tensor var_1095 = const()[name = tensor("op_1095"), val = tensor([1, 20, 64, -1])]; - tensor var_1096_cast_fp16 = reshape(shape = var_1095, x = key_9_cast_fp16)[name = tensor("op_1096_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_1094_cast_fp16, y = var_1096_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; - tensor mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; - tensor var_1104_cast_fp16 = softmax(axis = var_964, x = mh_w_15_cast_fp16)[name = tensor("op_1104_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 = value_9_cast_fp16)[name = tensor("op_1106_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_1106_cast_fp16, y = var_1104_cast_fp16)[name = tensor("attn_9_cast_fp16")]; - tensor var_1109 = const()[name = tensor("op_1109"), val = tensor([1, 1280, 1, -1])]; - tensor input_67_cast_fp16 = reshape(shape = var_1109, x = attn_9_cast_fp16)[name = tensor("input_67_cast_fp16")]; - tensor var_1116 = const()[name = tensor("op_1116"), val = tensor([1, 1])]; - tensor var_1118 = const()[name = tensor("op_1118"), val = tensor([1, 1])]; - tensor pretrained_out_47_pad_type_0 = const()[name = tensor("pretrained_out_47_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_47_pad_0 = const()[name = tensor("pretrained_out_47_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165091392))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165910656))), name = tensor("layers_2_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_2_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165910784)))]; - tensor pretrained_out_47_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_1118, groups = var_971, pad = pretrained_out_47_pad_0, pad_type = pretrained_out_47_pad_type_0, strides = var_1116, weight = layers_2_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_67_cast_fp16)[name = tensor("pretrained_out_47_cast_fp16")]; - tensor var_1122 = const()[name = tensor("op_1122"), val = tensor([1, 1])]; - tensor var_1124 = const()[name = tensor("op_1124"), val = tensor([1, 1])]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("custom")]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165913408)))]; - tensor input_69_cast_fp16 = conv(dilations = var_1124, groups = var_971, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = var_1122, weight = layers_2_self_attn_o_proj_loraA_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; - tensor var_1128 = const()[name = tensor("op_1128"), val = tensor([1, 1])]; - tensor var_1130 = const()[name = tensor("op_1130"), val = tensor([1, 1])]; - tensor lora_out_93_pad_type_0 = const()[name = tensor("lora_out_93_pad_type_0"), val = tensor("custom")]; - tensor lora_out_93_pad_0 = const()[name = tensor("lora_out_93_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_95_weight_0_to_fp16 = const()[name = tensor("lora_out_95_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165954432)))]; - tensor lora_out_95_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1130, groups = var_971, pad = lora_out_93_pad_0, pad_type = lora_out_93_pad_type_0, strides = var_1128, weight = lora_out_95_weight_0_to_fp16, x = input_69_cast_fp16)[name = tensor("lora_out_95_cast_fp16")]; - tensor obj_31_cast_fp16 = add(x = pretrained_out_47_cast_fp16, y = lora_out_95_cast_fp16)[name = tensor("obj_31_cast_fp16")]; - tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_31_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; - tensor var_1143 = const()[name = tensor("op_1143"), val = tensor([1])]; - tensor channels_mean_15_cast_fp16 = reduce_mean(axes = var_1143, keep_dims = var_972, 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_1147 = const()[name = tensor("op_1147"), val = tensor([1])]; - tensor var_1148_cast_fp16 = reduce_mean(axes = var_1147, keep_dims = var_972, x = zero_mean_sq_15_cast_fp16)[name = tensor("op_1148_cast_fp16")]; - tensor var_1149_to_fp16 = const()[name = tensor("op_1149_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_1150_cast_fp16 = add(x = var_1148_cast_fp16, y = var_1149_to_fp16)[name = tensor("op_1150_cast_fp16")]; - tensor denom_15_epsilon_0 = const()[name = tensor("denom_15_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_15_cast_fp16 = rsqrt(epsilon = denom_15_epsilon_0, x = var_1150_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_33_gamma_0_to_fp16 = const()[name = tensor("obj_33_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165995456)))]; - tensor obj_33_beta_0_to_fp16 = const()[name = tensor("obj_33_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165998080)))]; - tensor obj_33_epsilon_0_to_fp16 = const()[name = tensor("obj_33_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_33_cast_fp16 = batch_norm(beta = obj_33_beta_0_to_fp16, epsilon = obj_33_epsilon_0_to_fp16, gamma = obj_33_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_33_cast_fp16")]; - tensor var_1168 = const()[name = tensor("op_1168"), val = tensor([1, 1])]; - tensor var_1170 = const()[name = tensor("op_1170"), val = tensor([1, 1])]; - tensor pretrained_out_49_pad_type_0 = const()[name = tensor("pretrained_out_49_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_49_pad_0 = const()[name = tensor("pretrained_out_49_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166000704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166819968))), name = tensor("layers_2_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_2_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166820096)))]; - tensor pretrained_out_49_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_1170, groups = var_971, pad = pretrained_out_49_pad_0, pad_type = pretrained_out_49_pad_type_0, strides = var_1168, weight = layers_2_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_33_cast_fp16)[name = tensor("pretrained_out_49_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 input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("custom")]; - tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166822720)))]; - tensor input_71_cast_fp16 = conv(dilations = var_1176, groups = var_971, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = var_1174, weight = layers_2_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_33_cast_fp16)[name = tensor("input_71_cast_fp16")]; - tensor var_1180 = const()[name = tensor("op_1180"), val = tensor([1, 1])]; - tensor var_1182 = const()[name = tensor("op_1182"), val = tensor([1, 1])]; - tensor lora_out_97_pad_type_0 = const()[name = tensor("lora_out_97_pad_type_0"), val = tensor("custom")]; - tensor lora_out_97_pad_0 = const()[name = tensor("lora_out_97_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_99_weight_0_to_fp16 = const()[name = tensor("lora_out_99_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166863744)))]; - tensor lora_out_99_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1182, groups = var_971, pad = lora_out_97_pad_0, pad_type = lora_out_97_pad_type_0, strides = var_1180, weight = lora_out_99_weight_0_to_fp16, x = input_71_cast_fp16)[name = tensor("lora_out_99_cast_fp16")]; - tensor query_11_cast_fp16 = add(x = pretrained_out_49_cast_fp16, y = lora_out_99_cast_fp16)[name = tensor("query_11_cast_fp16")]; - tensor var_1192 = const()[name = tensor("op_1192"), val = tensor([1, 1])]; - tensor var_1194 = const()[name = tensor("op_1194"), val = tensor([1, 1])]; - tensor pretrained_out_51_pad_type_0 = const()[name = tensor("pretrained_out_51_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_51_pad_0 = const()[name = tensor("pretrained_out_51_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166904768))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167724032))), name = tensor("layers_2_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_51_cast_fp16 = conv(dilations = var_1194, groups = var_971, pad = pretrained_out_51_pad_0, pad_type = pretrained_out_51_pad_type_0, strides = var_1192, weight = layers_2_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_51_cast_fp16")]; - tensor var_1198 = const()[name = tensor("op_1198"), val = tensor([1, 1])]; - tensor var_1200 = const()[name = tensor("op_1200"), val = tensor([1, 1])]; - tensor input_73_pad_type_0 = const()[name = tensor("input_73_pad_type_0"), val = tensor("custom")]; - tensor input_73_pad_0 = const()[name = tensor("input_73_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167724160)))]; - tensor input_73_cast_fp16 = conv(dilations = var_1200, groups = var_971, pad = input_73_pad_0, pad_type = input_73_pad_type_0, strides = var_1198, weight = layers_2_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_73_cast_fp16")]; - tensor var_1204 = const()[name = tensor("op_1204"), val = tensor([1, 1])]; - tensor var_1206 = const()[name = tensor("op_1206"), val = tensor([1, 1])]; - tensor lora_out_101_pad_type_0 = const()[name = tensor("lora_out_101_pad_type_0"), val = tensor("custom")]; - tensor lora_out_101_pad_0 = const()[name = tensor("lora_out_101_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_103_weight_0_to_fp16 = const()[name = tensor("lora_out_103_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167765184)))]; - tensor lora_out_103_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1206, groups = var_971, pad = lora_out_101_pad_0, pad_type = lora_out_101_pad_type_0, strides = var_1204, weight = lora_out_103_weight_0_to_fp16, x = input_73_cast_fp16)[name = tensor("lora_out_103_cast_fp16")]; - tensor key_11_cast_fp16 = add(x = pretrained_out_51_cast_fp16, y = lora_out_103_cast_fp16)[name = tensor("key_11_cast_fp16")]; - tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([1, 1])]; - tensor var_1219 = const()[name = tensor("op_1219"), val = tensor([1, 1])]; - tensor pretrained_out_53_pad_type_0 = const()[name = tensor("pretrained_out_53_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_53_pad_0 = const()[name = tensor("pretrained_out_53_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167806208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168625472))), name = tensor("layers_2_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_2_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168625600)))]; - tensor pretrained_out_53_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_1219, groups = var_971, pad = pretrained_out_53_pad_0, pad_type = pretrained_out_53_pad_type_0, strides = var_1217, weight = layers_2_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_53_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_75_pad_type_0 = const()[name = tensor("input_75_pad_type_0"), val = tensor("custom")]; - tensor input_75_pad_0 = const()[name = tensor("input_75_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168628224)))]; - tensor input_75_cast_fp16 = conv(dilations = var_1225, groups = var_971, pad = input_75_pad_0, pad_type = input_75_pad_type_0, strides = var_1223, weight = layers_2_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_75_cast_fp16")]; - tensor var_1229 = const()[name = tensor("op_1229"), val = tensor([1, 1])]; - tensor var_1231 = const()[name = tensor("op_1231"), val = tensor([1, 1])]; - tensor lora_out_105_pad_type_0 = const()[name = tensor("lora_out_105_pad_type_0"), val = tensor("custom")]; - tensor lora_out_105_pad_0 = const()[name = tensor("lora_out_105_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_107_weight_0_to_fp16 = const()[name = tensor("lora_out_107_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168669248)))]; - tensor lora_out_107_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1231, groups = var_971, pad = lora_out_105_pad_0, pad_type = lora_out_105_pad_type_0, strides = var_1229, weight = lora_out_107_weight_0_to_fp16, x = input_75_cast_fp16)[name = tensor("lora_out_107_cast_fp16")]; - tensor value_11_cast_fp16 = add(x = pretrained_out_53_cast_fp16, y = lora_out_107_cast_fp16)[name = tensor("value_11_cast_fp16")]; - tensor var_1238 = const()[name = tensor("op_1238"), val = tensor([1, 20, 64, -1])]; - tensor var_1239_cast_fp16 = reshape(shape = var_1238, x = query_11_cast_fp16)[name = tensor("op_1239_cast_fp16")]; - tensor var_1240_to_fp16 = const()[name = tensor("op_1240_to_fp16"), val = tensor(0x1p-3)]; - tensor var_1241_cast_fp16 = mul(x = var_1239_cast_fp16, y = var_1240_to_fp16)[name = tensor("op_1241_cast_fp16")]; - tensor var_1242 = const()[name = tensor("op_1242"), val = tensor([1, 20, 64, -1])]; - tensor var_1243_cast_fp16 = reshape(shape = var_1242, x = key_11_cast_fp16)[name = tensor("op_1243_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_1241_cast_fp16, y = var_1243_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; - tensor var_1246_cast_fp16 = softmax(axis = var_964, x = mh_w_17_cast_fp16)[name = tensor("op_1246_cast_fp16")]; - tensor var_1247 = const()[name = tensor("op_1247"), val = tensor([1, 20, 64, -1])]; - tensor var_1248_cast_fp16 = reshape(shape = var_1247, x = value_11_cast_fp16)[name = tensor("op_1248_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_1248_cast_fp16, y = var_1246_cast_fp16)[name = tensor("attn_11_cast_fp16")]; - tensor var_1251 = const()[name = tensor("op_1251"), val = tensor([1, 1280, 1, -1])]; - tensor input_77_cast_fp16 = reshape(shape = var_1251, x = attn_11_cast_fp16)[name = tensor("input_77_cast_fp16")]; - tensor var_1258 = const()[name = tensor("op_1258"), val = tensor([1, 1])]; - tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([1, 1])]; - tensor pretrained_out_55_pad_type_0 = const()[name = tensor("pretrained_out_55_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_55_pad_0 = const()[name = tensor("pretrained_out_55_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168710272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169529536))), name = tensor("layers_2_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_2_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169529664)))]; - tensor pretrained_out_55_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_1260, groups = var_971, pad = pretrained_out_55_pad_0, pad_type = pretrained_out_55_pad_type_0, strides = var_1258, weight = layers_2_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_77_cast_fp16)[name = tensor("pretrained_out_55_cast_fp16")]; - tensor var_1264 = const()[name = tensor("op_1264"), val = tensor([1, 1])]; - tensor var_1266 = const()[name = tensor("op_1266"), val = tensor([1, 1])]; - tensor input_79_pad_type_0 = const()[name = tensor("input_79_pad_type_0"), val = tensor("custom")]; - tensor input_79_pad_0 = const()[name = tensor("input_79_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169532288)))]; - tensor input_79_cast_fp16 = conv(dilations = var_1266, groups = var_971, pad = input_79_pad_0, pad_type = input_79_pad_type_0, strides = var_1264, weight = layers_2_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; - tensor var_1270 = const()[name = tensor("op_1270"), val = tensor([1, 1])]; - tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([1, 1])]; - tensor lora_out_109_pad_type_0 = const()[name = tensor("lora_out_109_pad_type_0"), val = tensor("custom")]; - tensor lora_out_109_pad_0 = const()[name = tensor("lora_out_109_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_111_weight_0_to_fp16 = const()[name = tensor("lora_out_111_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169573312)))]; - tensor lora_out_111_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1272, groups = var_971, pad = lora_out_109_pad_0, pad_type = lora_out_109_pad_type_0, strides = var_1270, weight = lora_out_111_weight_0_to_fp16, x = input_79_cast_fp16)[name = tensor("lora_out_111_cast_fp16")]; - tensor obj_35_cast_fp16 = add(x = pretrained_out_55_cast_fp16, y = lora_out_111_cast_fp16)[name = tensor("obj_35_cast_fp16")]; - tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_35_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; - tensor var_1281 = const()[name = tensor("op_1281"), val = tensor([1])]; - tensor channels_mean_17_cast_fp16 = reduce_mean(axes = var_1281, keep_dims = var_972, 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_1285 = const()[name = tensor("op_1285"), val = tensor([1])]; - tensor var_1286_cast_fp16 = reduce_mean(axes = var_1285, keep_dims = var_972, x = zero_mean_sq_17_cast_fp16)[name = tensor("op_1286_cast_fp16")]; - tensor var_1287_to_fp16 = const()[name = tensor("op_1287_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_1288_cast_fp16 = add(x = var_1286_cast_fp16, y = var_1287_to_fp16)[name = tensor("op_1288_cast_fp16")]; - tensor denom_17_epsilon_0 = const()[name = tensor("denom_17_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_17_cast_fp16 = rsqrt(epsilon = denom_17_epsilon_0, x = var_1288_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_81_gamma_0_to_fp16 = const()[name = tensor("input_81_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169614336)))]; - tensor input_81_beta_0_to_fp16 = const()[name = tensor("input_81_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169616960)))]; - tensor input_81_epsilon_0_to_fp16 = const()[name = tensor("input_81_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_81_cast_fp16 = batch_norm(beta = input_81_beta_0_to_fp16, epsilon = input_81_epsilon_0_to_fp16, gamma = input_81_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_81_cast_fp16")]; - tensor var_1302 = const()[name = tensor("op_1302"), val = tensor([1, 1])]; - tensor var_1304 = const()[name = tensor("op_1304"), val = tensor([1, 1])]; - tensor pretrained_out_57_pad_type_0 = const()[name = tensor("pretrained_out_57_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_57_pad_0 = const()[name = tensor("pretrained_out_57_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169619584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172896448))), name = tensor("layers_2_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_2_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172896576)))]; - tensor pretrained_out_57_cast_fp16 = conv(bias = layers_2_fc1_pretrained_bias_to_fp16, dilations = var_1304, groups = var_971, pad = pretrained_out_57_pad_0, pad_type = pretrained_out_57_pad_type_0, strides = var_1302, weight = layers_2_fc1_pretrained_weight_to_fp16_palettized, x = input_81_cast_fp16)[name = tensor("pretrained_out_57_cast_fp16")]; - tensor var_1308 = const()[name = tensor("op_1308"), val = tensor([1, 1])]; - tensor var_1310 = const()[name = tensor("op_1310"), val = tensor([1, 1])]; - tensor input_83_pad_type_0 = const()[name = tensor("input_83_pad_type_0"), val = tensor("custom")]; - tensor input_83_pad_0 = const()[name = tensor("input_83_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_2_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172906880)))]; - tensor input_83_cast_fp16 = conv(dilations = var_1310, groups = var_971, pad = input_83_pad_0, pad_type = input_83_pad_type_0, strides = var_1308, weight = layers_2_fc1_loraA_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; - tensor var_1314 = const()[name = tensor("op_1314"), val = tensor([1, 1])]; - tensor var_1316 = const()[name = tensor("op_1316"), val = tensor([1, 1])]; - tensor lora_out_113_pad_type_0 = const()[name = tensor("lora_out_113_pad_type_0"), val = tensor("custom")]; - tensor lora_out_113_pad_0 = const()[name = tensor("lora_out_113_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_115_weight_0_to_fp16 = const()[name = tensor("lora_out_115_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172947904)))]; - tensor lora_out_115_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_1316, groups = var_971, pad = lora_out_113_pad_0, pad_type = lora_out_113_pad_type_0, strides = var_1314, weight = lora_out_115_weight_0_to_fp16, x = input_83_cast_fp16)[name = tensor("lora_out_115_cast_fp16")]; - tensor input_85_cast_fp16 = add(x = pretrained_out_57_cast_fp16, y = lora_out_115_cast_fp16)[name = tensor("input_85_cast_fp16")]; - tensor input_87_mode_0 = const()[name = tensor("input_87_mode_0"), val = tensor("EXACT")]; - tensor input_87_cast_fp16 = gelu(mode = input_87_mode_0, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; - tensor var_1328 = const()[name = tensor("op_1328"), val = tensor([1, 1])]; - tensor var_1330 = const()[name = tensor("op_1330"), val = tensor([1, 1])]; - tensor pretrained_out_59_pad_type_0 = const()[name = tensor("pretrained_out_59_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_59_pad_0 = const()[name = tensor("pretrained_out_59_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173111808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176388672))), name = tensor("layers_2_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_2_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176388800)))]; - tensor pretrained_out_59_cast_fp16 = conv(bias = layers_2_fc2_pretrained_bias_to_fp16, dilations = var_1330, groups = var_971, pad = pretrained_out_59_pad_0, pad_type = pretrained_out_59_pad_type_0, strides = var_1328, weight = layers_2_fc2_pretrained_weight_to_fp16_palettized, x = input_87_cast_fp16)[name = tensor("pretrained_out_59_cast_fp16")]; - tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 1])]; - tensor var_1336 = const()[name = tensor("op_1336"), val = tensor([1, 1])]; - tensor input_89_pad_type_0 = const()[name = tensor("input_89_pad_type_0"), val = tensor("custom")]; - tensor input_89_pad_0 = const()[name = tensor("input_89_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_2_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_2_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176391424)))]; - tensor input_89_cast_fp16 = conv(dilations = var_1336, groups = var_971, pad = input_89_pad_0, pad_type = input_89_pad_type_0, strides = var_1334, weight = layers_2_fc2_loraA_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; - tensor var_1340 = const()[name = tensor("op_1340"), val = tensor([1, 1])]; - tensor var_1342 = const()[name = tensor("op_1342"), val = tensor([1, 1])]; - tensor lora_out_117_pad_type_0 = const()[name = tensor("lora_out_117_pad_type_0"), val = tensor("custom")]; - tensor lora_out_117_pad_0 = const()[name = tensor("lora_out_117_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_119_weight_0_to_fp16 = const()[name = tensor("lora_out_119_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176555328)))]; - tensor lora_out_119_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1342, groups = var_971, pad = lora_out_117_pad_0, pad_type = lora_out_117_pad_type_0, strides = var_1340, weight = lora_out_119_weight_0_to_fp16, x = input_89_cast_fp16)[name = tensor("lora_out_119_cast_fp16")]; - tensor hidden_states_7_cast_fp16 = add(x = pretrained_out_59_cast_fp16, y = lora_out_119_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_1358 = const()[name = tensor("op_1358"), val = tensor(3)]; - tensor var_1365 = const()[name = tensor("op_1365"), val = tensor(1)]; - tensor var_1366 = const()[name = tensor("op_1366"), val = tensor(true)]; - tensor var_1378 = const()[name = tensor("op_1378"), val = tensor([1])]; - tensor channels_mean_19_cast_fp16 = reduce_mean(axes = var_1378, keep_dims = var_1366, 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_1382 = const()[name = tensor("op_1382"), val = tensor([1])]; - tensor var_1383_cast_fp16 = reduce_mean(axes = var_1382, keep_dims = var_1366, x = zero_mean_sq_19_cast_fp16)[name = tensor("op_1383_cast_fp16")]; - tensor var_1384_to_fp16 = const()[name = tensor("op_1384_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_1385_cast_fp16 = add(x = var_1383_cast_fp16, y = var_1384_to_fp16)[name = tensor("op_1385_cast_fp16")]; - tensor denom_19_epsilon_0 = const()[name = tensor("denom_19_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_19_cast_fp16 = rsqrt(epsilon = denom_19_epsilon_0, x = var_1385_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_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(176596352)))]; - 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(176598976)))]; - 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_19_cast_fp16)[name = tensor("obj_37_cast_fp16")]; - tensor var_1403 = const()[name = tensor("op_1403"), val = tensor([1, 1])]; - tensor var_1405 = const()[name = tensor("op_1405"), val = tensor([1, 1])]; - tensor pretrained_out_61_pad_type_0 = const()[name = tensor("pretrained_out_61_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_61_pad_0 = const()[name = tensor("pretrained_out_61_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176601600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177420864))), name = tensor("layers_3_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_3_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177420992)))]; - tensor pretrained_out_61_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_1405, groups = var_1365, pad = pretrained_out_61_pad_0, pad_type = pretrained_out_61_pad_type_0, strides = var_1403, weight = layers_3_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor("pretrained_out_61_cast_fp16")]; - tensor var_1409 = const()[name = tensor("op_1409"), val = tensor([1, 1])]; - tensor var_1411 = const()[name = tensor("op_1411"), val = tensor([1, 1])]; - tensor input_91_pad_type_0 = const()[name = tensor("input_91_pad_type_0"), val = tensor("custom")]; - tensor input_91_pad_0 = const()[name = tensor("input_91_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177423616)))]; - tensor input_91_cast_fp16 = conv(dilations = var_1411, groups = var_1365, pad = input_91_pad_0, pad_type = input_91_pad_type_0, strides = var_1409, weight = layers_3_self_attn_q_proj_loraA_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("input_91_cast_fp16")]; - tensor var_1415 = const()[name = tensor("op_1415"), val = tensor([1, 1])]; - tensor var_1417 = const()[name = tensor("op_1417"), val = tensor([1, 1])]; - tensor lora_out_121_pad_type_0 = const()[name = tensor("lora_out_121_pad_type_0"), val = tensor("custom")]; - tensor lora_out_121_pad_0 = const()[name = tensor("lora_out_121_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_123_weight_0_to_fp16 = const()[name = tensor("lora_out_123_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177464640)))]; - tensor lora_out_123_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1417, groups = var_1365, pad = lora_out_121_pad_0, pad_type = lora_out_121_pad_type_0, strides = var_1415, weight = lora_out_123_weight_0_to_fp16, x = input_91_cast_fp16)[name = tensor("lora_out_123_cast_fp16")]; - tensor query_13_cast_fp16 = add(x = pretrained_out_61_cast_fp16, y = lora_out_123_cast_fp16)[name = tensor("query_13_cast_fp16")]; - tensor var_1427 = const()[name = tensor("op_1427"), val = tensor([1, 1])]; - tensor var_1429 = const()[name = tensor("op_1429"), val = tensor([1, 1])]; - tensor pretrained_out_63_pad_type_0 = const()[name = tensor("pretrained_out_63_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_63_pad_0 = const()[name = tensor("pretrained_out_63_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177505664))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178324928))), name = tensor("layers_3_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_63_cast_fp16 = conv(dilations = var_1429, groups = var_1365, pad = pretrained_out_63_pad_0, pad_type = pretrained_out_63_pad_type_0, strides = var_1427, weight = layers_3_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor("pretrained_out_63_cast_fp16")]; - tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([1, 1])]; - tensor var_1435 = const()[name = tensor("op_1435"), val = tensor([1, 1])]; - tensor input_93_pad_type_0 = const()[name = tensor("input_93_pad_type_0"), val = tensor("custom")]; - tensor input_93_pad_0 = const()[name = tensor("input_93_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178325056)))]; - tensor input_93_cast_fp16 = conv(dilations = var_1435, groups = var_1365, pad = input_93_pad_0, pad_type = input_93_pad_type_0, strides = var_1433, weight = layers_3_self_attn_k_proj_loraA_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("input_93_cast_fp16")]; - tensor var_1439 = const()[name = tensor("op_1439"), val = tensor([1, 1])]; - tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([1, 1])]; - tensor lora_out_125_pad_type_0 = const()[name = tensor("lora_out_125_pad_type_0"), val = tensor("custom")]; - tensor lora_out_125_pad_0 = const()[name = tensor("lora_out_125_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_127_weight_0_to_fp16 = const()[name = tensor("lora_out_127_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178366080)))]; - tensor lora_out_127_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1441, groups = var_1365, pad = lora_out_125_pad_0, pad_type = lora_out_125_pad_type_0, strides = var_1439, weight = lora_out_127_weight_0_to_fp16, x = input_93_cast_fp16)[name = tensor("lora_out_127_cast_fp16")]; - tensor current_key_7_cast_fp16 = add(x = pretrained_out_63_cast_fp16, y = lora_out_127_cast_fp16)[name = tensor("current_key_7_cast_fp16")]; - tensor var_1452 = const()[name = tensor("op_1452"), val = tensor([1, 1])]; - tensor var_1454 = const()[name = tensor("op_1454"), val = tensor([1, 1])]; - tensor pretrained_out_65_pad_type_0 = const()[name = tensor("pretrained_out_65_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_65_pad_0 = const()[name = tensor("pretrained_out_65_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178407104))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179226368))), name = tensor("layers_3_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_3_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179226496)))]; - tensor pretrained_out_65_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_1454, groups = var_1365, pad = pretrained_out_65_pad_0, pad_type = pretrained_out_65_pad_type_0, strides = var_1452, weight = layers_3_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor("pretrained_out_65_cast_fp16")]; - tensor var_1458 = const()[name = tensor("op_1458"), val = tensor([1, 1])]; - tensor var_1460 = const()[name = tensor("op_1460"), val = tensor([1, 1])]; - tensor input_95_pad_type_0 = const()[name = tensor("input_95_pad_type_0"), val = tensor("custom")]; - tensor input_95_pad_0 = const()[name = tensor("input_95_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179229120)))]; - tensor input_95_cast_fp16 = conv(dilations = var_1460, groups = var_1365, pad = input_95_pad_0, pad_type = input_95_pad_type_0, strides = var_1458, weight = layers_3_self_attn_v_proj_loraA_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("input_95_cast_fp16")]; - tensor var_1464 = const()[name = tensor("op_1464"), val = tensor([1, 1])]; - tensor var_1466 = const()[name = tensor("op_1466"), val = tensor([1, 1])]; - tensor lora_out_129_pad_type_0 = const()[name = tensor("lora_out_129_pad_type_0"), val = tensor("custom")]; - tensor lora_out_129_pad_0 = const()[name = tensor("lora_out_129_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_131_weight_0_to_fp16 = const()[name = tensor("lora_out_131_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179270144)))]; - tensor lora_out_131_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1466, groups = var_1365, pad = lora_out_129_pad_0, pad_type = lora_out_129_pad_type_0, strides = var_1464, weight = lora_out_131_weight_0_to_fp16, x = input_95_cast_fp16)[name = tensor("lora_out_131_cast_fp16")]; - tensor current_value_7_cast_fp16 = add(x = pretrained_out_65_cast_fp16, y = lora_out_131_cast_fp16)[name = tensor("current_value_7_cast_fp16")]; - tensor var_1476_cast_fp16 = mul(x = current_key_7_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_1476_cast_fp16")]; - tensor var_1478_cast_fp16 = mul(x = var_103_cast_fp16_3, y = var_295_cast_fp16)[name = tensor("op_1478_cast_fp16")]; - tensor key_13_cast_fp16 = add(x = var_1476_cast_fp16, y = var_1478_cast_fp16)[name = tensor("key_13_cast_fp16")]; - tensor var_1480_cast_fp16 = mul(x = current_value_7_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_1480_cast_fp16")]; - tensor var_1482_cast_fp16 = mul(x = var_138_cast_fp16_3, y = var_295_cast_fp16)[name = tensor("op_1482_cast_fp16")]; - tensor value_13_cast_fp16 = add(x = var_1480_cast_fp16, y = var_1482_cast_fp16)[name = tensor("value_13_cast_fp16")]; - tensor var_1485 = const()[name = tensor("op_1485"), val = tensor([1, 20, 64, -1])]; - tensor var_1486_cast_fp16 = reshape(shape = var_1485, x = query_13_cast_fp16)[name = tensor("op_1486_cast_fp16")]; - tensor var_1487_to_fp16 = const()[name = tensor("op_1487_to_fp16"), val = tensor(0x1p-3)]; - tensor var_1488_cast_fp16 = mul(x = var_1486_cast_fp16, y = var_1487_to_fp16)[name = tensor("op_1488_cast_fp16")]; - tensor var_1489 = const()[name = tensor("op_1489"), val = tensor([1, 20, 64, -1])]; - tensor var_1490_cast_fp16 = reshape(shape = var_1489, x = key_13_cast_fp16)[name = tensor("op_1490_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_1488_cast_fp16, y = var_1490_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; - tensor mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; - tensor var_1498_cast_fp16 = softmax(axis = var_1358, x = mh_w_21_cast_fp16)[name = tensor("op_1498_cast_fp16")]; - tensor var_1499 = const()[name = tensor("op_1499"), val = tensor([1, 20, 64, -1])]; - tensor var_1500_cast_fp16 = reshape(shape = var_1499, x = value_13_cast_fp16)[name = tensor("op_1500_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_1500_cast_fp16, y = var_1498_cast_fp16)[name = tensor("attn_13_cast_fp16")]; - tensor var_1503 = const()[name = tensor("op_1503"), val = tensor([1, 1280, 1, -1])]; - tensor input_97_cast_fp16 = reshape(shape = var_1503, x = attn_13_cast_fp16)[name = tensor("input_97_cast_fp16")]; - tensor var_1510 = const()[name = tensor("op_1510"), val = tensor([1, 1])]; - tensor var_1512 = const()[name = tensor("op_1512"), val = tensor([1, 1])]; - tensor pretrained_out_67_pad_type_0 = const()[name = tensor("pretrained_out_67_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_67_pad_0 = const()[name = tensor("pretrained_out_67_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179311168))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180130432))), name = tensor("layers_3_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_3_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180130560)))]; - tensor pretrained_out_67_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_1512, groups = var_1365, pad = pretrained_out_67_pad_0, pad_type = pretrained_out_67_pad_type_0, strides = var_1510, weight = layers_3_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_97_cast_fp16)[name = tensor("pretrained_out_67_cast_fp16")]; - tensor var_1516 = const()[name = tensor("op_1516"), val = tensor([1, 1])]; - tensor var_1518 = const()[name = tensor("op_1518"), val = tensor([1, 1])]; - tensor input_99_pad_type_0 = const()[name = tensor("input_99_pad_type_0"), val = tensor("custom")]; - tensor input_99_pad_0 = const()[name = tensor("input_99_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180133184)))]; - tensor input_99_cast_fp16 = conv(dilations = var_1518, groups = var_1365, pad = input_99_pad_0, pad_type = input_99_pad_type_0, strides = var_1516, weight = layers_3_self_attn_o_proj_loraA_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; - tensor var_1522 = const()[name = tensor("op_1522"), val = tensor([1, 1])]; - tensor var_1524 = const()[name = tensor("op_1524"), val = tensor([1, 1])]; - tensor lora_out_133_pad_type_0 = const()[name = tensor("lora_out_133_pad_type_0"), val = tensor("custom")]; - tensor lora_out_133_pad_0 = const()[name = tensor("lora_out_133_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_135_weight_0_to_fp16 = const()[name = tensor("lora_out_135_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180174208)))]; - tensor lora_out_135_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1524, groups = var_1365, pad = lora_out_133_pad_0, pad_type = lora_out_133_pad_type_0, strides = var_1522, weight = lora_out_135_weight_0_to_fp16, x = input_99_cast_fp16)[name = tensor("lora_out_135_cast_fp16")]; - tensor obj_43_cast_fp16 = add(x = pretrained_out_67_cast_fp16, y = lora_out_135_cast_fp16)[name = tensor("obj_43_cast_fp16")]; - tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_43_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; - tensor var_1537 = const()[name = tensor("op_1537"), val = tensor([1])]; - tensor channels_mean_21_cast_fp16 = reduce_mean(axes = var_1537, keep_dims = var_1366, 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_1541 = const()[name = tensor("op_1541"), val = tensor([1])]; - tensor var_1542_cast_fp16 = reduce_mean(axes = var_1541, keep_dims = var_1366, x = zero_mean_sq_21_cast_fp16)[name = tensor("op_1542_cast_fp16")]; - tensor var_1543_to_fp16 = const()[name = tensor("op_1543_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_1544_cast_fp16 = add(x = var_1542_cast_fp16, y = var_1543_to_fp16)[name = tensor("op_1544_cast_fp16")]; - tensor denom_21_epsilon_0 = const()[name = tensor("denom_21_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_21_cast_fp16 = rsqrt(epsilon = denom_21_epsilon_0, x = var_1544_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_45_gamma_0_to_fp16 = const()[name = tensor("obj_45_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180215232)))]; - tensor obj_45_beta_0_to_fp16 = const()[name = tensor("obj_45_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180217856)))]; - tensor obj_45_epsilon_0_to_fp16 = const()[name = tensor("obj_45_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_45_cast_fp16 = batch_norm(beta = obj_45_beta_0_to_fp16, epsilon = obj_45_epsilon_0_to_fp16, gamma = obj_45_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_45_cast_fp16")]; - tensor var_1562 = const()[name = tensor("op_1562"), val = tensor([1, 1])]; - tensor var_1564 = const()[name = tensor("op_1564"), val = tensor([1, 1])]; - tensor pretrained_out_69_pad_type_0 = const()[name = tensor("pretrained_out_69_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_69_pad_0 = const()[name = tensor("pretrained_out_69_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180220480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181039744))), name = tensor("layers_3_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_3_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181039872)))]; - tensor pretrained_out_69_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_1564, groups = var_1365, pad = pretrained_out_69_pad_0, pad_type = pretrained_out_69_pad_type_0, strides = var_1562, weight = layers_3_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_45_cast_fp16)[name = tensor("pretrained_out_69_cast_fp16")]; - tensor var_1568 = const()[name = tensor("op_1568"), val = tensor([1, 1])]; - tensor var_1570 = const()[name = tensor("op_1570"), val = tensor([1, 1])]; - tensor input_101_pad_type_0 = const()[name = tensor("input_101_pad_type_0"), val = tensor("custom")]; - tensor input_101_pad_0 = const()[name = tensor("input_101_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181042496)))]; - tensor input_101_cast_fp16 = conv(dilations = var_1570, groups = var_1365, pad = input_101_pad_0, pad_type = input_101_pad_type_0, strides = var_1568, weight = layers_3_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_45_cast_fp16)[name = tensor("input_101_cast_fp16")]; - tensor var_1574 = const()[name = tensor("op_1574"), val = tensor([1, 1])]; - tensor var_1576 = const()[name = tensor("op_1576"), val = tensor([1, 1])]; - tensor lora_out_137_pad_type_0 = const()[name = tensor("lora_out_137_pad_type_0"), val = tensor("custom")]; - tensor lora_out_137_pad_0 = const()[name = tensor("lora_out_137_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_139_weight_0_to_fp16 = const()[name = tensor("lora_out_139_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181083520)))]; - tensor lora_out_139_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1576, groups = var_1365, pad = lora_out_137_pad_0, pad_type = lora_out_137_pad_type_0, strides = var_1574, weight = lora_out_139_weight_0_to_fp16, x = input_101_cast_fp16)[name = tensor("lora_out_139_cast_fp16")]; - tensor query_15_cast_fp16 = add(x = pretrained_out_69_cast_fp16, y = lora_out_139_cast_fp16)[name = tensor("query_15_cast_fp16")]; - tensor var_1586 = const()[name = tensor("op_1586"), val = tensor([1, 1])]; - tensor var_1588 = const()[name = tensor("op_1588"), val = tensor([1, 1])]; - tensor pretrained_out_71_pad_type_0 = const()[name = tensor("pretrained_out_71_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_71_pad_0 = const()[name = tensor("pretrained_out_71_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181124544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181943808))), name = tensor("layers_3_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_71_cast_fp16 = conv(dilations = var_1588, groups = var_1365, pad = pretrained_out_71_pad_0, pad_type = pretrained_out_71_pad_type_0, strides = var_1586, weight = layers_3_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_71_cast_fp16")]; - tensor var_1592 = const()[name = tensor("op_1592"), val = tensor([1, 1])]; - tensor var_1594 = const()[name = tensor("op_1594"), val = tensor([1, 1])]; - tensor input_103_pad_type_0 = const()[name = tensor("input_103_pad_type_0"), val = tensor("custom")]; - tensor input_103_pad_0 = const()[name = tensor("input_103_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181943936)))]; - tensor input_103_cast_fp16 = conv(dilations = var_1594, groups = var_1365, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = var_1592, weight = layers_3_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_103_cast_fp16")]; - tensor var_1598 = const()[name = tensor("op_1598"), val = tensor([1, 1])]; - tensor var_1600 = const()[name = tensor("op_1600"), val = tensor([1, 1])]; - tensor lora_out_141_pad_type_0 = const()[name = tensor("lora_out_141_pad_type_0"), val = tensor("custom")]; - tensor lora_out_141_pad_0 = const()[name = tensor("lora_out_141_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_143_weight_0_to_fp16 = const()[name = tensor("lora_out_143_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181984960)))]; - tensor lora_out_143_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1600, groups = var_1365, pad = lora_out_141_pad_0, pad_type = lora_out_141_pad_type_0, strides = var_1598, weight = lora_out_143_weight_0_to_fp16, x = input_103_cast_fp16)[name = tensor("lora_out_143_cast_fp16")]; - tensor key_15_cast_fp16 = add(x = pretrained_out_71_cast_fp16, y = lora_out_143_cast_fp16)[name = tensor("key_15_cast_fp16")]; - tensor var_1611 = const()[name = tensor("op_1611"), val = tensor([1, 1])]; - tensor var_1613 = const()[name = tensor("op_1613"), val = tensor([1, 1])]; - tensor pretrained_out_73_pad_type_0 = const()[name = tensor("pretrained_out_73_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_73_pad_0 = const()[name = tensor("pretrained_out_73_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182025984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182845248))), name = tensor("layers_3_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_3_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182845376)))]; - tensor pretrained_out_73_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_1613, groups = var_1365, pad = pretrained_out_73_pad_0, pad_type = pretrained_out_73_pad_type_0, strides = var_1611, weight = layers_3_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_73_cast_fp16")]; - tensor var_1617 = const()[name = tensor("op_1617"), val = tensor([1, 1])]; - tensor var_1619 = const()[name = tensor("op_1619"), val = tensor([1, 1])]; - tensor input_105_pad_type_0 = const()[name = tensor("input_105_pad_type_0"), val = tensor("custom")]; - tensor input_105_pad_0 = const()[name = tensor("input_105_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182848000)))]; - tensor input_105_cast_fp16 = conv(dilations = var_1619, groups = var_1365, pad = input_105_pad_0, pad_type = input_105_pad_type_0, strides = var_1617, weight = layers_3_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_105_cast_fp16")]; - tensor var_1623 = const()[name = tensor("op_1623"), val = tensor([1, 1])]; - tensor var_1625 = const()[name = tensor("op_1625"), val = tensor([1, 1])]; - tensor lora_out_145_pad_type_0 = const()[name = tensor("lora_out_145_pad_type_0"), val = tensor("custom")]; - tensor lora_out_145_pad_0 = const()[name = tensor("lora_out_145_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_147_weight_0_to_fp16 = const()[name = tensor("lora_out_147_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182889024)))]; - tensor lora_out_147_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1625, groups = var_1365, pad = lora_out_145_pad_0, pad_type = lora_out_145_pad_type_0, strides = var_1623, weight = lora_out_147_weight_0_to_fp16, x = input_105_cast_fp16)[name = tensor("lora_out_147_cast_fp16")]; - tensor value_15_cast_fp16 = add(x = pretrained_out_73_cast_fp16, y = lora_out_147_cast_fp16)[name = tensor("value_15_cast_fp16")]; - tensor var_1632 = const()[name = tensor("op_1632"), val = tensor([1, 20, 64, -1])]; - tensor var_1633_cast_fp16 = reshape(shape = var_1632, x = query_15_cast_fp16)[name = tensor("op_1633_cast_fp16")]; - tensor var_1634_to_fp16 = const()[name = tensor("op_1634_to_fp16"), val = tensor(0x1p-3)]; - tensor var_1635_cast_fp16 = mul(x = var_1633_cast_fp16, y = var_1634_to_fp16)[name = tensor("op_1635_cast_fp16")]; - tensor var_1636 = const()[name = tensor("op_1636"), val = tensor([1, 20, 64, -1])]; - tensor var_1637_cast_fp16 = reshape(shape = var_1636, x = key_15_cast_fp16)[name = tensor("op_1637_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_1635_cast_fp16, y = var_1637_cast_fp16)[name = tensor("mh_w_23_cast_fp16")]; - tensor var_1640_cast_fp16 = softmax(axis = var_1358, x = mh_w_23_cast_fp16)[name = tensor("op_1640_cast_fp16")]; - tensor var_1641 = const()[name = tensor("op_1641"), val = tensor([1, 20, 64, -1])]; - tensor var_1642_cast_fp16 = reshape(shape = var_1641, x = value_15_cast_fp16)[name = tensor("op_1642_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_1642_cast_fp16, y = var_1640_cast_fp16)[name = tensor("attn_15_cast_fp16")]; - tensor var_1645 = const()[name = tensor("op_1645"), val = tensor([1, 1280, 1, -1])]; - tensor input_107_cast_fp16 = reshape(shape = var_1645, x = attn_15_cast_fp16)[name = tensor("input_107_cast_fp16")]; - tensor var_1652 = const()[name = tensor("op_1652"), val = tensor([1, 1])]; - tensor var_1654 = const()[name = tensor("op_1654"), val = tensor([1, 1])]; - tensor pretrained_out_75_pad_type_0 = const()[name = tensor("pretrained_out_75_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_75_pad_0 = const()[name = tensor("pretrained_out_75_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182930048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183749312))), name = tensor("layers_3_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_3_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183749440)))]; - tensor pretrained_out_75_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_1654, groups = var_1365, pad = pretrained_out_75_pad_0, pad_type = pretrained_out_75_pad_type_0, strides = var_1652, weight = layers_3_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_107_cast_fp16)[name = tensor("pretrained_out_75_cast_fp16")]; - tensor var_1658 = const()[name = tensor("op_1658"), val = tensor([1, 1])]; - tensor var_1660 = const()[name = tensor("op_1660"), val = tensor([1, 1])]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("custom")]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183752064)))]; - tensor input_109_cast_fp16 = conv(dilations = var_1660, groups = var_1365, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = var_1658, weight = layers_3_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; - tensor var_1664 = const()[name = tensor("op_1664"), val = tensor([1, 1])]; - tensor var_1666 = const()[name = tensor("op_1666"), val = tensor([1, 1])]; - tensor lora_out_149_pad_type_0 = const()[name = tensor("lora_out_149_pad_type_0"), val = tensor("custom")]; - tensor lora_out_149_pad_0 = const()[name = tensor("lora_out_149_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_151_weight_0_to_fp16 = const()[name = tensor("lora_out_151_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183793088)))]; - tensor lora_out_151_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1666, groups = var_1365, pad = lora_out_149_pad_0, pad_type = lora_out_149_pad_type_0, strides = var_1664, weight = lora_out_151_weight_0_to_fp16, x = input_109_cast_fp16)[name = tensor("lora_out_151_cast_fp16")]; - tensor obj_47_cast_fp16 = add(x = pretrained_out_75_cast_fp16, y = lora_out_151_cast_fp16)[name = tensor("obj_47_cast_fp16")]; - tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_47_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; - tensor var_1675 = const()[name = tensor("op_1675"), val = tensor([1])]; - tensor channels_mean_23_cast_fp16 = reduce_mean(axes = var_1675, keep_dims = var_1366, 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_1679 = const()[name = tensor("op_1679"), val = tensor([1])]; - tensor var_1680_cast_fp16 = reduce_mean(axes = var_1679, keep_dims = var_1366, x = zero_mean_sq_23_cast_fp16)[name = tensor("op_1680_cast_fp16")]; - tensor var_1681_to_fp16 = const()[name = tensor("op_1681_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_1682_cast_fp16 = add(x = var_1680_cast_fp16, y = var_1681_to_fp16)[name = tensor("op_1682_cast_fp16")]; - tensor denom_23_epsilon_0 = const()[name = tensor("denom_23_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_23_cast_fp16 = rsqrt(epsilon = denom_23_epsilon_0, x = var_1682_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_111_gamma_0_to_fp16 = const()[name = tensor("input_111_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183834112)))]; - tensor input_111_beta_0_to_fp16 = const()[name = tensor("input_111_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183836736)))]; - tensor input_111_epsilon_0_to_fp16 = const()[name = tensor("input_111_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_111_cast_fp16 = batch_norm(beta = input_111_beta_0_to_fp16, epsilon = input_111_epsilon_0_to_fp16, gamma = input_111_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_111_cast_fp16")]; - tensor var_1696 = const()[name = tensor("op_1696"), val = tensor([1, 1])]; - tensor var_1698 = const()[name = tensor("op_1698"), val = tensor([1, 1])]; - tensor pretrained_out_77_pad_type_0 = const()[name = tensor("pretrained_out_77_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_77_pad_0 = const()[name = tensor("pretrained_out_77_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183839360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187116224))), name = tensor("layers_3_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_3_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187116352)))]; - tensor pretrained_out_77_cast_fp16 = conv(bias = layers_3_fc1_pretrained_bias_to_fp16, dilations = var_1698, groups = var_1365, pad = pretrained_out_77_pad_0, pad_type = pretrained_out_77_pad_type_0, strides = var_1696, weight = layers_3_fc1_pretrained_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = tensor("pretrained_out_77_cast_fp16")]; - tensor var_1702 = const()[name = tensor("op_1702"), val = tensor([1, 1])]; - tensor var_1704 = const()[name = tensor("op_1704"), val = tensor([1, 1])]; - tensor input_113_pad_type_0 = const()[name = tensor("input_113_pad_type_0"), val = tensor("custom")]; - tensor input_113_pad_0 = const()[name = tensor("input_113_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_3_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187126656)))]; - tensor input_113_cast_fp16 = conv(dilations = var_1704, groups = var_1365, pad = input_113_pad_0, pad_type = input_113_pad_type_0, strides = var_1702, weight = layers_3_fc1_loraA_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("input_113_cast_fp16")]; - tensor var_1708 = const()[name = tensor("op_1708"), val = tensor([1, 1])]; - tensor var_1710 = const()[name = tensor("op_1710"), val = tensor([1, 1])]; - tensor lora_out_153_pad_type_0 = const()[name = tensor("lora_out_153_pad_type_0"), val = tensor("custom")]; - tensor lora_out_153_pad_0 = const()[name = tensor("lora_out_153_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_155_weight_0_to_fp16 = const()[name = tensor("lora_out_155_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187167680)))]; - tensor lora_out_155_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_1710, groups = var_1365, pad = lora_out_153_pad_0, pad_type = lora_out_153_pad_type_0, strides = var_1708, weight = lora_out_155_weight_0_to_fp16, x = input_113_cast_fp16)[name = tensor("lora_out_155_cast_fp16")]; - tensor input_115_cast_fp16 = add(x = pretrained_out_77_cast_fp16, y = lora_out_155_cast_fp16)[name = tensor("input_115_cast_fp16")]; - tensor input_117_mode_0 = const()[name = tensor("input_117_mode_0"), val = tensor("EXACT")]; - tensor input_117_cast_fp16 = gelu(mode = input_117_mode_0, x = input_115_cast_fp16)[name = tensor("input_117_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 pretrained_out_79_pad_type_0 = const()[name = tensor("pretrained_out_79_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_79_pad_0 = const()[name = tensor("pretrained_out_79_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187331584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190608448))), name = tensor("layers_3_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_3_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190608576)))]; - tensor pretrained_out_79_cast_fp16 = conv(bias = layers_3_fc2_pretrained_bias_to_fp16, dilations = var_1724, groups = var_1365, pad = pretrained_out_79_pad_0, pad_type = pretrained_out_79_pad_type_0, strides = var_1722, weight = layers_3_fc2_pretrained_weight_to_fp16_palettized, x = input_117_cast_fp16)[name = tensor("pretrained_out_79_cast_fp16")]; - tensor var_1728 = const()[name = tensor("op_1728"), val = tensor([1, 1])]; - tensor var_1730 = const()[name = tensor("op_1730"), val = tensor([1, 1])]; - tensor input_119_pad_type_0 = const()[name = tensor("input_119_pad_type_0"), val = tensor("custom")]; - tensor input_119_pad_0 = const()[name = tensor("input_119_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_3_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_3_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190611200)))]; - tensor input_119_cast_fp16 = conv(dilations = var_1730, groups = var_1365, pad = input_119_pad_0, pad_type = input_119_pad_type_0, strides = var_1728, weight = layers_3_fc2_loraA_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; - tensor var_1734 = const()[name = tensor("op_1734"), val = tensor([1, 1])]; - tensor var_1736 = const()[name = tensor("op_1736"), val = tensor([1, 1])]; - tensor lora_out_157_pad_type_0 = const()[name = tensor("lora_out_157_pad_type_0"), val = tensor("custom")]; - tensor lora_out_157_pad_0 = const()[name = tensor("lora_out_157_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_159_weight_0_to_fp16 = const()[name = tensor("lora_out_159_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190775104)))]; - tensor lora_out_159_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1736, groups = var_1365, pad = lora_out_157_pad_0, pad_type = lora_out_157_pad_type_0, strides = var_1734, weight = lora_out_159_weight_0_to_fp16, x = input_119_cast_fp16)[name = tensor("lora_out_159_cast_fp16")]; - tensor hidden_states_9_cast_fp16 = add(x = pretrained_out_79_cast_fp16, y = lora_out_159_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_1752 = const()[name = tensor("op_1752"), val = tensor(3)]; - tensor var_1759 = const()[name = tensor("op_1759"), val = tensor(1)]; - tensor var_1760 = const()[name = tensor("op_1760"), val = tensor(true)]; - tensor var_1772 = const()[name = tensor("op_1772"), val = tensor([1])]; - tensor channels_mean_25_cast_fp16 = reduce_mean(axes = var_1772, keep_dims = var_1760, 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_1776 = const()[name = tensor("op_1776"), val = tensor([1])]; - tensor var_1777_cast_fp16 = reduce_mean(axes = var_1776, keep_dims = var_1760, x = zero_mean_sq_25_cast_fp16)[name = tensor("op_1777_cast_fp16")]; - tensor var_1778_to_fp16 = const()[name = tensor("op_1778_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_1779_cast_fp16 = add(x = var_1777_cast_fp16, y = var_1778_to_fp16)[name = tensor("op_1779_cast_fp16")]; - tensor denom_25_epsilon_0 = const()[name = tensor("denom_25_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_25_cast_fp16 = rsqrt(epsilon = denom_25_epsilon_0, x = var_1779_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_49_gamma_0_to_fp16 = const()[name = tensor("obj_49_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190816128)))]; - tensor obj_49_beta_0_to_fp16 = const()[name = tensor("obj_49_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190818752)))]; - tensor obj_49_epsilon_0_to_fp16 = const()[name = tensor("obj_49_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_49_cast_fp16 = batch_norm(beta = obj_49_beta_0_to_fp16, epsilon = obj_49_epsilon_0_to_fp16, gamma = obj_49_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_49_cast_fp16")]; - tensor var_1797 = const()[name = tensor("op_1797"), val = tensor([1, 1])]; - tensor var_1799 = const()[name = tensor("op_1799"), val = tensor([1, 1])]; - tensor pretrained_out_81_pad_type_0 = const()[name = tensor("pretrained_out_81_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_81_pad_0 = const()[name = tensor("pretrained_out_81_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190821376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191640640))), name = tensor("layers_4_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_4_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191640768)))]; - tensor pretrained_out_81_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_1799, groups = var_1759, pad = pretrained_out_81_pad_0, pad_type = pretrained_out_81_pad_type_0, strides = var_1797, weight = layers_4_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_49_cast_fp16)[name = tensor("pretrained_out_81_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 input_121_pad_type_0 = const()[name = tensor("input_121_pad_type_0"), val = tensor("custom")]; - tensor input_121_pad_0 = const()[name = tensor("input_121_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191643392)))]; - tensor input_121_cast_fp16 = conv(dilations = var_1805, groups = var_1759, pad = input_121_pad_0, pad_type = input_121_pad_type_0, strides = var_1803, weight = layers_4_self_attn_q_proj_loraA_weight_to_fp16, x = obj_49_cast_fp16)[name = tensor("input_121_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 lora_out_161_pad_type_0 = const()[name = tensor("lora_out_161_pad_type_0"), val = tensor("custom")]; - tensor lora_out_161_pad_0 = const()[name = tensor("lora_out_161_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_163_weight_0_to_fp16 = const()[name = tensor("lora_out_163_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191684416)))]; - tensor lora_out_163_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1811, groups = var_1759, pad = lora_out_161_pad_0, pad_type = lora_out_161_pad_type_0, strides = var_1809, weight = lora_out_163_weight_0_to_fp16, x = input_121_cast_fp16)[name = tensor("lora_out_163_cast_fp16")]; - tensor query_17_cast_fp16 = add(x = pretrained_out_81_cast_fp16, y = lora_out_163_cast_fp16)[name = tensor("query_17_cast_fp16")]; - tensor var_1821 = const()[name = tensor("op_1821"), val = tensor([1, 1])]; - tensor var_1823 = const()[name = tensor("op_1823"), val = tensor([1, 1])]; - tensor pretrained_out_83_pad_type_0 = const()[name = tensor("pretrained_out_83_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_83_pad_0 = const()[name = tensor("pretrained_out_83_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191725440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192544704))), name = tensor("layers_4_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_83_cast_fp16 = conv(dilations = var_1823, groups = var_1759, pad = pretrained_out_83_pad_0, pad_type = pretrained_out_83_pad_type_0, strides = var_1821, weight = layers_4_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_49_cast_fp16)[name = tensor("pretrained_out_83_cast_fp16")]; - tensor var_1827 = const()[name = tensor("op_1827"), val = tensor([1, 1])]; - tensor var_1829 = const()[name = tensor("op_1829"), val = tensor([1, 1])]; - tensor input_123_pad_type_0 = const()[name = tensor("input_123_pad_type_0"), val = tensor("custom")]; - tensor input_123_pad_0 = const()[name = tensor("input_123_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192544832)))]; - tensor input_123_cast_fp16 = conv(dilations = var_1829, groups = var_1759, pad = input_123_pad_0, pad_type = input_123_pad_type_0, strides = var_1827, weight = layers_4_self_attn_k_proj_loraA_weight_to_fp16, x = obj_49_cast_fp16)[name = tensor("input_123_cast_fp16")]; - tensor var_1833 = const()[name = tensor("op_1833"), val = tensor([1, 1])]; - tensor var_1835 = const()[name = tensor("op_1835"), val = tensor([1, 1])]; - tensor lora_out_165_pad_type_0 = const()[name = tensor("lora_out_165_pad_type_0"), val = tensor("custom")]; - tensor lora_out_165_pad_0 = const()[name = tensor("lora_out_165_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_167_weight_0_to_fp16 = const()[name = tensor("lora_out_167_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192585856)))]; - tensor lora_out_167_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1835, groups = var_1759, pad = lora_out_165_pad_0, pad_type = lora_out_165_pad_type_0, strides = var_1833, weight = lora_out_167_weight_0_to_fp16, x = input_123_cast_fp16)[name = tensor("lora_out_167_cast_fp16")]; - tensor current_key_9_cast_fp16 = add(x = pretrained_out_83_cast_fp16, y = lora_out_167_cast_fp16)[name = tensor("current_key_9_cast_fp16")]; - tensor var_1846 = const()[name = tensor("op_1846"), val = tensor([1, 1])]; - tensor var_1848 = const()[name = tensor("op_1848"), val = tensor([1, 1])]; - tensor pretrained_out_85_pad_type_0 = const()[name = tensor("pretrained_out_85_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_85_pad_0 = const()[name = tensor("pretrained_out_85_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192626880))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193446144))), name = tensor("layers_4_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_4_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193446272)))]; - tensor pretrained_out_85_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_1848, groups = var_1759, pad = pretrained_out_85_pad_0, pad_type = pretrained_out_85_pad_type_0, strides = var_1846, weight = layers_4_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_49_cast_fp16)[name = tensor("pretrained_out_85_cast_fp16")]; - tensor var_1852 = const()[name = tensor("op_1852"), val = tensor([1, 1])]; - tensor var_1854 = const()[name = tensor("op_1854"), val = tensor([1, 1])]; - tensor input_125_pad_type_0 = const()[name = tensor("input_125_pad_type_0"), val = tensor("custom")]; - tensor input_125_pad_0 = const()[name = tensor("input_125_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193448896)))]; - tensor input_125_cast_fp16 = conv(dilations = var_1854, groups = var_1759, pad = input_125_pad_0, pad_type = input_125_pad_type_0, strides = var_1852, weight = layers_4_self_attn_v_proj_loraA_weight_to_fp16, x = obj_49_cast_fp16)[name = tensor("input_125_cast_fp16")]; - tensor var_1858 = const()[name = tensor("op_1858"), val = tensor([1, 1])]; - tensor var_1860 = const()[name = tensor("op_1860"), val = tensor([1, 1])]; - tensor lora_out_169_pad_type_0 = const()[name = tensor("lora_out_169_pad_type_0"), val = tensor("custom")]; - tensor lora_out_169_pad_0 = const()[name = tensor("lora_out_169_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_171_weight_0_to_fp16 = const()[name = tensor("lora_out_171_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193489920)))]; - tensor lora_out_171_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1860, groups = var_1759, pad = lora_out_169_pad_0, pad_type = lora_out_169_pad_type_0, strides = var_1858, weight = lora_out_171_weight_0_to_fp16, x = input_125_cast_fp16)[name = tensor("lora_out_171_cast_fp16")]; - tensor current_value_9_cast_fp16 = add(x = pretrained_out_85_cast_fp16, y = lora_out_171_cast_fp16)[name = tensor("current_value_9_cast_fp16")]; - tensor var_1870_cast_fp16 = mul(x = current_key_9_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_1870_cast_fp16")]; - tensor var_1872_cast_fp16 = mul(x = var_103_cast_fp16_4, y = var_295_cast_fp16)[name = tensor("op_1872_cast_fp16")]; - tensor key_17_cast_fp16 = add(x = var_1870_cast_fp16, y = var_1872_cast_fp16)[name = tensor("key_17_cast_fp16")]; - tensor var_1874_cast_fp16 = mul(x = current_value_9_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_1874_cast_fp16")]; - tensor var_1876_cast_fp16 = mul(x = var_138_cast_fp16_4, y = var_295_cast_fp16)[name = tensor("op_1876_cast_fp16")]; - tensor value_17_cast_fp16 = add(x = var_1874_cast_fp16, y = var_1876_cast_fp16)[name = tensor("value_17_cast_fp16")]; - tensor var_1879 = const()[name = tensor("op_1879"), val = tensor([1, 20, 64, -1])]; - tensor var_1880_cast_fp16 = reshape(shape = var_1879, x = query_17_cast_fp16)[name = tensor("op_1880_cast_fp16")]; - tensor var_1881_to_fp16 = const()[name = tensor("op_1881_to_fp16"), val = tensor(0x1p-3)]; - tensor var_1882_cast_fp16 = mul(x = var_1880_cast_fp16, y = var_1881_to_fp16)[name = tensor("op_1882_cast_fp16")]; - tensor var_1883 = const()[name = tensor("op_1883"), val = tensor([1, 20, 64, -1])]; - tensor var_1884_cast_fp16 = reshape(shape = var_1883, x = key_17_cast_fp16)[name = tensor("op_1884_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_1882_cast_fp16, y = var_1884_cast_fp16)[name = tensor("mh_w_25_cast_fp16")]; - tensor mh_w_27_cast_fp16 = add(x = mh_w_25_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_27_cast_fp16")]; - tensor var_1892_cast_fp16 = softmax(axis = var_1752, x = mh_w_27_cast_fp16)[name = tensor("op_1892_cast_fp16")]; - tensor var_1893 = const()[name = tensor("op_1893"), val = tensor([1, 20, 64, -1])]; - tensor var_1894_cast_fp16 = reshape(shape = var_1893, x = value_17_cast_fp16)[name = tensor("op_1894_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_1894_cast_fp16, y = var_1892_cast_fp16)[name = tensor("attn_17_cast_fp16")]; - tensor var_1897 = const()[name = tensor("op_1897"), val = tensor([1, 1280, 1, -1])]; - tensor input_127_cast_fp16 = reshape(shape = var_1897, x = attn_17_cast_fp16)[name = tensor("input_127_cast_fp16")]; - tensor var_1904 = const()[name = tensor("op_1904"), val = tensor([1, 1])]; - tensor var_1906 = const()[name = tensor("op_1906"), val = tensor([1, 1])]; - tensor pretrained_out_87_pad_type_0 = const()[name = tensor("pretrained_out_87_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_87_pad_0 = const()[name = tensor("pretrained_out_87_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193530944))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194350208))), name = tensor("layers_4_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_4_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194350336)))]; - tensor pretrained_out_87_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_1906, groups = var_1759, pad = pretrained_out_87_pad_0, pad_type = pretrained_out_87_pad_type_0, strides = var_1904, weight = layers_4_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_127_cast_fp16)[name = tensor("pretrained_out_87_cast_fp16")]; - tensor var_1910 = const()[name = tensor("op_1910"), val = tensor([1, 1])]; - tensor var_1912 = const()[name = tensor("op_1912"), val = tensor([1, 1])]; - tensor input_129_pad_type_0 = const()[name = tensor("input_129_pad_type_0"), val = tensor("custom")]; - tensor input_129_pad_0 = const()[name = tensor("input_129_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194352960)))]; - tensor input_129_cast_fp16 = conv(dilations = var_1912, groups = var_1759, pad = input_129_pad_0, pad_type = input_129_pad_type_0, strides = var_1910, weight = layers_4_self_attn_o_proj_loraA_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; - tensor var_1916 = const()[name = tensor("op_1916"), val = tensor([1, 1])]; - tensor var_1918 = const()[name = tensor("op_1918"), val = tensor([1, 1])]; - tensor lora_out_173_pad_type_0 = const()[name = tensor("lora_out_173_pad_type_0"), val = tensor("custom")]; - tensor lora_out_173_pad_0 = const()[name = tensor("lora_out_173_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_175_weight_0_to_fp16 = const()[name = tensor("lora_out_175_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194393984)))]; - tensor lora_out_175_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1918, groups = var_1759, pad = lora_out_173_pad_0, pad_type = lora_out_173_pad_type_0, strides = var_1916, weight = lora_out_175_weight_0_to_fp16, x = input_129_cast_fp16)[name = tensor("lora_out_175_cast_fp16")]; - tensor obj_55_cast_fp16 = add(x = pretrained_out_87_cast_fp16, y = lora_out_175_cast_fp16)[name = tensor("obj_55_cast_fp16")]; - tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_55_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; - tensor var_1931 = const()[name = tensor("op_1931"), val = tensor([1])]; - tensor channels_mean_27_cast_fp16 = reduce_mean(axes = var_1931, keep_dims = var_1760, 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_1935 = const()[name = tensor("op_1935"), val = tensor([1])]; - tensor var_1936_cast_fp16 = reduce_mean(axes = var_1935, keep_dims = var_1760, x = zero_mean_sq_27_cast_fp16)[name = tensor("op_1936_cast_fp16")]; - tensor var_1937_to_fp16 = const()[name = tensor("op_1937_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_1938_cast_fp16 = add(x = var_1936_cast_fp16, y = var_1937_to_fp16)[name = tensor("op_1938_cast_fp16")]; - tensor denom_27_epsilon_0 = const()[name = tensor("denom_27_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_27_cast_fp16 = rsqrt(epsilon = denom_27_epsilon_0, x = var_1938_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_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(194435008)))]; - 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(194437632)))]; - 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_27_cast_fp16)[name = tensor("obj_57_cast_fp16")]; - tensor var_1956 = const()[name = tensor("op_1956"), val = tensor([1, 1])]; - tensor var_1958 = const()[name = tensor("op_1958"), val = tensor([1, 1])]; - tensor pretrained_out_89_pad_type_0 = const()[name = tensor("pretrained_out_89_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_89_pad_0 = const()[name = tensor("pretrained_out_89_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194440256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195259520))), name = tensor("layers_4_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_4_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195259648)))]; - tensor pretrained_out_89_cast_fp16 = conv(bias = layers_4_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_1958, groups = var_1759, pad = pretrained_out_89_pad_0, pad_type = pretrained_out_89_pad_type_0, strides = var_1956, weight = layers_4_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_57_cast_fp16)[name = tensor("pretrained_out_89_cast_fp16")]; - tensor var_1962 = const()[name = tensor("op_1962"), val = tensor([1, 1])]; - tensor var_1964 = const()[name = tensor("op_1964"), val = tensor([1, 1])]; - tensor input_131_pad_type_0 = const()[name = tensor("input_131_pad_type_0"), val = tensor("custom")]; - tensor input_131_pad_0 = const()[name = tensor("input_131_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195262272)))]; - tensor input_131_cast_fp16 = conv(dilations = var_1964, groups = var_1759, pad = input_131_pad_0, pad_type = input_131_pad_type_0, strides = var_1962, weight = layers_4_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("input_131_cast_fp16")]; - tensor var_1968 = const()[name = tensor("op_1968"), val = tensor([1, 1])]; - tensor var_1970 = const()[name = tensor("op_1970"), val = tensor([1, 1])]; - tensor lora_out_177_pad_type_0 = const()[name = tensor("lora_out_177_pad_type_0"), val = tensor("custom")]; - tensor lora_out_177_pad_0 = const()[name = tensor("lora_out_177_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_179_weight_0_to_fp16 = const()[name = tensor("lora_out_179_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195303296)))]; - tensor lora_out_179_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1970, groups = var_1759, pad = lora_out_177_pad_0, pad_type = lora_out_177_pad_type_0, strides = var_1968, weight = lora_out_179_weight_0_to_fp16, x = input_131_cast_fp16)[name = tensor("lora_out_179_cast_fp16")]; - tensor query_19_cast_fp16 = add(x = pretrained_out_89_cast_fp16, y = lora_out_179_cast_fp16)[name = tensor("query_19_cast_fp16")]; - tensor var_1980 = const()[name = tensor("op_1980"), val = tensor([1, 1])]; - tensor var_1982 = const()[name = tensor("op_1982"), val = tensor([1, 1])]; - tensor pretrained_out_91_pad_type_0 = const()[name = tensor("pretrained_out_91_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_91_pad_0 = const()[name = tensor("pretrained_out_91_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195344320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196163584))), name = tensor("layers_4_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_91_cast_fp16 = conv(dilations = var_1982, groups = var_1759, pad = pretrained_out_91_pad_0, pad_type = pretrained_out_91_pad_type_0, strides = var_1980, weight = layers_4_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_91_cast_fp16")]; - tensor var_1986 = const()[name = tensor("op_1986"), val = tensor([1, 1])]; - tensor var_1988 = const()[name = tensor("op_1988"), val = tensor([1, 1])]; - tensor input_133_pad_type_0 = const()[name = tensor("input_133_pad_type_0"), val = tensor("custom")]; - tensor input_133_pad_0 = const()[name = tensor("input_133_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196163712)))]; - tensor input_133_cast_fp16 = conv(dilations = var_1988, groups = var_1759, pad = input_133_pad_0, pad_type = input_133_pad_type_0, strides = var_1986, weight = layers_4_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_133_cast_fp16")]; - tensor var_1992 = const()[name = tensor("op_1992"), val = tensor([1, 1])]; - tensor var_1994 = const()[name = tensor("op_1994"), val = tensor([1, 1])]; - tensor lora_out_181_pad_type_0 = const()[name = tensor("lora_out_181_pad_type_0"), val = tensor("custom")]; - tensor lora_out_181_pad_0 = const()[name = tensor("lora_out_181_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_183_weight_0_to_fp16 = const()[name = tensor("lora_out_183_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196204736)))]; - tensor lora_out_183_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1994, groups = var_1759, pad = lora_out_181_pad_0, pad_type = lora_out_181_pad_type_0, strides = var_1992, weight = lora_out_183_weight_0_to_fp16, x = input_133_cast_fp16)[name = tensor("lora_out_183_cast_fp16")]; - tensor key_19_cast_fp16 = add(x = pretrained_out_91_cast_fp16, y = lora_out_183_cast_fp16)[name = tensor("key_19_cast_fp16")]; - tensor var_2005 = const()[name = tensor("op_2005"), val = tensor([1, 1])]; - tensor var_2007 = const()[name = tensor("op_2007"), val = tensor([1, 1])]; - tensor pretrained_out_93_pad_type_0 = const()[name = tensor("pretrained_out_93_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_93_pad_0 = const()[name = tensor("pretrained_out_93_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196245760))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197065024))), name = tensor("layers_4_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_4_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197065152)))]; - tensor pretrained_out_93_cast_fp16 = conv(bias = layers_4_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_2007, groups = var_1759, pad = pretrained_out_93_pad_0, pad_type = pretrained_out_93_pad_type_0, strides = var_2005, weight = layers_4_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_93_cast_fp16")]; - tensor var_2011 = const()[name = tensor("op_2011"), val = tensor([1, 1])]; - tensor var_2013 = const()[name = tensor("op_2013"), val = tensor([1, 1])]; - tensor input_135_pad_type_0 = const()[name = tensor("input_135_pad_type_0"), val = tensor("custom")]; - tensor input_135_pad_0 = const()[name = tensor("input_135_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197067776)))]; - tensor input_135_cast_fp16 = conv(dilations = var_2013, groups = var_1759, pad = input_135_pad_0, pad_type = input_135_pad_type_0, strides = var_2011, weight = layers_4_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_135_cast_fp16")]; - tensor var_2017 = const()[name = tensor("op_2017"), val = tensor([1, 1])]; - tensor var_2019 = const()[name = tensor("op_2019"), val = tensor([1, 1])]; - tensor lora_out_185_pad_type_0 = const()[name = tensor("lora_out_185_pad_type_0"), val = tensor("custom")]; - tensor lora_out_185_pad_0 = const()[name = tensor("lora_out_185_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_187_weight_0_to_fp16 = const()[name = tensor("lora_out_187_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197108800)))]; - tensor lora_out_187_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2019, groups = var_1759, pad = lora_out_185_pad_0, pad_type = lora_out_185_pad_type_0, strides = var_2017, weight = lora_out_187_weight_0_to_fp16, x = input_135_cast_fp16)[name = tensor("lora_out_187_cast_fp16")]; - tensor value_19_cast_fp16 = add(x = pretrained_out_93_cast_fp16, y = lora_out_187_cast_fp16)[name = tensor("value_19_cast_fp16")]; - tensor var_2026 = const()[name = tensor("op_2026"), val = tensor([1, 20, 64, -1])]; - tensor var_2027_cast_fp16 = reshape(shape = var_2026, x = query_19_cast_fp16)[name = tensor("op_2027_cast_fp16")]; - tensor var_2028_to_fp16 = const()[name = tensor("op_2028_to_fp16"), val = tensor(0x1p-3)]; - tensor var_2029_cast_fp16 = mul(x = var_2027_cast_fp16, y = var_2028_to_fp16)[name = tensor("op_2029_cast_fp16")]; - tensor var_2030 = const()[name = tensor("op_2030"), val = tensor([1, 20, 64, -1])]; - tensor var_2031_cast_fp16 = reshape(shape = var_2030, x = key_19_cast_fp16)[name = tensor("op_2031_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_2029_cast_fp16, y = var_2031_cast_fp16)[name = tensor("mh_w_29_cast_fp16")]; - tensor var_2034_cast_fp16 = softmax(axis = var_1752, x = mh_w_29_cast_fp16)[name = tensor("op_2034_cast_fp16")]; - tensor var_2035 = const()[name = tensor("op_2035"), val = tensor([1, 20, 64, -1])]; - tensor var_2036_cast_fp16 = reshape(shape = var_2035, x = value_19_cast_fp16)[name = tensor("op_2036_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_2036_cast_fp16, y = var_2034_cast_fp16)[name = tensor("attn_19_cast_fp16")]; - tensor var_2039 = const()[name = tensor("op_2039"), val = tensor([1, 1280, 1, -1])]; - tensor input_137_cast_fp16 = reshape(shape = var_2039, x = attn_19_cast_fp16)[name = tensor("input_137_cast_fp16")]; - tensor var_2046 = const()[name = tensor("op_2046"), val = tensor([1, 1])]; - tensor var_2048 = const()[name = tensor("op_2048"), val = tensor([1, 1])]; - tensor pretrained_out_95_pad_type_0 = const()[name = tensor("pretrained_out_95_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_95_pad_0 = const()[name = tensor("pretrained_out_95_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197149824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197969088))), name = tensor("layers_4_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_4_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197969216)))]; - tensor pretrained_out_95_cast_fp16 = conv(bias = layers_4_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_2048, groups = var_1759, pad = pretrained_out_95_pad_0, pad_type = pretrained_out_95_pad_type_0, strides = var_2046, weight = layers_4_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_137_cast_fp16)[name = tensor("pretrained_out_95_cast_fp16")]; - tensor var_2052 = const()[name = tensor("op_2052"), val = tensor([1, 1])]; - tensor var_2054 = const()[name = tensor("op_2054"), val = tensor([1, 1])]; - tensor input_139_pad_type_0 = const()[name = tensor("input_139_pad_type_0"), val = tensor("custom")]; - tensor input_139_pad_0 = const()[name = tensor("input_139_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197971840)))]; - tensor input_139_cast_fp16 = conv(dilations = var_2054, groups = var_1759, pad = input_139_pad_0, pad_type = input_139_pad_type_0, strides = var_2052, weight = layers_4_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("input_139_cast_fp16")]; - tensor var_2058 = const()[name = tensor("op_2058"), val = tensor([1, 1])]; - tensor var_2060 = const()[name = tensor("op_2060"), val = tensor([1, 1])]; - tensor lora_out_189_pad_type_0 = const()[name = tensor("lora_out_189_pad_type_0"), val = tensor("custom")]; - tensor lora_out_189_pad_0 = const()[name = tensor("lora_out_189_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_191_weight_0_to_fp16 = const()[name = tensor("lora_out_191_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198012864)))]; - tensor lora_out_191_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2060, groups = var_1759, pad = lora_out_189_pad_0, pad_type = lora_out_189_pad_type_0, strides = var_2058, weight = lora_out_191_weight_0_to_fp16, x = input_139_cast_fp16)[name = tensor("lora_out_191_cast_fp16")]; - tensor obj_59_cast_fp16 = add(x = pretrained_out_95_cast_fp16, y = lora_out_191_cast_fp16)[name = tensor("obj_59_cast_fp16")]; - tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = obj_59_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; - tensor var_2069 = const()[name = tensor("op_2069"), val = tensor([1])]; - tensor channels_mean_29_cast_fp16 = reduce_mean(axes = var_2069, keep_dims = var_1760, 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_2073 = const()[name = tensor("op_2073"), val = tensor([1])]; - tensor var_2074_cast_fp16 = reduce_mean(axes = var_2073, keep_dims = var_1760, x = zero_mean_sq_29_cast_fp16)[name = tensor("op_2074_cast_fp16")]; - tensor var_2075_to_fp16 = const()[name = tensor("op_2075_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_2076_cast_fp16 = add(x = var_2074_cast_fp16, y = var_2075_to_fp16)[name = tensor("op_2076_cast_fp16")]; - tensor denom_29_epsilon_0 = const()[name = tensor("denom_29_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_29_cast_fp16 = rsqrt(epsilon = denom_29_epsilon_0, x = var_2076_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_141_gamma_0_to_fp16 = const()[name = tensor("input_141_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198053888)))]; - tensor input_141_beta_0_to_fp16 = const()[name = tensor("input_141_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198056512)))]; - tensor input_141_epsilon_0_to_fp16 = const()[name = tensor("input_141_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_141_cast_fp16 = batch_norm(beta = input_141_beta_0_to_fp16, epsilon = input_141_epsilon_0_to_fp16, gamma = input_141_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_141_cast_fp16")]; - tensor var_2090 = const()[name = tensor("op_2090"), val = tensor([1, 1])]; - tensor var_2092 = const()[name = tensor("op_2092"), val = tensor([1, 1])]; - tensor pretrained_out_97_pad_type_0 = const()[name = tensor("pretrained_out_97_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_97_pad_0 = const()[name = tensor("pretrained_out_97_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198059136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201336000))), name = tensor("layers_4_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_4_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_4_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201336128)))]; - tensor pretrained_out_97_cast_fp16 = conv(bias = layers_4_fc1_pretrained_bias_to_fp16, dilations = var_2092, groups = var_1759, pad = pretrained_out_97_pad_0, pad_type = pretrained_out_97_pad_type_0, strides = var_2090, weight = layers_4_fc1_pretrained_weight_to_fp16_palettized, x = input_141_cast_fp16)[name = tensor("pretrained_out_97_cast_fp16")]; - tensor var_2096 = const()[name = tensor("op_2096"), val = tensor([1, 1])]; - tensor var_2098 = const()[name = tensor("op_2098"), val = tensor([1, 1])]; - tensor input_143_pad_type_0 = const()[name = tensor("input_143_pad_type_0"), val = tensor("custom")]; - tensor input_143_pad_0 = const()[name = tensor("input_143_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_4_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201346432)))]; - tensor input_143_cast_fp16 = conv(dilations = var_2098, groups = var_1759, pad = input_143_pad_0, pad_type = input_143_pad_type_0, strides = var_2096, weight = layers_4_fc1_loraA_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("input_143_cast_fp16")]; - tensor var_2102 = const()[name = tensor("op_2102"), val = tensor([1, 1])]; - tensor var_2104 = const()[name = tensor("op_2104"), val = tensor([1, 1])]; - tensor lora_out_193_pad_type_0 = const()[name = tensor("lora_out_193_pad_type_0"), val = tensor("custom")]; - tensor lora_out_193_pad_0 = const()[name = tensor("lora_out_193_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_195_weight_0_to_fp16 = const()[name = tensor("lora_out_195_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201387456)))]; - tensor lora_out_195_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_2104, groups = var_1759, pad = lora_out_193_pad_0, pad_type = lora_out_193_pad_type_0, strides = var_2102, weight = lora_out_195_weight_0_to_fp16, x = input_143_cast_fp16)[name = tensor("lora_out_195_cast_fp16")]; - tensor input_145_cast_fp16 = add(x = pretrained_out_97_cast_fp16, y = lora_out_195_cast_fp16)[name = tensor("input_145_cast_fp16")]; - tensor input_147_mode_0 = const()[name = tensor("input_147_mode_0"), val = tensor("EXACT")]; - tensor input_147_cast_fp16 = gelu(mode = input_147_mode_0, x = input_145_cast_fp16)[name = tensor("input_147_cast_fp16")]; - tensor var_2116 = const()[name = tensor("op_2116"), val = tensor([1, 1])]; - tensor var_2118 = const()[name = tensor("op_2118"), val = tensor([1, 1])]; - tensor pretrained_out_99_pad_type_0 = const()[name = tensor("pretrained_out_99_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_99_pad_0 = const()[name = tensor("pretrained_out_99_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201551360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204828224))), name = tensor("layers_4_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_4_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_4_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204828352)))]; - tensor pretrained_out_99_cast_fp16 = conv(bias = layers_4_fc2_pretrained_bias_to_fp16, dilations = var_2118, groups = var_1759, pad = pretrained_out_99_pad_0, pad_type = pretrained_out_99_pad_type_0, strides = var_2116, weight = layers_4_fc2_pretrained_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = tensor("pretrained_out_99_cast_fp16")]; - tensor var_2122 = const()[name = tensor("op_2122"), val = tensor([1, 1])]; - tensor var_2124 = const()[name = tensor("op_2124"), val = tensor([1, 1])]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("custom")]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_4_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_4_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204830976)))]; - tensor input_149_cast_fp16 = conv(dilations = var_2124, groups = var_1759, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = var_2122, weight = layers_4_fc2_loraA_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; - tensor var_2128 = const()[name = tensor("op_2128"), val = tensor([1, 1])]; - tensor var_2130 = const()[name = tensor("op_2130"), val = tensor([1, 1])]; - tensor lora_out_197_pad_type_0 = const()[name = tensor("lora_out_197_pad_type_0"), val = tensor("custom")]; - tensor lora_out_197_pad_0 = const()[name = tensor("lora_out_197_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_199_weight_0_to_fp16 = const()[name = tensor("lora_out_199_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204994880)))]; - tensor lora_out_199_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2130, groups = var_1759, pad = lora_out_197_pad_0, pad_type = lora_out_197_pad_type_0, strides = var_2128, weight = lora_out_199_weight_0_to_fp16, x = input_149_cast_fp16)[name = tensor("lora_out_199_cast_fp16")]; - tensor hidden_states_11_cast_fp16 = add(x = pretrained_out_99_cast_fp16, y = lora_out_199_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_2146 = const()[name = tensor("op_2146"), val = tensor(3)]; - tensor var_2153 = const()[name = tensor("op_2153"), val = tensor(1)]; - tensor var_2154 = const()[name = tensor("op_2154"), val = tensor(true)]; - tensor var_2166 = const()[name = tensor("op_2166"), val = tensor([1])]; - tensor channels_mean_31_cast_fp16 = reduce_mean(axes = var_2166, keep_dims = var_2154, 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_2170 = const()[name = tensor("op_2170"), val = tensor([1])]; - tensor var_2171_cast_fp16 = reduce_mean(axes = var_2170, keep_dims = var_2154, x = zero_mean_sq_31_cast_fp16)[name = tensor("op_2171_cast_fp16")]; - tensor var_2172_to_fp16 = const()[name = tensor("op_2172_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_2173_cast_fp16 = add(x = var_2171_cast_fp16, y = var_2172_to_fp16)[name = tensor("op_2173_cast_fp16")]; - tensor denom_31_epsilon_0 = const()[name = tensor("denom_31_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_31_cast_fp16 = rsqrt(epsilon = denom_31_epsilon_0, x = var_2173_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_61_gamma_0_to_fp16 = const()[name = tensor("obj_61_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205035904)))]; - tensor obj_61_beta_0_to_fp16 = const()[name = tensor("obj_61_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205038528)))]; - tensor obj_61_epsilon_0_to_fp16 = const()[name = tensor("obj_61_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_61_cast_fp16 = batch_norm(beta = obj_61_beta_0_to_fp16, epsilon = obj_61_epsilon_0_to_fp16, gamma = obj_61_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_61_cast_fp16")]; - tensor var_2191 = const()[name = tensor("op_2191"), val = tensor([1, 1])]; - tensor var_2193 = const()[name = tensor("op_2193"), val = tensor([1, 1])]; - tensor pretrained_out_101_pad_type_0 = const()[name = tensor("pretrained_out_101_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_101_pad_0 = const()[name = tensor("pretrained_out_101_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205041152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205860416))), name = tensor("layers_5_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_5_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205860544)))]; - tensor pretrained_out_101_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_2193, groups = var_2153, pad = pretrained_out_101_pad_0, pad_type = pretrained_out_101_pad_type_0, strides = var_2191, weight = layers_5_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_61_cast_fp16)[name = tensor("pretrained_out_101_cast_fp16")]; - tensor var_2197 = const()[name = tensor("op_2197"), val = tensor([1, 1])]; - tensor var_2199 = const()[name = tensor("op_2199"), val = tensor([1, 1])]; - tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("custom")]; - tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205863168)))]; - tensor input_151_cast_fp16 = conv(dilations = var_2199, groups = var_2153, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = var_2197, weight = layers_5_self_attn_q_proj_loraA_weight_to_fp16, x = obj_61_cast_fp16)[name = tensor("input_151_cast_fp16")]; - tensor var_2203 = const()[name = tensor("op_2203"), val = tensor([1, 1])]; - tensor var_2205 = const()[name = tensor("op_2205"), val = tensor([1, 1])]; - tensor lora_out_201_pad_type_0 = const()[name = tensor("lora_out_201_pad_type_0"), val = tensor("custom")]; - tensor lora_out_201_pad_0 = const()[name = tensor("lora_out_201_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_203_weight_0_to_fp16 = const()[name = tensor("lora_out_203_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205904192)))]; - tensor lora_out_203_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2205, groups = var_2153, pad = lora_out_201_pad_0, pad_type = lora_out_201_pad_type_0, strides = var_2203, weight = lora_out_203_weight_0_to_fp16, x = input_151_cast_fp16)[name = tensor("lora_out_203_cast_fp16")]; - tensor query_21_cast_fp16 = add(x = pretrained_out_101_cast_fp16, y = lora_out_203_cast_fp16)[name = tensor("query_21_cast_fp16")]; - tensor var_2215 = const()[name = tensor("op_2215"), val = tensor([1, 1])]; - tensor var_2217 = const()[name = tensor("op_2217"), val = tensor([1, 1])]; - tensor pretrained_out_103_pad_type_0 = const()[name = tensor("pretrained_out_103_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_103_pad_0 = const()[name = tensor("pretrained_out_103_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205945216))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206764480))), name = tensor("layers_5_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_103_cast_fp16 = conv(dilations = var_2217, groups = var_2153, pad = pretrained_out_103_pad_0, pad_type = pretrained_out_103_pad_type_0, strides = var_2215, weight = layers_5_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_61_cast_fp16)[name = tensor("pretrained_out_103_cast_fp16")]; - tensor var_2221 = const()[name = tensor("op_2221"), val = tensor([1, 1])]; - tensor var_2223 = const()[name = tensor("op_2223"), val = tensor([1, 1])]; - tensor input_153_pad_type_0 = const()[name = tensor("input_153_pad_type_0"), val = tensor("custom")]; - tensor input_153_pad_0 = const()[name = tensor("input_153_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206764608)))]; - tensor input_153_cast_fp16 = conv(dilations = var_2223, groups = var_2153, pad = input_153_pad_0, pad_type = input_153_pad_type_0, strides = var_2221, weight = layers_5_self_attn_k_proj_loraA_weight_to_fp16, x = obj_61_cast_fp16)[name = tensor("input_153_cast_fp16")]; - tensor var_2227 = const()[name = tensor("op_2227"), val = tensor([1, 1])]; - tensor var_2229 = const()[name = tensor("op_2229"), val = tensor([1, 1])]; - tensor lora_out_205_pad_type_0 = const()[name = tensor("lora_out_205_pad_type_0"), val = tensor("custom")]; - tensor lora_out_205_pad_0 = const()[name = tensor("lora_out_205_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_207_weight_0_to_fp16 = const()[name = tensor("lora_out_207_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206805632)))]; - tensor lora_out_207_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2229, groups = var_2153, pad = lora_out_205_pad_0, pad_type = lora_out_205_pad_type_0, strides = var_2227, weight = lora_out_207_weight_0_to_fp16, x = input_153_cast_fp16)[name = tensor("lora_out_207_cast_fp16")]; - tensor current_key_11_cast_fp16 = add(x = pretrained_out_103_cast_fp16, y = lora_out_207_cast_fp16)[name = tensor("current_key_11_cast_fp16")]; - tensor var_2240 = const()[name = tensor("op_2240"), val = tensor([1, 1])]; - tensor var_2242 = const()[name = tensor("op_2242"), val = tensor([1, 1])]; - tensor pretrained_out_105_pad_type_0 = const()[name = tensor("pretrained_out_105_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_105_pad_0 = const()[name = tensor("pretrained_out_105_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206846656))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207665920))), name = tensor("layers_5_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_5_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207666048)))]; - tensor pretrained_out_105_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_2242, groups = var_2153, pad = pretrained_out_105_pad_0, pad_type = pretrained_out_105_pad_type_0, strides = var_2240, weight = layers_5_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_61_cast_fp16)[name = tensor("pretrained_out_105_cast_fp16")]; - tensor var_2246 = const()[name = tensor("op_2246"), val = tensor([1, 1])]; - tensor var_2248 = const()[name = tensor("op_2248"), val = tensor([1, 1])]; - tensor input_155_pad_type_0 = const()[name = tensor("input_155_pad_type_0"), val = tensor("custom")]; - tensor input_155_pad_0 = const()[name = tensor("input_155_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207668672)))]; - tensor input_155_cast_fp16 = conv(dilations = var_2248, groups = var_2153, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = var_2246, weight = layers_5_self_attn_v_proj_loraA_weight_to_fp16, x = obj_61_cast_fp16)[name = tensor("input_155_cast_fp16")]; - tensor var_2252 = const()[name = tensor("op_2252"), val = tensor([1, 1])]; - tensor var_2254 = const()[name = tensor("op_2254"), val = tensor([1, 1])]; - tensor lora_out_209_pad_type_0 = const()[name = tensor("lora_out_209_pad_type_0"), val = tensor("custom")]; - tensor lora_out_209_pad_0 = const()[name = tensor("lora_out_209_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_211_weight_0_to_fp16 = const()[name = tensor("lora_out_211_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207709696)))]; - tensor lora_out_211_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2254, groups = var_2153, pad = lora_out_209_pad_0, pad_type = lora_out_209_pad_type_0, strides = var_2252, weight = lora_out_211_weight_0_to_fp16, x = input_155_cast_fp16)[name = tensor("lora_out_211_cast_fp16")]; - tensor current_value_11_cast_fp16 = add(x = pretrained_out_105_cast_fp16, y = lora_out_211_cast_fp16)[name = tensor("current_value_11_cast_fp16")]; - tensor var_2264_cast_fp16 = mul(x = current_key_11_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_2264_cast_fp16")]; - tensor var_2266_cast_fp16 = mul(x = var_103_cast_fp16_5, y = var_295_cast_fp16)[name = tensor("op_2266_cast_fp16")]; - tensor key_21_cast_fp16 = add(x = var_2264_cast_fp16, y = var_2266_cast_fp16)[name = tensor("key_21_cast_fp16")]; - tensor var_2268_cast_fp16 = mul(x = current_value_11_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_2268_cast_fp16")]; - tensor var_2270_cast_fp16 = mul(x = var_138_cast_fp16_5, y = var_295_cast_fp16)[name = tensor("op_2270_cast_fp16")]; - tensor value_21_cast_fp16 = add(x = var_2268_cast_fp16, y = var_2270_cast_fp16)[name = tensor("value_21_cast_fp16")]; - tensor var_2273 = const()[name = tensor("op_2273"), val = tensor([1, 20, 64, -1])]; - tensor var_2274_cast_fp16 = reshape(shape = var_2273, x = query_21_cast_fp16)[name = tensor("op_2274_cast_fp16")]; - tensor var_2275_to_fp16 = const()[name = tensor("op_2275_to_fp16"), val = tensor(0x1p-3)]; - tensor var_2276_cast_fp16 = mul(x = var_2274_cast_fp16, y = var_2275_to_fp16)[name = tensor("op_2276_cast_fp16")]; - tensor var_2277 = const()[name = tensor("op_2277"), val = tensor([1, 20, 64, -1])]; - tensor var_2278_cast_fp16 = reshape(shape = var_2277, x = key_21_cast_fp16)[name = tensor("op_2278_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_2276_cast_fp16, y = var_2278_cast_fp16)[name = tensor("mh_w_31_cast_fp16")]; - tensor mh_w_33_cast_fp16 = add(x = mh_w_31_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_33_cast_fp16")]; - tensor var_2286_cast_fp16 = softmax(axis = var_2146, x = mh_w_33_cast_fp16)[name = tensor("op_2286_cast_fp16")]; - tensor var_2287 = const()[name = tensor("op_2287"), val = tensor([1, 20, 64, -1])]; - tensor var_2288_cast_fp16 = reshape(shape = var_2287, x = value_21_cast_fp16)[name = tensor("op_2288_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_2288_cast_fp16, y = var_2286_cast_fp16)[name = tensor("attn_21_cast_fp16")]; - tensor var_2291 = const()[name = tensor("op_2291"), val = tensor([1, 1280, 1, -1])]; - tensor input_157_cast_fp16 = reshape(shape = var_2291, x = attn_21_cast_fp16)[name = tensor("input_157_cast_fp16")]; - tensor var_2298 = const()[name = tensor("op_2298"), val = tensor([1, 1])]; - tensor var_2300 = const()[name = tensor("op_2300"), val = tensor([1, 1])]; - tensor pretrained_out_107_pad_type_0 = const()[name = tensor("pretrained_out_107_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_107_pad_0 = const()[name = tensor("pretrained_out_107_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207750720))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208569984))), name = tensor("layers_5_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_5_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208570112)))]; - tensor pretrained_out_107_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_2300, groups = var_2153, pad = pretrained_out_107_pad_0, pad_type = pretrained_out_107_pad_type_0, strides = var_2298, weight = layers_5_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_157_cast_fp16)[name = tensor("pretrained_out_107_cast_fp16")]; - tensor var_2304 = const()[name = tensor("op_2304"), val = tensor([1, 1])]; - tensor var_2306 = const()[name = tensor("op_2306"), val = tensor([1, 1])]; - tensor input_159_pad_type_0 = const()[name = tensor("input_159_pad_type_0"), val = tensor("custom")]; - tensor input_159_pad_0 = const()[name = tensor("input_159_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208572736)))]; - tensor input_159_cast_fp16 = conv(dilations = var_2306, groups = var_2153, pad = input_159_pad_0, pad_type = input_159_pad_type_0, strides = var_2304, weight = layers_5_self_attn_o_proj_loraA_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; - tensor var_2310 = const()[name = tensor("op_2310"), val = tensor([1, 1])]; - tensor var_2312 = const()[name = tensor("op_2312"), val = tensor([1, 1])]; - tensor lora_out_213_pad_type_0 = const()[name = tensor("lora_out_213_pad_type_0"), val = tensor("custom")]; - tensor lora_out_213_pad_0 = const()[name = tensor("lora_out_213_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_215_weight_0_to_fp16 = const()[name = tensor("lora_out_215_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208613760)))]; - tensor lora_out_215_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2312, groups = var_2153, pad = lora_out_213_pad_0, pad_type = lora_out_213_pad_type_0, strides = var_2310, weight = lora_out_215_weight_0_to_fp16, x = input_159_cast_fp16)[name = tensor("lora_out_215_cast_fp16")]; - tensor obj_67_cast_fp16 = add(x = pretrained_out_107_cast_fp16, y = lora_out_215_cast_fp16)[name = tensor("obj_67_cast_fp16")]; - tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = obj_67_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; - tensor var_2325 = const()[name = tensor("op_2325"), val = tensor([1])]; - tensor channels_mean_33_cast_fp16 = reduce_mean(axes = var_2325, keep_dims = var_2154, 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_2329 = const()[name = tensor("op_2329"), val = tensor([1])]; - tensor var_2330_cast_fp16 = reduce_mean(axes = var_2329, keep_dims = var_2154, x = zero_mean_sq_33_cast_fp16)[name = tensor("op_2330_cast_fp16")]; - tensor var_2331_to_fp16 = const()[name = tensor("op_2331_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_2332_cast_fp16 = add(x = var_2330_cast_fp16, y = var_2331_to_fp16)[name = tensor("op_2332_cast_fp16")]; - tensor denom_33_epsilon_0 = const()[name = tensor("denom_33_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_33_cast_fp16 = rsqrt(epsilon = denom_33_epsilon_0, x = var_2332_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_69_gamma_0_to_fp16 = const()[name = tensor("obj_69_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208654784)))]; - tensor obj_69_beta_0_to_fp16 = const()[name = tensor("obj_69_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208657408)))]; - tensor obj_69_epsilon_0_to_fp16 = const()[name = tensor("obj_69_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_69_cast_fp16 = batch_norm(beta = obj_69_beta_0_to_fp16, epsilon = obj_69_epsilon_0_to_fp16, gamma = obj_69_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_69_cast_fp16")]; - tensor var_2350 = const()[name = tensor("op_2350"), val = tensor([1, 1])]; - tensor var_2352 = const()[name = tensor("op_2352"), val = tensor([1, 1])]; - tensor pretrained_out_109_pad_type_0 = const()[name = tensor("pretrained_out_109_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_109_pad_0 = const()[name = tensor("pretrained_out_109_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208660032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209479296))), name = tensor("layers_5_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_5_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209479424)))]; - tensor pretrained_out_109_cast_fp16 = conv(bias = layers_5_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_2352, groups = var_2153, pad = pretrained_out_109_pad_0, pad_type = pretrained_out_109_pad_type_0, strides = var_2350, weight = layers_5_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_69_cast_fp16)[name = tensor("pretrained_out_109_cast_fp16")]; - tensor var_2356 = const()[name = tensor("op_2356"), val = tensor([1, 1])]; - tensor var_2358 = const()[name = tensor("op_2358"), val = tensor([1, 1])]; - tensor input_161_pad_type_0 = const()[name = tensor("input_161_pad_type_0"), val = tensor("custom")]; - tensor input_161_pad_0 = const()[name = tensor("input_161_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209482048)))]; - tensor input_161_cast_fp16 = conv(dilations = var_2358, groups = var_2153, pad = input_161_pad_0, pad_type = input_161_pad_type_0, strides = var_2356, weight = layers_5_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_69_cast_fp16)[name = tensor("input_161_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 lora_out_217_pad_type_0 = const()[name = tensor("lora_out_217_pad_type_0"), val = tensor("custom")]; - tensor lora_out_217_pad_0 = const()[name = tensor("lora_out_217_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_219_weight_0_to_fp16 = const()[name = tensor("lora_out_219_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209523072)))]; - tensor lora_out_219_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2364, groups = var_2153, pad = lora_out_217_pad_0, pad_type = lora_out_217_pad_type_0, strides = var_2362, weight = lora_out_219_weight_0_to_fp16, x = input_161_cast_fp16)[name = tensor("lora_out_219_cast_fp16")]; - tensor query_23_cast_fp16 = add(x = pretrained_out_109_cast_fp16, y = lora_out_219_cast_fp16)[name = tensor("query_23_cast_fp16")]; - tensor var_2374 = const()[name = tensor("op_2374"), val = tensor([1, 1])]; - tensor var_2376 = const()[name = tensor("op_2376"), val = tensor([1, 1])]; - tensor pretrained_out_111_pad_type_0 = const()[name = tensor("pretrained_out_111_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_111_pad_0 = const()[name = tensor("pretrained_out_111_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209564096))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210383360))), name = tensor("layers_5_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_111_cast_fp16 = conv(dilations = var_2376, groups = var_2153, pad = pretrained_out_111_pad_0, pad_type = pretrained_out_111_pad_type_0, strides = var_2374, weight = layers_5_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_111_cast_fp16")]; - tensor var_2380 = const()[name = tensor("op_2380"), val = tensor([1, 1])]; - tensor var_2382 = const()[name = tensor("op_2382"), val = tensor([1, 1])]; - tensor input_163_pad_type_0 = const()[name = tensor("input_163_pad_type_0"), val = tensor("custom")]; - tensor input_163_pad_0 = const()[name = tensor("input_163_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210383488)))]; - tensor input_163_cast_fp16 = conv(dilations = var_2382, groups = var_2153, pad = input_163_pad_0, pad_type = input_163_pad_type_0, strides = var_2380, weight = layers_5_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_163_cast_fp16")]; - tensor var_2386 = const()[name = tensor("op_2386"), val = tensor([1, 1])]; - tensor var_2388 = const()[name = tensor("op_2388"), val = tensor([1, 1])]; - tensor lora_out_221_pad_type_0 = const()[name = tensor("lora_out_221_pad_type_0"), val = tensor("custom")]; - tensor lora_out_221_pad_0 = const()[name = tensor("lora_out_221_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_223_weight_0_to_fp16 = const()[name = tensor("lora_out_223_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210424512)))]; - tensor lora_out_223_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2388, groups = var_2153, pad = lora_out_221_pad_0, pad_type = lora_out_221_pad_type_0, strides = var_2386, weight = lora_out_223_weight_0_to_fp16, x = input_163_cast_fp16)[name = tensor("lora_out_223_cast_fp16")]; - tensor key_23_cast_fp16 = add(x = pretrained_out_111_cast_fp16, y = lora_out_223_cast_fp16)[name = tensor("key_23_cast_fp16")]; - tensor var_2399 = const()[name = tensor("op_2399"), val = tensor([1, 1])]; - tensor var_2401 = const()[name = tensor("op_2401"), val = tensor([1, 1])]; - tensor pretrained_out_113_pad_type_0 = const()[name = tensor("pretrained_out_113_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_113_pad_0 = const()[name = tensor("pretrained_out_113_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210465536))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211284800))), name = tensor("layers_5_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_5_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211284928)))]; - tensor pretrained_out_113_cast_fp16 = conv(bias = layers_5_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_2401, groups = var_2153, pad = pretrained_out_113_pad_0, pad_type = pretrained_out_113_pad_type_0, strides = var_2399, weight = layers_5_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_113_cast_fp16")]; - tensor var_2405 = const()[name = tensor("op_2405"), val = tensor([1, 1])]; - tensor var_2407 = const()[name = tensor("op_2407"), val = tensor([1, 1])]; - tensor input_165_pad_type_0 = const()[name = tensor("input_165_pad_type_0"), val = tensor("custom")]; - tensor input_165_pad_0 = const()[name = tensor("input_165_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211287552)))]; - tensor input_165_cast_fp16 = conv(dilations = var_2407, groups = var_2153, pad = input_165_pad_0, pad_type = input_165_pad_type_0, strides = var_2405, weight = layers_5_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_165_cast_fp16")]; - tensor var_2411 = const()[name = tensor("op_2411"), val = tensor([1, 1])]; - tensor var_2413 = const()[name = tensor("op_2413"), val = tensor([1, 1])]; - tensor lora_out_225_pad_type_0 = const()[name = tensor("lora_out_225_pad_type_0"), val = tensor("custom")]; - tensor lora_out_225_pad_0 = const()[name = tensor("lora_out_225_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_227_weight_0_to_fp16 = const()[name = tensor("lora_out_227_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211328576)))]; - tensor lora_out_227_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2413, groups = var_2153, pad = lora_out_225_pad_0, pad_type = lora_out_225_pad_type_0, strides = var_2411, weight = lora_out_227_weight_0_to_fp16, x = input_165_cast_fp16)[name = tensor("lora_out_227_cast_fp16")]; - tensor value_23_cast_fp16 = add(x = pretrained_out_113_cast_fp16, y = lora_out_227_cast_fp16)[name = tensor("value_23_cast_fp16")]; - tensor var_2420 = const()[name = tensor("op_2420"), val = tensor([1, 20, 64, -1])]; - tensor var_2421_cast_fp16 = reshape(shape = var_2420, x = query_23_cast_fp16)[name = tensor("op_2421_cast_fp16")]; - tensor var_2422_to_fp16 = const()[name = tensor("op_2422_to_fp16"), val = tensor(0x1p-3)]; - tensor var_2423_cast_fp16 = mul(x = var_2421_cast_fp16, y = var_2422_to_fp16)[name = tensor("op_2423_cast_fp16")]; - tensor var_2424 = const()[name = tensor("op_2424"), val = tensor([1, 20, 64, -1])]; - tensor var_2425_cast_fp16 = reshape(shape = var_2424, x = key_23_cast_fp16)[name = tensor("op_2425_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_2423_cast_fp16, y = var_2425_cast_fp16)[name = tensor("mh_w_35_cast_fp16")]; - tensor var_2428_cast_fp16 = softmax(axis = var_2146, x = mh_w_35_cast_fp16)[name = tensor("op_2428_cast_fp16")]; - tensor var_2429 = const()[name = tensor("op_2429"), val = tensor([1, 20, 64, -1])]; - tensor var_2430_cast_fp16 = reshape(shape = var_2429, x = value_23_cast_fp16)[name = tensor("op_2430_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_2430_cast_fp16, y = var_2428_cast_fp16)[name = tensor("attn_23_cast_fp16")]; - tensor var_2433 = const()[name = tensor("op_2433"), val = tensor([1, 1280, 1, -1])]; - tensor input_167_cast_fp16 = reshape(shape = var_2433, x = attn_23_cast_fp16)[name = tensor("input_167_cast_fp16")]; - tensor var_2440 = const()[name = tensor("op_2440"), val = tensor([1, 1])]; - tensor var_2442 = const()[name = tensor("op_2442"), val = tensor([1, 1])]; - tensor pretrained_out_115_pad_type_0 = const()[name = tensor("pretrained_out_115_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_115_pad_0 = const()[name = tensor("pretrained_out_115_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211369600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212188864))), name = tensor("layers_5_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_5_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212188992)))]; - tensor pretrained_out_115_cast_fp16 = conv(bias = layers_5_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_2442, groups = var_2153, pad = pretrained_out_115_pad_0, pad_type = pretrained_out_115_pad_type_0, strides = var_2440, weight = layers_5_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = tensor("pretrained_out_115_cast_fp16")]; - tensor var_2446 = const()[name = tensor("op_2446"), val = tensor([1, 1])]; - tensor var_2448 = const()[name = tensor("op_2448"), val = tensor([1, 1])]; - tensor input_169_pad_type_0 = const()[name = tensor("input_169_pad_type_0"), val = tensor("custom")]; - tensor input_169_pad_0 = const()[name = tensor("input_169_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212191616)))]; - tensor input_169_cast_fp16 = conv(dilations = var_2448, groups = var_2153, pad = input_169_pad_0, pad_type = input_169_pad_type_0, strides = var_2446, weight = layers_5_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; - tensor var_2452 = const()[name = tensor("op_2452"), val = tensor([1, 1])]; - tensor var_2454 = const()[name = tensor("op_2454"), val = tensor([1, 1])]; - tensor lora_out_229_pad_type_0 = const()[name = tensor("lora_out_229_pad_type_0"), val = tensor("custom")]; - tensor lora_out_229_pad_0 = const()[name = tensor("lora_out_229_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_231_weight_0_to_fp16 = const()[name = tensor("lora_out_231_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212232640)))]; - tensor lora_out_231_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2454, groups = var_2153, pad = lora_out_229_pad_0, pad_type = lora_out_229_pad_type_0, strides = var_2452, weight = lora_out_231_weight_0_to_fp16, x = input_169_cast_fp16)[name = tensor("lora_out_231_cast_fp16")]; - tensor obj_71_cast_fp16 = add(x = pretrained_out_115_cast_fp16, y = lora_out_231_cast_fp16)[name = tensor("obj_71_cast_fp16")]; - tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_71_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; - tensor var_2463 = const()[name = tensor("op_2463"), val = tensor([1])]; - tensor channels_mean_35_cast_fp16 = reduce_mean(axes = var_2463, keep_dims = var_2154, 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_2467 = const()[name = tensor("op_2467"), val = tensor([1])]; - tensor var_2468_cast_fp16 = reduce_mean(axes = var_2467, keep_dims = var_2154, x = zero_mean_sq_35_cast_fp16)[name = tensor("op_2468_cast_fp16")]; - tensor var_2469_to_fp16 = const()[name = tensor("op_2469_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_2470_cast_fp16 = add(x = var_2468_cast_fp16, y = var_2469_to_fp16)[name = tensor("op_2470_cast_fp16")]; - tensor denom_35_epsilon_0 = const()[name = tensor("denom_35_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_35_cast_fp16 = rsqrt(epsilon = denom_35_epsilon_0, x = var_2470_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_171_gamma_0_to_fp16 = const()[name = tensor("input_171_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212273664)))]; - tensor input_171_beta_0_to_fp16 = const()[name = tensor("input_171_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212276288)))]; - tensor input_171_epsilon_0_to_fp16 = const()[name = tensor("input_171_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_171_cast_fp16 = batch_norm(beta = input_171_beta_0_to_fp16, epsilon = input_171_epsilon_0_to_fp16, gamma = input_171_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_171_cast_fp16")]; - tensor var_2484 = const()[name = tensor("op_2484"), val = tensor([1, 1])]; - tensor var_2486 = const()[name = tensor("op_2486"), val = tensor([1, 1])]; - tensor pretrained_out_117_pad_type_0 = const()[name = tensor("pretrained_out_117_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_117_pad_0 = const()[name = tensor("pretrained_out_117_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212278912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215555776))), name = tensor("layers_5_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_5_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_5_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215555904)))]; - tensor pretrained_out_117_cast_fp16 = conv(bias = layers_5_fc1_pretrained_bias_to_fp16, dilations = var_2486, groups = var_2153, pad = pretrained_out_117_pad_0, pad_type = pretrained_out_117_pad_type_0, strides = var_2484, weight = layers_5_fc1_pretrained_weight_to_fp16_palettized, x = input_171_cast_fp16)[name = tensor("pretrained_out_117_cast_fp16")]; - tensor var_2490 = const()[name = tensor("op_2490"), val = tensor([1, 1])]; - tensor var_2492 = const()[name = tensor("op_2492"), val = tensor([1, 1])]; - tensor input_173_pad_type_0 = const()[name = tensor("input_173_pad_type_0"), val = tensor("custom")]; - tensor input_173_pad_0 = const()[name = tensor("input_173_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_5_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215566208)))]; - tensor input_173_cast_fp16 = conv(dilations = var_2492, groups = var_2153, pad = input_173_pad_0, pad_type = input_173_pad_type_0, strides = var_2490, weight = layers_5_fc1_loraA_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("input_173_cast_fp16")]; - tensor var_2496 = const()[name = tensor("op_2496"), val = tensor([1, 1])]; - tensor var_2498 = const()[name = tensor("op_2498"), val = tensor([1, 1])]; - tensor lora_out_233_pad_type_0 = const()[name = tensor("lora_out_233_pad_type_0"), val = tensor("custom")]; - tensor lora_out_233_pad_0 = const()[name = tensor("lora_out_233_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_235_weight_0_to_fp16 = const()[name = tensor("lora_out_235_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215607232)))]; - tensor lora_out_235_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_2498, groups = var_2153, pad = lora_out_233_pad_0, pad_type = lora_out_233_pad_type_0, strides = var_2496, weight = lora_out_235_weight_0_to_fp16, x = input_173_cast_fp16)[name = tensor("lora_out_235_cast_fp16")]; - tensor input_175_cast_fp16 = add(x = pretrained_out_117_cast_fp16, y = lora_out_235_cast_fp16)[name = tensor("input_175_cast_fp16")]; - tensor input_177_mode_0 = const()[name = tensor("input_177_mode_0"), val = tensor("EXACT")]; - tensor input_177_cast_fp16 = gelu(mode = input_177_mode_0, x = input_175_cast_fp16)[name = tensor("input_177_cast_fp16")]; - tensor var_2510 = const()[name = tensor("op_2510"), val = tensor([1, 1])]; - tensor var_2512 = const()[name = tensor("op_2512"), val = tensor([1, 1])]; - tensor pretrained_out_119_pad_type_0 = const()[name = tensor("pretrained_out_119_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_119_pad_0 = const()[name = tensor("pretrained_out_119_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215771136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219048000))), name = tensor("layers_5_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_5_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_5_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219048128)))]; - tensor pretrained_out_119_cast_fp16 = conv(bias = layers_5_fc2_pretrained_bias_to_fp16, dilations = var_2512, groups = var_2153, pad = pretrained_out_119_pad_0, pad_type = pretrained_out_119_pad_type_0, strides = var_2510, weight = layers_5_fc2_pretrained_weight_to_fp16_palettized, x = input_177_cast_fp16)[name = tensor("pretrained_out_119_cast_fp16")]; - tensor var_2516 = const()[name = tensor("op_2516"), val = tensor([1, 1])]; - tensor var_2518 = const()[name = tensor("op_2518"), val = tensor([1, 1])]; - tensor input_179_pad_type_0 = const()[name = tensor("input_179_pad_type_0"), val = tensor("custom")]; - tensor input_179_pad_0 = const()[name = tensor("input_179_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_5_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_5_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219050752)))]; - tensor input_179_cast_fp16 = conv(dilations = var_2518, groups = var_2153, pad = input_179_pad_0, pad_type = input_179_pad_type_0, strides = var_2516, weight = layers_5_fc2_loraA_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_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 lora_out_237_pad_type_0 = const()[name = tensor("lora_out_237_pad_type_0"), val = tensor("custom")]; - tensor lora_out_237_pad_0 = const()[name = tensor("lora_out_237_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_239_weight_0_to_fp16 = const()[name = tensor("lora_out_239_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219214656)))]; - tensor lora_out_239_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2524, groups = var_2153, pad = lora_out_237_pad_0, pad_type = lora_out_237_pad_type_0, strides = var_2522, weight = lora_out_239_weight_0_to_fp16, x = input_179_cast_fp16)[name = tensor("lora_out_239_cast_fp16")]; - tensor hidden_states_13_cast_fp16 = add(x = pretrained_out_119_cast_fp16, y = lora_out_239_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_2540 = const()[name = tensor("op_2540"), val = tensor(3)]; - tensor var_2547 = const()[name = tensor("op_2547"), val = tensor(1)]; - tensor var_2548 = const()[name = tensor("op_2548"), val = tensor(true)]; - tensor var_2560 = const()[name = tensor("op_2560"), val = tensor([1])]; - tensor channels_mean_37_cast_fp16 = reduce_mean(axes = var_2560, keep_dims = var_2548, 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_2564 = const()[name = tensor("op_2564"), val = tensor([1])]; - tensor var_2565_cast_fp16 = reduce_mean(axes = var_2564, keep_dims = var_2548, x = zero_mean_sq_37_cast_fp16)[name = tensor("op_2565_cast_fp16")]; - tensor var_2566_to_fp16 = const()[name = tensor("op_2566_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_2567_cast_fp16 = add(x = var_2565_cast_fp16, y = var_2566_to_fp16)[name = tensor("op_2567_cast_fp16")]; - tensor denom_37_epsilon_0 = const()[name = tensor("denom_37_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_37_cast_fp16 = rsqrt(epsilon = denom_37_epsilon_0, x = var_2567_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_73_gamma_0_to_fp16 = const()[name = tensor("obj_73_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219255680)))]; - tensor obj_73_beta_0_to_fp16 = const()[name = tensor("obj_73_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219258304)))]; - tensor obj_73_epsilon_0_to_fp16 = const()[name = tensor("obj_73_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_73_cast_fp16 = batch_norm(beta = obj_73_beta_0_to_fp16, epsilon = obj_73_epsilon_0_to_fp16, gamma = obj_73_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_73_cast_fp16")]; - tensor var_2585 = const()[name = tensor("op_2585"), val = tensor([1, 1])]; - tensor var_2587 = const()[name = tensor("op_2587"), val = tensor([1, 1])]; - tensor pretrained_out_121_pad_type_0 = const()[name = tensor("pretrained_out_121_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_121_pad_0 = const()[name = tensor("pretrained_out_121_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219260928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220080192))), name = tensor("layers_6_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_6_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220080320)))]; - tensor pretrained_out_121_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_2587, groups = var_2547, pad = pretrained_out_121_pad_0, pad_type = pretrained_out_121_pad_type_0, strides = var_2585, weight = layers_6_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_73_cast_fp16)[name = tensor("pretrained_out_121_cast_fp16")]; - tensor var_2591 = const()[name = tensor("op_2591"), val = tensor([1, 1])]; - tensor var_2593 = const()[name = tensor("op_2593"), val = tensor([1, 1])]; - tensor input_181_pad_type_0 = const()[name = tensor("input_181_pad_type_0"), val = tensor("custom")]; - tensor input_181_pad_0 = const()[name = tensor("input_181_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220082944)))]; - tensor input_181_cast_fp16 = conv(dilations = var_2593, groups = var_2547, pad = input_181_pad_0, pad_type = input_181_pad_type_0, strides = var_2591, weight = layers_6_self_attn_q_proj_loraA_weight_to_fp16, x = obj_73_cast_fp16)[name = tensor("input_181_cast_fp16")]; - tensor var_2597 = const()[name = tensor("op_2597"), val = tensor([1, 1])]; - tensor var_2599 = const()[name = tensor("op_2599"), val = tensor([1, 1])]; - tensor lora_out_241_pad_type_0 = const()[name = tensor("lora_out_241_pad_type_0"), val = tensor("custom")]; - tensor lora_out_241_pad_0 = const()[name = tensor("lora_out_241_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_243_weight_0_to_fp16 = const()[name = tensor("lora_out_243_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220123968)))]; - tensor lora_out_243_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2599, groups = var_2547, pad = lora_out_241_pad_0, pad_type = lora_out_241_pad_type_0, strides = var_2597, weight = lora_out_243_weight_0_to_fp16, x = input_181_cast_fp16)[name = tensor("lora_out_243_cast_fp16")]; - tensor query_25_cast_fp16 = add(x = pretrained_out_121_cast_fp16, y = lora_out_243_cast_fp16)[name = tensor("query_25_cast_fp16")]; - tensor var_2609 = const()[name = tensor("op_2609"), val = tensor([1, 1])]; - tensor var_2611 = const()[name = tensor("op_2611"), val = tensor([1, 1])]; - tensor pretrained_out_123_pad_type_0 = const()[name = tensor("pretrained_out_123_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_123_pad_0 = const()[name = tensor("pretrained_out_123_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220164992))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220984256))), name = tensor("layers_6_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_123_cast_fp16 = conv(dilations = var_2611, groups = var_2547, pad = pretrained_out_123_pad_0, pad_type = pretrained_out_123_pad_type_0, strides = var_2609, weight = layers_6_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_73_cast_fp16)[name = tensor("pretrained_out_123_cast_fp16")]; - tensor var_2615 = const()[name = tensor("op_2615"), val = tensor([1, 1])]; - tensor var_2617 = const()[name = tensor("op_2617"), val = tensor([1, 1])]; - tensor input_183_pad_type_0 = const()[name = tensor("input_183_pad_type_0"), val = tensor("custom")]; - tensor input_183_pad_0 = const()[name = tensor("input_183_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220984384)))]; - tensor input_183_cast_fp16 = conv(dilations = var_2617, groups = var_2547, pad = input_183_pad_0, pad_type = input_183_pad_type_0, strides = var_2615, weight = layers_6_self_attn_k_proj_loraA_weight_to_fp16, x = obj_73_cast_fp16)[name = tensor("input_183_cast_fp16")]; - tensor var_2621 = const()[name = tensor("op_2621"), val = tensor([1, 1])]; - tensor var_2623 = const()[name = tensor("op_2623"), val = tensor([1, 1])]; - tensor lora_out_245_pad_type_0 = const()[name = tensor("lora_out_245_pad_type_0"), val = tensor("custom")]; - tensor lora_out_245_pad_0 = const()[name = tensor("lora_out_245_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_247_weight_0_to_fp16 = const()[name = tensor("lora_out_247_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221025408)))]; - tensor lora_out_247_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2623, groups = var_2547, pad = lora_out_245_pad_0, pad_type = lora_out_245_pad_type_0, strides = var_2621, weight = lora_out_247_weight_0_to_fp16, x = input_183_cast_fp16)[name = tensor("lora_out_247_cast_fp16")]; - tensor current_key_13_cast_fp16 = add(x = pretrained_out_123_cast_fp16, y = lora_out_247_cast_fp16)[name = tensor("current_key_13_cast_fp16")]; - tensor var_2634 = const()[name = tensor("op_2634"), val = tensor([1, 1])]; - tensor var_2636 = const()[name = tensor("op_2636"), val = tensor([1, 1])]; - tensor pretrained_out_125_pad_type_0 = const()[name = tensor("pretrained_out_125_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_125_pad_0 = const()[name = tensor("pretrained_out_125_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221066432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221885696))), name = tensor("layers_6_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_6_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221885824)))]; - tensor pretrained_out_125_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_2636, groups = var_2547, pad = pretrained_out_125_pad_0, pad_type = pretrained_out_125_pad_type_0, strides = var_2634, weight = layers_6_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_73_cast_fp16)[name = tensor("pretrained_out_125_cast_fp16")]; - tensor var_2640 = const()[name = tensor("op_2640"), val = tensor([1, 1])]; - tensor var_2642 = const()[name = tensor("op_2642"), val = tensor([1, 1])]; - tensor input_185_pad_type_0 = const()[name = tensor("input_185_pad_type_0"), val = tensor("custom")]; - tensor input_185_pad_0 = const()[name = tensor("input_185_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221888448)))]; - tensor input_185_cast_fp16 = conv(dilations = var_2642, groups = var_2547, pad = input_185_pad_0, pad_type = input_185_pad_type_0, strides = var_2640, weight = layers_6_self_attn_v_proj_loraA_weight_to_fp16, x = obj_73_cast_fp16)[name = tensor("input_185_cast_fp16")]; - tensor var_2646 = const()[name = tensor("op_2646"), val = tensor([1, 1])]; - tensor var_2648 = const()[name = tensor("op_2648"), val = tensor([1, 1])]; - tensor lora_out_249_pad_type_0 = const()[name = tensor("lora_out_249_pad_type_0"), val = tensor("custom")]; - tensor lora_out_249_pad_0 = const()[name = tensor("lora_out_249_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_251_weight_0_to_fp16 = const()[name = tensor("lora_out_251_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221929472)))]; - tensor lora_out_251_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2648, groups = var_2547, pad = lora_out_249_pad_0, pad_type = lora_out_249_pad_type_0, strides = var_2646, weight = lora_out_251_weight_0_to_fp16, x = input_185_cast_fp16)[name = tensor("lora_out_251_cast_fp16")]; - tensor current_value_13_cast_fp16 = add(x = pretrained_out_125_cast_fp16, y = lora_out_251_cast_fp16)[name = tensor("current_value_13_cast_fp16")]; - tensor var_2658_cast_fp16 = mul(x = current_key_13_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_2658_cast_fp16")]; - tensor var_2660_cast_fp16 = mul(x = var_103_cast_fp16_6, y = var_295_cast_fp16)[name = tensor("op_2660_cast_fp16")]; - tensor key_25_cast_fp16 = add(x = var_2658_cast_fp16, y = var_2660_cast_fp16)[name = tensor("key_25_cast_fp16")]; - tensor var_2662_cast_fp16 = mul(x = current_value_13_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_2662_cast_fp16")]; - tensor var_2664_cast_fp16 = mul(x = var_138_cast_fp16_6, y = var_295_cast_fp16)[name = tensor("op_2664_cast_fp16")]; - tensor value_25_cast_fp16 = add(x = var_2662_cast_fp16, y = var_2664_cast_fp16)[name = tensor("value_25_cast_fp16")]; - tensor var_2667 = const()[name = tensor("op_2667"), val = tensor([1, 20, 64, -1])]; - tensor var_2668_cast_fp16 = reshape(shape = var_2667, x = query_25_cast_fp16)[name = tensor("op_2668_cast_fp16")]; - tensor var_2669_to_fp16 = const()[name = tensor("op_2669_to_fp16"), val = tensor(0x1p-3)]; - tensor var_2670_cast_fp16 = mul(x = var_2668_cast_fp16, y = var_2669_to_fp16)[name = tensor("op_2670_cast_fp16")]; - tensor var_2671 = const()[name = tensor("op_2671"), val = tensor([1, 20, 64, -1])]; - tensor var_2672_cast_fp16 = reshape(shape = var_2671, x = key_25_cast_fp16)[name = tensor("op_2672_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_2670_cast_fp16, y = var_2672_cast_fp16)[name = tensor("mh_w_37_cast_fp16")]; - tensor mh_w_39_cast_fp16 = add(x = mh_w_37_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_39_cast_fp16")]; - tensor var_2680_cast_fp16 = softmax(axis = var_2540, x = mh_w_39_cast_fp16)[name = tensor("op_2680_cast_fp16")]; - tensor var_2681 = const()[name = tensor("op_2681"), val = tensor([1, 20, 64, -1])]; - tensor var_2682_cast_fp16 = reshape(shape = var_2681, x = value_25_cast_fp16)[name = tensor("op_2682_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_2682_cast_fp16, y = var_2680_cast_fp16)[name = tensor("attn_25_cast_fp16")]; - tensor var_2685 = const()[name = tensor("op_2685"), val = tensor([1, 1280, 1, -1])]; - tensor input_187_cast_fp16 = reshape(shape = var_2685, x = attn_25_cast_fp16)[name = tensor("input_187_cast_fp16")]; - tensor var_2692 = const()[name = tensor("op_2692"), val = tensor([1, 1])]; - tensor var_2694 = const()[name = tensor("op_2694"), val = tensor([1, 1])]; - tensor pretrained_out_127_pad_type_0 = const()[name = tensor("pretrained_out_127_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_127_pad_0 = const()[name = tensor("pretrained_out_127_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221970496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222789760))), name = tensor("layers_6_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_6_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222789888)))]; - tensor pretrained_out_127_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_2694, groups = var_2547, pad = pretrained_out_127_pad_0, pad_type = pretrained_out_127_pad_type_0, strides = var_2692, weight = layers_6_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = tensor("pretrained_out_127_cast_fp16")]; - tensor var_2698 = const()[name = tensor("op_2698"), val = tensor([1, 1])]; - tensor var_2700 = const()[name = tensor("op_2700"), val = tensor([1, 1])]; - tensor input_189_pad_type_0 = const()[name = tensor("input_189_pad_type_0"), val = tensor("custom")]; - tensor input_189_pad_0 = const()[name = tensor("input_189_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222792512)))]; - tensor input_189_cast_fp16 = conv(dilations = var_2700, groups = var_2547, pad = input_189_pad_0, pad_type = input_189_pad_type_0, strides = var_2698, weight = layers_6_self_attn_o_proj_loraA_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; - tensor var_2704 = const()[name = tensor("op_2704"), val = tensor([1, 1])]; - tensor var_2706 = const()[name = tensor("op_2706"), val = tensor([1, 1])]; - tensor lora_out_253_pad_type_0 = const()[name = tensor("lora_out_253_pad_type_0"), val = tensor("custom")]; - tensor lora_out_253_pad_0 = const()[name = tensor("lora_out_253_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_255_weight_0_to_fp16 = const()[name = tensor("lora_out_255_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222833536)))]; - tensor lora_out_255_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2706, groups = var_2547, pad = lora_out_253_pad_0, pad_type = lora_out_253_pad_type_0, strides = var_2704, weight = lora_out_255_weight_0_to_fp16, x = input_189_cast_fp16)[name = tensor("lora_out_255_cast_fp16")]; - tensor obj_79_cast_fp16 = add(x = pretrained_out_127_cast_fp16, y = lora_out_255_cast_fp16)[name = tensor("obj_79_cast_fp16")]; - tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_79_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; - tensor var_2719 = const()[name = tensor("op_2719"), val = tensor([1])]; - tensor channels_mean_39_cast_fp16 = reduce_mean(axes = var_2719, keep_dims = var_2548, 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_2723 = const()[name = tensor("op_2723"), val = tensor([1])]; - tensor var_2724_cast_fp16 = reduce_mean(axes = var_2723, keep_dims = var_2548, x = zero_mean_sq_39_cast_fp16)[name = tensor("op_2724_cast_fp16")]; - tensor var_2725_to_fp16 = const()[name = tensor("op_2725_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_2726_cast_fp16 = add(x = var_2724_cast_fp16, y = var_2725_to_fp16)[name = tensor("op_2726_cast_fp16")]; - tensor denom_39_epsilon_0 = const()[name = tensor("denom_39_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_39_cast_fp16 = rsqrt(epsilon = denom_39_epsilon_0, x = var_2726_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_81_gamma_0_to_fp16 = const()[name = tensor("obj_81_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222874560)))]; - tensor obj_81_beta_0_to_fp16 = const()[name = tensor("obj_81_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222877184)))]; - tensor obj_81_epsilon_0_to_fp16 = const()[name = tensor("obj_81_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_81_cast_fp16 = batch_norm(beta = obj_81_beta_0_to_fp16, epsilon = obj_81_epsilon_0_to_fp16, gamma = obj_81_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_81_cast_fp16")]; - tensor var_2744 = const()[name = tensor("op_2744"), val = tensor([1, 1])]; - tensor var_2746 = const()[name = tensor("op_2746"), val = tensor([1, 1])]; - tensor pretrained_out_129_pad_type_0 = const()[name = tensor("pretrained_out_129_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_129_pad_0 = const()[name = tensor("pretrained_out_129_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222879808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223699072))), name = tensor("layers_6_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_6_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223699200)))]; - tensor pretrained_out_129_cast_fp16 = conv(bias = layers_6_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_2746, groups = var_2547, pad = pretrained_out_129_pad_0, pad_type = pretrained_out_129_pad_type_0, strides = var_2744, weight = layers_6_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_81_cast_fp16)[name = tensor("pretrained_out_129_cast_fp16")]; - tensor var_2750 = const()[name = tensor("op_2750"), val = tensor([1, 1])]; - tensor var_2752 = const()[name = tensor("op_2752"), val = tensor([1, 1])]; - tensor input_191_pad_type_0 = const()[name = tensor("input_191_pad_type_0"), val = tensor("custom")]; - tensor input_191_pad_0 = const()[name = tensor("input_191_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223701824)))]; - tensor input_191_cast_fp16 = conv(dilations = var_2752, groups = var_2547, pad = input_191_pad_0, pad_type = input_191_pad_type_0, strides = var_2750, weight = layers_6_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_81_cast_fp16)[name = tensor("input_191_cast_fp16")]; - tensor var_2756 = const()[name = tensor("op_2756"), val = tensor([1, 1])]; - tensor var_2758 = const()[name = tensor("op_2758"), val = tensor([1, 1])]; - tensor lora_out_257_pad_type_0 = const()[name = tensor("lora_out_257_pad_type_0"), val = tensor("custom")]; - tensor lora_out_257_pad_0 = const()[name = tensor("lora_out_257_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_259_weight_0_to_fp16 = const()[name = tensor("lora_out_259_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223742848)))]; - tensor lora_out_259_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2758, groups = var_2547, pad = lora_out_257_pad_0, pad_type = lora_out_257_pad_type_0, strides = var_2756, weight = lora_out_259_weight_0_to_fp16, x = input_191_cast_fp16)[name = tensor("lora_out_259_cast_fp16")]; - tensor query_27_cast_fp16 = add(x = pretrained_out_129_cast_fp16, y = lora_out_259_cast_fp16)[name = tensor("query_27_cast_fp16")]; - tensor var_2768 = const()[name = tensor("op_2768"), val = tensor([1, 1])]; - tensor var_2770 = const()[name = tensor("op_2770"), val = tensor([1, 1])]; - tensor pretrained_out_131_pad_type_0 = const()[name = tensor("pretrained_out_131_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_131_pad_0 = const()[name = tensor("pretrained_out_131_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223783872))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224603136))), name = tensor("layers_6_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_131_cast_fp16 = conv(dilations = var_2770, groups = var_2547, pad = pretrained_out_131_pad_0, pad_type = pretrained_out_131_pad_type_0, strides = var_2768, weight = layers_6_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_131_cast_fp16")]; - tensor var_2774 = const()[name = tensor("op_2774"), val = tensor([1, 1])]; - tensor var_2776 = const()[name = tensor("op_2776"), val = tensor([1, 1])]; - tensor input_193_pad_type_0 = const()[name = tensor("input_193_pad_type_0"), val = tensor("custom")]; - tensor input_193_pad_0 = const()[name = tensor("input_193_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224603264)))]; - tensor input_193_cast_fp16 = conv(dilations = var_2776, groups = var_2547, pad = input_193_pad_0, pad_type = input_193_pad_type_0, strides = var_2774, weight = layers_6_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_193_cast_fp16")]; - tensor var_2780 = const()[name = tensor("op_2780"), val = tensor([1, 1])]; - tensor var_2782 = const()[name = tensor("op_2782"), val = tensor([1, 1])]; - tensor lora_out_261_pad_type_0 = const()[name = tensor("lora_out_261_pad_type_0"), val = tensor("custom")]; - tensor lora_out_261_pad_0 = const()[name = tensor("lora_out_261_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_263_weight_0_to_fp16 = const()[name = tensor("lora_out_263_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224644288)))]; - tensor lora_out_263_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2782, groups = var_2547, pad = lora_out_261_pad_0, pad_type = lora_out_261_pad_type_0, strides = var_2780, weight = lora_out_263_weight_0_to_fp16, x = input_193_cast_fp16)[name = tensor("lora_out_263_cast_fp16")]; - tensor key_27_cast_fp16 = add(x = pretrained_out_131_cast_fp16, y = lora_out_263_cast_fp16)[name = tensor("key_27_cast_fp16")]; - tensor var_2793 = const()[name = tensor("op_2793"), val = tensor([1, 1])]; - tensor var_2795 = const()[name = tensor("op_2795"), val = tensor([1, 1])]; - tensor pretrained_out_133_pad_type_0 = const()[name = tensor("pretrained_out_133_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_133_pad_0 = const()[name = tensor("pretrained_out_133_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224685312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(225504576))), name = tensor("layers_6_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_6_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(225504704)))]; - tensor pretrained_out_133_cast_fp16 = conv(bias = layers_6_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_2795, groups = var_2547, pad = pretrained_out_133_pad_0, pad_type = pretrained_out_133_pad_type_0, strides = var_2793, weight = layers_6_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_133_cast_fp16")]; - tensor var_2799 = const()[name = tensor("op_2799"), val = tensor([1, 1])]; - tensor var_2801 = const()[name = tensor("op_2801"), val = tensor([1, 1])]; - tensor input_195_pad_type_0 = const()[name = tensor("input_195_pad_type_0"), val = tensor("custom")]; - tensor input_195_pad_0 = const()[name = tensor("input_195_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(225507328)))]; - tensor input_195_cast_fp16 = conv(dilations = var_2801, groups = var_2547, pad = input_195_pad_0, pad_type = input_195_pad_type_0, strides = var_2799, weight = layers_6_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_195_cast_fp16")]; - tensor var_2805 = const()[name = tensor("op_2805"), val = tensor([1, 1])]; - tensor var_2807 = const()[name = tensor("op_2807"), val = tensor([1, 1])]; - tensor lora_out_265_pad_type_0 = const()[name = tensor("lora_out_265_pad_type_0"), val = tensor("custom")]; - tensor lora_out_265_pad_0 = const()[name = tensor("lora_out_265_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_267_weight_0_to_fp16 = const()[name = tensor("lora_out_267_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(225548352)))]; - tensor lora_out_267_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2807, groups = var_2547, pad = lora_out_265_pad_0, pad_type = lora_out_265_pad_type_0, strides = var_2805, weight = lora_out_267_weight_0_to_fp16, x = input_195_cast_fp16)[name = tensor("lora_out_267_cast_fp16")]; - tensor value_27_cast_fp16 = add(x = pretrained_out_133_cast_fp16, y = lora_out_267_cast_fp16)[name = tensor("value_27_cast_fp16")]; - tensor var_2814 = const()[name = tensor("op_2814"), val = tensor([1, 20, 64, -1])]; - tensor var_2815_cast_fp16 = reshape(shape = var_2814, x = query_27_cast_fp16)[name = tensor("op_2815_cast_fp16")]; - tensor var_2816_to_fp16 = const()[name = tensor("op_2816_to_fp16"), val = tensor(0x1p-3)]; - tensor var_2817_cast_fp16 = mul(x = var_2815_cast_fp16, y = var_2816_to_fp16)[name = tensor("op_2817_cast_fp16")]; - tensor var_2818 = const()[name = tensor("op_2818"), val = tensor([1, 20, 64, -1])]; - tensor var_2819_cast_fp16 = reshape(shape = var_2818, x = key_27_cast_fp16)[name = tensor("op_2819_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_2817_cast_fp16, y = var_2819_cast_fp16)[name = tensor("mh_w_41_cast_fp16")]; - tensor var_2822_cast_fp16 = softmax(axis = var_2540, x = mh_w_41_cast_fp16)[name = tensor("op_2822_cast_fp16")]; - tensor var_2823 = const()[name = tensor("op_2823"), val = tensor([1, 20, 64, -1])]; - tensor var_2824_cast_fp16 = reshape(shape = var_2823, x = value_27_cast_fp16)[name = tensor("op_2824_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_2824_cast_fp16, y = var_2822_cast_fp16)[name = tensor("attn_27_cast_fp16")]; - tensor var_2827 = const()[name = tensor("op_2827"), val = tensor([1, 1280, 1, -1])]; - tensor input_197_cast_fp16 = reshape(shape = var_2827, x = attn_27_cast_fp16)[name = tensor("input_197_cast_fp16")]; - tensor var_2834 = const()[name = tensor("op_2834"), val = tensor([1, 1])]; - tensor var_2836 = const()[name = tensor("op_2836"), val = tensor([1, 1])]; - tensor pretrained_out_135_pad_type_0 = const()[name = tensor("pretrained_out_135_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_135_pad_0 = const()[name = tensor("pretrained_out_135_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(225589376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226408640))), name = tensor("layers_6_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_6_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226408768)))]; - tensor pretrained_out_135_cast_fp16 = conv(bias = layers_6_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_2836, groups = var_2547, pad = pretrained_out_135_pad_0, pad_type = pretrained_out_135_pad_type_0, strides = var_2834, weight = layers_6_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_197_cast_fp16)[name = tensor("pretrained_out_135_cast_fp16")]; - tensor var_2840 = const()[name = tensor("op_2840"), val = tensor([1, 1])]; - tensor var_2842 = const()[name = tensor("op_2842"), val = tensor([1, 1])]; - tensor input_199_pad_type_0 = const()[name = tensor("input_199_pad_type_0"), val = tensor("custom")]; - tensor input_199_pad_0 = const()[name = tensor("input_199_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226411392)))]; - tensor input_199_cast_fp16 = conv(dilations = var_2842, groups = var_2547, pad = input_199_pad_0, pad_type = input_199_pad_type_0, strides = var_2840, weight = layers_6_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; - tensor var_2846 = const()[name = tensor("op_2846"), val = tensor([1, 1])]; - tensor var_2848 = const()[name = tensor("op_2848"), val = tensor([1, 1])]; - tensor lora_out_269_pad_type_0 = const()[name = tensor("lora_out_269_pad_type_0"), val = tensor("custom")]; - tensor lora_out_269_pad_0 = const()[name = tensor("lora_out_269_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_271_weight_0_to_fp16 = const()[name = tensor("lora_out_271_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226452416)))]; - tensor lora_out_271_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2848, groups = var_2547, pad = lora_out_269_pad_0, pad_type = lora_out_269_pad_type_0, strides = var_2846, weight = lora_out_271_weight_0_to_fp16, x = input_199_cast_fp16)[name = tensor("lora_out_271_cast_fp16")]; - tensor obj_83_cast_fp16 = add(x = pretrained_out_135_cast_fp16, y = lora_out_271_cast_fp16)[name = tensor("obj_83_cast_fp16")]; - tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = obj_83_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; - tensor var_2857 = const()[name = tensor("op_2857"), val = tensor([1])]; - tensor channels_mean_41_cast_fp16 = reduce_mean(axes = var_2857, keep_dims = var_2548, 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_2861 = const()[name = tensor("op_2861"), val = tensor([1])]; - tensor var_2862_cast_fp16 = reduce_mean(axes = var_2861, keep_dims = var_2548, x = zero_mean_sq_41_cast_fp16)[name = tensor("op_2862_cast_fp16")]; - tensor var_2863_to_fp16 = const()[name = tensor("op_2863_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_2864_cast_fp16 = add(x = var_2862_cast_fp16, y = var_2863_to_fp16)[name = tensor("op_2864_cast_fp16")]; - tensor denom_41_epsilon_0 = const()[name = tensor("denom_41_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_41_cast_fp16 = rsqrt(epsilon = denom_41_epsilon_0, x = var_2864_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_201_gamma_0_to_fp16 = const()[name = tensor("input_201_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226493440)))]; - tensor input_201_beta_0_to_fp16 = const()[name = tensor("input_201_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226496064)))]; - tensor input_201_epsilon_0_to_fp16 = const()[name = tensor("input_201_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_201_cast_fp16 = batch_norm(beta = input_201_beta_0_to_fp16, epsilon = input_201_epsilon_0_to_fp16, gamma = input_201_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_201_cast_fp16")]; - tensor var_2878 = const()[name = tensor("op_2878"), val = tensor([1, 1])]; - tensor var_2880 = const()[name = tensor("op_2880"), val = tensor([1, 1])]; - tensor pretrained_out_137_pad_type_0 = const()[name = tensor("pretrained_out_137_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_137_pad_0 = const()[name = tensor("pretrained_out_137_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226498688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229775552))), name = tensor("layers_6_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_6_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_6_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229775680)))]; - tensor pretrained_out_137_cast_fp16 = conv(bias = layers_6_fc1_pretrained_bias_to_fp16, dilations = var_2880, groups = var_2547, pad = pretrained_out_137_pad_0, pad_type = pretrained_out_137_pad_type_0, strides = var_2878, weight = layers_6_fc1_pretrained_weight_to_fp16_palettized, x = input_201_cast_fp16)[name = tensor("pretrained_out_137_cast_fp16")]; - tensor var_2884 = const()[name = tensor("op_2884"), val = tensor([1, 1])]; - tensor var_2886 = const()[name = tensor("op_2886"), val = tensor([1, 1])]; - tensor input_203_pad_type_0 = const()[name = tensor("input_203_pad_type_0"), val = tensor("custom")]; - tensor input_203_pad_0 = const()[name = tensor("input_203_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_6_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229785984)))]; - tensor input_203_cast_fp16 = conv(dilations = var_2886, groups = var_2547, pad = input_203_pad_0, pad_type = input_203_pad_type_0, strides = var_2884, weight = layers_6_fc1_loraA_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("input_203_cast_fp16")]; - tensor var_2890 = const()[name = tensor("op_2890"), val = tensor([1, 1])]; - tensor var_2892 = const()[name = tensor("op_2892"), val = tensor([1, 1])]; - tensor lora_out_273_pad_type_0 = const()[name = tensor("lora_out_273_pad_type_0"), val = tensor("custom")]; - tensor lora_out_273_pad_0 = const()[name = tensor("lora_out_273_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_275_weight_0_to_fp16 = const()[name = tensor("lora_out_275_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229827008)))]; - tensor lora_out_275_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_2892, groups = var_2547, pad = lora_out_273_pad_0, pad_type = lora_out_273_pad_type_0, strides = var_2890, weight = lora_out_275_weight_0_to_fp16, x = input_203_cast_fp16)[name = tensor("lora_out_275_cast_fp16")]; - tensor input_205_cast_fp16 = add(x = pretrained_out_137_cast_fp16, y = lora_out_275_cast_fp16)[name = tensor("input_205_cast_fp16")]; - tensor input_207_mode_0 = const()[name = tensor("input_207_mode_0"), val = tensor("EXACT")]; - tensor input_207_cast_fp16 = gelu(mode = input_207_mode_0, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; - tensor var_2904 = const()[name = tensor("op_2904"), val = tensor([1, 1])]; - tensor var_2906 = const()[name = tensor("op_2906"), val = tensor([1, 1])]; - tensor pretrained_out_139_pad_type_0 = const()[name = tensor("pretrained_out_139_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_139_pad_0 = const()[name = tensor("pretrained_out_139_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229990912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233267776))), name = tensor("layers_6_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_6_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_6_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233267904)))]; - tensor pretrained_out_139_cast_fp16 = conv(bias = layers_6_fc2_pretrained_bias_to_fp16, dilations = var_2906, groups = var_2547, pad = pretrained_out_139_pad_0, pad_type = pretrained_out_139_pad_type_0, strides = var_2904, weight = layers_6_fc2_pretrained_weight_to_fp16_palettized, x = input_207_cast_fp16)[name = tensor("pretrained_out_139_cast_fp16")]; - tensor var_2910 = const()[name = tensor("op_2910"), val = tensor([1, 1])]; - tensor var_2912 = const()[name = tensor("op_2912"), val = tensor([1, 1])]; - tensor input_209_pad_type_0 = const()[name = tensor("input_209_pad_type_0"), val = tensor("custom")]; - tensor input_209_pad_0 = const()[name = tensor("input_209_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_6_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_6_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233270528)))]; - tensor input_209_cast_fp16 = conv(dilations = var_2912, groups = var_2547, pad = input_209_pad_0, pad_type = input_209_pad_type_0, strides = var_2910, weight = layers_6_fc2_loraA_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; - tensor var_2916 = const()[name = tensor("op_2916"), val = tensor([1, 1])]; - tensor var_2918 = const()[name = tensor("op_2918"), val = tensor([1, 1])]; - tensor lora_out_277_pad_type_0 = const()[name = tensor("lora_out_277_pad_type_0"), val = tensor("custom")]; - tensor lora_out_277_pad_0 = const()[name = tensor("lora_out_277_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_279_weight_0_to_fp16 = const()[name = tensor("lora_out_279_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233434432)))]; - tensor lora_out_279_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2918, groups = var_2547, pad = lora_out_277_pad_0, pad_type = lora_out_277_pad_type_0, strides = var_2916, weight = lora_out_279_weight_0_to_fp16, x = input_209_cast_fp16)[name = tensor("lora_out_279_cast_fp16")]; - tensor hidden_states_15_cast_fp16 = add(x = pretrained_out_139_cast_fp16, y = lora_out_279_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_2934 = const()[name = tensor("op_2934"), val = tensor(3)]; - tensor var_2941 = const()[name = tensor("op_2941"), val = tensor(1)]; - tensor var_2942 = const()[name = tensor("op_2942"), val = tensor(true)]; - tensor var_2954 = const()[name = tensor("op_2954"), val = tensor([1])]; - tensor channels_mean_43_cast_fp16 = reduce_mean(axes = var_2954, keep_dims = var_2942, 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_2958 = const()[name = tensor("op_2958"), val = tensor([1])]; - tensor var_2959_cast_fp16 = reduce_mean(axes = var_2958, keep_dims = var_2942, x = zero_mean_sq_43_cast_fp16)[name = tensor("op_2959_cast_fp16")]; - tensor var_2960_to_fp16 = const()[name = tensor("op_2960_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_2961_cast_fp16 = add(x = var_2959_cast_fp16, y = var_2960_to_fp16)[name = tensor("op_2961_cast_fp16")]; - tensor denom_43_epsilon_0 = const()[name = tensor("denom_43_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_43_cast_fp16 = rsqrt(epsilon = denom_43_epsilon_0, x = var_2961_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_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(233475456)))]; - 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(233478080)))]; - 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_43_cast_fp16)[name = tensor("obj_85_cast_fp16")]; - tensor var_2979 = const()[name = tensor("op_2979"), val = tensor([1, 1])]; - tensor var_2981 = const()[name = tensor("op_2981"), val = tensor([1, 1])]; - tensor pretrained_out_141_pad_type_0 = const()[name = tensor("pretrained_out_141_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_141_pad_0 = const()[name = tensor("pretrained_out_141_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233480704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234299968))), name = tensor("layers_7_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_7_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234300096)))]; - tensor pretrained_out_141_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_2981, groups = var_2941, pad = pretrained_out_141_pad_0, pad_type = pretrained_out_141_pad_type_0, strides = var_2979, weight = layers_7_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_85_cast_fp16)[name = tensor("pretrained_out_141_cast_fp16")]; - tensor var_2985 = const()[name = tensor("op_2985"), val = tensor([1, 1])]; - tensor var_2987 = const()[name = tensor("op_2987"), val = tensor([1, 1])]; - tensor input_211_pad_type_0 = const()[name = tensor("input_211_pad_type_0"), val = tensor("custom")]; - tensor input_211_pad_0 = const()[name = tensor("input_211_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234302720)))]; - tensor input_211_cast_fp16 = conv(dilations = var_2987, groups = var_2941, pad = input_211_pad_0, pad_type = input_211_pad_type_0, strides = var_2985, weight = layers_7_self_attn_q_proj_loraA_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("input_211_cast_fp16")]; - tensor var_2991 = const()[name = tensor("op_2991"), val = tensor([1, 1])]; - tensor var_2993 = const()[name = tensor("op_2993"), val = tensor([1, 1])]; - tensor lora_out_281_pad_type_0 = const()[name = tensor("lora_out_281_pad_type_0"), val = tensor("custom")]; - tensor lora_out_281_pad_0 = const()[name = tensor("lora_out_281_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_283_weight_0_to_fp16 = const()[name = tensor("lora_out_283_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234343744)))]; - tensor lora_out_283_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2993, groups = var_2941, pad = lora_out_281_pad_0, pad_type = lora_out_281_pad_type_0, strides = var_2991, weight = lora_out_283_weight_0_to_fp16, x = input_211_cast_fp16)[name = tensor("lora_out_283_cast_fp16")]; - tensor query_29_cast_fp16 = add(x = pretrained_out_141_cast_fp16, y = lora_out_283_cast_fp16)[name = tensor("query_29_cast_fp16")]; - tensor var_3003 = const()[name = tensor("op_3003"), val = tensor([1, 1])]; - tensor var_3005 = const()[name = tensor("op_3005"), val = tensor([1, 1])]; - tensor pretrained_out_143_pad_type_0 = const()[name = tensor("pretrained_out_143_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_143_pad_0 = const()[name = tensor("pretrained_out_143_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234384768))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235204032))), name = tensor("layers_7_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_143_cast_fp16 = conv(dilations = var_3005, groups = var_2941, pad = pretrained_out_143_pad_0, pad_type = pretrained_out_143_pad_type_0, strides = var_3003, weight = layers_7_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_85_cast_fp16)[name = tensor("pretrained_out_143_cast_fp16")]; - tensor var_3009 = const()[name = tensor("op_3009"), val = tensor([1, 1])]; - tensor var_3011 = const()[name = tensor("op_3011"), val = tensor([1, 1])]; - tensor input_213_pad_type_0 = const()[name = tensor("input_213_pad_type_0"), val = tensor("custom")]; - tensor input_213_pad_0 = const()[name = tensor("input_213_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235204160)))]; - tensor input_213_cast_fp16 = conv(dilations = var_3011, groups = var_2941, pad = input_213_pad_0, pad_type = input_213_pad_type_0, strides = var_3009, weight = layers_7_self_attn_k_proj_loraA_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("input_213_cast_fp16")]; - tensor var_3015 = const()[name = tensor("op_3015"), val = tensor([1, 1])]; - tensor var_3017 = const()[name = tensor("op_3017"), val = tensor([1, 1])]; - tensor lora_out_285_pad_type_0 = const()[name = tensor("lora_out_285_pad_type_0"), val = tensor("custom")]; - tensor lora_out_285_pad_0 = const()[name = tensor("lora_out_285_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_287_weight_0_to_fp16 = const()[name = tensor("lora_out_287_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235245184)))]; - tensor lora_out_287_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3017, groups = var_2941, pad = lora_out_285_pad_0, pad_type = lora_out_285_pad_type_0, strides = var_3015, weight = lora_out_287_weight_0_to_fp16, x = input_213_cast_fp16)[name = tensor("lora_out_287_cast_fp16")]; - tensor current_key_15_cast_fp16 = add(x = pretrained_out_143_cast_fp16, y = lora_out_287_cast_fp16)[name = tensor("current_key_15_cast_fp16")]; - tensor var_3028 = const()[name = tensor("op_3028"), val = tensor([1, 1])]; - tensor var_3030 = const()[name = tensor("op_3030"), val = tensor([1, 1])]; - tensor pretrained_out_145_pad_type_0 = const()[name = tensor("pretrained_out_145_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_145_pad_0 = const()[name = tensor("pretrained_out_145_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235286208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236105472))), name = tensor("layers_7_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_7_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236105600)))]; - tensor pretrained_out_145_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_3030, groups = var_2941, pad = pretrained_out_145_pad_0, pad_type = pretrained_out_145_pad_type_0, strides = var_3028, weight = layers_7_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_85_cast_fp16)[name = tensor("pretrained_out_145_cast_fp16")]; - tensor var_3034 = const()[name = tensor("op_3034"), val = tensor([1, 1])]; - tensor var_3036 = const()[name = tensor("op_3036"), val = tensor([1, 1])]; - tensor input_215_pad_type_0 = const()[name = tensor("input_215_pad_type_0"), val = tensor("custom")]; - tensor input_215_pad_0 = const()[name = tensor("input_215_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236108224)))]; - tensor input_215_cast_fp16 = conv(dilations = var_3036, groups = var_2941, pad = input_215_pad_0, pad_type = input_215_pad_type_0, strides = var_3034, weight = layers_7_self_attn_v_proj_loraA_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("input_215_cast_fp16")]; - tensor var_3040 = const()[name = tensor("op_3040"), val = tensor([1, 1])]; - tensor var_3042 = const()[name = tensor("op_3042"), val = tensor([1, 1])]; - tensor lora_out_289_pad_type_0 = const()[name = tensor("lora_out_289_pad_type_0"), val = tensor("custom")]; - tensor lora_out_289_pad_0 = const()[name = tensor("lora_out_289_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_291_weight_0_to_fp16 = const()[name = tensor("lora_out_291_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236149248)))]; - tensor lora_out_291_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3042, groups = var_2941, pad = lora_out_289_pad_0, pad_type = lora_out_289_pad_type_0, strides = var_3040, weight = lora_out_291_weight_0_to_fp16, x = input_215_cast_fp16)[name = tensor("lora_out_291_cast_fp16")]; - tensor current_value_15_cast_fp16 = add(x = pretrained_out_145_cast_fp16, y = lora_out_291_cast_fp16)[name = tensor("current_value_15_cast_fp16")]; - tensor var_3052_cast_fp16 = mul(x = current_key_15_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_3052_cast_fp16")]; - tensor var_3054_cast_fp16 = mul(x = var_103_cast_fp16_7, y = var_295_cast_fp16)[name = tensor("op_3054_cast_fp16")]; - tensor key_29_cast_fp16 = add(x = var_3052_cast_fp16, y = var_3054_cast_fp16)[name = tensor("key_29_cast_fp16")]; - tensor var_3056_cast_fp16 = mul(x = current_value_15_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_3056_cast_fp16")]; - tensor var_3058_cast_fp16 = mul(x = var_138_cast_fp16_7, y = var_295_cast_fp16)[name = tensor("op_3058_cast_fp16")]; - tensor value_29_cast_fp16 = add(x = var_3056_cast_fp16, y = var_3058_cast_fp16)[name = tensor("value_29_cast_fp16")]; - tensor var_3061 = const()[name = tensor("op_3061"), val = tensor([1, 20, 64, -1])]; - tensor var_3062_cast_fp16 = reshape(shape = var_3061, x = query_29_cast_fp16)[name = tensor("op_3062_cast_fp16")]; - tensor var_3063_to_fp16 = const()[name = tensor("op_3063_to_fp16"), val = tensor(0x1p-3)]; - tensor var_3064_cast_fp16 = mul(x = var_3062_cast_fp16, y = var_3063_to_fp16)[name = tensor("op_3064_cast_fp16")]; - tensor var_3065 = const()[name = tensor("op_3065"), val = tensor([1, 20, 64, -1])]; - tensor var_3066_cast_fp16 = reshape(shape = var_3065, x = key_29_cast_fp16)[name = tensor("op_3066_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_3064_cast_fp16, y = var_3066_cast_fp16)[name = tensor("mh_w_43_cast_fp16")]; - tensor mh_w_45_cast_fp16 = add(x = mh_w_43_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_45_cast_fp16")]; - tensor var_3074_cast_fp16 = softmax(axis = var_2934, x = mh_w_45_cast_fp16)[name = tensor("op_3074_cast_fp16")]; - tensor var_3075 = const()[name = tensor("op_3075"), val = tensor([1, 20, 64, -1])]; - tensor var_3076_cast_fp16 = reshape(shape = var_3075, x = value_29_cast_fp16)[name = tensor("op_3076_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_3076_cast_fp16, y = var_3074_cast_fp16)[name = tensor("attn_29_cast_fp16")]; - tensor var_3079 = const()[name = tensor("op_3079"), val = tensor([1, 1280, 1, -1])]; - tensor input_217_cast_fp16 = reshape(shape = var_3079, x = attn_29_cast_fp16)[name = tensor("input_217_cast_fp16")]; - tensor var_3086 = const()[name = tensor("op_3086"), val = tensor([1, 1])]; - tensor var_3088 = const()[name = tensor("op_3088"), val = tensor([1, 1])]; - tensor pretrained_out_147_pad_type_0 = const()[name = tensor("pretrained_out_147_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_147_pad_0 = const()[name = tensor("pretrained_out_147_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236190272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237009536))), name = tensor("layers_7_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_7_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237009664)))]; - tensor pretrained_out_147_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_3088, groups = var_2941, pad = pretrained_out_147_pad_0, pad_type = pretrained_out_147_pad_type_0, strides = var_3086, weight = layers_7_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_217_cast_fp16)[name = tensor("pretrained_out_147_cast_fp16")]; - tensor var_3092 = const()[name = tensor("op_3092"), val = tensor([1, 1])]; - tensor var_3094 = const()[name = tensor("op_3094"), val = tensor([1, 1])]; - tensor input_219_pad_type_0 = const()[name = tensor("input_219_pad_type_0"), val = tensor("custom")]; - tensor input_219_pad_0 = const()[name = tensor("input_219_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237012288)))]; - tensor input_219_cast_fp16 = conv(dilations = var_3094, groups = var_2941, pad = input_219_pad_0, pad_type = input_219_pad_type_0, strides = var_3092, weight = layers_7_self_attn_o_proj_loraA_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; - tensor var_3098 = const()[name = tensor("op_3098"), val = tensor([1, 1])]; - tensor var_3100 = const()[name = tensor("op_3100"), val = tensor([1, 1])]; - tensor lora_out_293_pad_type_0 = const()[name = tensor("lora_out_293_pad_type_0"), val = tensor("custom")]; - tensor lora_out_293_pad_0 = const()[name = tensor("lora_out_293_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_295_weight_0_to_fp16 = const()[name = tensor("lora_out_295_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237053312)))]; - tensor lora_out_295_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3100, groups = var_2941, pad = lora_out_293_pad_0, pad_type = lora_out_293_pad_type_0, strides = var_3098, weight = lora_out_295_weight_0_to_fp16, x = input_219_cast_fp16)[name = tensor("lora_out_295_cast_fp16")]; - tensor obj_91_cast_fp16 = add(x = pretrained_out_147_cast_fp16, y = lora_out_295_cast_fp16)[name = tensor("obj_91_cast_fp16")]; - tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = obj_91_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; - tensor var_3113 = const()[name = tensor("op_3113"), val = tensor([1])]; - tensor channels_mean_45_cast_fp16 = reduce_mean(axes = var_3113, keep_dims = var_2942, 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_3117 = const()[name = tensor("op_3117"), val = tensor([1])]; - tensor var_3118_cast_fp16 = reduce_mean(axes = var_3117, keep_dims = var_2942, x = zero_mean_sq_45_cast_fp16)[name = tensor("op_3118_cast_fp16")]; - tensor var_3119_to_fp16 = const()[name = tensor("op_3119_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_3120_cast_fp16 = add(x = var_3118_cast_fp16, y = var_3119_to_fp16)[name = tensor("op_3120_cast_fp16")]; - tensor denom_45_epsilon_0 = const()[name = tensor("denom_45_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_45_cast_fp16 = rsqrt(epsilon = denom_45_epsilon_0, x = var_3120_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_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(237094336)))]; - 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(237096960)))]; - 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_45_cast_fp16)[name = tensor("obj_93_cast_fp16")]; - tensor var_3138 = const()[name = tensor("op_3138"), val = tensor([1, 1])]; - tensor var_3140 = const()[name = tensor("op_3140"), val = tensor([1, 1])]; - tensor pretrained_out_149_pad_type_0 = const()[name = tensor("pretrained_out_149_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_149_pad_0 = const()[name = tensor("pretrained_out_149_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237099584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237918848))), name = tensor("layers_7_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_7_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237918976)))]; - tensor pretrained_out_149_cast_fp16 = conv(bias = layers_7_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_3140, groups = var_2941, pad = pretrained_out_149_pad_0, pad_type = pretrained_out_149_pad_type_0, strides = var_3138, weight = layers_7_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_93_cast_fp16)[name = tensor("pretrained_out_149_cast_fp16")]; - tensor var_3144 = const()[name = tensor("op_3144"), val = tensor([1, 1])]; - tensor var_3146 = const()[name = tensor("op_3146"), val = tensor([1, 1])]; - tensor input_221_pad_type_0 = const()[name = tensor("input_221_pad_type_0"), val = tensor("custom")]; - tensor input_221_pad_0 = const()[name = tensor("input_221_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237921600)))]; - tensor input_221_cast_fp16 = conv(dilations = var_3146, groups = var_2941, pad = input_221_pad_0, pad_type = input_221_pad_type_0, strides = var_3144, weight = layers_7_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor("input_221_cast_fp16")]; - tensor var_3150 = const()[name = tensor("op_3150"), val = tensor([1, 1])]; - tensor var_3152 = const()[name = tensor("op_3152"), val = tensor([1, 1])]; - tensor lora_out_297_pad_type_0 = const()[name = tensor("lora_out_297_pad_type_0"), val = tensor("custom")]; - tensor lora_out_297_pad_0 = const()[name = tensor("lora_out_297_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_299_weight_0_to_fp16 = const()[name = tensor("lora_out_299_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237962624)))]; - tensor lora_out_299_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3152, groups = var_2941, pad = lora_out_297_pad_0, pad_type = lora_out_297_pad_type_0, strides = var_3150, weight = lora_out_299_weight_0_to_fp16, x = input_221_cast_fp16)[name = tensor("lora_out_299_cast_fp16")]; - tensor query_31_cast_fp16 = add(x = pretrained_out_149_cast_fp16, y = lora_out_299_cast_fp16)[name = tensor("query_31_cast_fp16")]; - tensor var_3162 = const()[name = tensor("op_3162"), val = tensor([1, 1])]; - tensor var_3164 = const()[name = tensor("op_3164"), val = tensor([1, 1])]; - tensor pretrained_out_151_pad_type_0 = const()[name = tensor("pretrained_out_151_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_151_pad_0 = const()[name = tensor("pretrained_out_151_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238003648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238822912))), name = tensor("layers_7_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_151_cast_fp16 = conv(dilations = var_3164, groups = var_2941, pad = pretrained_out_151_pad_0, pad_type = pretrained_out_151_pad_type_0, strides = var_3162, weight = layers_7_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_151_cast_fp16")]; - tensor var_3168 = const()[name = tensor("op_3168"), val = tensor([1, 1])]; - tensor var_3170 = const()[name = tensor("op_3170"), val = tensor([1, 1])]; - tensor input_223_pad_type_0 = const()[name = tensor("input_223_pad_type_0"), val = tensor("custom")]; - tensor input_223_pad_0 = const()[name = tensor("input_223_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238823040)))]; - tensor input_223_cast_fp16 = conv(dilations = var_3170, groups = var_2941, pad = input_223_pad_0, pad_type = input_223_pad_type_0, strides = var_3168, weight = layers_7_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_223_cast_fp16")]; - tensor var_3174 = const()[name = tensor("op_3174"), val = tensor([1, 1])]; - tensor var_3176 = const()[name = tensor("op_3176"), val = tensor([1, 1])]; - tensor lora_out_301_pad_type_0 = const()[name = tensor("lora_out_301_pad_type_0"), val = tensor("custom")]; - tensor lora_out_301_pad_0 = const()[name = tensor("lora_out_301_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_303_weight_0_to_fp16 = const()[name = tensor("lora_out_303_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238864064)))]; - tensor lora_out_303_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3176, groups = var_2941, pad = lora_out_301_pad_0, pad_type = lora_out_301_pad_type_0, strides = var_3174, weight = lora_out_303_weight_0_to_fp16, x = input_223_cast_fp16)[name = tensor("lora_out_303_cast_fp16")]; - tensor key_31_cast_fp16 = add(x = pretrained_out_151_cast_fp16, y = lora_out_303_cast_fp16)[name = tensor("key_31_cast_fp16")]; - tensor var_3187 = const()[name = tensor("op_3187"), val = tensor([1, 1])]; - tensor var_3189 = const()[name = tensor("op_3189"), val = tensor([1, 1])]; - tensor pretrained_out_153_pad_type_0 = const()[name = tensor("pretrained_out_153_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_153_pad_0 = const()[name = tensor("pretrained_out_153_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238905088))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239724352))), name = tensor("layers_7_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_7_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239724480)))]; - tensor pretrained_out_153_cast_fp16 = conv(bias = layers_7_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_3189, groups = var_2941, pad = pretrained_out_153_pad_0, pad_type = pretrained_out_153_pad_type_0, strides = var_3187, weight = layers_7_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_153_cast_fp16")]; - tensor var_3193 = const()[name = tensor("op_3193"), val = tensor([1, 1])]; - tensor var_3195 = const()[name = tensor("op_3195"), val = tensor([1, 1])]; - tensor input_225_pad_type_0 = const()[name = tensor("input_225_pad_type_0"), val = tensor("custom")]; - tensor input_225_pad_0 = const()[name = tensor("input_225_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239727104)))]; - tensor input_225_cast_fp16 = conv(dilations = var_3195, groups = var_2941, pad = input_225_pad_0, pad_type = input_225_pad_type_0, strides = var_3193, weight = layers_7_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_225_cast_fp16")]; - tensor var_3199 = const()[name = tensor("op_3199"), val = tensor([1, 1])]; - tensor var_3201 = const()[name = tensor("op_3201"), val = tensor([1, 1])]; - tensor lora_out_305_pad_type_0 = const()[name = tensor("lora_out_305_pad_type_0"), val = tensor("custom")]; - tensor lora_out_305_pad_0 = const()[name = tensor("lora_out_305_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_307_weight_0_to_fp16 = const()[name = tensor("lora_out_307_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239768128)))]; - tensor lora_out_307_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3201, groups = var_2941, pad = lora_out_305_pad_0, pad_type = lora_out_305_pad_type_0, strides = var_3199, weight = lora_out_307_weight_0_to_fp16, x = input_225_cast_fp16)[name = tensor("lora_out_307_cast_fp16")]; - tensor value_31_cast_fp16 = add(x = pretrained_out_153_cast_fp16, y = lora_out_307_cast_fp16)[name = tensor("value_31_cast_fp16")]; - tensor var_3208 = const()[name = tensor("op_3208"), val = tensor([1, 20, 64, -1])]; - tensor var_3209_cast_fp16 = reshape(shape = var_3208, x = query_31_cast_fp16)[name = tensor("op_3209_cast_fp16")]; - tensor var_3210_to_fp16 = const()[name = tensor("op_3210_to_fp16"), val = tensor(0x1p-3)]; - tensor var_3211_cast_fp16 = mul(x = var_3209_cast_fp16, y = var_3210_to_fp16)[name = tensor("op_3211_cast_fp16")]; - tensor var_3212 = const()[name = tensor("op_3212"), val = tensor([1, 20, 64, -1])]; - tensor var_3213_cast_fp16 = reshape(shape = var_3212, x = key_31_cast_fp16)[name = tensor("op_3213_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_3211_cast_fp16, y = var_3213_cast_fp16)[name = tensor("mh_w_47_cast_fp16")]; - tensor var_3216_cast_fp16 = softmax(axis = var_2934, x = mh_w_47_cast_fp16)[name = tensor("op_3216_cast_fp16")]; - tensor var_3217 = const()[name = tensor("op_3217"), val = tensor([1, 20, 64, -1])]; - tensor var_3218_cast_fp16 = reshape(shape = var_3217, x = value_31_cast_fp16)[name = tensor("op_3218_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_3218_cast_fp16, y = var_3216_cast_fp16)[name = tensor("attn_31_cast_fp16")]; - tensor var_3221 = const()[name = tensor("op_3221"), val = tensor([1, 1280, 1, -1])]; - tensor input_227_cast_fp16 = reshape(shape = var_3221, x = attn_31_cast_fp16)[name = tensor("input_227_cast_fp16")]; - tensor var_3228 = const()[name = tensor("op_3228"), val = tensor([1, 1])]; - tensor var_3230 = const()[name = tensor("op_3230"), val = tensor([1, 1])]; - tensor pretrained_out_155_pad_type_0 = const()[name = tensor("pretrained_out_155_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_155_pad_0 = const()[name = tensor("pretrained_out_155_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239809152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240628416))), name = tensor("layers_7_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_7_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240628544)))]; - tensor pretrained_out_155_cast_fp16 = conv(bias = layers_7_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_3230, groups = var_2941, pad = pretrained_out_155_pad_0, pad_type = pretrained_out_155_pad_type_0, strides = var_3228, weight = layers_7_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_227_cast_fp16)[name = tensor("pretrained_out_155_cast_fp16")]; - tensor var_3234 = const()[name = tensor("op_3234"), val = tensor([1, 1])]; - tensor var_3236 = const()[name = tensor("op_3236"), val = tensor([1, 1])]; - tensor input_229_pad_type_0 = const()[name = tensor("input_229_pad_type_0"), val = tensor("custom")]; - tensor input_229_pad_0 = const()[name = tensor("input_229_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240631168)))]; - tensor input_229_cast_fp16 = conv(dilations = var_3236, groups = var_2941, pad = input_229_pad_0, pad_type = input_229_pad_type_0, strides = var_3234, weight = layers_7_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; - tensor var_3240 = const()[name = tensor("op_3240"), val = tensor([1, 1])]; - tensor var_3242 = const()[name = tensor("op_3242"), val = tensor([1, 1])]; - tensor lora_out_309_pad_type_0 = const()[name = tensor("lora_out_309_pad_type_0"), val = tensor("custom")]; - tensor lora_out_309_pad_0 = const()[name = tensor("lora_out_309_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_311_weight_0_to_fp16 = const()[name = tensor("lora_out_311_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240672192)))]; - tensor lora_out_311_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3242, groups = var_2941, pad = lora_out_309_pad_0, pad_type = lora_out_309_pad_type_0, strides = var_3240, weight = lora_out_311_weight_0_to_fp16, x = input_229_cast_fp16)[name = tensor("lora_out_311_cast_fp16")]; - tensor obj_95_cast_fp16 = add(x = pretrained_out_155_cast_fp16, y = lora_out_311_cast_fp16)[name = tensor("obj_95_cast_fp16")]; - tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_95_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; - tensor var_3251 = const()[name = tensor("op_3251"), val = tensor([1])]; - tensor channels_mean_47_cast_fp16 = reduce_mean(axes = var_3251, keep_dims = var_2942, 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_3255 = const()[name = tensor("op_3255"), val = tensor([1])]; - tensor var_3256_cast_fp16 = reduce_mean(axes = var_3255, keep_dims = var_2942, x = zero_mean_sq_47_cast_fp16)[name = tensor("op_3256_cast_fp16")]; - tensor var_3257_to_fp16 = const()[name = tensor("op_3257_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_3258_cast_fp16 = add(x = var_3256_cast_fp16, y = var_3257_to_fp16)[name = tensor("op_3258_cast_fp16")]; - tensor denom_47_epsilon_0 = const()[name = tensor("denom_47_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_47_cast_fp16 = rsqrt(epsilon = denom_47_epsilon_0, x = var_3258_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_231_gamma_0_to_fp16 = const()[name = tensor("input_231_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240713216)))]; - tensor input_231_beta_0_to_fp16 = const()[name = tensor("input_231_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240715840)))]; - tensor input_231_epsilon_0_to_fp16 = const()[name = tensor("input_231_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_231_cast_fp16 = batch_norm(beta = input_231_beta_0_to_fp16, epsilon = input_231_epsilon_0_to_fp16, gamma = input_231_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_231_cast_fp16")]; - tensor var_3272 = const()[name = tensor("op_3272"), val = tensor([1, 1])]; - tensor var_3274 = const()[name = tensor("op_3274"), val = tensor([1, 1])]; - tensor pretrained_out_157_pad_type_0 = const()[name = tensor("pretrained_out_157_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_157_pad_0 = const()[name = tensor("pretrained_out_157_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240718464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243995328))), name = tensor("layers_7_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_7_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_7_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243995456)))]; - tensor pretrained_out_157_cast_fp16 = conv(bias = layers_7_fc1_pretrained_bias_to_fp16, dilations = var_3274, groups = var_2941, pad = pretrained_out_157_pad_0, pad_type = pretrained_out_157_pad_type_0, strides = var_3272, weight = layers_7_fc1_pretrained_weight_to_fp16_palettized, x = input_231_cast_fp16)[name = tensor("pretrained_out_157_cast_fp16")]; - tensor var_3278 = const()[name = tensor("op_3278"), val = tensor([1, 1])]; - tensor var_3280 = const()[name = tensor("op_3280"), val = tensor([1, 1])]; - tensor input_233_pad_type_0 = const()[name = tensor("input_233_pad_type_0"), val = tensor("custom")]; - tensor input_233_pad_0 = const()[name = tensor("input_233_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_7_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244005760)))]; - tensor input_233_cast_fp16 = conv(dilations = var_3280, groups = var_2941, pad = input_233_pad_0, pad_type = input_233_pad_type_0, strides = var_3278, weight = layers_7_fc1_loraA_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("input_233_cast_fp16")]; - tensor var_3284 = const()[name = tensor("op_3284"), val = tensor([1, 1])]; - tensor var_3286 = const()[name = tensor("op_3286"), val = tensor([1, 1])]; - tensor lora_out_313_pad_type_0 = const()[name = tensor("lora_out_313_pad_type_0"), val = tensor("custom")]; - tensor lora_out_313_pad_0 = const()[name = tensor("lora_out_313_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_315_weight_0_to_fp16 = const()[name = tensor("lora_out_315_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244046784)))]; - tensor lora_out_315_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_3286, groups = var_2941, pad = lora_out_313_pad_0, pad_type = lora_out_313_pad_type_0, strides = var_3284, weight = lora_out_315_weight_0_to_fp16, x = input_233_cast_fp16)[name = tensor("lora_out_315_cast_fp16")]; - tensor input_235_cast_fp16 = add(x = pretrained_out_157_cast_fp16, y = lora_out_315_cast_fp16)[name = tensor("input_235_cast_fp16")]; - tensor input_237_mode_0 = const()[name = tensor("input_237_mode_0"), val = tensor("EXACT")]; - tensor input_237_cast_fp16 = gelu(mode = input_237_mode_0, x = input_235_cast_fp16)[name = tensor("input_237_cast_fp16")]; - tensor var_3298 = const()[name = tensor("op_3298"), val = tensor([1, 1])]; - tensor var_3300 = const()[name = tensor("op_3300"), val = tensor([1, 1])]; - tensor pretrained_out_159_pad_type_0 = const()[name = tensor("pretrained_out_159_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_159_pad_0 = const()[name = tensor("pretrained_out_159_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244210688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247487552))), name = tensor("layers_7_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_7_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_7_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247487680)))]; - tensor pretrained_out_159_cast_fp16 = conv(bias = layers_7_fc2_pretrained_bias_to_fp16, dilations = var_3300, groups = var_2941, pad = pretrained_out_159_pad_0, pad_type = pretrained_out_159_pad_type_0, strides = var_3298, weight = layers_7_fc2_pretrained_weight_to_fp16_palettized, x = input_237_cast_fp16)[name = tensor("pretrained_out_159_cast_fp16")]; - tensor var_3304 = const()[name = tensor("op_3304"), val = tensor([1, 1])]; - tensor var_3306 = const()[name = tensor("op_3306"), val = tensor([1, 1])]; - tensor input_239_pad_type_0 = const()[name = tensor("input_239_pad_type_0"), val = tensor("custom")]; - tensor input_239_pad_0 = const()[name = tensor("input_239_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_7_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_7_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247490304)))]; - tensor input_239_cast_fp16 = conv(dilations = var_3306, groups = var_2941, pad = input_239_pad_0, pad_type = input_239_pad_type_0, strides = var_3304, weight = layers_7_fc2_loraA_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; - tensor var_3310 = const()[name = tensor("op_3310"), val = tensor([1, 1])]; - tensor var_3312 = const()[name = tensor("op_3312"), val = tensor([1, 1])]; - tensor lora_out_317_pad_type_0 = const()[name = tensor("lora_out_317_pad_type_0"), val = tensor("custom")]; - tensor lora_out_317_pad_0 = const()[name = tensor("lora_out_317_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_319_weight_0_to_fp16 = const()[name = tensor("lora_out_319_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247654208)))]; - tensor lora_out_319_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3312, groups = var_2941, pad = lora_out_317_pad_0, pad_type = lora_out_317_pad_type_0, strides = var_3310, weight = lora_out_319_weight_0_to_fp16, x = input_239_cast_fp16)[name = tensor("lora_out_319_cast_fp16")]; - tensor hidden_states_17_cast_fp16 = add(x = pretrained_out_159_cast_fp16, y = lora_out_319_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_3328 = const()[name = tensor("op_3328"), val = tensor(3)]; - tensor var_3335 = const()[name = tensor("op_3335"), val = tensor(1)]; - tensor var_3336 = const()[name = tensor("op_3336"), val = tensor(true)]; - tensor var_3348 = const()[name = tensor("op_3348"), val = tensor([1])]; - tensor channels_mean_49_cast_fp16 = reduce_mean(axes = var_3348, keep_dims = var_3336, 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_3352 = const()[name = tensor("op_3352"), val = tensor([1])]; - tensor var_3353_cast_fp16 = reduce_mean(axes = var_3352, keep_dims = var_3336, x = zero_mean_sq_49_cast_fp16)[name = tensor("op_3353_cast_fp16")]; - tensor var_3354_to_fp16 = const()[name = tensor("op_3354_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_3355_cast_fp16 = add(x = var_3353_cast_fp16, y = var_3354_to_fp16)[name = tensor("op_3355_cast_fp16")]; - tensor denom_49_epsilon_0 = const()[name = tensor("denom_49_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_49_cast_fp16 = rsqrt(epsilon = denom_49_epsilon_0, x = var_3355_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_97_gamma_0_to_fp16 = const()[name = tensor("obj_97_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247695232)))]; - tensor obj_97_beta_0_to_fp16 = const()[name = tensor("obj_97_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247697856)))]; - tensor obj_97_epsilon_0_to_fp16 = const()[name = tensor("obj_97_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_97_cast_fp16 = batch_norm(beta = obj_97_beta_0_to_fp16, epsilon = obj_97_epsilon_0_to_fp16, gamma = obj_97_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_97_cast_fp16")]; - tensor var_3373 = const()[name = tensor("op_3373"), val = tensor([1, 1])]; - tensor var_3375 = const()[name = tensor("op_3375"), val = tensor([1, 1])]; - tensor pretrained_out_161_pad_type_0 = const()[name = tensor("pretrained_out_161_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_161_pad_0 = const()[name = tensor("pretrained_out_161_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247700480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248519744))), name = tensor("layers_8_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_8_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248519872)))]; - tensor pretrained_out_161_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_3375, groups = var_3335, pad = pretrained_out_161_pad_0, pad_type = pretrained_out_161_pad_type_0, strides = var_3373, weight = layers_8_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_97_cast_fp16)[name = tensor("pretrained_out_161_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_241_pad_type_0 = const()[name = tensor("input_241_pad_type_0"), val = tensor("custom")]; - tensor input_241_pad_0 = const()[name = tensor("input_241_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248522496)))]; - tensor input_241_cast_fp16 = conv(dilations = var_3381, groups = var_3335, pad = input_241_pad_0, pad_type = input_241_pad_type_0, strides = var_3379, weight = layers_8_self_attn_q_proj_loraA_weight_to_fp16, x = obj_97_cast_fp16)[name = tensor("input_241_cast_fp16")]; - tensor var_3385 = const()[name = tensor("op_3385"), val = tensor([1, 1])]; - tensor var_3387 = const()[name = tensor("op_3387"), val = tensor([1, 1])]; - tensor lora_out_321_pad_type_0 = const()[name = tensor("lora_out_321_pad_type_0"), val = tensor("custom")]; - tensor lora_out_321_pad_0 = const()[name = tensor("lora_out_321_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_323_weight_0_to_fp16 = const()[name = tensor("lora_out_323_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248563520)))]; - tensor lora_out_323_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3387, groups = var_3335, pad = lora_out_321_pad_0, pad_type = lora_out_321_pad_type_0, strides = var_3385, weight = lora_out_323_weight_0_to_fp16, x = input_241_cast_fp16)[name = tensor("lora_out_323_cast_fp16")]; - tensor query_33_cast_fp16 = add(x = pretrained_out_161_cast_fp16, y = lora_out_323_cast_fp16)[name = tensor("query_33_cast_fp16")]; - tensor var_3397 = const()[name = tensor("op_3397"), val = tensor([1, 1])]; - tensor var_3399 = const()[name = tensor("op_3399"), val = tensor([1, 1])]; - tensor pretrained_out_163_pad_type_0 = const()[name = tensor("pretrained_out_163_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_163_pad_0 = const()[name = tensor("pretrained_out_163_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248604544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249423808))), name = tensor("layers_8_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_163_cast_fp16 = conv(dilations = var_3399, groups = var_3335, pad = pretrained_out_163_pad_0, pad_type = pretrained_out_163_pad_type_0, strides = var_3397, weight = layers_8_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_97_cast_fp16)[name = tensor("pretrained_out_163_cast_fp16")]; - tensor var_3403 = const()[name = tensor("op_3403"), val = tensor([1, 1])]; - tensor var_3405 = const()[name = tensor("op_3405"), val = tensor([1, 1])]; - tensor input_243_pad_type_0 = const()[name = tensor("input_243_pad_type_0"), val = tensor("custom")]; - tensor input_243_pad_0 = const()[name = tensor("input_243_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249423936)))]; - tensor input_243_cast_fp16 = conv(dilations = var_3405, groups = var_3335, pad = input_243_pad_0, pad_type = input_243_pad_type_0, strides = var_3403, weight = layers_8_self_attn_k_proj_loraA_weight_to_fp16, x = obj_97_cast_fp16)[name = tensor("input_243_cast_fp16")]; - tensor var_3409 = const()[name = tensor("op_3409"), val = tensor([1, 1])]; - tensor var_3411 = const()[name = tensor("op_3411"), val = tensor([1, 1])]; - tensor lora_out_325_pad_type_0 = const()[name = tensor("lora_out_325_pad_type_0"), val = tensor("custom")]; - tensor lora_out_325_pad_0 = const()[name = tensor("lora_out_325_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_327_weight_0_to_fp16 = const()[name = tensor("lora_out_327_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249464960)))]; - tensor lora_out_327_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3411, groups = var_3335, pad = lora_out_325_pad_0, pad_type = lora_out_325_pad_type_0, strides = var_3409, weight = lora_out_327_weight_0_to_fp16, x = input_243_cast_fp16)[name = tensor("lora_out_327_cast_fp16")]; - tensor current_key_17_cast_fp16 = add(x = pretrained_out_163_cast_fp16, y = lora_out_327_cast_fp16)[name = tensor("current_key_17_cast_fp16")]; - tensor var_3422 = const()[name = tensor("op_3422"), val = tensor([1, 1])]; - tensor var_3424 = const()[name = tensor("op_3424"), val = tensor([1, 1])]; - tensor pretrained_out_165_pad_type_0 = const()[name = tensor("pretrained_out_165_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_165_pad_0 = const()[name = tensor("pretrained_out_165_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249505984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250325248))), name = tensor("layers_8_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_8_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250325376)))]; - tensor pretrained_out_165_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_3424, groups = var_3335, pad = pretrained_out_165_pad_0, pad_type = pretrained_out_165_pad_type_0, strides = var_3422, weight = layers_8_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_97_cast_fp16)[name = tensor("pretrained_out_165_cast_fp16")]; - tensor var_3428 = const()[name = tensor("op_3428"), val = tensor([1, 1])]; - tensor var_3430 = const()[name = tensor("op_3430"), val = tensor([1, 1])]; - tensor input_245_pad_type_0 = const()[name = tensor("input_245_pad_type_0"), val = tensor("custom")]; - tensor input_245_pad_0 = const()[name = tensor("input_245_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250328000)))]; - tensor input_245_cast_fp16 = conv(dilations = var_3430, groups = var_3335, pad = input_245_pad_0, pad_type = input_245_pad_type_0, strides = var_3428, weight = layers_8_self_attn_v_proj_loraA_weight_to_fp16, x = obj_97_cast_fp16)[name = tensor("input_245_cast_fp16")]; - tensor var_3434 = const()[name = tensor("op_3434"), val = tensor([1, 1])]; - tensor var_3436 = const()[name = tensor("op_3436"), val = tensor([1, 1])]; - tensor lora_out_329_pad_type_0 = const()[name = tensor("lora_out_329_pad_type_0"), val = tensor("custom")]; - tensor lora_out_329_pad_0 = const()[name = tensor("lora_out_329_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_331_weight_0_to_fp16 = const()[name = tensor("lora_out_331_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250369024)))]; - tensor lora_out_331_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3436, groups = var_3335, pad = lora_out_329_pad_0, pad_type = lora_out_329_pad_type_0, strides = var_3434, weight = lora_out_331_weight_0_to_fp16, x = input_245_cast_fp16)[name = tensor("lora_out_331_cast_fp16")]; - tensor current_value_17_cast_fp16 = add(x = pretrained_out_165_cast_fp16, y = lora_out_331_cast_fp16)[name = tensor("current_value_17_cast_fp16")]; - tensor var_3446_cast_fp16 = mul(x = current_key_17_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_3446_cast_fp16")]; - tensor var_3448_cast_fp16 = mul(x = var_103_cast_fp16_8, y = var_295_cast_fp16)[name = tensor("op_3448_cast_fp16")]; - tensor key_33_cast_fp16 = add(x = var_3446_cast_fp16, y = var_3448_cast_fp16)[name = tensor("key_33_cast_fp16")]; - tensor var_3450_cast_fp16 = mul(x = current_value_17_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_3450_cast_fp16")]; - tensor var_3452_cast_fp16 = mul(x = var_138_cast_fp16_8, y = var_295_cast_fp16)[name = tensor("op_3452_cast_fp16")]; - tensor value_33_cast_fp16 = add(x = var_3450_cast_fp16, y = var_3452_cast_fp16)[name = tensor("value_33_cast_fp16")]; - tensor var_3455 = const()[name = tensor("op_3455"), val = tensor([1, 20, 64, -1])]; - tensor var_3456_cast_fp16 = reshape(shape = var_3455, x = query_33_cast_fp16)[name = tensor("op_3456_cast_fp16")]; - tensor var_3457_to_fp16 = const()[name = tensor("op_3457_to_fp16"), val = tensor(0x1p-3)]; - tensor var_3458_cast_fp16 = mul(x = var_3456_cast_fp16, y = var_3457_to_fp16)[name = tensor("op_3458_cast_fp16")]; - tensor var_3459 = const()[name = tensor("op_3459"), val = tensor([1, 20, 64, -1])]; - tensor var_3460_cast_fp16 = reshape(shape = var_3459, x = key_33_cast_fp16)[name = tensor("op_3460_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_3458_cast_fp16, y = var_3460_cast_fp16)[name = tensor("mh_w_49_cast_fp16")]; - tensor mh_w_51_cast_fp16 = add(x = mh_w_49_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_51_cast_fp16")]; - tensor var_3468_cast_fp16 = softmax(axis = var_3328, x = mh_w_51_cast_fp16)[name = tensor("op_3468_cast_fp16")]; - tensor var_3469 = const()[name = tensor("op_3469"), val = tensor([1, 20, 64, -1])]; - tensor var_3470_cast_fp16 = reshape(shape = var_3469, x = value_33_cast_fp16)[name = tensor("op_3470_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_3470_cast_fp16, y = var_3468_cast_fp16)[name = tensor("attn_33_cast_fp16")]; - tensor var_3473 = const()[name = tensor("op_3473"), val = tensor([1, 1280, 1, -1])]; - tensor input_247_cast_fp16 = reshape(shape = var_3473, x = attn_33_cast_fp16)[name = tensor("input_247_cast_fp16")]; - tensor var_3480 = const()[name = tensor("op_3480"), val = tensor([1, 1])]; - tensor var_3482 = const()[name = tensor("op_3482"), val = tensor([1, 1])]; - tensor pretrained_out_167_pad_type_0 = const()[name = tensor("pretrained_out_167_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_167_pad_0 = const()[name = tensor("pretrained_out_167_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250410048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251229312))), name = tensor("layers_8_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_8_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251229440)))]; - tensor pretrained_out_167_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_3482, groups = var_3335, pad = pretrained_out_167_pad_0, pad_type = pretrained_out_167_pad_type_0, strides = var_3480, weight = layers_8_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = tensor("pretrained_out_167_cast_fp16")]; - tensor var_3486 = const()[name = tensor("op_3486"), val = tensor([1, 1])]; - tensor var_3488 = const()[name = tensor("op_3488"), val = tensor([1, 1])]; - tensor input_249_pad_type_0 = const()[name = tensor("input_249_pad_type_0"), val = tensor("custom")]; - tensor input_249_pad_0 = const()[name = tensor("input_249_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251232064)))]; - tensor input_249_cast_fp16 = conv(dilations = var_3488, groups = var_3335, pad = input_249_pad_0, pad_type = input_249_pad_type_0, strides = var_3486, weight = layers_8_self_attn_o_proj_loraA_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; - tensor var_3492 = const()[name = tensor("op_3492"), val = tensor([1, 1])]; - tensor var_3494 = const()[name = tensor("op_3494"), val = tensor([1, 1])]; - tensor lora_out_333_pad_type_0 = const()[name = tensor("lora_out_333_pad_type_0"), val = tensor("custom")]; - tensor lora_out_333_pad_0 = const()[name = tensor("lora_out_333_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_335_weight_0_to_fp16 = const()[name = tensor("lora_out_335_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251273088)))]; - tensor lora_out_335_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3494, groups = var_3335, pad = lora_out_333_pad_0, pad_type = lora_out_333_pad_type_0, strides = var_3492, weight = lora_out_335_weight_0_to_fp16, x = input_249_cast_fp16)[name = tensor("lora_out_335_cast_fp16")]; - tensor obj_103_cast_fp16 = add(x = pretrained_out_167_cast_fp16, y = lora_out_335_cast_fp16)[name = tensor("obj_103_cast_fp16")]; - tensor inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_103_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; - tensor var_3507 = const()[name = tensor("op_3507"), val = tensor([1])]; - tensor channels_mean_51_cast_fp16 = reduce_mean(axes = var_3507, keep_dims = var_3336, 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_3511 = const()[name = tensor("op_3511"), val = tensor([1])]; - tensor var_3512_cast_fp16 = reduce_mean(axes = var_3511, keep_dims = var_3336, x = zero_mean_sq_51_cast_fp16)[name = tensor("op_3512_cast_fp16")]; - tensor var_3513_to_fp16 = const()[name = tensor("op_3513_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_3514_cast_fp16 = add(x = var_3512_cast_fp16, y = var_3513_to_fp16)[name = tensor("op_3514_cast_fp16")]; - tensor denom_51_epsilon_0 = const()[name = tensor("denom_51_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_51_cast_fp16 = rsqrt(epsilon = denom_51_epsilon_0, x = var_3514_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_105_gamma_0_to_fp16 = const()[name = tensor("obj_105_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251314112)))]; - tensor obj_105_beta_0_to_fp16 = const()[name = tensor("obj_105_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251316736)))]; - tensor obj_105_epsilon_0_to_fp16 = const()[name = tensor("obj_105_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_105_cast_fp16 = batch_norm(beta = obj_105_beta_0_to_fp16, epsilon = obj_105_epsilon_0_to_fp16, gamma = obj_105_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_105_cast_fp16")]; - tensor var_3532 = const()[name = tensor("op_3532"), val = tensor([1, 1])]; - tensor var_3534 = const()[name = tensor("op_3534"), val = tensor([1, 1])]; - tensor pretrained_out_169_pad_type_0 = const()[name = tensor("pretrained_out_169_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_169_pad_0 = const()[name = tensor("pretrained_out_169_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251319360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252138624))), name = tensor("layers_8_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_8_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252138752)))]; - tensor pretrained_out_169_cast_fp16 = conv(bias = layers_8_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_3534, groups = var_3335, pad = pretrained_out_169_pad_0, pad_type = pretrained_out_169_pad_type_0, strides = var_3532, weight = layers_8_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_105_cast_fp16)[name = tensor("pretrained_out_169_cast_fp16")]; - tensor var_3538 = const()[name = tensor("op_3538"), val = tensor([1, 1])]; - tensor var_3540 = const()[name = tensor("op_3540"), val = tensor([1, 1])]; - tensor input_251_pad_type_0 = const()[name = tensor("input_251_pad_type_0"), val = tensor("custom")]; - tensor input_251_pad_0 = const()[name = tensor("input_251_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252141376)))]; - tensor input_251_cast_fp16 = conv(dilations = var_3540, groups = var_3335, pad = input_251_pad_0, pad_type = input_251_pad_type_0, strides = var_3538, weight = layers_8_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_105_cast_fp16)[name = tensor("input_251_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 lora_out_337_pad_type_0 = const()[name = tensor("lora_out_337_pad_type_0"), val = tensor("custom")]; - tensor lora_out_337_pad_0 = const()[name = tensor("lora_out_337_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_339_weight_0_to_fp16 = const()[name = tensor("lora_out_339_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252182400)))]; - tensor lora_out_339_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3546, groups = var_3335, pad = lora_out_337_pad_0, pad_type = lora_out_337_pad_type_0, strides = var_3544, weight = lora_out_339_weight_0_to_fp16, x = input_251_cast_fp16)[name = tensor("lora_out_339_cast_fp16")]; - tensor query_35_cast_fp16 = add(x = pretrained_out_169_cast_fp16, y = lora_out_339_cast_fp16)[name = tensor("query_35_cast_fp16")]; - tensor var_3556 = const()[name = tensor("op_3556"), val = tensor([1, 1])]; - tensor var_3558 = const()[name = tensor("op_3558"), val = tensor([1, 1])]; - tensor pretrained_out_171_pad_type_0 = const()[name = tensor("pretrained_out_171_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_171_pad_0 = const()[name = tensor("pretrained_out_171_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252223424))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253042688))), name = tensor("layers_8_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_171_cast_fp16 = conv(dilations = var_3558, groups = var_3335, pad = pretrained_out_171_pad_0, pad_type = pretrained_out_171_pad_type_0, strides = var_3556, weight = layers_8_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_171_cast_fp16")]; - tensor var_3562 = const()[name = tensor("op_3562"), val = tensor([1, 1])]; - tensor var_3564 = const()[name = tensor("op_3564"), val = tensor([1, 1])]; - tensor input_253_pad_type_0 = const()[name = tensor("input_253_pad_type_0"), val = tensor("custom")]; - tensor input_253_pad_0 = const()[name = tensor("input_253_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253042816)))]; - tensor input_253_cast_fp16 = conv(dilations = var_3564, groups = var_3335, pad = input_253_pad_0, pad_type = input_253_pad_type_0, strides = var_3562, weight = layers_8_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_253_cast_fp16")]; - tensor var_3568 = const()[name = tensor("op_3568"), val = tensor([1, 1])]; - tensor var_3570 = const()[name = tensor("op_3570"), val = tensor([1, 1])]; - tensor lora_out_341_pad_type_0 = const()[name = tensor("lora_out_341_pad_type_0"), val = tensor("custom")]; - tensor lora_out_341_pad_0 = const()[name = tensor("lora_out_341_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_343_weight_0_to_fp16 = const()[name = tensor("lora_out_343_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253083840)))]; - tensor lora_out_343_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3570, groups = var_3335, pad = lora_out_341_pad_0, pad_type = lora_out_341_pad_type_0, strides = var_3568, weight = lora_out_343_weight_0_to_fp16, x = input_253_cast_fp16)[name = tensor("lora_out_343_cast_fp16")]; - tensor key_35_cast_fp16 = add(x = pretrained_out_171_cast_fp16, y = lora_out_343_cast_fp16)[name = tensor("key_35_cast_fp16")]; - tensor var_3581 = const()[name = tensor("op_3581"), val = tensor([1, 1])]; - tensor var_3583 = const()[name = tensor("op_3583"), val = tensor([1, 1])]; - tensor pretrained_out_173_pad_type_0 = const()[name = tensor("pretrained_out_173_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_173_pad_0 = const()[name = tensor("pretrained_out_173_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253124864))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253944128))), name = tensor("layers_8_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_8_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253944256)))]; - tensor pretrained_out_173_cast_fp16 = conv(bias = layers_8_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_3583, groups = var_3335, pad = pretrained_out_173_pad_0, pad_type = pretrained_out_173_pad_type_0, strides = var_3581, weight = layers_8_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_173_cast_fp16")]; - tensor var_3587 = const()[name = tensor("op_3587"), val = tensor([1, 1])]; - tensor var_3589 = const()[name = tensor("op_3589"), val = tensor([1, 1])]; - tensor input_255_pad_type_0 = const()[name = tensor("input_255_pad_type_0"), val = tensor("custom")]; - tensor input_255_pad_0 = const()[name = tensor("input_255_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253946880)))]; - tensor input_255_cast_fp16 = conv(dilations = var_3589, groups = var_3335, pad = input_255_pad_0, pad_type = input_255_pad_type_0, strides = var_3587, weight = layers_8_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_255_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 lora_out_345_pad_type_0 = const()[name = tensor("lora_out_345_pad_type_0"), val = tensor("custom")]; - tensor lora_out_345_pad_0 = const()[name = tensor("lora_out_345_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_347_weight_0_to_fp16 = const()[name = tensor("lora_out_347_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253987904)))]; - tensor lora_out_347_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3595, groups = var_3335, pad = lora_out_345_pad_0, pad_type = lora_out_345_pad_type_0, strides = var_3593, weight = lora_out_347_weight_0_to_fp16, x = input_255_cast_fp16)[name = tensor("lora_out_347_cast_fp16")]; - tensor value_35_cast_fp16 = add(x = pretrained_out_173_cast_fp16, y = lora_out_347_cast_fp16)[name = tensor("value_35_cast_fp16")]; - tensor var_3602 = const()[name = tensor("op_3602"), val = tensor([1, 20, 64, -1])]; - tensor var_3603_cast_fp16 = reshape(shape = var_3602, x = query_35_cast_fp16)[name = tensor("op_3603_cast_fp16")]; - tensor var_3604_to_fp16 = const()[name = tensor("op_3604_to_fp16"), val = tensor(0x1p-3)]; - tensor var_3605_cast_fp16 = mul(x = var_3603_cast_fp16, y = var_3604_to_fp16)[name = tensor("op_3605_cast_fp16")]; - tensor var_3606 = const()[name = tensor("op_3606"), val = tensor([1, 20, 64, -1])]; - tensor var_3607_cast_fp16 = reshape(shape = var_3606, x = key_35_cast_fp16)[name = tensor("op_3607_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_3605_cast_fp16, y = var_3607_cast_fp16)[name = tensor("mh_w_53_cast_fp16")]; - tensor var_3610_cast_fp16 = softmax(axis = var_3328, x = mh_w_53_cast_fp16)[name = tensor("op_3610_cast_fp16")]; - tensor var_3611 = const()[name = tensor("op_3611"), val = tensor([1, 20, 64, -1])]; - tensor var_3612_cast_fp16 = reshape(shape = var_3611, x = value_35_cast_fp16)[name = tensor("op_3612_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_3612_cast_fp16, y = var_3610_cast_fp16)[name = tensor("attn_35_cast_fp16")]; - tensor var_3615 = const()[name = tensor("op_3615"), val = tensor([1, 1280, 1, -1])]; - tensor input_257_cast_fp16 = reshape(shape = var_3615, x = attn_35_cast_fp16)[name = tensor("input_257_cast_fp16")]; - tensor var_3622 = const()[name = tensor("op_3622"), val = tensor([1, 1])]; - tensor var_3624 = const()[name = tensor("op_3624"), val = tensor([1, 1])]; - tensor pretrained_out_175_pad_type_0 = const()[name = tensor("pretrained_out_175_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_175_pad_0 = const()[name = tensor("pretrained_out_175_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254028928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254848192))), name = tensor("layers_8_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_8_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254848320)))]; - tensor pretrained_out_175_cast_fp16 = conv(bias = layers_8_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_3624, groups = var_3335, pad = pretrained_out_175_pad_0, pad_type = pretrained_out_175_pad_type_0, strides = var_3622, weight = layers_8_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_257_cast_fp16)[name = tensor("pretrained_out_175_cast_fp16")]; - tensor var_3628 = const()[name = tensor("op_3628"), val = tensor([1, 1])]; - tensor var_3630 = const()[name = tensor("op_3630"), val = tensor([1, 1])]; - tensor input_259_pad_type_0 = const()[name = tensor("input_259_pad_type_0"), val = tensor("custom")]; - tensor input_259_pad_0 = const()[name = tensor("input_259_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254850944)))]; - tensor input_259_cast_fp16 = conv(dilations = var_3630, groups = var_3335, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = var_3628, weight = layers_8_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("input_259_cast_fp16")]; - tensor var_3634 = const()[name = tensor("op_3634"), val = tensor([1, 1])]; - tensor var_3636 = const()[name = tensor("op_3636"), val = tensor([1, 1])]; - tensor lora_out_349_pad_type_0 = const()[name = tensor("lora_out_349_pad_type_0"), val = tensor("custom")]; - tensor lora_out_349_pad_0 = const()[name = tensor("lora_out_349_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_351_weight_0_to_fp16 = const()[name = tensor("lora_out_351_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254891968)))]; - tensor lora_out_351_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3636, groups = var_3335, pad = lora_out_349_pad_0, pad_type = lora_out_349_pad_type_0, strides = var_3634, weight = lora_out_351_weight_0_to_fp16, x = input_259_cast_fp16)[name = tensor("lora_out_351_cast_fp16")]; - tensor obj_107_cast_fp16 = add(x = pretrained_out_175_cast_fp16, y = lora_out_351_cast_fp16)[name = tensor("obj_107_cast_fp16")]; - tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = obj_107_cast_fp16)[name = tensor("inputs_53_cast_fp16")]; - tensor var_3645 = const()[name = tensor("op_3645"), val = tensor([1])]; - tensor channels_mean_53_cast_fp16 = reduce_mean(axes = var_3645, keep_dims = var_3336, 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_3649 = const()[name = tensor("op_3649"), val = tensor([1])]; - tensor var_3650_cast_fp16 = reduce_mean(axes = var_3649, keep_dims = var_3336, x = zero_mean_sq_53_cast_fp16)[name = tensor("op_3650_cast_fp16")]; - tensor var_3651_to_fp16 = const()[name = tensor("op_3651_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_3652_cast_fp16 = add(x = var_3650_cast_fp16, y = var_3651_to_fp16)[name = tensor("op_3652_cast_fp16")]; - tensor denom_53_epsilon_0 = const()[name = tensor("denom_53_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_53_cast_fp16 = rsqrt(epsilon = denom_53_epsilon_0, x = var_3652_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_261_gamma_0_to_fp16 = const()[name = tensor("input_261_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254932992)))]; - tensor input_261_beta_0_to_fp16 = const()[name = tensor("input_261_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254935616)))]; - tensor input_261_epsilon_0_to_fp16 = const()[name = tensor("input_261_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_261_cast_fp16 = batch_norm(beta = input_261_beta_0_to_fp16, epsilon = input_261_epsilon_0_to_fp16, gamma = input_261_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_261_cast_fp16")]; - tensor var_3666 = const()[name = tensor("op_3666"), val = tensor([1, 1])]; - tensor var_3668 = const()[name = tensor("op_3668"), val = tensor([1, 1])]; - tensor pretrained_out_177_pad_type_0 = const()[name = tensor("pretrained_out_177_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_177_pad_0 = const()[name = tensor("pretrained_out_177_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254938240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258215104))), name = tensor("layers_8_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_8_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_8_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258215232)))]; - tensor pretrained_out_177_cast_fp16 = conv(bias = layers_8_fc1_pretrained_bias_to_fp16, dilations = var_3668, groups = var_3335, pad = pretrained_out_177_pad_0, pad_type = pretrained_out_177_pad_type_0, strides = var_3666, weight = layers_8_fc1_pretrained_weight_to_fp16_palettized, x = input_261_cast_fp16)[name = tensor("pretrained_out_177_cast_fp16")]; - tensor var_3672 = const()[name = tensor("op_3672"), val = tensor([1, 1])]; - tensor var_3674 = const()[name = tensor("op_3674"), val = tensor([1, 1])]; - tensor input_263_pad_type_0 = const()[name = tensor("input_263_pad_type_0"), val = tensor("custom")]; - tensor input_263_pad_0 = const()[name = tensor("input_263_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_8_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258225536)))]; - tensor input_263_cast_fp16 = conv(dilations = var_3674, groups = var_3335, pad = input_263_pad_0, pad_type = input_263_pad_type_0, strides = var_3672, weight = layers_8_fc1_loraA_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("input_263_cast_fp16")]; - tensor var_3678 = const()[name = tensor("op_3678"), val = tensor([1, 1])]; - tensor var_3680 = const()[name = tensor("op_3680"), val = tensor([1, 1])]; - tensor lora_out_353_pad_type_0 = const()[name = tensor("lora_out_353_pad_type_0"), val = tensor("custom")]; - tensor lora_out_353_pad_0 = const()[name = tensor("lora_out_353_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_355_weight_0_to_fp16 = const()[name = tensor("lora_out_355_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258266560)))]; - tensor lora_out_355_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_3680, groups = var_3335, pad = lora_out_353_pad_0, pad_type = lora_out_353_pad_type_0, strides = var_3678, weight = lora_out_355_weight_0_to_fp16, x = input_263_cast_fp16)[name = tensor("lora_out_355_cast_fp16")]; - tensor input_265_cast_fp16 = add(x = pretrained_out_177_cast_fp16, y = lora_out_355_cast_fp16)[name = tensor("input_265_cast_fp16")]; - tensor input_267_mode_0 = const()[name = tensor("input_267_mode_0"), val = tensor("EXACT")]; - tensor input_267_cast_fp16 = gelu(mode = input_267_mode_0, x = input_265_cast_fp16)[name = tensor("input_267_cast_fp16")]; - tensor var_3692 = const()[name = tensor("op_3692"), val = tensor([1, 1])]; - tensor var_3694 = const()[name = tensor("op_3694"), val = tensor([1, 1])]; - tensor pretrained_out_179_pad_type_0 = const()[name = tensor("pretrained_out_179_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_179_pad_0 = const()[name = tensor("pretrained_out_179_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258430464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261707328))), name = tensor("layers_8_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_8_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_8_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261707456)))]; - tensor pretrained_out_179_cast_fp16 = conv(bias = layers_8_fc2_pretrained_bias_to_fp16, dilations = var_3694, groups = var_3335, pad = pretrained_out_179_pad_0, pad_type = pretrained_out_179_pad_type_0, strides = var_3692, weight = layers_8_fc2_pretrained_weight_to_fp16_palettized, x = input_267_cast_fp16)[name = tensor("pretrained_out_179_cast_fp16")]; - tensor var_3698 = const()[name = tensor("op_3698"), val = tensor([1, 1])]; - tensor var_3700 = const()[name = tensor("op_3700"), val = tensor([1, 1])]; - tensor input_269_pad_type_0 = const()[name = tensor("input_269_pad_type_0"), val = tensor("custom")]; - tensor input_269_pad_0 = const()[name = tensor("input_269_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_8_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_8_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261710080)))]; - tensor input_269_cast_fp16 = conv(dilations = var_3700, groups = var_3335, pad = input_269_pad_0, pad_type = input_269_pad_type_0, strides = var_3698, weight = layers_8_fc2_loraA_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; - tensor var_3704 = const()[name = tensor("op_3704"), val = tensor([1, 1])]; - tensor var_3706 = const()[name = tensor("op_3706"), val = tensor([1, 1])]; - tensor lora_out_357_pad_type_0 = const()[name = tensor("lora_out_357_pad_type_0"), val = tensor("custom")]; - tensor lora_out_357_pad_0 = const()[name = tensor("lora_out_357_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_359_weight_0_to_fp16 = const()[name = tensor("lora_out_359_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261873984)))]; - tensor lora_out_359_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3706, groups = var_3335, pad = lora_out_357_pad_0, pad_type = lora_out_357_pad_type_0, strides = var_3704, weight = lora_out_359_weight_0_to_fp16, x = input_269_cast_fp16)[name = tensor("lora_out_359_cast_fp16")]; - tensor hidden_states_19_cast_fp16 = add(x = pretrained_out_179_cast_fp16, y = lora_out_359_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_3722 = const()[name = tensor("op_3722"), val = tensor(3)]; - tensor var_3729 = const()[name = tensor("op_3729"), val = tensor(1)]; - tensor var_3730 = const()[name = tensor("op_3730"), val = tensor(true)]; - tensor var_3742 = const()[name = tensor("op_3742"), val = tensor([1])]; - tensor channels_mean_55_cast_fp16 = reduce_mean(axes = var_3742, keep_dims = var_3730, 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_3746 = const()[name = tensor("op_3746"), val = tensor([1])]; - tensor var_3747_cast_fp16 = reduce_mean(axes = var_3746, keep_dims = var_3730, x = zero_mean_sq_55_cast_fp16)[name = tensor("op_3747_cast_fp16")]; - tensor var_3748_to_fp16 = const()[name = tensor("op_3748_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_3749_cast_fp16 = add(x = var_3747_cast_fp16, y = var_3748_to_fp16)[name = tensor("op_3749_cast_fp16")]; - tensor denom_55_epsilon_0 = const()[name = tensor("denom_55_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_55_cast_fp16 = rsqrt(epsilon = denom_55_epsilon_0, x = var_3749_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_109_gamma_0_to_fp16 = const()[name = tensor("obj_109_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261915008)))]; - tensor obj_109_beta_0_to_fp16 = const()[name = tensor("obj_109_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261917632)))]; - tensor obj_109_epsilon_0_to_fp16 = const()[name = tensor("obj_109_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_109_cast_fp16 = batch_norm(beta = obj_109_beta_0_to_fp16, epsilon = obj_109_epsilon_0_to_fp16, gamma = obj_109_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_109_cast_fp16")]; - tensor var_3767 = const()[name = tensor("op_3767"), val = tensor([1, 1])]; - tensor var_3769 = const()[name = tensor("op_3769"), val = tensor([1, 1])]; - tensor pretrained_out_181_pad_type_0 = const()[name = tensor("pretrained_out_181_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_181_pad_0 = const()[name = tensor("pretrained_out_181_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261920256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262739520))), name = tensor("layers_9_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_9_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262739648)))]; - tensor pretrained_out_181_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_3769, groups = var_3729, pad = pretrained_out_181_pad_0, pad_type = pretrained_out_181_pad_type_0, strides = var_3767, weight = layers_9_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_109_cast_fp16)[name = tensor("pretrained_out_181_cast_fp16")]; - tensor var_3773 = const()[name = tensor("op_3773"), val = tensor([1, 1])]; - tensor var_3775 = const()[name = tensor("op_3775"), val = tensor([1, 1])]; - tensor input_271_pad_type_0 = const()[name = tensor("input_271_pad_type_0"), val = tensor("custom")]; - tensor input_271_pad_0 = const()[name = tensor("input_271_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262742272)))]; - tensor input_271_cast_fp16 = conv(dilations = var_3775, groups = var_3729, pad = input_271_pad_0, pad_type = input_271_pad_type_0, strides = var_3773, weight = layers_9_self_attn_q_proj_loraA_weight_to_fp16, x = obj_109_cast_fp16)[name = tensor("input_271_cast_fp16")]; - tensor var_3779 = const()[name = tensor("op_3779"), val = tensor([1, 1])]; - tensor var_3781 = const()[name = tensor("op_3781"), val = tensor([1, 1])]; - tensor lora_out_361_pad_type_0 = const()[name = tensor("lora_out_361_pad_type_0"), val = tensor("custom")]; - tensor lora_out_361_pad_0 = const()[name = tensor("lora_out_361_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_363_weight_0_to_fp16 = const()[name = tensor("lora_out_363_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262783296)))]; - tensor lora_out_363_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3781, groups = var_3729, pad = lora_out_361_pad_0, pad_type = lora_out_361_pad_type_0, strides = var_3779, weight = lora_out_363_weight_0_to_fp16, x = input_271_cast_fp16)[name = tensor("lora_out_363_cast_fp16")]; - tensor query_37_cast_fp16 = add(x = pretrained_out_181_cast_fp16, y = lora_out_363_cast_fp16)[name = tensor("query_37_cast_fp16")]; - tensor var_3791 = const()[name = tensor("op_3791"), val = tensor([1, 1])]; - tensor var_3793 = const()[name = tensor("op_3793"), val = tensor([1, 1])]; - tensor pretrained_out_183_pad_type_0 = const()[name = tensor("pretrained_out_183_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_183_pad_0 = const()[name = tensor("pretrained_out_183_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262824320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263643584))), name = tensor("layers_9_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_183_cast_fp16 = conv(dilations = var_3793, groups = var_3729, pad = pretrained_out_183_pad_0, pad_type = pretrained_out_183_pad_type_0, strides = var_3791, weight = layers_9_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_109_cast_fp16)[name = tensor("pretrained_out_183_cast_fp16")]; - tensor var_3797 = const()[name = tensor("op_3797"), val = tensor([1, 1])]; - tensor var_3799 = const()[name = tensor("op_3799"), val = tensor([1, 1])]; - tensor input_273_pad_type_0 = const()[name = tensor("input_273_pad_type_0"), val = tensor("custom")]; - tensor input_273_pad_0 = const()[name = tensor("input_273_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263643712)))]; - tensor input_273_cast_fp16 = conv(dilations = var_3799, groups = var_3729, pad = input_273_pad_0, pad_type = input_273_pad_type_0, strides = var_3797, weight = layers_9_self_attn_k_proj_loraA_weight_to_fp16, x = obj_109_cast_fp16)[name = tensor("input_273_cast_fp16")]; - tensor var_3803 = const()[name = tensor("op_3803"), val = tensor([1, 1])]; - tensor var_3805 = const()[name = tensor("op_3805"), val = tensor([1, 1])]; - tensor lora_out_365_pad_type_0 = const()[name = tensor("lora_out_365_pad_type_0"), val = tensor("custom")]; - tensor lora_out_365_pad_0 = const()[name = tensor("lora_out_365_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_367_weight_0_to_fp16 = const()[name = tensor("lora_out_367_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263684736)))]; - tensor lora_out_367_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3805, groups = var_3729, pad = lora_out_365_pad_0, pad_type = lora_out_365_pad_type_0, strides = var_3803, weight = lora_out_367_weight_0_to_fp16, x = input_273_cast_fp16)[name = tensor("lora_out_367_cast_fp16")]; - tensor current_key_19_cast_fp16 = add(x = pretrained_out_183_cast_fp16, y = lora_out_367_cast_fp16)[name = tensor("current_key_19_cast_fp16")]; - tensor var_3816 = const()[name = tensor("op_3816"), val = tensor([1, 1])]; - tensor var_3818 = const()[name = tensor("op_3818"), val = tensor([1, 1])]; - tensor pretrained_out_185_pad_type_0 = const()[name = tensor("pretrained_out_185_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_185_pad_0 = const()[name = tensor("pretrained_out_185_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263725760))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264545024))), name = tensor("layers_9_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_9_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264545152)))]; - tensor pretrained_out_185_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_3818, groups = var_3729, pad = pretrained_out_185_pad_0, pad_type = pretrained_out_185_pad_type_0, strides = var_3816, weight = layers_9_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_109_cast_fp16)[name = tensor("pretrained_out_185_cast_fp16")]; - tensor var_3822 = const()[name = tensor("op_3822"), val = tensor([1, 1])]; - tensor var_3824 = const()[name = tensor("op_3824"), val = tensor([1, 1])]; - tensor input_275_pad_type_0 = const()[name = tensor("input_275_pad_type_0"), val = tensor("custom")]; - tensor input_275_pad_0 = const()[name = tensor("input_275_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264547776)))]; - tensor input_275_cast_fp16 = conv(dilations = var_3824, groups = var_3729, pad = input_275_pad_0, pad_type = input_275_pad_type_0, strides = var_3822, weight = layers_9_self_attn_v_proj_loraA_weight_to_fp16, x = obj_109_cast_fp16)[name = tensor("input_275_cast_fp16")]; - tensor var_3828 = const()[name = tensor("op_3828"), val = tensor([1, 1])]; - tensor var_3830 = const()[name = tensor("op_3830"), val = tensor([1, 1])]; - tensor lora_out_369_pad_type_0 = const()[name = tensor("lora_out_369_pad_type_0"), val = tensor("custom")]; - tensor lora_out_369_pad_0 = const()[name = tensor("lora_out_369_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_371_weight_0_to_fp16 = const()[name = tensor("lora_out_371_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264588800)))]; - tensor lora_out_371_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3830, groups = var_3729, pad = lora_out_369_pad_0, pad_type = lora_out_369_pad_type_0, strides = var_3828, weight = lora_out_371_weight_0_to_fp16, x = input_275_cast_fp16)[name = tensor("lora_out_371_cast_fp16")]; - tensor current_value_19_cast_fp16 = add(x = pretrained_out_185_cast_fp16, y = lora_out_371_cast_fp16)[name = tensor("current_value_19_cast_fp16")]; - tensor var_3840_cast_fp16 = mul(x = current_key_19_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_3840_cast_fp16")]; - tensor var_3842_cast_fp16 = mul(x = var_103_cast_fp16_9, y = var_295_cast_fp16)[name = tensor("op_3842_cast_fp16")]; - tensor key_37_cast_fp16 = add(x = var_3840_cast_fp16, y = var_3842_cast_fp16)[name = tensor("key_37_cast_fp16")]; - tensor var_3844_cast_fp16 = mul(x = current_value_19_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_3844_cast_fp16")]; - tensor var_3846_cast_fp16 = mul(x = var_138_cast_fp16_9, y = var_295_cast_fp16)[name = tensor("op_3846_cast_fp16")]; - tensor value_37_cast_fp16 = add(x = var_3844_cast_fp16, y = var_3846_cast_fp16)[name = tensor("value_37_cast_fp16")]; - tensor var_3849 = const()[name = tensor("op_3849"), val = tensor([1, 20, 64, -1])]; - tensor var_3850_cast_fp16 = reshape(shape = var_3849, x = query_37_cast_fp16)[name = tensor("op_3850_cast_fp16")]; - tensor var_3851_to_fp16 = const()[name = tensor("op_3851_to_fp16"), val = tensor(0x1p-3)]; - tensor var_3852_cast_fp16 = mul(x = var_3850_cast_fp16, y = var_3851_to_fp16)[name = tensor("op_3852_cast_fp16")]; - tensor var_3853 = const()[name = tensor("op_3853"), val = tensor([1, 20, 64, -1])]; - tensor var_3854_cast_fp16 = reshape(shape = var_3853, x = key_37_cast_fp16)[name = tensor("op_3854_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_3852_cast_fp16, y = var_3854_cast_fp16)[name = tensor("mh_w_55_cast_fp16")]; - tensor mh_w_57_cast_fp16 = add(x = mh_w_55_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_57_cast_fp16")]; - tensor var_3862_cast_fp16 = softmax(axis = var_3722, x = mh_w_57_cast_fp16)[name = tensor("op_3862_cast_fp16")]; - tensor var_3863 = const()[name = tensor("op_3863"), val = tensor([1, 20, 64, -1])]; - tensor var_3864_cast_fp16 = reshape(shape = var_3863, x = value_37_cast_fp16)[name = tensor("op_3864_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_3864_cast_fp16, y = var_3862_cast_fp16)[name = tensor("attn_37_cast_fp16")]; - tensor var_3867 = const()[name = tensor("op_3867"), val = tensor([1, 1280, 1, -1])]; - tensor input_277_cast_fp16 = reshape(shape = var_3867, x = attn_37_cast_fp16)[name = tensor("input_277_cast_fp16")]; - tensor var_3874 = const()[name = tensor("op_3874"), val = tensor([1, 1])]; - tensor var_3876 = const()[name = tensor("op_3876"), val = tensor([1, 1])]; - tensor pretrained_out_187_pad_type_0 = const()[name = tensor("pretrained_out_187_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_187_pad_0 = const()[name = tensor("pretrained_out_187_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264629824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265449088))), name = tensor("layers_9_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_9_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265449216)))]; - tensor pretrained_out_187_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_3876, groups = var_3729, pad = pretrained_out_187_pad_0, pad_type = pretrained_out_187_pad_type_0, strides = var_3874, weight = layers_9_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_277_cast_fp16)[name = tensor("pretrained_out_187_cast_fp16")]; - tensor var_3880 = const()[name = tensor("op_3880"), val = tensor([1, 1])]; - tensor var_3882 = const()[name = tensor("op_3882"), val = tensor([1, 1])]; - tensor input_279_pad_type_0 = const()[name = tensor("input_279_pad_type_0"), val = tensor("custom")]; - tensor input_279_pad_0 = const()[name = tensor("input_279_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265451840)))]; - tensor input_279_cast_fp16 = conv(dilations = var_3882, groups = var_3729, pad = input_279_pad_0, pad_type = input_279_pad_type_0, strides = var_3880, weight = layers_9_self_attn_o_proj_loraA_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; - tensor var_3886 = const()[name = tensor("op_3886"), val = tensor([1, 1])]; - tensor var_3888 = const()[name = tensor("op_3888"), val = tensor([1, 1])]; - tensor lora_out_373_pad_type_0 = const()[name = tensor("lora_out_373_pad_type_0"), val = tensor("custom")]; - tensor lora_out_373_pad_0 = const()[name = tensor("lora_out_373_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_375_weight_0_to_fp16 = const()[name = tensor("lora_out_375_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265492864)))]; - tensor lora_out_375_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3888, groups = var_3729, pad = lora_out_373_pad_0, pad_type = lora_out_373_pad_type_0, strides = var_3886, weight = lora_out_375_weight_0_to_fp16, x = input_279_cast_fp16)[name = tensor("lora_out_375_cast_fp16")]; - tensor obj_115_cast_fp16 = add(x = pretrained_out_187_cast_fp16, y = lora_out_375_cast_fp16)[name = tensor("obj_115_cast_fp16")]; - tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = obj_115_cast_fp16)[name = tensor("inputs_57_cast_fp16")]; - tensor var_3901 = const()[name = tensor("op_3901"), val = tensor([1])]; - tensor channels_mean_57_cast_fp16 = reduce_mean(axes = var_3901, keep_dims = var_3730, 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_3905 = const()[name = tensor("op_3905"), val = tensor([1])]; - tensor var_3906_cast_fp16 = reduce_mean(axes = var_3905, keep_dims = var_3730, x = zero_mean_sq_57_cast_fp16)[name = tensor("op_3906_cast_fp16")]; - tensor var_3907_to_fp16 = const()[name = tensor("op_3907_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_3908_cast_fp16 = add(x = var_3906_cast_fp16, y = var_3907_to_fp16)[name = tensor("op_3908_cast_fp16")]; - tensor denom_57_epsilon_0 = const()[name = tensor("denom_57_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_57_cast_fp16 = rsqrt(epsilon = denom_57_epsilon_0, x = var_3908_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_117_gamma_0_to_fp16 = const()[name = tensor("obj_117_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265533888)))]; - tensor obj_117_beta_0_to_fp16 = const()[name = tensor("obj_117_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265536512)))]; - tensor obj_117_epsilon_0_to_fp16 = const()[name = tensor("obj_117_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_117_cast_fp16 = batch_norm(beta = obj_117_beta_0_to_fp16, epsilon = obj_117_epsilon_0_to_fp16, gamma = obj_117_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_117_cast_fp16")]; - tensor var_3926 = const()[name = tensor("op_3926"), val = tensor([1, 1])]; - tensor var_3928 = const()[name = tensor("op_3928"), val = tensor([1, 1])]; - tensor pretrained_out_189_pad_type_0 = const()[name = tensor("pretrained_out_189_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_189_pad_0 = const()[name = tensor("pretrained_out_189_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265539136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266358400))), name = tensor("layers_9_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_9_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266358528)))]; - tensor pretrained_out_189_cast_fp16 = conv(bias = layers_9_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_3928, groups = var_3729, pad = pretrained_out_189_pad_0, pad_type = pretrained_out_189_pad_type_0, strides = var_3926, weight = layers_9_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_117_cast_fp16)[name = tensor("pretrained_out_189_cast_fp16")]; - tensor var_3932 = const()[name = tensor("op_3932"), val = tensor([1, 1])]; - tensor var_3934 = const()[name = tensor("op_3934"), val = tensor([1, 1])]; - tensor input_281_pad_type_0 = const()[name = tensor("input_281_pad_type_0"), val = tensor("custom")]; - tensor input_281_pad_0 = const()[name = tensor("input_281_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266361152)))]; - tensor input_281_cast_fp16 = conv(dilations = var_3934, groups = var_3729, pad = input_281_pad_0, pad_type = input_281_pad_type_0, strides = var_3932, weight = layers_9_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_117_cast_fp16)[name = tensor("input_281_cast_fp16")]; - tensor var_3938 = const()[name = tensor("op_3938"), val = tensor([1, 1])]; - tensor var_3940 = const()[name = tensor("op_3940"), val = tensor([1, 1])]; - tensor lora_out_377_pad_type_0 = const()[name = tensor("lora_out_377_pad_type_0"), val = tensor("custom")]; - tensor lora_out_377_pad_0 = const()[name = tensor("lora_out_377_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_379_weight_0_to_fp16 = const()[name = tensor("lora_out_379_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266402176)))]; - tensor lora_out_379_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3940, groups = var_3729, pad = lora_out_377_pad_0, pad_type = lora_out_377_pad_type_0, strides = var_3938, weight = lora_out_379_weight_0_to_fp16, x = input_281_cast_fp16)[name = tensor("lora_out_379_cast_fp16")]; - tensor query_39_cast_fp16 = add(x = pretrained_out_189_cast_fp16, y = lora_out_379_cast_fp16)[name = tensor("query_39_cast_fp16")]; - tensor var_3950 = const()[name = tensor("op_3950"), val = tensor([1, 1])]; - tensor var_3952 = const()[name = tensor("op_3952"), val = tensor([1, 1])]; - tensor pretrained_out_191_pad_type_0 = const()[name = tensor("pretrained_out_191_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_191_pad_0 = const()[name = tensor("pretrained_out_191_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266443200))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267262464))), name = tensor("layers_9_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_191_cast_fp16 = conv(dilations = var_3952, groups = var_3729, pad = pretrained_out_191_pad_0, pad_type = pretrained_out_191_pad_type_0, strides = var_3950, weight = layers_9_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_191_cast_fp16")]; - tensor var_3956 = const()[name = tensor("op_3956"), val = tensor([1, 1])]; - tensor var_3958 = const()[name = tensor("op_3958"), val = tensor([1, 1])]; - tensor input_283_pad_type_0 = const()[name = tensor("input_283_pad_type_0"), val = tensor("custom")]; - tensor input_283_pad_0 = const()[name = tensor("input_283_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267262592)))]; - tensor input_283_cast_fp16 = conv(dilations = var_3958, groups = var_3729, pad = input_283_pad_0, pad_type = input_283_pad_type_0, strides = var_3956, weight = layers_9_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_283_cast_fp16")]; - tensor var_3962 = const()[name = tensor("op_3962"), val = tensor([1, 1])]; - tensor var_3964 = const()[name = tensor("op_3964"), val = tensor([1, 1])]; - tensor lora_out_381_pad_type_0 = const()[name = tensor("lora_out_381_pad_type_0"), val = tensor("custom")]; - tensor lora_out_381_pad_0 = const()[name = tensor("lora_out_381_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_383_weight_0_to_fp16 = const()[name = tensor("lora_out_383_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267303616)))]; - tensor lora_out_383_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3964, groups = var_3729, pad = lora_out_381_pad_0, pad_type = lora_out_381_pad_type_0, strides = var_3962, weight = lora_out_383_weight_0_to_fp16, x = input_283_cast_fp16)[name = tensor("lora_out_383_cast_fp16")]; - tensor key_39_cast_fp16 = add(x = pretrained_out_191_cast_fp16, y = lora_out_383_cast_fp16)[name = tensor("key_39_cast_fp16")]; - tensor var_3975 = const()[name = tensor("op_3975"), val = tensor([1, 1])]; - tensor var_3977 = const()[name = tensor("op_3977"), val = tensor([1, 1])]; - tensor pretrained_out_193_pad_type_0 = const()[name = tensor("pretrained_out_193_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_193_pad_0 = const()[name = tensor("pretrained_out_193_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267344640))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268163904))), name = tensor("layers_9_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_9_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268164032)))]; - tensor pretrained_out_193_cast_fp16 = conv(bias = layers_9_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_3977, groups = var_3729, pad = pretrained_out_193_pad_0, pad_type = pretrained_out_193_pad_type_0, strides = var_3975, weight = layers_9_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_193_cast_fp16")]; - tensor var_3981 = const()[name = tensor("op_3981"), val = tensor([1, 1])]; - tensor var_3983 = const()[name = tensor("op_3983"), val = tensor([1, 1])]; - tensor input_285_pad_type_0 = const()[name = tensor("input_285_pad_type_0"), val = tensor("custom")]; - tensor input_285_pad_0 = const()[name = tensor("input_285_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268166656)))]; - tensor input_285_cast_fp16 = conv(dilations = var_3983, groups = var_3729, pad = input_285_pad_0, pad_type = input_285_pad_type_0, strides = var_3981, weight = layers_9_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_285_cast_fp16")]; - tensor var_3987 = const()[name = tensor("op_3987"), val = tensor([1, 1])]; - tensor var_3989 = const()[name = tensor("op_3989"), val = tensor([1, 1])]; - tensor lora_out_385_pad_type_0 = const()[name = tensor("lora_out_385_pad_type_0"), val = tensor("custom")]; - tensor lora_out_385_pad_0 = const()[name = tensor("lora_out_385_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_387_weight_0_to_fp16 = const()[name = tensor("lora_out_387_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268207680)))]; - tensor lora_out_387_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3989, groups = var_3729, pad = lora_out_385_pad_0, pad_type = lora_out_385_pad_type_0, strides = var_3987, weight = lora_out_387_weight_0_to_fp16, x = input_285_cast_fp16)[name = tensor("lora_out_387_cast_fp16")]; - tensor value_39_cast_fp16 = add(x = pretrained_out_193_cast_fp16, y = lora_out_387_cast_fp16)[name = tensor("value_39_cast_fp16")]; - tensor var_3996 = const()[name = tensor("op_3996"), val = tensor([1, 20, 64, -1])]; - tensor var_3997_cast_fp16 = reshape(shape = var_3996, x = query_39_cast_fp16)[name = tensor("op_3997_cast_fp16")]; - tensor var_3998_to_fp16 = const()[name = tensor("op_3998_to_fp16"), val = tensor(0x1p-3)]; - tensor var_3999_cast_fp16 = mul(x = var_3997_cast_fp16, y = var_3998_to_fp16)[name = tensor("op_3999_cast_fp16")]; - tensor var_4000 = const()[name = tensor("op_4000"), val = tensor([1, 20, 64, -1])]; - tensor var_4001_cast_fp16 = reshape(shape = var_4000, x = key_39_cast_fp16)[name = tensor("op_4001_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_3999_cast_fp16, y = var_4001_cast_fp16)[name = tensor("mh_w_59_cast_fp16")]; - tensor var_4004_cast_fp16 = softmax(axis = var_3722, x = mh_w_59_cast_fp16)[name = tensor("op_4004_cast_fp16")]; - tensor var_4005 = const()[name = tensor("op_4005"), val = tensor([1, 20, 64, -1])]; - tensor var_4006_cast_fp16 = reshape(shape = var_4005, x = value_39_cast_fp16)[name = tensor("op_4006_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_4006_cast_fp16, y = var_4004_cast_fp16)[name = tensor("attn_39_cast_fp16")]; - tensor var_4009 = const()[name = tensor("op_4009"), val = tensor([1, 1280, 1, -1])]; - tensor input_287_cast_fp16 = reshape(shape = var_4009, x = attn_39_cast_fp16)[name = tensor("input_287_cast_fp16")]; - tensor var_4016 = const()[name = tensor("op_4016"), val = tensor([1, 1])]; - tensor var_4018 = const()[name = tensor("op_4018"), val = tensor([1, 1])]; - tensor pretrained_out_195_pad_type_0 = const()[name = tensor("pretrained_out_195_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_195_pad_0 = const()[name = tensor("pretrained_out_195_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268248704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269067968))), name = tensor("layers_9_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_9_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269068096)))]; - tensor pretrained_out_195_cast_fp16 = conv(bias = layers_9_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_4018, groups = var_3729, pad = pretrained_out_195_pad_0, pad_type = pretrained_out_195_pad_type_0, strides = var_4016, weight = layers_9_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = tensor("pretrained_out_195_cast_fp16")]; - tensor var_4022 = const()[name = tensor("op_4022"), val = tensor([1, 1])]; - tensor var_4024 = const()[name = tensor("op_4024"), val = tensor([1, 1])]; - tensor input_289_pad_type_0 = const()[name = tensor("input_289_pad_type_0"), val = tensor("custom")]; - tensor input_289_pad_0 = const()[name = tensor("input_289_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269070720)))]; - tensor input_289_cast_fp16 = conv(dilations = var_4024, groups = var_3729, pad = input_289_pad_0, pad_type = input_289_pad_type_0, strides = var_4022, weight = layers_9_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("input_289_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 lora_out_389_pad_type_0 = const()[name = tensor("lora_out_389_pad_type_0"), val = tensor("custom")]; - tensor lora_out_389_pad_0 = const()[name = tensor("lora_out_389_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_391_weight_0_to_fp16 = const()[name = tensor("lora_out_391_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269111744)))]; - tensor lora_out_391_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4030, groups = var_3729, pad = lora_out_389_pad_0, pad_type = lora_out_389_pad_type_0, strides = var_4028, weight = lora_out_391_weight_0_to_fp16, x = input_289_cast_fp16)[name = tensor("lora_out_391_cast_fp16")]; - tensor obj_119_cast_fp16 = add(x = pretrained_out_195_cast_fp16, y = lora_out_391_cast_fp16)[name = tensor("obj_119_cast_fp16")]; - tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_119_cast_fp16)[name = tensor("inputs_59_cast_fp16")]; - tensor var_4039 = const()[name = tensor("op_4039"), val = tensor([1])]; - tensor channels_mean_59_cast_fp16 = reduce_mean(axes = var_4039, keep_dims = var_3730, 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_4043 = const()[name = tensor("op_4043"), val = tensor([1])]; - tensor var_4044_cast_fp16 = reduce_mean(axes = var_4043, keep_dims = var_3730, x = zero_mean_sq_59_cast_fp16)[name = tensor("op_4044_cast_fp16")]; - tensor var_4045_to_fp16 = const()[name = tensor("op_4045_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_4046_cast_fp16 = add(x = var_4044_cast_fp16, y = var_4045_to_fp16)[name = tensor("op_4046_cast_fp16")]; - tensor denom_59_epsilon_0 = const()[name = tensor("denom_59_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_59_cast_fp16 = rsqrt(epsilon = denom_59_epsilon_0, x = var_4046_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_291_gamma_0_to_fp16 = const()[name = tensor("input_291_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269152768)))]; - tensor input_291_beta_0_to_fp16 = const()[name = tensor("input_291_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269155392)))]; - tensor input_291_epsilon_0_to_fp16 = const()[name = tensor("input_291_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_291_cast_fp16 = batch_norm(beta = input_291_beta_0_to_fp16, epsilon = input_291_epsilon_0_to_fp16, gamma = input_291_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_291_cast_fp16")]; - tensor var_4060 = const()[name = tensor("op_4060"), val = tensor([1, 1])]; - tensor var_4062 = const()[name = tensor("op_4062"), val = tensor([1, 1])]; - tensor pretrained_out_197_pad_type_0 = const()[name = tensor("pretrained_out_197_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_197_pad_0 = const()[name = tensor("pretrained_out_197_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269158016))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272434880))), name = tensor("layers_9_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_9_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_9_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272435008)))]; - tensor pretrained_out_197_cast_fp16 = conv(bias = layers_9_fc1_pretrained_bias_to_fp16, dilations = var_4062, groups = var_3729, pad = pretrained_out_197_pad_0, pad_type = pretrained_out_197_pad_type_0, strides = var_4060, weight = layers_9_fc1_pretrained_weight_to_fp16_palettized, x = input_291_cast_fp16)[name = tensor("pretrained_out_197_cast_fp16")]; - tensor var_4066 = const()[name = tensor("op_4066"), val = tensor([1, 1])]; - tensor var_4068 = const()[name = tensor("op_4068"), val = tensor([1, 1])]; - tensor input_293_pad_type_0 = const()[name = tensor("input_293_pad_type_0"), val = tensor("custom")]; - tensor input_293_pad_0 = const()[name = tensor("input_293_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_9_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272445312)))]; - tensor input_293_cast_fp16 = conv(dilations = var_4068, groups = var_3729, pad = input_293_pad_0, pad_type = input_293_pad_type_0, strides = var_4066, weight = layers_9_fc1_loraA_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("input_293_cast_fp16")]; - tensor var_4072 = const()[name = tensor("op_4072"), val = tensor([1, 1])]; - tensor var_4074 = const()[name = tensor("op_4074"), val = tensor([1, 1])]; - tensor lora_out_393_pad_type_0 = const()[name = tensor("lora_out_393_pad_type_0"), val = tensor("custom")]; - tensor lora_out_393_pad_0 = const()[name = tensor("lora_out_393_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_395_weight_0_to_fp16 = const()[name = tensor("lora_out_395_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272486336)))]; - tensor lora_out_395_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_4074, groups = var_3729, pad = lora_out_393_pad_0, pad_type = lora_out_393_pad_type_0, strides = var_4072, weight = lora_out_395_weight_0_to_fp16, x = input_293_cast_fp16)[name = tensor("lora_out_395_cast_fp16")]; - tensor input_295_cast_fp16 = add(x = pretrained_out_197_cast_fp16, y = lora_out_395_cast_fp16)[name = tensor("input_295_cast_fp16")]; - tensor input_297_mode_0 = const()[name = tensor("input_297_mode_0"), val = tensor("EXACT")]; - tensor input_297_cast_fp16 = gelu(mode = input_297_mode_0, x = input_295_cast_fp16)[name = tensor("input_297_cast_fp16")]; - tensor var_4086 = const()[name = tensor("op_4086"), val = tensor([1, 1])]; - tensor var_4088 = const()[name = tensor("op_4088"), val = tensor([1, 1])]; - tensor pretrained_out_199_pad_type_0 = const()[name = tensor("pretrained_out_199_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_199_pad_0 = const()[name = tensor("pretrained_out_199_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272650240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277565504))), name = tensor("layers_9_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_9_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_9_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277565696)))]; - tensor pretrained_out_199_cast_fp16 = conv(bias = layers_9_fc2_pretrained_bias_to_fp16, dilations = var_4088, groups = var_3729, pad = pretrained_out_199_pad_0, pad_type = pretrained_out_199_pad_type_0, strides = var_4086, weight = layers_9_fc2_pretrained_weight_to_fp16_palettized, x = input_297_cast_fp16)[name = tensor("pretrained_out_199_cast_fp16")]; - tensor var_4092 = const()[name = tensor("op_4092"), val = tensor([1, 1])]; - tensor var_4094 = const()[name = tensor("op_4094"), val = tensor([1, 1])]; - tensor input_299_pad_type_0 = const()[name = tensor("input_299_pad_type_0"), val = tensor("custom")]; - tensor input_299_pad_0 = const()[name = tensor("input_299_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_9_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_9_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277568320)))]; - tensor input_299_cast_fp16 = conv(dilations = var_4094, groups = var_3729, pad = input_299_pad_0, pad_type = input_299_pad_type_0, strides = var_4092, weight = layers_9_fc2_loraA_weight_to_fp16, x = input_297_cast_fp16)[name = tensor("input_299_cast_fp16")]; - tensor var_4098 = const()[name = tensor("op_4098"), val = tensor([1, 1])]; - tensor var_4100 = const()[name = tensor("op_4100"), val = tensor([1, 1])]; - tensor lora_out_397_pad_type_0 = const()[name = tensor("lora_out_397_pad_type_0"), val = tensor("custom")]; - tensor lora_out_397_pad_0 = const()[name = tensor("lora_out_397_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_399_weight_0_to_fp16 = const()[name = tensor("lora_out_399_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277732224)))]; - tensor lora_out_399_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4100, groups = var_3729, pad = lora_out_397_pad_0, pad_type = lora_out_397_pad_type_0, strides = var_4098, weight = lora_out_399_weight_0_to_fp16, x = input_299_cast_fp16)[name = tensor("lora_out_399_cast_fp16")]; - tensor hidden_states_21_cast_fp16 = add(x = pretrained_out_199_cast_fp16, y = lora_out_399_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_4116 = const()[name = tensor("op_4116"), val = tensor(3)]; - tensor var_4123 = const()[name = tensor("op_4123"), val = tensor(1)]; - tensor var_4124 = const()[name = tensor("op_4124"), val = tensor(true)]; - tensor var_4136 = const()[name = tensor("op_4136"), val = tensor([1])]; - tensor channels_mean_61_cast_fp16 = reduce_mean(axes = var_4136, keep_dims = var_4124, 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_4140 = const()[name = tensor("op_4140"), val = tensor([1])]; - tensor var_4141_cast_fp16 = reduce_mean(axes = var_4140, keep_dims = var_4124, x = zero_mean_sq_61_cast_fp16)[name = tensor("op_4141_cast_fp16")]; - tensor var_4142_to_fp16 = const()[name = tensor("op_4142_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_4143_cast_fp16 = add(x = var_4141_cast_fp16, y = var_4142_to_fp16)[name = tensor("op_4143_cast_fp16")]; - tensor denom_61_epsilon_0 = const()[name = tensor("denom_61_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_61_cast_fp16 = rsqrt(epsilon = denom_61_epsilon_0, x = var_4143_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_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(277773248)))]; - 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(277775872)))]; - 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_61_cast_fp16)[name = tensor("obj_121_cast_fp16")]; - tensor var_4161 = const()[name = tensor("op_4161"), val = tensor([1, 1])]; - tensor var_4163 = const()[name = tensor("op_4163"), val = tensor([1, 1])]; - tensor pretrained_out_201_pad_type_0 = const()[name = tensor("pretrained_out_201_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_201_pad_0 = const()[name = tensor("pretrained_out_201_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277778496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278597760))), name = tensor("layers_10_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_10_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278597888)))]; - tensor pretrained_out_201_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_4163, groups = var_4123, pad = pretrained_out_201_pad_0, pad_type = pretrained_out_201_pad_type_0, strides = var_4161, weight = layers_10_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_121_cast_fp16)[name = tensor("pretrained_out_201_cast_fp16")]; - tensor var_4167 = const()[name = tensor("op_4167"), val = tensor([1, 1])]; - tensor var_4169 = const()[name = tensor("op_4169"), val = tensor([1, 1])]; - tensor input_301_pad_type_0 = const()[name = tensor("input_301_pad_type_0"), val = tensor("custom")]; - tensor input_301_pad_0 = const()[name = tensor("input_301_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278600512)))]; - tensor input_301_cast_fp16 = conv(dilations = var_4169, groups = var_4123, pad = input_301_pad_0, pad_type = input_301_pad_type_0, strides = var_4167, weight = layers_10_self_attn_q_proj_loraA_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("input_301_cast_fp16")]; - tensor var_4173 = const()[name = tensor("op_4173"), val = tensor([1, 1])]; - tensor var_4175 = const()[name = tensor("op_4175"), val = tensor([1, 1])]; - tensor lora_out_401_pad_type_0 = const()[name = tensor("lora_out_401_pad_type_0"), val = tensor("custom")]; - tensor lora_out_401_pad_0 = const()[name = tensor("lora_out_401_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_403_weight_0_to_fp16 = const()[name = tensor("lora_out_403_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278641536)))]; - tensor lora_out_403_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4175, groups = var_4123, pad = lora_out_401_pad_0, pad_type = lora_out_401_pad_type_0, strides = var_4173, weight = lora_out_403_weight_0_to_fp16, x = input_301_cast_fp16)[name = tensor("lora_out_403_cast_fp16")]; - tensor query_41_cast_fp16 = add(x = pretrained_out_201_cast_fp16, y = lora_out_403_cast_fp16)[name = tensor("query_41_cast_fp16")]; - tensor var_4185 = const()[name = tensor("op_4185"), val = tensor([1, 1])]; - tensor var_4187 = const()[name = tensor("op_4187"), val = tensor([1, 1])]; - tensor pretrained_out_203_pad_type_0 = const()[name = tensor("pretrained_out_203_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_203_pad_0 = const()[name = tensor("pretrained_out_203_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278682560))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279501824))), name = tensor("layers_10_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_203_cast_fp16 = conv(dilations = var_4187, groups = var_4123, pad = pretrained_out_203_pad_0, pad_type = pretrained_out_203_pad_type_0, strides = var_4185, weight = layers_10_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_121_cast_fp16)[name = tensor("pretrained_out_203_cast_fp16")]; - tensor var_4191 = const()[name = tensor("op_4191"), val = tensor([1, 1])]; - tensor var_4193 = const()[name = tensor("op_4193"), val = tensor([1, 1])]; - tensor input_303_pad_type_0 = const()[name = tensor("input_303_pad_type_0"), val = tensor("custom")]; - tensor input_303_pad_0 = const()[name = tensor("input_303_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279501952)))]; - tensor input_303_cast_fp16 = conv(dilations = var_4193, groups = var_4123, pad = input_303_pad_0, pad_type = input_303_pad_type_0, strides = var_4191, weight = layers_10_self_attn_k_proj_loraA_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("input_303_cast_fp16")]; - tensor var_4197 = const()[name = tensor("op_4197"), val = tensor([1, 1])]; - tensor var_4199 = const()[name = tensor("op_4199"), val = tensor([1, 1])]; - tensor lora_out_405_pad_type_0 = const()[name = tensor("lora_out_405_pad_type_0"), val = tensor("custom")]; - tensor lora_out_405_pad_0 = const()[name = tensor("lora_out_405_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_407_weight_0_to_fp16 = const()[name = tensor("lora_out_407_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279542976)))]; - tensor lora_out_407_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4199, groups = var_4123, pad = lora_out_405_pad_0, pad_type = lora_out_405_pad_type_0, strides = var_4197, weight = lora_out_407_weight_0_to_fp16, x = input_303_cast_fp16)[name = tensor("lora_out_407_cast_fp16")]; - tensor current_key_21_cast_fp16 = add(x = pretrained_out_203_cast_fp16, y = lora_out_407_cast_fp16)[name = tensor("current_key_21_cast_fp16")]; - tensor var_4210 = const()[name = tensor("op_4210"), val = tensor([1, 1])]; - tensor var_4212 = const()[name = tensor("op_4212"), val = tensor([1, 1])]; - tensor pretrained_out_205_pad_type_0 = const()[name = tensor("pretrained_out_205_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_205_pad_0 = const()[name = tensor("pretrained_out_205_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279584000))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280403264))), name = tensor("layers_10_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_10_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280403392)))]; - tensor pretrained_out_205_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_4212, groups = var_4123, pad = pretrained_out_205_pad_0, pad_type = pretrained_out_205_pad_type_0, strides = var_4210, weight = layers_10_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_121_cast_fp16)[name = tensor("pretrained_out_205_cast_fp16")]; - tensor var_4216 = const()[name = tensor("op_4216"), val = tensor([1, 1])]; - tensor var_4218 = const()[name = tensor("op_4218"), val = tensor([1, 1])]; - tensor input_305_pad_type_0 = const()[name = tensor("input_305_pad_type_0"), val = tensor("custom")]; - tensor input_305_pad_0 = const()[name = tensor("input_305_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280406016)))]; - tensor input_305_cast_fp16 = conv(dilations = var_4218, groups = var_4123, pad = input_305_pad_0, pad_type = input_305_pad_type_0, strides = var_4216, weight = layers_10_self_attn_v_proj_loraA_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("input_305_cast_fp16")]; - tensor var_4222 = const()[name = tensor("op_4222"), val = tensor([1, 1])]; - tensor var_4224 = const()[name = tensor("op_4224"), val = tensor([1, 1])]; - tensor lora_out_409_pad_type_0 = const()[name = tensor("lora_out_409_pad_type_0"), val = tensor("custom")]; - tensor lora_out_409_pad_0 = const()[name = tensor("lora_out_409_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_411_weight_0_to_fp16 = const()[name = tensor("lora_out_411_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280447040)))]; - tensor lora_out_411_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4224, groups = var_4123, pad = lora_out_409_pad_0, pad_type = lora_out_409_pad_type_0, strides = var_4222, weight = lora_out_411_weight_0_to_fp16, x = input_305_cast_fp16)[name = tensor("lora_out_411_cast_fp16")]; - tensor current_value_21_cast_fp16 = add(x = pretrained_out_205_cast_fp16, y = lora_out_411_cast_fp16)[name = tensor("current_value_21_cast_fp16")]; - tensor var_4234_cast_fp16 = mul(x = current_key_21_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_4234_cast_fp16")]; - tensor var_4236_cast_fp16 = mul(x = var_103_cast_fp16_10, y = var_295_cast_fp16)[name = tensor("op_4236_cast_fp16")]; - tensor key_41_cast_fp16 = add(x = var_4234_cast_fp16, y = var_4236_cast_fp16)[name = tensor("key_41_cast_fp16")]; - tensor var_4238_cast_fp16 = mul(x = current_value_21_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_4238_cast_fp16")]; - tensor var_4240_cast_fp16 = mul(x = var_138_cast_fp16_10, y = var_295_cast_fp16)[name = tensor("op_4240_cast_fp16")]; - tensor value_41_cast_fp16 = add(x = var_4238_cast_fp16, y = var_4240_cast_fp16)[name = tensor("value_41_cast_fp16")]; - tensor var_4243 = const()[name = tensor("op_4243"), val = tensor([1, 20, 64, -1])]; - tensor var_4244_cast_fp16 = reshape(shape = var_4243, x = query_41_cast_fp16)[name = tensor("op_4244_cast_fp16")]; - tensor var_4245_to_fp16 = const()[name = tensor("op_4245_to_fp16"), val = tensor(0x1p-3)]; - tensor var_4246_cast_fp16 = mul(x = var_4244_cast_fp16, y = var_4245_to_fp16)[name = tensor("op_4246_cast_fp16")]; - tensor var_4247 = const()[name = tensor("op_4247"), val = tensor([1, 20, 64, -1])]; - tensor var_4248_cast_fp16 = reshape(shape = var_4247, x = key_41_cast_fp16)[name = tensor("op_4248_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_4246_cast_fp16, y = var_4248_cast_fp16)[name = tensor("mh_w_61_cast_fp16")]; - tensor mh_w_63_cast_fp16 = add(x = mh_w_61_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_63_cast_fp16")]; - tensor var_4256_cast_fp16 = softmax(axis = var_4116, x = mh_w_63_cast_fp16)[name = tensor("op_4256_cast_fp16")]; - tensor var_4257 = const()[name = tensor("op_4257"), val = tensor([1, 20, 64, -1])]; - tensor var_4258_cast_fp16 = reshape(shape = var_4257, x = value_41_cast_fp16)[name = tensor("op_4258_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_4258_cast_fp16, y = var_4256_cast_fp16)[name = tensor("attn_41_cast_fp16")]; - tensor var_4261 = const()[name = tensor("op_4261"), val = tensor([1, 1280, 1, -1])]; - tensor input_307_cast_fp16 = reshape(shape = var_4261, x = attn_41_cast_fp16)[name = tensor("input_307_cast_fp16")]; - tensor var_4268 = const()[name = tensor("op_4268"), val = tensor([1, 1])]; - tensor var_4270 = const()[name = tensor("op_4270"), val = tensor([1, 1])]; - tensor pretrained_out_207_pad_type_0 = const()[name = tensor("pretrained_out_207_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_207_pad_0 = const()[name = tensor("pretrained_out_207_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280488064))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281307328))), name = tensor("layers_10_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_10_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281307456)))]; - tensor pretrained_out_207_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_4270, groups = var_4123, pad = pretrained_out_207_pad_0, pad_type = pretrained_out_207_pad_type_0, strides = var_4268, weight = layers_10_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_307_cast_fp16)[name = tensor("pretrained_out_207_cast_fp16")]; - tensor var_4274 = const()[name = tensor("op_4274"), val = tensor([1, 1])]; - tensor var_4276 = const()[name = tensor("op_4276"), val = tensor([1, 1])]; - tensor input_309_pad_type_0 = const()[name = tensor("input_309_pad_type_0"), val = tensor("custom")]; - tensor input_309_pad_0 = const()[name = tensor("input_309_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281310080)))]; - tensor input_309_cast_fp16 = conv(dilations = var_4276, groups = var_4123, pad = input_309_pad_0, pad_type = input_309_pad_type_0, strides = var_4274, weight = layers_10_self_attn_o_proj_loraA_weight_to_fp16, x = input_307_cast_fp16)[name = tensor("input_309_cast_fp16")]; - tensor var_4280 = const()[name = tensor("op_4280"), val = tensor([1, 1])]; - tensor var_4282 = const()[name = tensor("op_4282"), val = tensor([1, 1])]; - tensor lora_out_413_pad_type_0 = const()[name = tensor("lora_out_413_pad_type_0"), val = tensor("custom")]; - tensor lora_out_413_pad_0 = const()[name = tensor("lora_out_413_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_415_weight_0_to_fp16 = const()[name = tensor("lora_out_415_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281351104)))]; - tensor lora_out_415_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4282, groups = var_4123, pad = lora_out_413_pad_0, pad_type = lora_out_413_pad_type_0, strides = var_4280, weight = lora_out_415_weight_0_to_fp16, x = input_309_cast_fp16)[name = tensor("lora_out_415_cast_fp16")]; - tensor obj_127_cast_fp16 = add(x = pretrained_out_207_cast_fp16, y = lora_out_415_cast_fp16)[name = tensor("obj_127_cast_fp16")]; - tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_127_cast_fp16)[name = tensor("inputs_63_cast_fp16")]; - tensor var_4295 = const()[name = tensor("op_4295"), val = tensor([1])]; - tensor channels_mean_63_cast_fp16 = reduce_mean(axes = var_4295, keep_dims = var_4124, 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_4299 = const()[name = tensor("op_4299"), val = tensor([1])]; - tensor var_4300_cast_fp16 = reduce_mean(axes = var_4299, keep_dims = var_4124, x = zero_mean_sq_63_cast_fp16)[name = tensor("op_4300_cast_fp16")]; - tensor var_4301_to_fp16 = const()[name = tensor("op_4301_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_4302_cast_fp16 = add(x = var_4300_cast_fp16, y = var_4301_to_fp16)[name = tensor("op_4302_cast_fp16")]; - tensor denom_63_epsilon_0 = const()[name = tensor("denom_63_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_63_cast_fp16 = rsqrt(epsilon = denom_63_epsilon_0, x = var_4302_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_129_gamma_0_to_fp16 = const()[name = tensor("obj_129_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281392128)))]; - tensor obj_129_beta_0_to_fp16 = const()[name = tensor("obj_129_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281394752)))]; - tensor obj_129_epsilon_0_to_fp16 = const()[name = tensor("obj_129_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_129_cast_fp16 = batch_norm(beta = obj_129_beta_0_to_fp16, epsilon = obj_129_epsilon_0_to_fp16, gamma = obj_129_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_129_cast_fp16")]; - tensor var_4320 = const()[name = tensor("op_4320"), val = tensor([1, 1])]; - tensor var_4322 = const()[name = tensor("op_4322"), val = tensor([1, 1])]; - tensor pretrained_out_209_pad_type_0 = const()[name = tensor("pretrained_out_209_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_209_pad_0 = const()[name = tensor("pretrained_out_209_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281397376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282216640))), name = tensor("layers_10_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_10_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282216768)))]; - tensor pretrained_out_209_cast_fp16 = conv(bias = layers_10_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_4322, groups = var_4123, pad = pretrained_out_209_pad_0, pad_type = pretrained_out_209_pad_type_0, strides = var_4320, weight = layers_10_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_129_cast_fp16)[name = tensor("pretrained_out_209_cast_fp16")]; - tensor var_4326 = const()[name = tensor("op_4326"), val = tensor([1, 1])]; - tensor var_4328 = const()[name = tensor("op_4328"), val = tensor([1, 1])]; - tensor input_311_pad_type_0 = const()[name = tensor("input_311_pad_type_0"), val = tensor("custom")]; - tensor input_311_pad_0 = const()[name = tensor("input_311_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282219392)))]; - tensor input_311_cast_fp16 = conv(dilations = var_4328, groups = var_4123, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = var_4326, weight = layers_10_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_129_cast_fp16)[name = tensor("input_311_cast_fp16")]; - tensor var_4332 = const()[name = tensor("op_4332"), val = tensor([1, 1])]; - tensor var_4334 = const()[name = tensor("op_4334"), val = tensor([1, 1])]; - tensor lora_out_417_pad_type_0 = const()[name = tensor("lora_out_417_pad_type_0"), val = tensor("custom")]; - tensor lora_out_417_pad_0 = const()[name = tensor("lora_out_417_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_419_weight_0_to_fp16 = const()[name = tensor("lora_out_419_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282260416)))]; - tensor lora_out_419_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4334, groups = var_4123, pad = lora_out_417_pad_0, pad_type = lora_out_417_pad_type_0, strides = var_4332, weight = lora_out_419_weight_0_to_fp16, x = input_311_cast_fp16)[name = tensor("lora_out_419_cast_fp16")]; - tensor query_43_cast_fp16 = add(x = pretrained_out_209_cast_fp16, y = lora_out_419_cast_fp16)[name = tensor("query_43_cast_fp16")]; - tensor var_4344 = const()[name = tensor("op_4344"), val = tensor([1, 1])]; - tensor var_4346 = const()[name = tensor("op_4346"), val = tensor([1, 1])]; - tensor pretrained_out_211_pad_type_0 = const()[name = tensor("pretrained_out_211_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_211_pad_0 = const()[name = tensor("pretrained_out_211_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282301440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283120704))), name = tensor("layers_10_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_211_cast_fp16 = conv(dilations = var_4346, groups = var_4123, pad = pretrained_out_211_pad_0, pad_type = pretrained_out_211_pad_type_0, strides = var_4344, weight = layers_10_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_211_cast_fp16")]; - tensor var_4350 = const()[name = tensor("op_4350"), val = tensor([1, 1])]; - tensor var_4352 = const()[name = tensor("op_4352"), val = tensor([1, 1])]; - tensor input_313_pad_type_0 = const()[name = tensor("input_313_pad_type_0"), val = tensor("custom")]; - tensor input_313_pad_0 = const()[name = tensor("input_313_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283120832)))]; - tensor input_313_cast_fp16 = conv(dilations = var_4352, groups = var_4123, pad = input_313_pad_0, pad_type = input_313_pad_type_0, strides = var_4350, weight = layers_10_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_313_cast_fp16")]; - tensor var_4356 = const()[name = tensor("op_4356"), val = tensor([1, 1])]; - tensor var_4358 = const()[name = tensor("op_4358"), val = tensor([1, 1])]; - tensor lora_out_421_pad_type_0 = const()[name = tensor("lora_out_421_pad_type_0"), val = tensor("custom")]; - tensor lora_out_421_pad_0 = const()[name = tensor("lora_out_421_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_423_weight_0_to_fp16 = const()[name = tensor("lora_out_423_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283161856)))]; - tensor lora_out_423_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4358, groups = var_4123, pad = lora_out_421_pad_0, pad_type = lora_out_421_pad_type_0, strides = var_4356, weight = lora_out_423_weight_0_to_fp16, x = input_313_cast_fp16)[name = tensor("lora_out_423_cast_fp16")]; - tensor key_43_cast_fp16 = add(x = pretrained_out_211_cast_fp16, y = lora_out_423_cast_fp16)[name = tensor("key_43_cast_fp16")]; - tensor var_4369 = const()[name = tensor("op_4369"), val = tensor([1, 1])]; - tensor var_4371 = const()[name = tensor("op_4371"), val = tensor([1, 1])]; - tensor pretrained_out_213_pad_type_0 = const()[name = tensor("pretrained_out_213_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_213_pad_0 = const()[name = tensor("pretrained_out_213_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283202880))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284022144))), name = tensor("layers_10_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_10_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284022272)))]; - tensor pretrained_out_213_cast_fp16 = conv(bias = layers_10_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_4371, groups = var_4123, pad = pretrained_out_213_pad_0, pad_type = pretrained_out_213_pad_type_0, strides = var_4369, weight = layers_10_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_213_cast_fp16")]; - tensor var_4375 = const()[name = tensor("op_4375"), val = tensor([1, 1])]; - tensor var_4377 = const()[name = tensor("op_4377"), val = tensor([1, 1])]; - tensor input_315_pad_type_0 = const()[name = tensor("input_315_pad_type_0"), val = tensor("custom")]; - tensor input_315_pad_0 = const()[name = tensor("input_315_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284024896)))]; - tensor input_315_cast_fp16 = conv(dilations = var_4377, groups = var_4123, pad = input_315_pad_0, pad_type = input_315_pad_type_0, strides = var_4375, weight = layers_10_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_315_cast_fp16")]; - tensor var_4381 = const()[name = tensor("op_4381"), val = tensor([1, 1])]; - tensor var_4383 = const()[name = tensor("op_4383"), val = tensor([1, 1])]; - tensor lora_out_425_pad_type_0 = const()[name = tensor("lora_out_425_pad_type_0"), val = tensor("custom")]; - tensor lora_out_425_pad_0 = const()[name = tensor("lora_out_425_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_427_weight_0_to_fp16 = const()[name = tensor("lora_out_427_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284065920)))]; - tensor lora_out_427_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4383, groups = var_4123, pad = lora_out_425_pad_0, pad_type = lora_out_425_pad_type_0, strides = var_4381, weight = lora_out_427_weight_0_to_fp16, x = input_315_cast_fp16)[name = tensor("lora_out_427_cast_fp16")]; - tensor value_43_cast_fp16 = add(x = pretrained_out_213_cast_fp16, y = lora_out_427_cast_fp16)[name = tensor("value_43_cast_fp16")]; - tensor var_4390 = const()[name = tensor("op_4390"), val = tensor([1, 20, 64, -1])]; - tensor var_4391_cast_fp16 = reshape(shape = var_4390, x = query_43_cast_fp16)[name = tensor("op_4391_cast_fp16")]; - tensor var_4392_to_fp16 = const()[name = tensor("op_4392_to_fp16"), val = tensor(0x1p-3)]; - tensor var_4393_cast_fp16 = mul(x = var_4391_cast_fp16, y = var_4392_to_fp16)[name = tensor("op_4393_cast_fp16")]; - tensor var_4394 = const()[name = tensor("op_4394"), val = tensor([1, 20, 64, -1])]; - tensor var_4395_cast_fp16 = reshape(shape = var_4394, x = key_43_cast_fp16)[name = tensor("op_4395_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_4393_cast_fp16, y = var_4395_cast_fp16)[name = tensor("mh_w_65_cast_fp16")]; - tensor var_4398_cast_fp16 = softmax(axis = var_4116, x = mh_w_65_cast_fp16)[name = tensor("op_4398_cast_fp16")]; - tensor var_4399 = const()[name = tensor("op_4399"), val = tensor([1, 20, 64, -1])]; - tensor var_4400_cast_fp16 = reshape(shape = var_4399, x = value_43_cast_fp16)[name = tensor("op_4400_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_4400_cast_fp16, y = var_4398_cast_fp16)[name = tensor("attn_43_cast_fp16")]; - tensor var_4403 = const()[name = tensor("op_4403"), val = tensor([1, 1280, 1, -1])]; - tensor input_317_cast_fp16 = reshape(shape = var_4403, x = attn_43_cast_fp16)[name = tensor("input_317_cast_fp16")]; - tensor var_4410 = const()[name = tensor("op_4410"), val = tensor([1, 1])]; - tensor var_4412 = const()[name = tensor("op_4412"), val = tensor([1, 1])]; - tensor pretrained_out_215_pad_type_0 = const()[name = tensor("pretrained_out_215_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_215_pad_0 = const()[name = tensor("pretrained_out_215_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284106944))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284926208))), name = tensor("layers_10_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_10_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284926336)))]; - tensor pretrained_out_215_cast_fp16 = conv(bias = layers_10_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_4412, groups = var_4123, pad = pretrained_out_215_pad_0, pad_type = pretrained_out_215_pad_type_0, strides = var_4410, weight = layers_10_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_317_cast_fp16)[name = tensor("pretrained_out_215_cast_fp16")]; - tensor var_4416 = const()[name = tensor("op_4416"), val = tensor([1, 1])]; - tensor var_4418 = const()[name = tensor("op_4418"), val = tensor([1, 1])]; - tensor input_319_pad_type_0 = const()[name = tensor("input_319_pad_type_0"), val = tensor("custom")]; - tensor input_319_pad_0 = const()[name = tensor("input_319_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284928960)))]; - tensor input_319_cast_fp16 = conv(dilations = var_4418, groups = var_4123, pad = input_319_pad_0, pad_type = input_319_pad_type_0, strides = var_4416, weight = layers_10_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_317_cast_fp16)[name = tensor("input_319_cast_fp16")]; - tensor var_4422 = const()[name = tensor("op_4422"), val = tensor([1, 1])]; - tensor var_4424 = const()[name = tensor("op_4424"), val = tensor([1, 1])]; - tensor lora_out_429_pad_type_0 = const()[name = tensor("lora_out_429_pad_type_0"), val = tensor("custom")]; - tensor lora_out_429_pad_0 = const()[name = tensor("lora_out_429_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_431_weight_0_to_fp16 = const()[name = tensor("lora_out_431_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284969984)))]; - tensor lora_out_431_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4424, groups = var_4123, pad = lora_out_429_pad_0, pad_type = lora_out_429_pad_type_0, strides = var_4422, weight = lora_out_431_weight_0_to_fp16, x = input_319_cast_fp16)[name = tensor("lora_out_431_cast_fp16")]; - tensor obj_131_cast_fp16 = add(x = pretrained_out_215_cast_fp16, y = lora_out_431_cast_fp16)[name = tensor("obj_131_cast_fp16")]; - tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = obj_131_cast_fp16)[name = tensor("inputs_65_cast_fp16")]; - tensor var_4433 = const()[name = tensor("op_4433"), val = tensor([1])]; - tensor channels_mean_65_cast_fp16 = reduce_mean(axes = var_4433, keep_dims = var_4124, 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_4437 = const()[name = tensor("op_4437"), val = tensor([1])]; - tensor var_4438_cast_fp16 = reduce_mean(axes = var_4437, keep_dims = var_4124, x = zero_mean_sq_65_cast_fp16)[name = tensor("op_4438_cast_fp16")]; - tensor var_4439_to_fp16 = const()[name = tensor("op_4439_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_4440_cast_fp16 = add(x = var_4438_cast_fp16, y = var_4439_to_fp16)[name = tensor("op_4440_cast_fp16")]; - tensor denom_65_epsilon_0 = const()[name = tensor("denom_65_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_65_cast_fp16 = rsqrt(epsilon = denom_65_epsilon_0, x = var_4440_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_321_gamma_0_to_fp16 = const()[name = tensor("input_321_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285011008)))]; - tensor input_321_beta_0_to_fp16 = const()[name = tensor("input_321_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285013632)))]; - tensor input_321_epsilon_0_to_fp16 = const()[name = tensor("input_321_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_321_cast_fp16 = batch_norm(beta = input_321_beta_0_to_fp16, epsilon = input_321_epsilon_0_to_fp16, gamma = input_321_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_321_cast_fp16")]; - tensor var_4454 = const()[name = tensor("op_4454"), val = tensor([1, 1])]; - tensor var_4456 = const()[name = tensor("op_4456"), val = tensor([1, 1])]; - tensor pretrained_out_217_pad_type_0 = const()[name = tensor("pretrained_out_217_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_217_pad_0 = const()[name = tensor("pretrained_out_217_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285016256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288293120))), name = tensor("layers_10_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_10_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_10_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288293248)))]; - tensor pretrained_out_217_cast_fp16 = conv(bias = layers_10_fc1_pretrained_bias_to_fp16, dilations = var_4456, groups = var_4123, pad = pretrained_out_217_pad_0, pad_type = pretrained_out_217_pad_type_0, strides = var_4454, weight = layers_10_fc1_pretrained_weight_to_fp16_palettized, x = input_321_cast_fp16)[name = tensor("pretrained_out_217_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_323_pad_type_0 = const()[name = tensor("input_323_pad_type_0"), val = tensor("custom")]; - tensor input_323_pad_0 = const()[name = tensor("input_323_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_10_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288303552)))]; - tensor input_323_cast_fp16 = conv(dilations = var_4462, groups = var_4123, pad = input_323_pad_0, pad_type = input_323_pad_type_0, strides = var_4460, weight = layers_10_fc1_loraA_weight_to_fp16, x = input_321_cast_fp16)[name = tensor("input_323_cast_fp16")]; - tensor var_4466 = const()[name = tensor("op_4466"), val = tensor([1, 1])]; - tensor var_4468 = const()[name = tensor("op_4468"), val = tensor([1, 1])]; - tensor lora_out_433_pad_type_0 = const()[name = tensor("lora_out_433_pad_type_0"), val = tensor("custom")]; - tensor lora_out_433_pad_0 = const()[name = tensor("lora_out_433_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_435_weight_0_to_fp16 = const()[name = tensor("lora_out_435_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288344576)))]; - tensor lora_out_435_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_4468, groups = var_4123, pad = lora_out_433_pad_0, pad_type = lora_out_433_pad_type_0, strides = var_4466, weight = lora_out_435_weight_0_to_fp16, x = input_323_cast_fp16)[name = tensor("lora_out_435_cast_fp16")]; - tensor input_325_cast_fp16 = add(x = pretrained_out_217_cast_fp16, y = lora_out_435_cast_fp16)[name = tensor("input_325_cast_fp16")]; - tensor input_327_mode_0 = const()[name = tensor("input_327_mode_0"), val = tensor("EXACT")]; - tensor input_327_cast_fp16 = gelu(mode = input_327_mode_0, x = input_325_cast_fp16)[name = tensor("input_327_cast_fp16")]; - tensor var_4480 = const()[name = tensor("op_4480"), val = tensor([1, 1])]; - tensor var_4482 = const()[name = tensor("op_4482"), val = tensor([1, 1])]; - tensor pretrained_out_219_pad_type_0 = const()[name = tensor("pretrained_out_219_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_219_pad_0 = const()[name = tensor("pretrained_out_219_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288508480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291785344))), name = tensor("layers_10_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_10_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_10_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291785472)))]; - tensor pretrained_out_219_cast_fp16 = conv(bias = layers_10_fc2_pretrained_bias_to_fp16, dilations = var_4482, groups = var_4123, pad = pretrained_out_219_pad_0, pad_type = pretrained_out_219_pad_type_0, strides = var_4480, weight = layers_10_fc2_pretrained_weight_to_fp16_palettized, x = input_327_cast_fp16)[name = tensor("pretrained_out_219_cast_fp16")]; - tensor var_4486 = const()[name = tensor("op_4486"), val = tensor([1, 1])]; - tensor var_4488 = const()[name = tensor("op_4488"), val = tensor([1, 1])]; - tensor input_329_pad_type_0 = const()[name = tensor("input_329_pad_type_0"), val = tensor("custom")]; - tensor input_329_pad_0 = const()[name = tensor("input_329_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_10_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_10_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291788096)))]; - tensor input_329_cast_fp16 = conv(dilations = var_4488, groups = var_4123, pad = input_329_pad_0, pad_type = input_329_pad_type_0, strides = var_4486, weight = layers_10_fc2_loraA_weight_to_fp16, x = input_327_cast_fp16)[name = tensor("input_329_cast_fp16")]; - tensor var_4492 = const()[name = tensor("op_4492"), val = tensor([1, 1])]; - tensor var_4494 = const()[name = tensor("op_4494"), val = tensor([1, 1])]; - tensor lora_out_437_pad_type_0 = const()[name = tensor("lora_out_437_pad_type_0"), val = tensor("custom")]; - tensor lora_out_437_pad_0 = const()[name = tensor("lora_out_437_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_439_weight_0_to_fp16 = const()[name = tensor("lora_out_439_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291952000)))]; - tensor lora_out_439_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4494, groups = var_4123, pad = lora_out_437_pad_0, pad_type = lora_out_437_pad_type_0, strides = var_4492, weight = lora_out_439_weight_0_to_fp16, x = input_329_cast_fp16)[name = tensor("lora_out_439_cast_fp16")]; - tensor hidden_states_23_cast_fp16 = add(x = pretrained_out_219_cast_fp16, y = lora_out_439_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_4510 = const()[name = tensor("op_4510"), val = tensor(3)]; - tensor var_4517 = const()[name = tensor("op_4517"), val = tensor(1)]; - tensor var_4518 = const()[name = tensor("op_4518"), val = tensor(true)]; - tensor var_4530 = const()[name = tensor("op_4530"), val = tensor([1])]; - tensor channels_mean_67_cast_fp16 = reduce_mean(axes = var_4530, keep_dims = var_4518, 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_4534 = const()[name = tensor("op_4534"), val = tensor([1])]; - tensor var_4535_cast_fp16 = reduce_mean(axes = var_4534, keep_dims = var_4518, x = zero_mean_sq_67_cast_fp16)[name = tensor("op_4535_cast_fp16")]; - tensor var_4536_to_fp16 = const()[name = tensor("op_4536_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_4537_cast_fp16 = add(x = var_4535_cast_fp16, y = var_4536_to_fp16)[name = tensor("op_4537_cast_fp16")]; - tensor denom_67_epsilon_0 = const()[name = tensor("denom_67_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_67_cast_fp16 = rsqrt(epsilon = denom_67_epsilon_0, x = var_4537_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_133_gamma_0_to_fp16 = const()[name = tensor("obj_133_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291993024)))]; - tensor obj_133_beta_0_to_fp16 = const()[name = tensor("obj_133_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291995648)))]; - tensor obj_133_epsilon_0_to_fp16 = const()[name = tensor("obj_133_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_133_cast_fp16 = batch_norm(beta = obj_133_beta_0_to_fp16, epsilon = obj_133_epsilon_0_to_fp16, gamma = obj_133_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_133_cast_fp16")]; - tensor var_4555 = const()[name = tensor("op_4555"), val = tensor([1, 1])]; - tensor var_4557 = const()[name = tensor("op_4557"), val = tensor([1, 1])]; - tensor pretrained_out_221_pad_type_0 = const()[name = tensor("pretrained_out_221_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_221_pad_0 = const()[name = tensor("pretrained_out_221_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291998272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292817536))), name = tensor("layers_11_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_11_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292817664)))]; - tensor pretrained_out_221_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_4557, groups = var_4517, pad = pretrained_out_221_pad_0, pad_type = pretrained_out_221_pad_type_0, strides = var_4555, weight = layers_11_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_133_cast_fp16)[name = tensor("pretrained_out_221_cast_fp16")]; - tensor var_4561 = const()[name = tensor("op_4561"), val = tensor([1, 1])]; - tensor var_4563 = const()[name = tensor("op_4563"), val = tensor([1, 1])]; - tensor input_331_pad_type_0 = const()[name = tensor("input_331_pad_type_0"), val = tensor("custom")]; - tensor input_331_pad_0 = const()[name = tensor("input_331_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292820288)))]; - tensor input_331_cast_fp16 = conv(dilations = var_4563, groups = var_4517, pad = input_331_pad_0, pad_type = input_331_pad_type_0, strides = var_4561, weight = layers_11_self_attn_q_proj_loraA_weight_to_fp16, x = obj_133_cast_fp16)[name = tensor("input_331_cast_fp16")]; - tensor var_4567 = const()[name = tensor("op_4567"), val = tensor([1, 1])]; - tensor var_4569 = const()[name = tensor("op_4569"), val = tensor([1, 1])]; - tensor lora_out_441_pad_type_0 = const()[name = tensor("lora_out_441_pad_type_0"), val = tensor("custom")]; - tensor lora_out_441_pad_0 = const()[name = tensor("lora_out_441_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_443_weight_0_to_fp16 = const()[name = tensor("lora_out_443_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292861312)))]; - tensor lora_out_443_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4569, groups = var_4517, pad = lora_out_441_pad_0, pad_type = lora_out_441_pad_type_0, strides = var_4567, weight = lora_out_443_weight_0_to_fp16, x = input_331_cast_fp16)[name = tensor("lora_out_443_cast_fp16")]; - tensor query_45_cast_fp16 = add(x = pretrained_out_221_cast_fp16, y = lora_out_443_cast_fp16)[name = tensor("query_45_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 pretrained_out_223_pad_type_0 = const()[name = tensor("pretrained_out_223_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_223_pad_0 = const()[name = tensor("pretrained_out_223_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292902336))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293721600))), name = tensor("layers_11_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_223_cast_fp16 = conv(dilations = var_4581, groups = var_4517, pad = pretrained_out_223_pad_0, pad_type = pretrained_out_223_pad_type_0, strides = var_4579, weight = layers_11_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_133_cast_fp16)[name = tensor("pretrained_out_223_cast_fp16")]; - tensor var_4585 = const()[name = tensor("op_4585"), val = tensor([1, 1])]; - tensor var_4587 = const()[name = tensor("op_4587"), val = tensor([1, 1])]; - tensor input_333_pad_type_0 = const()[name = tensor("input_333_pad_type_0"), val = tensor("custom")]; - tensor input_333_pad_0 = const()[name = tensor("input_333_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293721728)))]; - tensor input_333_cast_fp16 = conv(dilations = var_4587, groups = var_4517, pad = input_333_pad_0, pad_type = input_333_pad_type_0, strides = var_4585, weight = layers_11_self_attn_k_proj_loraA_weight_to_fp16, x = obj_133_cast_fp16)[name = tensor("input_333_cast_fp16")]; - tensor var_4591 = const()[name = tensor("op_4591"), val = tensor([1, 1])]; - tensor var_4593 = const()[name = tensor("op_4593"), val = tensor([1, 1])]; - tensor lora_out_445_pad_type_0 = const()[name = tensor("lora_out_445_pad_type_0"), val = tensor("custom")]; - tensor lora_out_445_pad_0 = const()[name = tensor("lora_out_445_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_447_weight_0_to_fp16 = const()[name = tensor("lora_out_447_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293762752)))]; - tensor lora_out_447_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4593, groups = var_4517, pad = lora_out_445_pad_0, pad_type = lora_out_445_pad_type_0, strides = var_4591, weight = lora_out_447_weight_0_to_fp16, x = input_333_cast_fp16)[name = tensor("lora_out_447_cast_fp16")]; - tensor current_key_23_cast_fp16 = add(x = pretrained_out_223_cast_fp16, y = lora_out_447_cast_fp16)[name = tensor("current_key_23_cast_fp16")]; - tensor var_4604 = const()[name = tensor("op_4604"), val = tensor([1, 1])]; - tensor var_4606 = const()[name = tensor("op_4606"), val = tensor([1, 1])]; - tensor pretrained_out_225_pad_type_0 = const()[name = tensor("pretrained_out_225_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_225_pad_0 = const()[name = tensor("pretrained_out_225_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293803776))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294623040))), name = tensor("layers_11_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_11_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294623168)))]; - tensor pretrained_out_225_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_4606, groups = var_4517, pad = pretrained_out_225_pad_0, pad_type = pretrained_out_225_pad_type_0, strides = var_4604, weight = layers_11_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_133_cast_fp16)[name = tensor("pretrained_out_225_cast_fp16")]; - tensor var_4610 = const()[name = tensor("op_4610"), val = tensor([1, 1])]; - tensor var_4612 = const()[name = tensor("op_4612"), val = tensor([1, 1])]; - tensor input_335_pad_type_0 = const()[name = tensor("input_335_pad_type_0"), val = tensor("custom")]; - tensor input_335_pad_0 = const()[name = tensor("input_335_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294625792)))]; - tensor input_335_cast_fp16 = conv(dilations = var_4612, groups = var_4517, pad = input_335_pad_0, pad_type = input_335_pad_type_0, strides = var_4610, weight = layers_11_self_attn_v_proj_loraA_weight_to_fp16, x = obj_133_cast_fp16)[name = tensor("input_335_cast_fp16")]; - tensor var_4616 = const()[name = tensor("op_4616"), val = tensor([1, 1])]; - tensor var_4618 = const()[name = tensor("op_4618"), val = tensor([1, 1])]; - tensor lora_out_449_pad_type_0 = const()[name = tensor("lora_out_449_pad_type_0"), val = tensor("custom")]; - tensor lora_out_449_pad_0 = const()[name = tensor("lora_out_449_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_451_weight_0_to_fp16 = const()[name = tensor("lora_out_451_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294666816)))]; - tensor lora_out_451_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4618, groups = var_4517, pad = lora_out_449_pad_0, pad_type = lora_out_449_pad_type_0, strides = var_4616, weight = lora_out_451_weight_0_to_fp16, x = input_335_cast_fp16)[name = tensor("lora_out_451_cast_fp16")]; - tensor current_value_23_cast_fp16 = add(x = pretrained_out_225_cast_fp16, y = lora_out_451_cast_fp16)[name = tensor("current_value_23_cast_fp16")]; - tensor var_4628_cast_fp16 = mul(x = current_key_23_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_4628_cast_fp16")]; - tensor var_4630_cast_fp16 = mul(x = var_103_cast_fp16_11, y = var_295_cast_fp16)[name = tensor("op_4630_cast_fp16")]; - tensor key_45_cast_fp16 = add(x = var_4628_cast_fp16, y = var_4630_cast_fp16)[name = tensor("key_45_cast_fp16")]; - tensor var_4632_cast_fp16 = mul(x = current_value_23_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_4632_cast_fp16")]; - tensor var_4634_cast_fp16 = mul(x = var_138_cast_fp16_11, y = var_295_cast_fp16)[name = tensor("op_4634_cast_fp16")]; - tensor value_45_cast_fp16 = add(x = var_4632_cast_fp16, y = var_4634_cast_fp16)[name = tensor("value_45_cast_fp16")]; - tensor var_4637 = const()[name = tensor("op_4637"), val = tensor([1, 20, 64, -1])]; - tensor var_4638_cast_fp16 = reshape(shape = var_4637, x = query_45_cast_fp16)[name = tensor("op_4638_cast_fp16")]; - tensor var_4639_to_fp16 = const()[name = tensor("op_4639_to_fp16"), val = tensor(0x1p-3)]; - tensor var_4640_cast_fp16 = mul(x = var_4638_cast_fp16, y = var_4639_to_fp16)[name = tensor("op_4640_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 = key_45_cast_fp16)[name = tensor("op_4642_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_4640_cast_fp16, y = var_4642_cast_fp16)[name = tensor("mh_w_67_cast_fp16")]; - tensor mh_w_69_cast_fp16 = add(x = mh_w_67_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_69_cast_fp16")]; - tensor var_4650_cast_fp16 = softmax(axis = var_4510, x = mh_w_69_cast_fp16)[name = tensor("op_4650_cast_fp16")]; - tensor var_4651 = const()[name = tensor("op_4651"), val = tensor([1, 20, 64, -1])]; - tensor var_4652_cast_fp16 = reshape(shape = var_4651, x = value_45_cast_fp16)[name = tensor("op_4652_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_4652_cast_fp16, y = var_4650_cast_fp16)[name = tensor("attn_45_cast_fp16")]; - tensor var_4655 = const()[name = tensor("op_4655"), val = tensor([1, 1280, 1, -1])]; - tensor input_337_cast_fp16 = reshape(shape = var_4655, x = attn_45_cast_fp16)[name = tensor("input_337_cast_fp16")]; - tensor var_4662 = const()[name = tensor("op_4662"), val = tensor([1, 1])]; - tensor var_4664 = const()[name = tensor("op_4664"), val = tensor([1, 1])]; - tensor pretrained_out_227_pad_type_0 = const()[name = tensor("pretrained_out_227_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_227_pad_0 = const()[name = tensor("pretrained_out_227_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294707840))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295527104))), name = tensor("layers_11_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_11_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295527232)))]; - tensor pretrained_out_227_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_4664, groups = var_4517, pad = pretrained_out_227_pad_0, pad_type = pretrained_out_227_pad_type_0, strides = var_4662, weight = layers_11_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_337_cast_fp16)[name = tensor("pretrained_out_227_cast_fp16")]; - tensor var_4668 = const()[name = tensor("op_4668"), val = tensor([1, 1])]; - tensor var_4670 = const()[name = tensor("op_4670"), val = tensor([1, 1])]; - tensor input_339_pad_type_0 = const()[name = tensor("input_339_pad_type_0"), val = tensor("custom")]; - tensor input_339_pad_0 = const()[name = tensor("input_339_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295529856)))]; - tensor input_339_cast_fp16 = conv(dilations = var_4670, groups = var_4517, pad = input_339_pad_0, pad_type = input_339_pad_type_0, strides = var_4668, weight = layers_11_self_attn_o_proj_loraA_weight_to_fp16, x = input_337_cast_fp16)[name = tensor("input_339_cast_fp16")]; - tensor var_4674 = const()[name = tensor("op_4674"), val = tensor([1, 1])]; - tensor var_4676 = const()[name = tensor("op_4676"), val = tensor([1, 1])]; - tensor lora_out_453_pad_type_0 = const()[name = tensor("lora_out_453_pad_type_0"), val = tensor("custom")]; - tensor lora_out_453_pad_0 = const()[name = tensor("lora_out_453_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_455_weight_0_to_fp16 = const()[name = tensor("lora_out_455_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295570880)))]; - tensor lora_out_455_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4676, groups = var_4517, pad = lora_out_453_pad_0, pad_type = lora_out_453_pad_type_0, strides = var_4674, weight = lora_out_455_weight_0_to_fp16, x = input_339_cast_fp16)[name = tensor("lora_out_455_cast_fp16")]; - tensor obj_139_cast_fp16 = add(x = pretrained_out_227_cast_fp16, y = lora_out_455_cast_fp16)[name = tensor("obj_139_cast_fp16")]; - tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = obj_139_cast_fp16)[name = tensor("inputs_69_cast_fp16")]; - tensor var_4689 = const()[name = tensor("op_4689"), val = tensor([1])]; - tensor channels_mean_69_cast_fp16 = reduce_mean(axes = var_4689, keep_dims = var_4518, 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_4693 = const()[name = tensor("op_4693"), val = tensor([1])]; - tensor var_4694_cast_fp16 = reduce_mean(axes = var_4693, keep_dims = var_4518, x = zero_mean_sq_69_cast_fp16)[name = tensor("op_4694_cast_fp16")]; - tensor var_4695_to_fp16 = const()[name = tensor("op_4695_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_4696_cast_fp16 = add(x = var_4694_cast_fp16, y = var_4695_to_fp16)[name = tensor("op_4696_cast_fp16")]; - tensor denom_69_epsilon_0 = const()[name = tensor("denom_69_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_69_cast_fp16 = rsqrt(epsilon = denom_69_epsilon_0, x = var_4696_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_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(295611904)))]; - 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(295614528)))]; - 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_69_cast_fp16)[name = tensor("obj_141_cast_fp16")]; - tensor var_4714 = const()[name = tensor("op_4714"), val = tensor([1, 1])]; - tensor var_4716 = const()[name = tensor("op_4716"), val = tensor([1, 1])]; - tensor pretrained_out_229_pad_type_0 = const()[name = tensor("pretrained_out_229_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_229_pad_0 = const()[name = tensor("pretrained_out_229_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295617152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296436416))), name = tensor("layers_11_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_11_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296436544)))]; - tensor pretrained_out_229_cast_fp16 = conv(bias = layers_11_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_4716, groups = var_4517, pad = pretrained_out_229_pad_0, pad_type = pretrained_out_229_pad_type_0, strides = var_4714, weight = layers_11_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_141_cast_fp16)[name = tensor("pretrained_out_229_cast_fp16")]; - tensor var_4720 = const()[name = tensor("op_4720"), val = tensor([1, 1])]; - tensor var_4722 = const()[name = tensor("op_4722"), val = tensor([1, 1])]; - tensor input_341_pad_type_0 = const()[name = tensor("input_341_pad_type_0"), val = tensor("custom")]; - tensor input_341_pad_0 = const()[name = tensor("input_341_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296439168)))]; - tensor input_341_cast_fp16 = conv(dilations = var_4722, groups = var_4517, pad = input_341_pad_0, pad_type = input_341_pad_type_0, strides = var_4720, weight = layers_11_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("input_341_cast_fp16")]; - tensor var_4726 = const()[name = tensor("op_4726"), val = tensor([1, 1])]; - tensor var_4728 = const()[name = tensor("op_4728"), val = tensor([1, 1])]; - tensor lora_out_457_pad_type_0 = const()[name = tensor("lora_out_457_pad_type_0"), val = tensor("custom")]; - tensor lora_out_457_pad_0 = const()[name = tensor("lora_out_457_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_459_weight_0_to_fp16 = const()[name = tensor("lora_out_459_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296480192)))]; - tensor lora_out_459_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4728, groups = var_4517, pad = lora_out_457_pad_0, pad_type = lora_out_457_pad_type_0, strides = var_4726, weight = lora_out_459_weight_0_to_fp16, x = input_341_cast_fp16)[name = tensor("lora_out_459_cast_fp16")]; - tensor query_47_cast_fp16 = add(x = pretrained_out_229_cast_fp16, y = lora_out_459_cast_fp16)[name = tensor("query_47_cast_fp16")]; - tensor var_4738 = const()[name = tensor("op_4738"), val = tensor([1, 1])]; - tensor var_4740 = const()[name = tensor("op_4740"), val = tensor([1, 1])]; - tensor pretrained_out_231_pad_type_0 = const()[name = tensor("pretrained_out_231_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_231_pad_0 = const()[name = tensor("pretrained_out_231_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296521216))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297340480))), name = tensor("layers_11_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_231_cast_fp16 = conv(dilations = var_4740, groups = var_4517, pad = pretrained_out_231_pad_0, pad_type = pretrained_out_231_pad_type_0, strides = var_4738, weight = layers_11_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_231_cast_fp16")]; - tensor var_4744 = const()[name = tensor("op_4744"), val = tensor([1, 1])]; - tensor var_4746 = const()[name = tensor("op_4746"), val = tensor([1, 1])]; - tensor input_343_pad_type_0 = const()[name = tensor("input_343_pad_type_0"), val = tensor("custom")]; - tensor input_343_pad_0 = const()[name = tensor("input_343_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297340608)))]; - tensor input_343_cast_fp16 = conv(dilations = var_4746, groups = var_4517, pad = input_343_pad_0, pad_type = input_343_pad_type_0, strides = var_4744, weight = layers_11_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_343_cast_fp16")]; - tensor var_4750 = const()[name = tensor("op_4750"), val = tensor([1, 1])]; - tensor var_4752 = const()[name = tensor("op_4752"), val = tensor([1, 1])]; - tensor lora_out_461_pad_type_0 = const()[name = tensor("lora_out_461_pad_type_0"), val = tensor("custom")]; - tensor lora_out_461_pad_0 = const()[name = tensor("lora_out_461_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_463_weight_0_to_fp16 = const()[name = tensor("lora_out_463_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297381632)))]; - tensor lora_out_463_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4752, groups = var_4517, pad = lora_out_461_pad_0, pad_type = lora_out_461_pad_type_0, strides = var_4750, weight = lora_out_463_weight_0_to_fp16, x = input_343_cast_fp16)[name = tensor("lora_out_463_cast_fp16")]; - tensor key_47_cast_fp16 = add(x = pretrained_out_231_cast_fp16, y = lora_out_463_cast_fp16)[name = tensor("key_47_cast_fp16")]; - tensor var_4763 = const()[name = tensor("op_4763"), val = tensor([1, 1])]; - tensor var_4765 = const()[name = tensor("op_4765"), val = tensor([1, 1])]; - tensor pretrained_out_233_pad_type_0 = const()[name = tensor("pretrained_out_233_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_233_pad_0 = const()[name = tensor("pretrained_out_233_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297422656))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298241920))), name = tensor("layers_11_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_11_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298242048)))]; - tensor pretrained_out_233_cast_fp16 = conv(bias = layers_11_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_4765, groups = var_4517, pad = pretrained_out_233_pad_0, pad_type = pretrained_out_233_pad_type_0, strides = var_4763, weight = layers_11_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_233_cast_fp16")]; - tensor var_4769 = const()[name = tensor("op_4769"), val = tensor([1, 1])]; - tensor var_4771 = const()[name = tensor("op_4771"), val = tensor([1, 1])]; - tensor input_345_pad_type_0 = const()[name = tensor("input_345_pad_type_0"), val = tensor("custom")]; - tensor input_345_pad_0 = const()[name = tensor("input_345_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298244672)))]; - tensor input_345_cast_fp16 = conv(dilations = var_4771, groups = var_4517, pad = input_345_pad_0, pad_type = input_345_pad_type_0, strides = var_4769, weight = layers_11_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_345_cast_fp16")]; - tensor var_4775 = const()[name = tensor("op_4775"), val = tensor([1, 1])]; - tensor var_4777 = const()[name = tensor("op_4777"), val = tensor([1, 1])]; - tensor lora_out_465_pad_type_0 = const()[name = tensor("lora_out_465_pad_type_0"), val = tensor("custom")]; - tensor lora_out_465_pad_0 = const()[name = tensor("lora_out_465_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_467_weight_0_to_fp16 = const()[name = tensor("lora_out_467_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298285696)))]; - tensor lora_out_467_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4777, groups = var_4517, pad = lora_out_465_pad_0, pad_type = lora_out_465_pad_type_0, strides = var_4775, weight = lora_out_467_weight_0_to_fp16, x = input_345_cast_fp16)[name = tensor("lora_out_467_cast_fp16")]; - tensor value_47_cast_fp16 = add(x = pretrained_out_233_cast_fp16, y = lora_out_467_cast_fp16)[name = tensor("value_47_cast_fp16")]; - tensor var_4784 = const()[name = tensor("op_4784"), val = tensor([1, 20, 64, -1])]; - tensor var_4785_cast_fp16 = reshape(shape = var_4784, x = query_47_cast_fp16)[name = tensor("op_4785_cast_fp16")]; - tensor var_4786_to_fp16 = const()[name = tensor("op_4786_to_fp16"), val = tensor(0x1p-3)]; - tensor var_4787_cast_fp16 = mul(x = var_4785_cast_fp16, y = var_4786_to_fp16)[name = tensor("op_4787_cast_fp16")]; - tensor var_4788 = const()[name = tensor("op_4788"), val = tensor([1, 20, 64, -1])]; - tensor var_4789_cast_fp16 = reshape(shape = var_4788, x = key_47_cast_fp16)[name = tensor("op_4789_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_4787_cast_fp16, y = var_4789_cast_fp16)[name = tensor("mh_w_71_cast_fp16")]; - tensor var_4792_cast_fp16 = softmax(axis = var_4510, x = mh_w_71_cast_fp16)[name = tensor("op_4792_cast_fp16")]; - tensor var_4793 = const()[name = tensor("op_4793"), val = tensor([1, 20, 64, -1])]; - tensor var_4794_cast_fp16 = reshape(shape = var_4793, x = value_47_cast_fp16)[name = tensor("op_4794_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_4794_cast_fp16, y = var_4792_cast_fp16)[name = tensor("attn_47_cast_fp16")]; - tensor var_4797 = const()[name = tensor("op_4797"), val = tensor([1, 1280, 1, -1])]; - tensor input_347_cast_fp16 = reshape(shape = var_4797, x = attn_47_cast_fp16)[name = tensor("input_347_cast_fp16")]; - tensor var_4804 = const()[name = tensor("op_4804"), val = tensor([1, 1])]; - tensor var_4806 = const()[name = tensor("op_4806"), val = tensor([1, 1])]; - tensor pretrained_out_235_pad_type_0 = const()[name = tensor("pretrained_out_235_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_235_pad_0 = const()[name = tensor("pretrained_out_235_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298326720))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299145984))), name = tensor("layers_11_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_11_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299146112)))]; - tensor pretrained_out_235_cast_fp16 = conv(bias = layers_11_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_4806, groups = var_4517, pad = pretrained_out_235_pad_0, pad_type = pretrained_out_235_pad_type_0, strides = var_4804, weight = layers_11_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_347_cast_fp16)[name = tensor("pretrained_out_235_cast_fp16")]; - tensor var_4810 = const()[name = tensor("op_4810"), val = tensor([1, 1])]; - tensor var_4812 = const()[name = tensor("op_4812"), val = tensor([1, 1])]; - tensor input_349_pad_type_0 = const()[name = tensor("input_349_pad_type_0"), val = tensor("custom")]; - tensor input_349_pad_0 = const()[name = tensor("input_349_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299148736)))]; - tensor input_349_cast_fp16 = conv(dilations = var_4812, groups = var_4517, pad = input_349_pad_0, pad_type = input_349_pad_type_0, strides = var_4810, weight = layers_11_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_347_cast_fp16)[name = tensor("input_349_cast_fp16")]; - tensor var_4816 = const()[name = tensor("op_4816"), val = tensor([1, 1])]; - tensor var_4818 = const()[name = tensor("op_4818"), val = tensor([1, 1])]; - tensor lora_out_469_pad_type_0 = const()[name = tensor("lora_out_469_pad_type_0"), val = tensor("custom")]; - tensor lora_out_469_pad_0 = const()[name = tensor("lora_out_469_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_471_weight_0_to_fp16 = const()[name = tensor("lora_out_471_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299189760)))]; - tensor lora_out_471_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4818, groups = var_4517, pad = lora_out_469_pad_0, pad_type = lora_out_469_pad_type_0, strides = var_4816, weight = lora_out_471_weight_0_to_fp16, x = input_349_cast_fp16)[name = tensor("lora_out_471_cast_fp16")]; - tensor obj_143_cast_fp16 = add(x = pretrained_out_235_cast_fp16, y = lora_out_471_cast_fp16)[name = tensor("obj_143_cast_fp16")]; - tensor inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_143_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; - tensor var_4827 = const()[name = tensor("op_4827"), val = tensor([1])]; - tensor channels_mean_71_cast_fp16 = reduce_mean(axes = var_4827, keep_dims = var_4518, 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_4831 = const()[name = tensor("op_4831"), val = tensor([1])]; - tensor var_4832_cast_fp16 = reduce_mean(axes = var_4831, keep_dims = var_4518, x = zero_mean_sq_71_cast_fp16)[name = tensor("op_4832_cast_fp16")]; - tensor var_4833_to_fp16 = const()[name = tensor("op_4833_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_4834_cast_fp16 = add(x = var_4832_cast_fp16, y = var_4833_to_fp16)[name = tensor("op_4834_cast_fp16")]; - tensor denom_71_epsilon_0 = const()[name = tensor("denom_71_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_71_cast_fp16 = rsqrt(epsilon = denom_71_epsilon_0, x = var_4834_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_351_gamma_0_to_fp16 = const()[name = tensor("input_351_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299230784)))]; - tensor input_351_beta_0_to_fp16 = const()[name = tensor("input_351_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299233408)))]; - tensor input_351_epsilon_0_to_fp16 = const()[name = tensor("input_351_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_351_cast_fp16 = batch_norm(beta = input_351_beta_0_to_fp16, epsilon = input_351_epsilon_0_to_fp16, gamma = input_351_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_351_cast_fp16")]; - tensor var_4848 = const()[name = tensor("op_4848"), val = tensor([1, 1])]; - tensor var_4850 = const()[name = tensor("op_4850"), val = tensor([1, 1])]; - tensor pretrained_out_237_pad_type_0 = const()[name = tensor("pretrained_out_237_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_237_pad_0 = const()[name = tensor("pretrained_out_237_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299236032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302512896))), name = tensor("layers_11_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_11_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_11_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302513024)))]; - tensor pretrained_out_237_cast_fp16 = conv(bias = layers_11_fc1_pretrained_bias_to_fp16, dilations = var_4850, groups = var_4517, pad = pretrained_out_237_pad_0, pad_type = pretrained_out_237_pad_type_0, strides = var_4848, weight = layers_11_fc1_pretrained_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = tensor("pretrained_out_237_cast_fp16")]; - tensor var_4854 = const()[name = tensor("op_4854"), val = tensor([1, 1])]; - tensor var_4856 = const()[name = tensor("op_4856"), val = tensor([1, 1])]; - tensor input_353_pad_type_0 = const()[name = tensor("input_353_pad_type_0"), val = tensor("custom")]; - tensor input_353_pad_0 = const()[name = tensor("input_353_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_11_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302523328)))]; - tensor input_353_cast_fp16 = conv(dilations = var_4856, groups = var_4517, pad = input_353_pad_0, pad_type = input_353_pad_type_0, strides = var_4854, weight = layers_11_fc1_loraA_weight_to_fp16, x = input_351_cast_fp16)[name = tensor("input_353_cast_fp16")]; - tensor var_4860 = const()[name = tensor("op_4860"), val = tensor([1, 1])]; - tensor var_4862 = const()[name = tensor("op_4862"), val = tensor([1, 1])]; - tensor lora_out_473_pad_type_0 = const()[name = tensor("lora_out_473_pad_type_0"), val = tensor("custom")]; - tensor lora_out_473_pad_0 = const()[name = tensor("lora_out_473_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_475_weight_0_to_fp16 = const()[name = tensor("lora_out_475_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302564352)))]; - tensor lora_out_475_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_4862, groups = var_4517, pad = lora_out_473_pad_0, pad_type = lora_out_473_pad_type_0, strides = var_4860, weight = lora_out_475_weight_0_to_fp16, x = input_353_cast_fp16)[name = tensor("lora_out_475_cast_fp16")]; - tensor input_355_cast_fp16 = add(x = pretrained_out_237_cast_fp16, y = lora_out_475_cast_fp16)[name = tensor("input_355_cast_fp16")]; - tensor input_357_mode_0 = const()[name = tensor("input_357_mode_0"), val = tensor("EXACT")]; - tensor input_357_cast_fp16 = gelu(mode = input_357_mode_0, x = input_355_cast_fp16)[name = tensor("input_357_cast_fp16")]; - tensor var_4874 = const()[name = tensor("op_4874"), val = tensor([1, 1])]; - tensor var_4876 = const()[name = tensor("op_4876"), val = tensor([1, 1])]; - tensor pretrained_out_239_pad_type_0 = const()[name = tensor("pretrained_out_239_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_239_pad_0 = const()[name = tensor("pretrained_out_239_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302728256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306005120))), name = tensor("layers_11_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_11_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_11_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306005248)))]; - tensor pretrained_out_239_cast_fp16 = conv(bias = layers_11_fc2_pretrained_bias_to_fp16, dilations = var_4876, groups = var_4517, pad = pretrained_out_239_pad_0, pad_type = pretrained_out_239_pad_type_0, strides = var_4874, weight = layers_11_fc2_pretrained_weight_to_fp16_palettized, x = input_357_cast_fp16)[name = tensor("pretrained_out_239_cast_fp16")]; - tensor var_4880 = const()[name = tensor("op_4880"), val = tensor([1, 1])]; - tensor var_4882 = const()[name = tensor("op_4882"), val = tensor([1, 1])]; - tensor input_359_pad_type_0 = const()[name = tensor("input_359_pad_type_0"), val = tensor("custom")]; - tensor input_359_pad_0 = const()[name = tensor("input_359_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_11_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_11_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306007872)))]; - tensor input_359_cast_fp16 = conv(dilations = var_4882, groups = var_4517, pad = input_359_pad_0, pad_type = input_359_pad_type_0, strides = var_4880, weight = layers_11_fc2_loraA_weight_to_fp16, x = input_357_cast_fp16)[name = tensor("input_359_cast_fp16")]; - tensor var_4886 = const()[name = tensor("op_4886"), val = tensor([1, 1])]; - tensor var_4888 = const()[name = tensor("op_4888"), val = tensor([1, 1])]; - tensor lora_out_477_pad_type_0 = const()[name = tensor("lora_out_477_pad_type_0"), val = tensor("custom")]; - tensor lora_out_477_pad_0 = const()[name = tensor("lora_out_477_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_479_weight_0_to_fp16 = const()[name = tensor("lora_out_479_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306171776)))]; - tensor lora_out_479_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4888, groups = var_4517, pad = lora_out_477_pad_0, pad_type = lora_out_477_pad_type_0, strides = var_4886, weight = lora_out_479_weight_0_to_fp16, x = input_359_cast_fp16)[name = tensor("lora_out_479_cast_fp16")]; - tensor hidden_states_25_cast_fp16 = add(x = pretrained_out_239_cast_fp16, y = lora_out_479_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_4904 = const()[name = tensor("op_4904"), val = tensor(3)]; - tensor var_4911 = const()[name = tensor("op_4911"), val = tensor(1)]; - tensor var_4912 = const()[name = tensor("op_4912"), val = tensor(true)]; - tensor var_4924 = const()[name = tensor("op_4924"), val = tensor([1])]; - tensor channels_mean_73_cast_fp16 = reduce_mean(axes = var_4924, keep_dims = var_4912, 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_4928 = const()[name = tensor("op_4928"), val = tensor([1])]; - tensor var_4929_cast_fp16 = reduce_mean(axes = var_4928, keep_dims = var_4912, x = zero_mean_sq_73_cast_fp16)[name = tensor("op_4929_cast_fp16")]; - tensor var_4930_to_fp16 = const()[name = tensor("op_4930_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_4931_cast_fp16 = add(x = var_4929_cast_fp16, y = var_4930_to_fp16)[name = tensor("op_4931_cast_fp16")]; - tensor denom_73_epsilon_0 = const()[name = tensor("denom_73_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_73_cast_fp16 = rsqrt(epsilon = denom_73_epsilon_0, x = var_4931_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_145_gamma_0_to_fp16 = const()[name = tensor("obj_145_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306212800)))]; - tensor obj_145_beta_0_to_fp16 = const()[name = tensor("obj_145_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306215424)))]; - tensor obj_145_epsilon_0_to_fp16 = const()[name = tensor("obj_145_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_145_cast_fp16 = batch_norm(beta = obj_145_beta_0_to_fp16, epsilon = obj_145_epsilon_0_to_fp16, gamma = obj_145_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_145_cast_fp16")]; - tensor var_4949 = const()[name = tensor("op_4949"), val = tensor([1, 1])]; - tensor var_4951 = const()[name = tensor("op_4951"), val = tensor([1, 1])]; - tensor pretrained_out_241_pad_type_0 = const()[name = tensor("pretrained_out_241_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_241_pad_0 = const()[name = tensor("pretrained_out_241_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306218048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307037312))), name = tensor("layers_12_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_12_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307037440)))]; - tensor pretrained_out_241_cast_fp16 = conv(bias = layers_12_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_4951, groups = var_4911, pad = pretrained_out_241_pad_0, pad_type = pretrained_out_241_pad_type_0, strides = var_4949, weight = layers_12_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_145_cast_fp16)[name = tensor("pretrained_out_241_cast_fp16")]; - tensor var_4955 = const()[name = tensor("op_4955"), val = tensor([1, 1])]; - tensor var_4957 = const()[name = tensor("op_4957"), val = tensor([1, 1])]; - tensor input_361_pad_type_0 = const()[name = tensor("input_361_pad_type_0"), val = tensor("custom")]; - tensor input_361_pad_0 = const()[name = tensor("input_361_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307040064)))]; - tensor input_361_cast_fp16 = conv(dilations = var_4957, groups = var_4911, pad = input_361_pad_0, pad_type = input_361_pad_type_0, strides = var_4955, weight = layers_12_self_attn_q_proj_loraA_weight_to_fp16, x = obj_145_cast_fp16)[name = tensor("input_361_cast_fp16")]; - tensor var_4961 = const()[name = tensor("op_4961"), val = tensor([1, 1])]; - tensor var_4963 = const()[name = tensor("op_4963"), val = tensor([1, 1])]; - tensor lora_out_481_pad_type_0 = const()[name = tensor("lora_out_481_pad_type_0"), val = tensor("custom")]; - tensor lora_out_481_pad_0 = const()[name = tensor("lora_out_481_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_483_weight_0_to_fp16 = const()[name = tensor("lora_out_483_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307081088)))]; - tensor lora_out_483_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4963, groups = var_4911, pad = lora_out_481_pad_0, pad_type = lora_out_481_pad_type_0, strides = var_4961, weight = lora_out_483_weight_0_to_fp16, x = input_361_cast_fp16)[name = tensor("lora_out_483_cast_fp16")]; - tensor query_49_cast_fp16 = add(x = pretrained_out_241_cast_fp16, y = lora_out_483_cast_fp16)[name = tensor("query_49_cast_fp16")]; - tensor var_4973 = const()[name = tensor("op_4973"), val = tensor([1, 1])]; - tensor var_4975 = const()[name = tensor("op_4975"), val = tensor([1, 1])]; - tensor pretrained_out_243_pad_type_0 = const()[name = tensor("pretrained_out_243_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_243_pad_0 = const()[name = tensor("pretrained_out_243_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307122112))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307941376))), name = tensor("layers_12_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_243_cast_fp16 = conv(dilations = var_4975, groups = var_4911, pad = pretrained_out_243_pad_0, pad_type = pretrained_out_243_pad_type_0, strides = var_4973, weight = layers_12_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_145_cast_fp16)[name = tensor("pretrained_out_243_cast_fp16")]; - tensor var_4979 = const()[name = tensor("op_4979"), val = tensor([1, 1])]; - tensor var_4981 = const()[name = tensor("op_4981"), val = tensor([1, 1])]; - tensor input_363_pad_type_0 = const()[name = tensor("input_363_pad_type_0"), val = tensor("custom")]; - tensor input_363_pad_0 = const()[name = tensor("input_363_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307941504)))]; - tensor input_363_cast_fp16 = conv(dilations = var_4981, groups = var_4911, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = var_4979, weight = layers_12_self_attn_k_proj_loraA_weight_to_fp16, x = obj_145_cast_fp16)[name = tensor("input_363_cast_fp16")]; - tensor var_4985 = const()[name = tensor("op_4985"), val = tensor([1, 1])]; - tensor var_4987 = const()[name = tensor("op_4987"), val = tensor([1, 1])]; - tensor lora_out_485_pad_type_0 = const()[name = tensor("lora_out_485_pad_type_0"), val = tensor("custom")]; - tensor lora_out_485_pad_0 = const()[name = tensor("lora_out_485_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_487_weight_0_to_fp16 = const()[name = tensor("lora_out_487_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307982528)))]; - tensor lora_out_487_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4987, groups = var_4911, pad = lora_out_485_pad_0, pad_type = lora_out_485_pad_type_0, strides = var_4985, weight = lora_out_487_weight_0_to_fp16, x = input_363_cast_fp16)[name = tensor("lora_out_487_cast_fp16")]; - tensor current_key_25_cast_fp16 = add(x = pretrained_out_243_cast_fp16, y = lora_out_487_cast_fp16)[name = tensor("current_key_25_cast_fp16")]; - tensor var_4998 = const()[name = tensor("op_4998"), val = tensor([1, 1])]; - tensor var_5000 = const()[name = tensor("op_5000"), val = tensor([1, 1])]; - tensor pretrained_out_245_pad_type_0 = const()[name = tensor("pretrained_out_245_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_245_pad_0 = const()[name = tensor("pretrained_out_245_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308023552))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308842816))), name = tensor("layers_12_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_12_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308842944)))]; - tensor pretrained_out_245_cast_fp16 = conv(bias = layers_12_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_5000, groups = var_4911, pad = pretrained_out_245_pad_0, pad_type = pretrained_out_245_pad_type_0, strides = var_4998, weight = layers_12_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_145_cast_fp16)[name = tensor("pretrained_out_245_cast_fp16")]; - tensor var_5004 = const()[name = tensor("op_5004"), val = tensor([1, 1])]; - tensor var_5006 = const()[name = tensor("op_5006"), val = tensor([1, 1])]; - tensor input_365_pad_type_0 = const()[name = tensor("input_365_pad_type_0"), val = tensor("custom")]; - tensor input_365_pad_0 = const()[name = tensor("input_365_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308845568)))]; - tensor input_365_cast_fp16 = conv(dilations = var_5006, groups = var_4911, pad = input_365_pad_0, pad_type = input_365_pad_type_0, strides = var_5004, weight = layers_12_self_attn_v_proj_loraA_weight_to_fp16, x = obj_145_cast_fp16)[name = tensor("input_365_cast_fp16")]; - tensor var_5010 = const()[name = tensor("op_5010"), val = tensor([1, 1])]; - tensor var_5012 = const()[name = tensor("op_5012"), val = tensor([1, 1])]; - tensor lora_out_489_pad_type_0 = const()[name = tensor("lora_out_489_pad_type_0"), val = tensor("custom")]; - tensor lora_out_489_pad_0 = const()[name = tensor("lora_out_489_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_491_weight_0_to_fp16 = const()[name = tensor("lora_out_491_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308886592)))]; - tensor lora_out_491_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5012, groups = var_4911, pad = lora_out_489_pad_0, pad_type = lora_out_489_pad_type_0, strides = var_5010, weight = lora_out_491_weight_0_to_fp16, x = input_365_cast_fp16)[name = tensor("lora_out_491_cast_fp16")]; - tensor current_value_25_cast_fp16 = add(x = pretrained_out_245_cast_fp16, y = lora_out_491_cast_fp16)[name = tensor("current_value_25_cast_fp16")]; - tensor var_5022_cast_fp16 = mul(x = current_key_25_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_5022_cast_fp16")]; - tensor var_5024_cast_fp16 = mul(x = var_103_cast_fp16_12, y = var_295_cast_fp16)[name = tensor("op_5024_cast_fp16")]; - tensor key_49_cast_fp16 = add(x = var_5022_cast_fp16, y = var_5024_cast_fp16)[name = tensor("key_49_cast_fp16")]; - tensor var_5026_cast_fp16 = mul(x = current_value_25_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_5026_cast_fp16")]; - tensor var_5028_cast_fp16 = mul(x = var_138_cast_fp16_12, y = var_295_cast_fp16)[name = tensor("op_5028_cast_fp16")]; - tensor value_49_cast_fp16 = add(x = var_5026_cast_fp16, y = var_5028_cast_fp16)[name = tensor("value_49_cast_fp16")]; - tensor var_5031 = const()[name = tensor("op_5031"), val = tensor([1, 20, 64, -1])]; - tensor var_5032_cast_fp16 = reshape(shape = var_5031, x = query_49_cast_fp16)[name = tensor("op_5032_cast_fp16")]; - tensor var_5033_to_fp16 = const()[name = tensor("op_5033_to_fp16"), val = tensor(0x1p-3)]; - tensor var_5034_cast_fp16 = mul(x = var_5032_cast_fp16, y = var_5033_to_fp16)[name = tensor("op_5034_cast_fp16")]; - tensor var_5035 = const()[name = tensor("op_5035"), val = tensor([1, 20, 64, -1])]; - tensor var_5036_cast_fp16 = reshape(shape = var_5035, x = key_49_cast_fp16)[name = tensor("op_5036_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_5034_cast_fp16, y = var_5036_cast_fp16)[name = tensor("mh_w_73_cast_fp16")]; - tensor mh_w_75_cast_fp16 = add(x = mh_w_73_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_75_cast_fp16")]; - tensor var_5044_cast_fp16 = softmax(axis = var_4904, x = mh_w_75_cast_fp16)[name = tensor("op_5044_cast_fp16")]; - tensor var_5045 = const()[name = tensor("op_5045"), val = tensor([1, 20, 64, -1])]; - tensor var_5046_cast_fp16 = reshape(shape = var_5045, x = value_49_cast_fp16)[name = tensor("op_5046_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_5046_cast_fp16, y = var_5044_cast_fp16)[name = tensor("attn_49_cast_fp16")]; - tensor var_5049 = const()[name = tensor("op_5049"), val = tensor([1, 1280, 1, -1])]; - tensor input_367_cast_fp16 = reshape(shape = var_5049, x = attn_49_cast_fp16)[name = tensor("input_367_cast_fp16")]; - tensor var_5056 = const()[name = tensor("op_5056"), val = tensor([1, 1])]; - tensor var_5058 = const()[name = tensor("op_5058"), val = tensor([1, 1])]; - tensor pretrained_out_247_pad_type_0 = const()[name = tensor("pretrained_out_247_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_247_pad_0 = const()[name = tensor("pretrained_out_247_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308927616))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309746880))), name = tensor("layers_12_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_12_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309747008)))]; - tensor pretrained_out_247_cast_fp16 = conv(bias = layers_12_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_5058, groups = var_4911, pad = pretrained_out_247_pad_0, pad_type = pretrained_out_247_pad_type_0, strides = var_5056, weight = layers_12_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_367_cast_fp16)[name = tensor("pretrained_out_247_cast_fp16")]; - tensor var_5062 = const()[name = tensor("op_5062"), val = tensor([1, 1])]; - tensor var_5064 = const()[name = tensor("op_5064"), val = tensor([1, 1])]; - tensor input_369_pad_type_0 = const()[name = tensor("input_369_pad_type_0"), val = tensor("custom")]; - tensor input_369_pad_0 = const()[name = tensor("input_369_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309749632)))]; - tensor input_369_cast_fp16 = conv(dilations = var_5064, groups = var_4911, pad = input_369_pad_0, pad_type = input_369_pad_type_0, strides = var_5062, weight = layers_12_self_attn_o_proj_loraA_weight_to_fp16, x = input_367_cast_fp16)[name = tensor("input_369_cast_fp16")]; - tensor var_5068 = const()[name = tensor("op_5068"), val = tensor([1, 1])]; - tensor var_5070 = const()[name = tensor("op_5070"), val = tensor([1, 1])]; - tensor lora_out_493_pad_type_0 = const()[name = tensor("lora_out_493_pad_type_0"), val = tensor("custom")]; - tensor lora_out_493_pad_0 = const()[name = tensor("lora_out_493_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_495_weight_0_to_fp16 = const()[name = tensor("lora_out_495_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309790656)))]; - tensor lora_out_495_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5070, groups = var_4911, pad = lora_out_493_pad_0, pad_type = lora_out_493_pad_type_0, strides = var_5068, weight = lora_out_495_weight_0_to_fp16, x = input_369_cast_fp16)[name = tensor("lora_out_495_cast_fp16")]; - tensor obj_151_cast_fp16 = add(x = pretrained_out_247_cast_fp16, y = lora_out_495_cast_fp16)[name = tensor("obj_151_cast_fp16")]; - tensor inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = obj_151_cast_fp16)[name = tensor("inputs_75_cast_fp16")]; - tensor var_5083 = const()[name = tensor("op_5083"), val = tensor([1])]; - tensor channels_mean_75_cast_fp16 = reduce_mean(axes = var_5083, keep_dims = var_4912, 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_5087 = const()[name = tensor("op_5087"), val = tensor([1])]; - tensor var_5088_cast_fp16 = reduce_mean(axes = var_5087, keep_dims = var_4912, x = zero_mean_sq_75_cast_fp16)[name = tensor("op_5088_cast_fp16")]; - tensor var_5089_to_fp16 = const()[name = tensor("op_5089_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_5090_cast_fp16 = add(x = var_5088_cast_fp16, y = var_5089_to_fp16)[name = tensor("op_5090_cast_fp16")]; - tensor denom_75_epsilon_0 = const()[name = tensor("denom_75_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_75_cast_fp16 = rsqrt(epsilon = denom_75_epsilon_0, x = var_5090_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_153_gamma_0_to_fp16 = const()[name = tensor("obj_153_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309831680)))]; - tensor obj_153_beta_0_to_fp16 = const()[name = tensor("obj_153_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309834304)))]; - tensor obj_153_epsilon_0_to_fp16 = const()[name = tensor("obj_153_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_153_cast_fp16 = batch_norm(beta = obj_153_beta_0_to_fp16, epsilon = obj_153_epsilon_0_to_fp16, gamma = obj_153_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_153_cast_fp16")]; - tensor var_5108 = const()[name = tensor("op_5108"), val = tensor([1, 1])]; - tensor var_5110 = const()[name = tensor("op_5110"), val = tensor([1, 1])]; - tensor pretrained_out_249_pad_type_0 = const()[name = tensor("pretrained_out_249_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_249_pad_0 = const()[name = tensor("pretrained_out_249_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309836928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310656192))), name = tensor("layers_12_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_12_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310656320)))]; - tensor pretrained_out_249_cast_fp16 = conv(bias = layers_12_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_5110, groups = var_4911, pad = pretrained_out_249_pad_0, pad_type = pretrained_out_249_pad_type_0, strides = var_5108, weight = layers_12_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_153_cast_fp16)[name = tensor("pretrained_out_249_cast_fp16")]; - tensor var_5114 = const()[name = tensor("op_5114"), val = tensor([1, 1])]; - tensor var_5116 = const()[name = tensor("op_5116"), val = tensor([1, 1])]; - tensor input_371_pad_type_0 = const()[name = tensor("input_371_pad_type_0"), val = tensor("custom")]; - tensor input_371_pad_0 = const()[name = tensor("input_371_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310658944)))]; - tensor input_371_cast_fp16 = conv(dilations = var_5116, groups = var_4911, pad = input_371_pad_0, pad_type = input_371_pad_type_0, strides = var_5114, weight = layers_12_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_153_cast_fp16)[name = tensor("input_371_cast_fp16")]; - tensor var_5120 = const()[name = tensor("op_5120"), val = tensor([1, 1])]; - tensor var_5122 = const()[name = tensor("op_5122"), val = tensor([1, 1])]; - tensor lora_out_497_pad_type_0 = const()[name = tensor("lora_out_497_pad_type_0"), val = tensor("custom")]; - tensor lora_out_497_pad_0 = const()[name = tensor("lora_out_497_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_499_weight_0_to_fp16 = const()[name = tensor("lora_out_499_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310699968)))]; - tensor lora_out_499_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5122, groups = var_4911, pad = lora_out_497_pad_0, pad_type = lora_out_497_pad_type_0, strides = var_5120, weight = lora_out_499_weight_0_to_fp16, x = input_371_cast_fp16)[name = tensor("lora_out_499_cast_fp16")]; - tensor query_51_cast_fp16 = add(x = pretrained_out_249_cast_fp16, y = lora_out_499_cast_fp16)[name = tensor("query_51_cast_fp16")]; - tensor var_5132 = const()[name = tensor("op_5132"), val = tensor([1, 1])]; - tensor var_5134 = const()[name = tensor("op_5134"), val = tensor([1, 1])]; - tensor pretrained_out_251_pad_type_0 = const()[name = tensor("pretrained_out_251_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_251_pad_0 = const()[name = tensor("pretrained_out_251_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310740992))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311560256))), name = tensor("layers_12_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_251_cast_fp16 = conv(dilations = var_5134, groups = var_4911, pad = pretrained_out_251_pad_0, pad_type = pretrained_out_251_pad_type_0, strides = var_5132, weight = layers_12_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_251_cast_fp16")]; - tensor var_5138 = const()[name = tensor("op_5138"), val = tensor([1, 1])]; - tensor var_5140 = const()[name = tensor("op_5140"), val = tensor([1, 1])]; - tensor input_373_pad_type_0 = const()[name = tensor("input_373_pad_type_0"), val = tensor("custom")]; - tensor input_373_pad_0 = const()[name = tensor("input_373_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311560384)))]; - tensor input_373_cast_fp16 = conv(dilations = var_5140, groups = var_4911, pad = input_373_pad_0, pad_type = input_373_pad_type_0, strides = var_5138, weight = layers_12_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_373_cast_fp16")]; - tensor var_5144 = const()[name = tensor("op_5144"), val = tensor([1, 1])]; - tensor var_5146 = const()[name = tensor("op_5146"), val = tensor([1, 1])]; - tensor lora_out_501_pad_type_0 = const()[name = tensor("lora_out_501_pad_type_0"), val = tensor("custom")]; - tensor lora_out_501_pad_0 = const()[name = tensor("lora_out_501_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_503_weight_0_to_fp16 = const()[name = tensor("lora_out_503_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311601408)))]; - tensor lora_out_503_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5146, groups = var_4911, pad = lora_out_501_pad_0, pad_type = lora_out_501_pad_type_0, strides = var_5144, weight = lora_out_503_weight_0_to_fp16, x = input_373_cast_fp16)[name = tensor("lora_out_503_cast_fp16")]; - tensor key_51_cast_fp16 = add(x = pretrained_out_251_cast_fp16, y = lora_out_503_cast_fp16)[name = tensor("key_51_cast_fp16")]; - tensor var_5157 = const()[name = tensor("op_5157"), val = tensor([1, 1])]; - tensor var_5159 = const()[name = tensor("op_5159"), val = tensor([1, 1])]; - tensor pretrained_out_253_pad_type_0 = const()[name = tensor("pretrained_out_253_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_253_pad_0 = const()[name = tensor("pretrained_out_253_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311642432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312461696))), name = tensor("layers_12_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_12_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312461824)))]; - tensor pretrained_out_253_cast_fp16 = conv(bias = layers_12_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_5159, groups = var_4911, pad = pretrained_out_253_pad_0, pad_type = pretrained_out_253_pad_type_0, strides = var_5157, weight = layers_12_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_253_cast_fp16")]; - tensor var_5163 = const()[name = tensor("op_5163"), val = tensor([1, 1])]; - tensor var_5165 = const()[name = tensor("op_5165"), val = tensor([1, 1])]; - tensor input_375_pad_type_0 = const()[name = tensor("input_375_pad_type_0"), val = tensor("custom")]; - tensor input_375_pad_0 = const()[name = tensor("input_375_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312464448)))]; - tensor input_375_cast_fp16 = conv(dilations = var_5165, groups = var_4911, pad = input_375_pad_0, pad_type = input_375_pad_type_0, strides = var_5163, weight = layers_12_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_375_cast_fp16")]; - tensor var_5169 = const()[name = tensor("op_5169"), val = tensor([1, 1])]; - tensor var_5171 = const()[name = tensor("op_5171"), val = tensor([1, 1])]; - tensor lora_out_505_pad_type_0 = const()[name = tensor("lora_out_505_pad_type_0"), val = tensor("custom")]; - tensor lora_out_505_pad_0 = const()[name = tensor("lora_out_505_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_507_weight_0_to_fp16 = const()[name = tensor("lora_out_507_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312505472)))]; - tensor lora_out_507_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5171, groups = var_4911, pad = lora_out_505_pad_0, pad_type = lora_out_505_pad_type_0, strides = var_5169, weight = lora_out_507_weight_0_to_fp16, x = input_375_cast_fp16)[name = tensor("lora_out_507_cast_fp16")]; - tensor value_51_cast_fp16 = add(x = pretrained_out_253_cast_fp16, y = lora_out_507_cast_fp16)[name = tensor("value_51_cast_fp16")]; - tensor var_5178 = const()[name = tensor("op_5178"), val = tensor([1, 20, 64, -1])]; - tensor var_5179_cast_fp16 = reshape(shape = var_5178, x = query_51_cast_fp16)[name = tensor("op_5179_cast_fp16")]; - tensor var_5180_to_fp16 = const()[name = tensor("op_5180_to_fp16"), val = tensor(0x1p-3)]; - tensor var_5181_cast_fp16 = mul(x = var_5179_cast_fp16, y = var_5180_to_fp16)[name = tensor("op_5181_cast_fp16")]; - tensor var_5182 = const()[name = tensor("op_5182"), val = tensor([1, 20, 64, -1])]; - tensor var_5183_cast_fp16 = reshape(shape = var_5182, x = key_51_cast_fp16)[name = tensor("op_5183_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_5181_cast_fp16, y = var_5183_cast_fp16)[name = tensor("mh_w_77_cast_fp16")]; - tensor var_5186_cast_fp16 = softmax(axis = var_4904, x = mh_w_77_cast_fp16)[name = tensor("op_5186_cast_fp16")]; - tensor var_5187 = const()[name = tensor("op_5187"), val = tensor([1, 20, 64, -1])]; - tensor var_5188_cast_fp16 = reshape(shape = var_5187, x = value_51_cast_fp16)[name = tensor("op_5188_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_5188_cast_fp16, y = var_5186_cast_fp16)[name = tensor("attn_51_cast_fp16")]; - tensor var_5191 = const()[name = tensor("op_5191"), val = tensor([1, 1280, 1, -1])]; - tensor input_377_cast_fp16 = reshape(shape = var_5191, x = attn_51_cast_fp16)[name = tensor("input_377_cast_fp16")]; - tensor var_5198 = const()[name = tensor("op_5198"), val = tensor([1, 1])]; - tensor var_5200 = const()[name = tensor("op_5200"), val = tensor([1, 1])]; - tensor pretrained_out_255_pad_type_0 = const()[name = tensor("pretrained_out_255_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_255_pad_0 = const()[name = tensor("pretrained_out_255_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312546496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313365760))), name = tensor("layers_12_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_12_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313365888)))]; - tensor pretrained_out_255_cast_fp16 = conv(bias = layers_12_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_5200, groups = var_4911, pad = pretrained_out_255_pad_0, pad_type = pretrained_out_255_pad_type_0, strides = var_5198, weight = layers_12_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_377_cast_fp16)[name = tensor("pretrained_out_255_cast_fp16")]; - tensor var_5204 = const()[name = tensor("op_5204"), val = tensor([1, 1])]; - tensor var_5206 = const()[name = tensor("op_5206"), val = tensor([1, 1])]; - tensor input_379_pad_type_0 = const()[name = tensor("input_379_pad_type_0"), val = tensor("custom")]; - tensor input_379_pad_0 = const()[name = tensor("input_379_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313368512)))]; - tensor input_379_cast_fp16 = conv(dilations = var_5206, groups = var_4911, pad = input_379_pad_0, pad_type = input_379_pad_type_0, strides = var_5204, weight = layers_12_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_377_cast_fp16)[name = tensor("input_379_cast_fp16")]; - tensor var_5210 = const()[name = tensor("op_5210"), val = tensor([1, 1])]; - tensor var_5212 = const()[name = tensor("op_5212"), val = tensor([1, 1])]; - tensor lora_out_509_pad_type_0 = const()[name = tensor("lora_out_509_pad_type_0"), val = tensor("custom")]; - tensor lora_out_509_pad_0 = const()[name = tensor("lora_out_509_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_511_weight_0_to_fp16 = const()[name = tensor("lora_out_511_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313409536)))]; - tensor lora_out_511_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5212, groups = var_4911, pad = lora_out_509_pad_0, pad_type = lora_out_509_pad_type_0, strides = var_5210, weight = lora_out_511_weight_0_to_fp16, x = input_379_cast_fp16)[name = tensor("lora_out_511_cast_fp16")]; - tensor obj_155_cast_fp16 = add(x = pretrained_out_255_cast_fp16, y = lora_out_511_cast_fp16)[name = tensor("obj_155_cast_fp16")]; - tensor inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = obj_155_cast_fp16)[name = tensor("inputs_77_cast_fp16")]; - tensor var_5221 = const()[name = tensor("op_5221"), val = tensor([1])]; - tensor channels_mean_77_cast_fp16 = reduce_mean(axes = var_5221, keep_dims = var_4912, 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_5225 = const()[name = tensor("op_5225"), val = tensor([1])]; - tensor var_5226_cast_fp16 = reduce_mean(axes = var_5225, keep_dims = var_4912, x = zero_mean_sq_77_cast_fp16)[name = tensor("op_5226_cast_fp16")]; - tensor var_5227_to_fp16 = const()[name = tensor("op_5227_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_5228_cast_fp16 = add(x = var_5226_cast_fp16, y = var_5227_to_fp16)[name = tensor("op_5228_cast_fp16")]; - tensor denom_77_epsilon_0 = const()[name = tensor("denom_77_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_77_cast_fp16 = rsqrt(epsilon = denom_77_epsilon_0, x = var_5228_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_381_gamma_0_to_fp16 = const()[name = tensor("input_381_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313450560)))]; - tensor input_381_beta_0_to_fp16 = const()[name = tensor("input_381_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313453184)))]; - tensor input_381_epsilon_0_to_fp16 = const()[name = tensor("input_381_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_381_cast_fp16 = batch_norm(beta = input_381_beta_0_to_fp16, epsilon = input_381_epsilon_0_to_fp16, gamma = input_381_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_381_cast_fp16")]; - tensor var_5242 = const()[name = tensor("op_5242"), val = tensor([1, 1])]; - tensor var_5244 = const()[name = tensor("op_5244"), val = tensor([1, 1])]; - tensor pretrained_out_257_pad_type_0 = const()[name = tensor("pretrained_out_257_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_257_pad_0 = const()[name = tensor("pretrained_out_257_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313455808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316732672))), name = tensor("layers_12_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_12_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_12_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316732800)))]; - tensor pretrained_out_257_cast_fp16 = conv(bias = layers_12_fc1_pretrained_bias_to_fp16, dilations = var_5244, groups = var_4911, pad = pretrained_out_257_pad_0, pad_type = pretrained_out_257_pad_type_0, strides = var_5242, weight = layers_12_fc1_pretrained_weight_to_fp16_palettized, x = input_381_cast_fp16)[name = tensor("pretrained_out_257_cast_fp16")]; - tensor var_5248 = const()[name = tensor("op_5248"), val = tensor([1, 1])]; - tensor var_5250 = const()[name = tensor("op_5250"), val = tensor([1, 1])]; - tensor input_383_pad_type_0 = const()[name = tensor("input_383_pad_type_0"), val = tensor("custom")]; - tensor input_383_pad_0 = const()[name = tensor("input_383_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_12_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316743104)))]; - tensor input_383_cast_fp16 = conv(dilations = var_5250, groups = var_4911, pad = input_383_pad_0, pad_type = input_383_pad_type_0, strides = var_5248, weight = layers_12_fc1_loraA_weight_to_fp16, x = input_381_cast_fp16)[name = tensor("input_383_cast_fp16")]; - tensor var_5254 = const()[name = tensor("op_5254"), val = tensor([1, 1])]; - tensor var_5256 = const()[name = tensor("op_5256"), val = tensor([1, 1])]; - tensor lora_out_513_pad_type_0 = const()[name = tensor("lora_out_513_pad_type_0"), val = tensor("custom")]; - tensor lora_out_513_pad_0 = const()[name = tensor("lora_out_513_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_515_weight_0_to_fp16 = const()[name = tensor("lora_out_515_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316784128)))]; - tensor lora_out_515_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_5256, groups = var_4911, pad = lora_out_513_pad_0, pad_type = lora_out_513_pad_type_0, strides = var_5254, weight = lora_out_515_weight_0_to_fp16, x = input_383_cast_fp16)[name = tensor("lora_out_515_cast_fp16")]; - tensor input_385_cast_fp16 = add(x = pretrained_out_257_cast_fp16, y = lora_out_515_cast_fp16)[name = tensor("input_385_cast_fp16")]; - tensor input_387_mode_0 = const()[name = tensor("input_387_mode_0"), val = tensor("EXACT")]; - tensor input_387_cast_fp16 = gelu(mode = input_387_mode_0, x = input_385_cast_fp16)[name = tensor("input_387_cast_fp16")]; - tensor var_5268 = const()[name = tensor("op_5268"), val = tensor([1, 1])]; - tensor var_5270 = const()[name = tensor("op_5270"), val = tensor([1, 1])]; - tensor pretrained_out_259_pad_type_0 = const()[name = tensor("pretrained_out_259_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_259_pad_0 = const()[name = tensor("pretrained_out_259_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316948032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320224896))), name = tensor("layers_12_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_12_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_12_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320225024)))]; - tensor pretrained_out_259_cast_fp16 = conv(bias = layers_12_fc2_pretrained_bias_to_fp16, dilations = var_5270, groups = var_4911, pad = pretrained_out_259_pad_0, pad_type = pretrained_out_259_pad_type_0, strides = var_5268, weight = layers_12_fc2_pretrained_weight_to_fp16_palettized, x = input_387_cast_fp16)[name = tensor("pretrained_out_259_cast_fp16")]; - tensor var_5274 = const()[name = tensor("op_5274"), val = tensor([1, 1])]; - tensor var_5276 = const()[name = tensor("op_5276"), val = tensor([1, 1])]; - tensor input_389_pad_type_0 = const()[name = tensor("input_389_pad_type_0"), val = tensor("custom")]; - tensor input_389_pad_0 = const()[name = tensor("input_389_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_12_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_12_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320227648)))]; - tensor input_389_cast_fp16 = conv(dilations = var_5276, groups = var_4911, pad = input_389_pad_0, pad_type = input_389_pad_type_0, strides = var_5274, weight = layers_12_fc2_loraA_weight_to_fp16, x = input_387_cast_fp16)[name = tensor("input_389_cast_fp16")]; - tensor var_5280 = const()[name = tensor("op_5280"), val = tensor([1, 1])]; - tensor var_5282 = const()[name = tensor("op_5282"), val = tensor([1, 1])]; - tensor lora_out_517_pad_type_0 = const()[name = tensor("lora_out_517_pad_type_0"), val = tensor("custom")]; - tensor lora_out_517_pad_0 = const()[name = tensor("lora_out_517_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_519_weight_0_to_fp16 = const()[name = tensor("lora_out_519_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320391552)))]; - tensor lora_out_519_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5282, groups = var_4911, pad = lora_out_517_pad_0, pad_type = lora_out_517_pad_type_0, strides = var_5280, weight = lora_out_519_weight_0_to_fp16, x = input_389_cast_fp16)[name = tensor("lora_out_519_cast_fp16")]; - tensor hidden_states_27_cast_fp16 = add(x = pretrained_out_259_cast_fp16, y = lora_out_519_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_5298 = const()[name = tensor("op_5298"), val = tensor(3)]; - tensor var_5305 = const()[name = tensor("op_5305"), val = tensor(1)]; - tensor var_5306 = const()[name = tensor("op_5306"), val = tensor(true)]; - tensor var_5318 = const()[name = tensor("op_5318"), val = tensor([1])]; - tensor channels_mean_79_cast_fp16 = reduce_mean(axes = var_5318, keep_dims = var_5306, 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_5322 = const()[name = tensor("op_5322"), val = tensor([1])]; - tensor var_5323_cast_fp16 = reduce_mean(axes = var_5322, keep_dims = var_5306, x = zero_mean_sq_79_cast_fp16)[name = tensor("op_5323_cast_fp16")]; - tensor var_5324_to_fp16 = const()[name = tensor("op_5324_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_5325_cast_fp16 = add(x = var_5323_cast_fp16, y = var_5324_to_fp16)[name = tensor("op_5325_cast_fp16")]; - tensor denom_79_epsilon_0 = const()[name = tensor("denom_79_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_79_cast_fp16 = rsqrt(epsilon = denom_79_epsilon_0, x = var_5325_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_157_gamma_0_to_fp16 = const()[name = tensor("obj_157_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320432576)))]; - tensor obj_157_beta_0_to_fp16 = const()[name = tensor("obj_157_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320435200)))]; - tensor obj_157_epsilon_0_to_fp16 = const()[name = tensor("obj_157_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_157_cast_fp16 = batch_norm(beta = obj_157_beta_0_to_fp16, epsilon = obj_157_epsilon_0_to_fp16, gamma = obj_157_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_157_cast_fp16")]; - tensor var_5343 = const()[name = tensor("op_5343"), val = tensor([1, 1])]; - tensor var_5345 = const()[name = tensor("op_5345"), val = tensor([1, 1])]; - tensor pretrained_out_261_pad_type_0 = const()[name = tensor("pretrained_out_261_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_261_pad_0 = const()[name = tensor("pretrained_out_261_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320437824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321257088))), name = tensor("layers_13_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_13_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321257216)))]; - tensor pretrained_out_261_cast_fp16 = conv(bias = layers_13_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_5345, groups = var_5305, pad = pretrained_out_261_pad_0, pad_type = pretrained_out_261_pad_type_0, strides = var_5343, weight = layers_13_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_157_cast_fp16)[name = tensor("pretrained_out_261_cast_fp16")]; - tensor var_5349 = const()[name = tensor("op_5349"), val = tensor([1, 1])]; - tensor var_5351 = const()[name = tensor("op_5351"), val = tensor([1, 1])]; - tensor input_391_pad_type_0 = const()[name = tensor("input_391_pad_type_0"), val = tensor("custom")]; - tensor input_391_pad_0 = const()[name = tensor("input_391_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321259840)))]; - tensor input_391_cast_fp16 = conv(dilations = var_5351, groups = var_5305, pad = input_391_pad_0, pad_type = input_391_pad_type_0, strides = var_5349, weight = layers_13_self_attn_q_proj_loraA_weight_to_fp16, x = obj_157_cast_fp16)[name = tensor("input_391_cast_fp16")]; - tensor var_5355 = const()[name = tensor("op_5355"), val = tensor([1, 1])]; - tensor var_5357 = const()[name = tensor("op_5357"), val = tensor([1, 1])]; - tensor lora_out_521_pad_type_0 = const()[name = tensor("lora_out_521_pad_type_0"), val = tensor("custom")]; - tensor lora_out_521_pad_0 = const()[name = tensor("lora_out_521_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_523_weight_0_to_fp16 = const()[name = tensor("lora_out_523_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321300864)))]; - tensor lora_out_523_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5357, groups = var_5305, pad = lora_out_521_pad_0, pad_type = lora_out_521_pad_type_0, strides = var_5355, weight = lora_out_523_weight_0_to_fp16, x = input_391_cast_fp16)[name = tensor("lora_out_523_cast_fp16")]; - tensor query_53_cast_fp16 = add(x = pretrained_out_261_cast_fp16, y = lora_out_523_cast_fp16)[name = tensor("query_53_cast_fp16")]; - tensor var_5367 = const()[name = tensor("op_5367"), val = tensor([1, 1])]; - tensor var_5369 = const()[name = tensor("op_5369"), val = tensor([1, 1])]; - tensor pretrained_out_263_pad_type_0 = const()[name = tensor("pretrained_out_263_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_263_pad_0 = const()[name = tensor("pretrained_out_263_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321341888))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322161152))), name = tensor("layers_13_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_263_cast_fp16 = conv(dilations = var_5369, groups = var_5305, pad = pretrained_out_263_pad_0, pad_type = pretrained_out_263_pad_type_0, strides = var_5367, weight = layers_13_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_157_cast_fp16)[name = tensor("pretrained_out_263_cast_fp16")]; - tensor var_5373 = const()[name = tensor("op_5373"), val = tensor([1, 1])]; - tensor var_5375 = const()[name = tensor("op_5375"), val = tensor([1, 1])]; - tensor input_393_pad_type_0 = const()[name = tensor("input_393_pad_type_0"), val = tensor("custom")]; - tensor input_393_pad_0 = const()[name = tensor("input_393_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322161280)))]; - tensor input_393_cast_fp16 = conv(dilations = var_5375, groups = var_5305, pad = input_393_pad_0, pad_type = input_393_pad_type_0, strides = var_5373, weight = layers_13_self_attn_k_proj_loraA_weight_to_fp16, x = obj_157_cast_fp16)[name = tensor("input_393_cast_fp16")]; - tensor var_5379 = const()[name = tensor("op_5379"), val = tensor([1, 1])]; - tensor var_5381 = const()[name = tensor("op_5381"), val = tensor([1, 1])]; - tensor lora_out_525_pad_type_0 = const()[name = tensor("lora_out_525_pad_type_0"), val = tensor("custom")]; - tensor lora_out_525_pad_0 = const()[name = tensor("lora_out_525_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_527_weight_0_to_fp16 = const()[name = tensor("lora_out_527_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322202304)))]; - tensor lora_out_527_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5381, groups = var_5305, pad = lora_out_525_pad_0, pad_type = lora_out_525_pad_type_0, strides = var_5379, weight = lora_out_527_weight_0_to_fp16, x = input_393_cast_fp16)[name = tensor("lora_out_527_cast_fp16")]; - tensor current_key_27_cast_fp16 = add(x = pretrained_out_263_cast_fp16, y = lora_out_527_cast_fp16)[name = tensor("current_key_27_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 pretrained_out_265_pad_type_0 = const()[name = tensor("pretrained_out_265_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_265_pad_0 = const()[name = tensor("pretrained_out_265_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322243328))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323062592))), name = tensor("layers_13_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_13_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323062720)))]; - tensor pretrained_out_265_cast_fp16 = conv(bias = layers_13_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_5394, groups = var_5305, pad = pretrained_out_265_pad_0, pad_type = pretrained_out_265_pad_type_0, strides = var_5392, weight = layers_13_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_157_cast_fp16)[name = tensor("pretrained_out_265_cast_fp16")]; - tensor var_5398 = const()[name = tensor("op_5398"), val = tensor([1, 1])]; - tensor var_5400 = const()[name = tensor("op_5400"), val = tensor([1, 1])]; - tensor input_395_pad_type_0 = const()[name = tensor("input_395_pad_type_0"), val = tensor("custom")]; - tensor input_395_pad_0 = const()[name = tensor("input_395_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323065344)))]; - tensor input_395_cast_fp16 = conv(dilations = var_5400, groups = var_5305, pad = input_395_pad_0, pad_type = input_395_pad_type_0, strides = var_5398, weight = layers_13_self_attn_v_proj_loraA_weight_to_fp16, x = obj_157_cast_fp16)[name = tensor("input_395_cast_fp16")]; - tensor var_5404 = const()[name = tensor("op_5404"), val = tensor([1, 1])]; - tensor var_5406 = const()[name = tensor("op_5406"), val = tensor([1, 1])]; - tensor lora_out_529_pad_type_0 = const()[name = tensor("lora_out_529_pad_type_0"), val = tensor("custom")]; - tensor lora_out_529_pad_0 = const()[name = tensor("lora_out_529_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_531_weight_0_to_fp16 = const()[name = tensor("lora_out_531_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323106368)))]; - tensor lora_out_531_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5406, groups = var_5305, pad = lora_out_529_pad_0, pad_type = lora_out_529_pad_type_0, strides = var_5404, weight = lora_out_531_weight_0_to_fp16, x = input_395_cast_fp16)[name = tensor("lora_out_531_cast_fp16")]; - tensor current_value_27_cast_fp16 = add(x = pretrained_out_265_cast_fp16, y = lora_out_531_cast_fp16)[name = tensor("current_value_27_cast_fp16")]; - tensor var_5416_cast_fp16 = mul(x = current_key_27_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_5416_cast_fp16")]; - tensor var_5418_cast_fp16 = mul(x = var_103_cast_fp16_13, y = var_295_cast_fp16)[name = tensor("op_5418_cast_fp16")]; - tensor key_53_cast_fp16 = add(x = var_5416_cast_fp16, y = var_5418_cast_fp16)[name = tensor("key_53_cast_fp16")]; - tensor var_5420_cast_fp16 = mul(x = current_value_27_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_5420_cast_fp16")]; - tensor var_5422_cast_fp16 = mul(x = var_138_cast_fp16_13, y = var_295_cast_fp16)[name = tensor("op_5422_cast_fp16")]; - tensor value_53_cast_fp16 = add(x = var_5420_cast_fp16, y = var_5422_cast_fp16)[name = tensor("value_53_cast_fp16")]; - tensor var_5425 = const()[name = tensor("op_5425"), val = tensor([1, 20, 64, -1])]; - tensor var_5426_cast_fp16 = reshape(shape = var_5425, x = query_53_cast_fp16)[name = tensor("op_5426_cast_fp16")]; - tensor var_5427_to_fp16 = const()[name = tensor("op_5427_to_fp16"), val = tensor(0x1p-3)]; - tensor var_5428_cast_fp16 = mul(x = var_5426_cast_fp16, y = var_5427_to_fp16)[name = tensor("op_5428_cast_fp16")]; - tensor var_5429 = const()[name = tensor("op_5429"), val = tensor([1, 20, 64, -1])]; - tensor var_5430_cast_fp16 = reshape(shape = var_5429, x = key_53_cast_fp16)[name = tensor("op_5430_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_5428_cast_fp16, y = var_5430_cast_fp16)[name = tensor("mh_w_79_cast_fp16")]; - tensor mh_w_81_cast_fp16 = add(x = mh_w_79_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_81_cast_fp16")]; - tensor var_5438_cast_fp16 = softmax(axis = var_5298, x = mh_w_81_cast_fp16)[name = tensor("op_5438_cast_fp16")]; - tensor var_5439 = const()[name = tensor("op_5439"), val = tensor([1, 20, 64, -1])]; - tensor var_5440_cast_fp16 = reshape(shape = var_5439, x = value_53_cast_fp16)[name = tensor("op_5440_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_5440_cast_fp16, y = var_5438_cast_fp16)[name = tensor("attn_53_cast_fp16")]; - tensor var_5443 = const()[name = tensor("op_5443"), val = tensor([1, 1280, 1, -1])]; - tensor input_397_cast_fp16 = reshape(shape = var_5443, x = attn_53_cast_fp16)[name = tensor("input_397_cast_fp16")]; - tensor var_5450 = const()[name = tensor("op_5450"), val = tensor([1, 1])]; - tensor var_5452 = const()[name = tensor("op_5452"), val = tensor([1, 1])]; - tensor pretrained_out_267_pad_type_0 = const()[name = tensor("pretrained_out_267_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_267_pad_0 = const()[name = tensor("pretrained_out_267_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323147392))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323966656))), name = tensor("layers_13_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_13_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323966784)))]; - tensor pretrained_out_267_cast_fp16 = conv(bias = layers_13_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_5452, groups = var_5305, pad = pretrained_out_267_pad_0, pad_type = pretrained_out_267_pad_type_0, strides = var_5450, weight = layers_13_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = tensor("pretrained_out_267_cast_fp16")]; - tensor var_5456 = const()[name = tensor("op_5456"), val = tensor([1, 1])]; - tensor var_5458 = const()[name = tensor("op_5458"), val = tensor([1, 1])]; - tensor input_399_pad_type_0 = const()[name = tensor("input_399_pad_type_0"), val = tensor("custom")]; - tensor input_399_pad_0 = const()[name = tensor("input_399_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323969408)))]; - tensor input_399_cast_fp16 = conv(dilations = var_5458, groups = var_5305, pad = input_399_pad_0, pad_type = input_399_pad_type_0, strides = var_5456, weight = layers_13_self_attn_o_proj_loraA_weight_to_fp16, x = input_397_cast_fp16)[name = tensor("input_399_cast_fp16")]; - tensor var_5462 = const()[name = tensor("op_5462"), val = tensor([1, 1])]; - tensor var_5464 = const()[name = tensor("op_5464"), val = tensor([1, 1])]; - tensor lora_out_533_pad_type_0 = const()[name = tensor("lora_out_533_pad_type_0"), val = tensor("custom")]; - tensor lora_out_533_pad_0 = const()[name = tensor("lora_out_533_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_535_weight_0_to_fp16 = const()[name = tensor("lora_out_535_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324010432)))]; - tensor lora_out_535_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5464, groups = var_5305, pad = lora_out_533_pad_0, pad_type = lora_out_533_pad_type_0, strides = var_5462, weight = lora_out_535_weight_0_to_fp16, x = input_399_cast_fp16)[name = tensor("lora_out_535_cast_fp16")]; - tensor obj_163_cast_fp16 = add(x = pretrained_out_267_cast_fp16, y = lora_out_535_cast_fp16)[name = tensor("obj_163_cast_fp16")]; - tensor inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = obj_163_cast_fp16)[name = tensor("inputs_81_cast_fp16")]; - tensor var_5477 = const()[name = tensor("op_5477"), val = tensor([1])]; - tensor channels_mean_81_cast_fp16 = reduce_mean(axes = var_5477, keep_dims = var_5306, 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_5481 = const()[name = tensor("op_5481"), val = tensor([1])]; - tensor var_5482_cast_fp16 = reduce_mean(axes = var_5481, keep_dims = var_5306, x = zero_mean_sq_81_cast_fp16)[name = tensor("op_5482_cast_fp16")]; - tensor var_5483_to_fp16 = const()[name = tensor("op_5483_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_5484_cast_fp16 = add(x = var_5482_cast_fp16, y = var_5483_to_fp16)[name = tensor("op_5484_cast_fp16")]; - tensor denom_81_epsilon_0 = const()[name = tensor("denom_81_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_81_cast_fp16 = rsqrt(epsilon = denom_81_epsilon_0, x = var_5484_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_165_gamma_0_to_fp16 = const()[name = tensor("obj_165_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324051456)))]; - tensor obj_165_beta_0_to_fp16 = const()[name = tensor("obj_165_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324054080)))]; - tensor obj_165_epsilon_0_to_fp16 = const()[name = tensor("obj_165_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_165_cast_fp16 = batch_norm(beta = obj_165_beta_0_to_fp16, epsilon = obj_165_epsilon_0_to_fp16, gamma = obj_165_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_165_cast_fp16")]; - tensor var_5502 = const()[name = tensor("op_5502"), val = tensor([1, 1])]; - tensor var_5504 = const()[name = tensor("op_5504"), val = tensor([1, 1])]; - tensor pretrained_out_269_pad_type_0 = const()[name = tensor("pretrained_out_269_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_269_pad_0 = const()[name = tensor("pretrained_out_269_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324056704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324875968))), name = tensor("layers_13_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_13_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324876096)))]; - tensor pretrained_out_269_cast_fp16 = conv(bias = layers_13_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_5504, groups = var_5305, pad = pretrained_out_269_pad_0, pad_type = pretrained_out_269_pad_type_0, strides = var_5502, weight = layers_13_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_165_cast_fp16)[name = tensor("pretrained_out_269_cast_fp16")]; - tensor var_5508 = const()[name = tensor("op_5508"), val = tensor([1, 1])]; - tensor var_5510 = const()[name = tensor("op_5510"), val = tensor([1, 1])]; - tensor input_401_pad_type_0 = const()[name = tensor("input_401_pad_type_0"), val = tensor("custom")]; - tensor input_401_pad_0 = const()[name = tensor("input_401_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324878720)))]; - tensor input_401_cast_fp16 = conv(dilations = var_5510, groups = var_5305, pad = input_401_pad_0, pad_type = input_401_pad_type_0, strides = var_5508, weight = layers_13_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_165_cast_fp16)[name = tensor("input_401_cast_fp16")]; - tensor var_5514 = const()[name = tensor("op_5514"), val = tensor([1, 1])]; - tensor var_5516 = const()[name = tensor("op_5516"), val = tensor([1, 1])]; - tensor lora_out_537_pad_type_0 = const()[name = tensor("lora_out_537_pad_type_0"), val = tensor("custom")]; - tensor lora_out_537_pad_0 = const()[name = tensor("lora_out_537_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_539_weight_0_to_fp16 = const()[name = tensor("lora_out_539_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324919744)))]; - tensor lora_out_539_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5516, groups = var_5305, pad = lora_out_537_pad_0, pad_type = lora_out_537_pad_type_0, strides = var_5514, weight = lora_out_539_weight_0_to_fp16, x = input_401_cast_fp16)[name = tensor("lora_out_539_cast_fp16")]; - tensor query_55_cast_fp16 = add(x = pretrained_out_269_cast_fp16, y = lora_out_539_cast_fp16)[name = tensor("query_55_cast_fp16")]; - tensor var_5526 = const()[name = tensor("op_5526"), val = tensor([1, 1])]; - tensor var_5528 = const()[name = tensor("op_5528"), val = tensor([1, 1])]; - tensor pretrained_out_271_pad_type_0 = const()[name = tensor("pretrained_out_271_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_271_pad_0 = const()[name = tensor("pretrained_out_271_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324960768))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325780032))), name = tensor("layers_13_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_271_cast_fp16 = conv(dilations = var_5528, groups = var_5305, pad = pretrained_out_271_pad_0, pad_type = pretrained_out_271_pad_type_0, strides = var_5526, weight = layers_13_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_271_cast_fp16")]; - tensor var_5532 = const()[name = tensor("op_5532"), val = tensor([1, 1])]; - tensor var_5534 = const()[name = tensor("op_5534"), val = tensor([1, 1])]; - tensor input_403_pad_type_0 = const()[name = tensor("input_403_pad_type_0"), val = tensor("custom")]; - tensor input_403_pad_0 = const()[name = tensor("input_403_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325780160)))]; - tensor input_403_cast_fp16 = conv(dilations = var_5534, groups = var_5305, pad = input_403_pad_0, pad_type = input_403_pad_type_0, strides = var_5532, weight = layers_13_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_403_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 lora_out_541_pad_type_0 = const()[name = tensor("lora_out_541_pad_type_0"), val = tensor("custom")]; - tensor lora_out_541_pad_0 = const()[name = tensor("lora_out_541_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_543_weight_0_to_fp16 = const()[name = tensor("lora_out_543_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325821184)))]; - tensor lora_out_543_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5540, groups = var_5305, pad = lora_out_541_pad_0, pad_type = lora_out_541_pad_type_0, strides = var_5538, weight = lora_out_543_weight_0_to_fp16, x = input_403_cast_fp16)[name = tensor("lora_out_543_cast_fp16")]; - tensor key_55_cast_fp16 = add(x = pretrained_out_271_cast_fp16, y = lora_out_543_cast_fp16)[name = tensor("key_55_cast_fp16")]; - tensor var_5551 = const()[name = tensor("op_5551"), val = tensor([1, 1])]; - tensor var_5553 = const()[name = tensor("op_5553"), val = tensor([1, 1])]; - tensor pretrained_out_273_pad_type_0 = const()[name = tensor("pretrained_out_273_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_273_pad_0 = const()[name = tensor("pretrained_out_273_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325862208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(326681472))), name = tensor("layers_13_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_13_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(326681600)))]; - tensor pretrained_out_273_cast_fp16 = conv(bias = layers_13_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_5553, groups = var_5305, pad = pretrained_out_273_pad_0, pad_type = pretrained_out_273_pad_type_0, strides = var_5551, weight = layers_13_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_273_cast_fp16")]; - tensor var_5557 = const()[name = tensor("op_5557"), val = tensor([1, 1])]; - tensor var_5559 = const()[name = tensor("op_5559"), val = tensor([1, 1])]; - tensor input_405_pad_type_0 = const()[name = tensor("input_405_pad_type_0"), val = tensor("custom")]; - tensor input_405_pad_0 = const()[name = tensor("input_405_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(326684224)))]; - tensor input_405_cast_fp16 = conv(dilations = var_5559, groups = var_5305, pad = input_405_pad_0, pad_type = input_405_pad_type_0, strides = var_5557, weight = layers_13_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_405_cast_fp16")]; - tensor var_5563 = const()[name = tensor("op_5563"), val = tensor([1, 1])]; - tensor var_5565 = const()[name = tensor("op_5565"), val = tensor([1, 1])]; - tensor lora_out_545_pad_type_0 = const()[name = tensor("lora_out_545_pad_type_0"), val = tensor("custom")]; - tensor lora_out_545_pad_0 = const()[name = tensor("lora_out_545_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_547_weight_0_to_fp16 = const()[name = tensor("lora_out_547_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(326725248)))]; - tensor lora_out_547_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5565, groups = var_5305, pad = lora_out_545_pad_0, pad_type = lora_out_545_pad_type_0, strides = var_5563, weight = lora_out_547_weight_0_to_fp16, x = input_405_cast_fp16)[name = tensor("lora_out_547_cast_fp16")]; - tensor value_55_cast_fp16 = add(x = pretrained_out_273_cast_fp16, y = lora_out_547_cast_fp16)[name = tensor("value_55_cast_fp16")]; - tensor var_5572 = const()[name = tensor("op_5572"), val = tensor([1, 20, 64, -1])]; - tensor var_5573_cast_fp16 = reshape(shape = var_5572, x = query_55_cast_fp16)[name = tensor("op_5573_cast_fp16")]; - tensor var_5574_to_fp16 = const()[name = tensor("op_5574_to_fp16"), val = tensor(0x1p-3)]; - tensor var_5575_cast_fp16 = mul(x = var_5573_cast_fp16, y = var_5574_to_fp16)[name = tensor("op_5575_cast_fp16")]; - tensor var_5576 = const()[name = tensor("op_5576"), val = tensor([1, 20, 64, -1])]; - tensor var_5577_cast_fp16 = reshape(shape = var_5576, x = key_55_cast_fp16)[name = tensor("op_5577_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_5575_cast_fp16, y = var_5577_cast_fp16)[name = tensor("mh_w_83_cast_fp16")]; - tensor var_5580_cast_fp16 = softmax(axis = var_5298, x = mh_w_83_cast_fp16)[name = tensor("op_5580_cast_fp16")]; - tensor var_5581 = const()[name = tensor("op_5581"), val = tensor([1, 20, 64, -1])]; - tensor var_5582_cast_fp16 = reshape(shape = var_5581, x = value_55_cast_fp16)[name = tensor("op_5582_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_5582_cast_fp16, y = var_5580_cast_fp16)[name = tensor("attn_55_cast_fp16")]; - tensor var_5585 = const()[name = tensor("op_5585"), val = tensor([1, 1280, 1, -1])]; - tensor input_407_cast_fp16 = reshape(shape = var_5585, x = attn_55_cast_fp16)[name = tensor("input_407_cast_fp16")]; - tensor var_5592 = const()[name = tensor("op_5592"), val = tensor([1, 1])]; - tensor var_5594 = const()[name = tensor("op_5594"), val = tensor([1, 1])]; - tensor pretrained_out_275_pad_type_0 = const()[name = tensor("pretrained_out_275_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_275_pad_0 = const()[name = tensor("pretrained_out_275_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(326766272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327585536))), name = tensor("layers_13_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_13_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327585664)))]; - tensor pretrained_out_275_cast_fp16 = conv(bias = layers_13_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_5594, groups = var_5305, pad = pretrained_out_275_pad_0, pad_type = pretrained_out_275_pad_type_0, strides = var_5592, weight = layers_13_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = tensor("pretrained_out_275_cast_fp16")]; - tensor var_5598 = const()[name = tensor("op_5598"), val = tensor([1, 1])]; - tensor var_5600 = const()[name = tensor("op_5600"), val = tensor([1, 1])]; - tensor input_409_pad_type_0 = const()[name = tensor("input_409_pad_type_0"), val = tensor("custom")]; - tensor input_409_pad_0 = const()[name = tensor("input_409_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327588288)))]; - tensor input_409_cast_fp16 = conv(dilations = var_5600, groups = var_5305, pad = input_409_pad_0, pad_type = input_409_pad_type_0, strides = var_5598, weight = layers_13_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_407_cast_fp16)[name = tensor("input_409_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 lora_out_549_pad_type_0 = const()[name = tensor("lora_out_549_pad_type_0"), val = tensor("custom")]; - tensor lora_out_549_pad_0 = const()[name = tensor("lora_out_549_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_551_weight_0_to_fp16 = const()[name = tensor("lora_out_551_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327629312)))]; - tensor lora_out_551_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5606, groups = var_5305, pad = lora_out_549_pad_0, pad_type = lora_out_549_pad_type_0, strides = var_5604, weight = lora_out_551_weight_0_to_fp16, x = input_409_cast_fp16)[name = tensor("lora_out_551_cast_fp16")]; - tensor obj_167_cast_fp16 = add(x = pretrained_out_275_cast_fp16, y = lora_out_551_cast_fp16)[name = tensor("obj_167_cast_fp16")]; - tensor inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = obj_167_cast_fp16)[name = tensor("inputs_83_cast_fp16")]; - tensor var_5615 = const()[name = tensor("op_5615"), val = tensor([1])]; - tensor channels_mean_83_cast_fp16 = reduce_mean(axes = var_5615, keep_dims = var_5306, 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_5619 = const()[name = tensor("op_5619"), val = tensor([1])]; - tensor var_5620_cast_fp16 = reduce_mean(axes = var_5619, keep_dims = var_5306, x = zero_mean_sq_83_cast_fp16)[name = tensor("op_5620_cast_fp16")]; - tensor var_5621_to_fp16 = const()[name = tensor("op_5621_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_5622_cast_fp16 = add(x = var_5620_cast_fp16, y = var_5621_to_fp16)[name = tensor("op_5622_cast_fp16")]; - tensor denom_83_epsilon_0 = const()[name = tensor("denom_83_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_83_cast_fp16 = rsqrt(epsilon = denom_83_epsilon_0, x = var_5622_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_411_gamma_0_to_fp16 = const()[name = tensor("input_411_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327670336)))]; - tensor input_411_beta_0_to_fp16 = const()[name = tensor("input_411_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327672960)))]; - tensor input_411_epsilon_0_to_fp16 = const()[name = tensor("input_411_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_411_cast_fp16 = batch_norm(beta = input_411_beta_0_to_fp16, epsilon = input_411_epsilon_0_to_fp16, gamma = input_411_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_411_cast_fp16")]; - tensor var_5636 = const()[name = tensor("op_5636"), val = tensor([1, 1])]; - tensor var_5638 = const()[name = tensor("op_5638"), val = tensor([1, 1])]; - tensor pretrained_out_277_pad_type_0 = const()[name = tensor("pretrained_out_277_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_277_pad_0 = const()[name = tensor("pretrained_out_277_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327675584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(330952448))), name = tensor("layers_13_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_13_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_13_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(330952576)))]; - tensor pretrained_out_277_cast_fp16 = conv(bias = layers_13_fc1_pretrained_bias_to_fp16, dilations = var_5638, groups = var_5305, pad = pretrained_out_277_pad_0, pad_type = pretrained_out_277_pad_type_0, strides = var_5636, weight = layers_13_fc1_pretrained_weight_to_fp16_palettized, x = input_411_cast_fp16)[name = tensor("pretrained_out_277_cast_fp16")]; - tensor var_5642 = const()[name = tensor("op_5642"), val = tensor([1, 1])]; - tensor var_5644 = const()[name = tensor("op_5644"), val = tensor([1, 1])]; - tensor input_413_pad_type_0 = const()[name = tensor("input_413_pad_type_0"), val = tensor("custom")]; - tensor input_413_pad_0 = const()[name = tensor("input_413_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_13_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(330962880)))]; - tensor input_413_cast_fp16 = conv(dilations = var_5644, groups = var_5305, pad = input_413_pad_0, pad_type = input_413_pad_type_0, strides = var_5642, weight = layers_13_fc1_loraA_weight_to_fp16, x = input_411_cast_fp16)[name = tensor("input_413_cast_fp16")]; - tensor var_5648 = const()[name = tensor("op_5648"), val = tensor([1, 1])]; - tensor var_5650 = const()[name = tensor("op_5650"), val = tensor([1, 1])]; - tensor lora_out_553_pad_type_0 = const()[name = tensor("lora_out_553_pad_type_0"), val = tensor("custom")]; - tensor lora_out_553_pad_0 = const()[name = tensor("lora_out_553_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_555_weight_0_to_fp16 = const()[name = tensor("lora_out_555_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(331003904)))]; - tensor lora_out_555_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_5650, groups = var_5305, pad = lora_out_553_pad_0, pad_type = lora_out_553_pad_type_0, strides = var_5648, weight = lora_out_555_weight_0_to_fp16, x = input_413_cast_fp16)[name = tensor("lora_out_555_cast_fp16")]; - tensor input_415_cast_fp16 = add(x = pretrained_out_277_cast_fp16, y = lora_out_555_cast_fp16)[name = tensor("input_415_cast_fp16")]; - tensor input_417_mode_0 = const()[name = tensor("input_417_mode_0"), val = tensor("EXACT")]; - tensor input_417_cast_fp16 = gelu(mode = input_417_mode_0, x = input_415_cast_fp16)[name = tensor("input_417_cast_fp16")]; - tensor var_5662 = const()[name = tensor("op_5662"), val = tensor([1, 1])]; - tensor var_5664 = const()[name = tensor("op_5664"), val = tensor([1, 1])]; - tensor pretrained_out_279_pad_type_0 = const()[name = tensor("pretrained_out_279_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_279_pad_0 = const()[name = tensor("pretrained_out_279_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(331167808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334444672))), name = tensor("layers_13_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_13_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_13_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334444800)))]; - tensor pretrained_out_279_cast_fp16 = conv(bias = layers_13_fc2_pretrained_bias_to_fp16, dilations = var_5664, groups = var_5305, pad = pretrained_out_279_pad_0, pad_type = pretrained_out_279_pad_type_0, strides = var_5662, weight = layers_13_fc2_pretrained_weight_to_fp16_palettized, x = input_417_cast_fp16)[name = tensor("pretrained_out_279_cast_fp16")]; - tensor var_5668 = const()[name = tensor("op_5668"), val = tensor([1, 1])]; - tensor var_5670 = const()[name = tensor("op_5670"), val = tensor([1, 1])]; - tensor input_419_pad_type_0 = const()[name = tensor("input_419_pad_type_0"), val = tensor("custom")]; - tensor input_419_pad_0 = const()[name = tensor("input_419_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_13_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_13_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334447424)))]; - tensor input_419_cast_fp16 = conv(dilations = var_5670, groups = var_5305, pad = input_419_pad_0, pad_type = input_419_pad_type_0, strides = var_5668, weight = layers_13_fc2_loraA_weight_to_fp16, x = input_417_cast_fp16)[name = tensor("input_419_cast_fp16")]; - tensor var_5674 = const()[name = tensor("op_5674"), val = tensor([1, 1])]; - tensor var_5676 = const()[name = tensor("op_5676"), val = tensor([1, 1])]; - tensor lora_out_557_pad_type_0 = const()[name = tensor("lora_out_557_pad_type_0"), val = tensor("custom")]; - tensor lora_out_557_pad_0 = const()[name = tensor("lora_out_557_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_559_weight_0_to_fp16 = const()[name = tensor("lora_out_559_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334611328)))]; - tensor lora_out_559_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5676, groups = var_5305, pad = lora_out_557_pad_0, pad_type = lora_out_557_pad_type_0, strides = var_5674, weight = lora_out_559_weight_0_to_fp16, x = input_419_cast_fp16)[name = tensor("lora_out_559_cast_fp16")]; - tensor hidden_states_29_cast_fp16 = add(x = pretrained_out_279_cast_fp16, y = lora_out_559_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_5692 = const()[name = tensor("op_5692"), val = tensor(3)]; - tensor var_5699 = const()[name = tensor("op_5699"), val = tensor(1)]; - tensor var_5700 = const()[name = tensor("op_5700"), val = tensor(true)]; - tensor var_5712 = const()[name = tensor("op_5712"), val = tensor([1])]; - tensor channels_mean_85_cast_fp16 = reduce_mean(axes = var_5712, keep_dims = var_5700, 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_5716 = const()[name = tensor("op_5716"), val = tensor([1])]; - tensor var_5717_cast_fp16 = reduce_mean(axes = var_5716, keep_dims = var_5700, x = zero_mean_sq_85_cast_fp16)[name = tensor("op_5717_cast_fp16")]; - tensor var_5718_to_fp16 = const()[name = tensor("op_5718_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_5719_cast_fp16 = add(x = var_5717_cast_fp16, y = var_5718_to_fp16)[name = tensor("op_5719_cast_fp16")]; - tensor denom_85_epsilon_0 = const()[name = tensor("denom_85_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_85_cast_fp16 = rsqrt(epsilon = denom_85_epsilon_0, x = var_5719_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_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(334652352)))]; - 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(334654976)))]; - 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_85_cast_fp16)[name = tensor("obj_169_cast_fp16")]; - tensor var_5737 = const()[name = tensor("op_5737"), val = tensor([1, 1])]; - tensor var_5739 = const()[name = tensor("op_5739"), val = tensor([1, 1])]; - tensor pretrained_out_281_pad_type_0 = const()[name = tensor("pretrained_out_281_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_281_pad_0 = const()[name = tensor("pretrained_out_281_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334657600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335476864))), name = tensor("layers_14_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_14_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335476992)))]; - tensor pretrained_out_281_cast_fp16 = conv(bias = layers_14_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_5739, groups = var_5699, pad = pretrained_out_281_pad_0, pad_type = pretrained_out_281_pad_type_0, strides = var_5737, weight = layers_14_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_169_cast_fp16)[name = tensor("pretrained_out_281_cast_fp16")]; - tensor var_5743 = const()[name = tensor("op_5743"), val = tensor([1, 1])]; - tensor var_5745 = const()[name = tensor("op_5745"), val = tensor([1, 1])]; - tensor input_421_pad_type_0 = const()[name = tensor("input_421_pad_type_0"), val = tensor("custom")]; - tensor input_421_pad_0 = const()[name = tensor("input_421_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335479616)))]; - tensor input_421_cast_fp16 = conv(dilations = var_5745, groups = var_5699, pad = input_421_pad_0, pad_type = input_421_pad_type_0, strides = var_5743, weight = layers_14_self_attn_q_proj_loraA_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("input_421_cast_fp16")]; - tensor var_5749 = const()[name = tensor("op_5749"), val = tensor([1, 1])]; - tensor var_5751 = const()[name = tensor("op_5751"), val = tensor([1, 1])]; - tensor lora_out_561_pad_type_0 = const()[name = tensor("lora_out_561_pad_type_0"), val = tensor("custom")]; - tensor lora_out_561_pad_0 = const()[name = tensor("lora_out_561_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_563_weight_0_to_fp16 = const()[name = tensor("lora_out_563_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335520640)))]; - tensor lora_out_563_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5751, groups = var_5699, pad = lora_out_561_pad_0, pad_type = lora_out_561_pad_type_0, strides = var_5749, weight = lora_out_563_weight_0_to_fp16, x = input_421_cast_fp16)[name = tensor("lora_out_563_cast_fp16")]; - tensor query_57_cast_fp16 = add(x = pretrained_out_281_cast_fp16, y = lora_out_563_cast_fp16)[name = tensor("query_57_cast_fp16")]; - tensor var_5761 = const()[name = tensor("op_5761"), val = tensor([1, 1])]; - tensor var_5763 = const()[name = tensor("op_5763"), val = tensor([1, 1])]; - tensor pretrained_out_283_pad_type_0 = const()[name = tensor("pretrained_out_283_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_283_pad_0 = const()[name = tensor("pretrained_out_283_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335561664))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336380928))), name = tensor("layers_14_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_283_cast_fp16 = conv(dilations = var_5763, groups = var_5699, pad = pretrained_out_283_pad_0, pad_type = pretrained_out_283_pad_type_0, strides = var_5761, weight = layers_14_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_169_cast_fp16)[name = tensor("pretrained_out_283_cast_fp16")]; - tensor var_5767 = const()[name = tensor("op_5767"), val = tensor([1, 1])]; - tensor var_5769 = const()[name = tensor("op_5769"), val = tensor([1, 1])]; - tensor input_423_pad_type_0 = const()[name = tensor("input_423_pad_type_0"), val = tensor("custom")]; - tensor input_423_pad_0 = const()[name = tensor("input_423_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336381056)))]; - tensor input_423_cast_fp16 = conv(dilations = var_5769, groups = var_5699, pad = input_423_pad_0, pad_type = input_423_pad_type_0, strides = var_5767, weight = layers_14_self_attn_k_proj_loraA_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("input_423_cast_fp16")]; - tensor var_5773 = const()[name = tensor("op_5773"), val = tensor([1, 1])]; - tensor var_5775 = const()[name = tensor("op_5775"), val = tensor([1, 1])]; - tensor lora_out_565_pad_type_0 = const()[name = tensor("lora_out_565_pad_type_0"), val = tensor("custom")]; - tensor lora_out_565_pad_0 = const()[name = tensor("lora_out_565_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_567_weight_0_to_fp16 = const()[name = tensor("lora_out_567_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336422080)))]; - tensor lora_out_567_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5775, groups = var_5699, pad = lora_out_565_pad_0, pad_type = lora_out_565_pad_type_0, strides = var_5773, weight = lora_out_567_weight_0_to_fp16, x = input_423_cast_fp16)[name = tensor("lora_out_567_cast_fp16")]; - tensor current_key_29_cast_fp16 = add(x = pretrained_out_283_cast_fp16, y = lora_out_567_cast_fp16)[name = tensor("current_key_29_cast_fp16")]; - tensor var_5786 = const()[name = tensor("op_5786"), val = tensor([1, 1])]; - tensor var_5788 = const()[name = tensor("op_5788"), val = tensor([1, 1])]; - tensor pretrained_out_285_pad_type_0 = const()[name = tensor("pretrained_out_285_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_285_pad_0 = const()[name = tensor("pretrained_out_285_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336463104))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337282368))), name = tensor("layers_14_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_14_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337282496)))]; - tensor pretrained_out_285_cast_fp16 = conv(bias = layers_14_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_5788, groups = var_5699, pad = pretrained_out_285_pad_0, pad_type = pretrained_out_285_pad_type_0, strides = var_5786, weight = layers_14_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_169_cast_fp16)[name = tensor("pretrained_out_285_cast_fp16")]; - tensor var_5792 = const()[name = tensor("op_5792"), val = tensor([1, 1])]; - tensor var_5794 = const()[name = tensor("op_5794"), val = tensor([1, 1])]; - tensor input_425_pad_type_0 = const()[name = tensor("input_425_pad_type_0"), val = tensor("custom")]; - tensor input_425_pad_0 = const()[name = tensor("input_425_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337285120)))]; - tensor input_425_cast_fp16 = conv(dilations = var_5794, groups = var_5699, pad = input_425_pad_0, pad_type = input_425_pad_type_0, strides = var_5792, weight = layers_14_self_attn_v_proj_loraA_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("input_425_cast_fp16")]; - tensor var_5798 = const()[name = tensor("op_5798"), val = tensor([1, 1])]; - tensor var_5800 = const()[name = tensor("op_5800"), val = tensor([1, 1])]; - tensor lora_out_569_pad_type_0 = const()[name = tensor("lora_out_569_pad_type_0"), val = tensor("custom")]; - tensor lora_out_569_pad_0 = const()[name = tensor("lora_out_569_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_571_weight_0_to_fp16 = const()[name = tensor("lora_out_571_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337326144)))]; - tensor lora_out_571_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5800, groups = var_5699, pad = lora_out_569_pad_0, pad_type = lora_out_569_pad_type_0, strides = var_5798, weight = lora_out_571_weight_0_to_fp16, x = input_425_cast_fp16)[name = tensor("lora_out_571_cast_fp16")]; - tensor current_value_29_cast_fp16 = add(x = pretrained_out_285_cast_fp16, y = lora_out_571_cast_fp16)[name = tensor("current_value_29_cast_fp16")]; - tensor var_5810_cast_fp16 = mul(x = current_key_29_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_5810_cast_fp16")]; - tensor var_5812_cast_fp16 = mul(x = var_103_cast_fp16_14, y = var_295_cast_fp16)[name = tensor("op_5812_cast_fp16")]; - tensor key_57_cast_fp16 = add(x = var_5810_cast_fp16, y = var_5812_cast_fp16)[name = tensor("key_57_cast_fp16")]; - tensor var_5814_cast_fp16 = mul(x = current_value_29_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_5814_cast_fp16")]; - tensor var_5816_cast_fp16 = mul(x = var_138_cast_fp16_14, y = var_295_cast_fp16)[name = tensor("op_5816_cast_fp16")]; - tensor value_57_cast_fp16 = add(x = var_5814_cast_fp16, y = var_5816_cast_fp16)[name = tensor("value_57_cast_fp16")]; - tensor var_5819 = const()[name = tensor("op_5819"), val = tensor([1, 20, 64, -1])]; - tensor var_5820_cast_fp16 = reshape(shape = var_5819, x = query_57_cast_fp16)[name = tensor("op_5820_cast_fp16")]; - tensor var_5821_to_fp16 = const()[name = tensor("op_5821_to_fp16"), val = tensor(0x1p-3)]; - tensor var_5822_cast_fp16 = mul(x = var_5820_cast_fp16, y = var_5821_to_fp16)[name = tensor("op_5822_cast_fp16")]; - tensor var_5823 = const()[name = tensor("op_5823"), val = tensor([1, 20, 64, -1])]; - tensor var_5824_cast_fp16 = reshape(shape = var_5823, x = key_57_cast_fp16)[name = tensor("op_5824_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_5822_cast_fp16, y = var_5824_cast_fp16)[name = tensor("mh_w_85_cast_fp16")]; - tensor mh_w_87_cast_fp16 = add(x = mh_w_85_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_87_cast_fp16")]; - tensor var_5832_cast_fp16 = softmax(axis = var_5692, x = mh_w_87_cast_fp16)[name = tensor("op_5832_cast_fp16")]; - tensor var_5833 = const()[name = tensor("op_5833"), val = tensor([1, 20, 64, -1])]; - tensor var_5834_cast_fp16 = reshape(shape = var_5833, x = value_57_cast_fp16)[name = tensor("op_5834_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_5834_cast_fp16, y = var_5832_cast_fp16)[name = tensor("attn_57_cast_fp16")]; - tensor var_5837 = const()[name = tensor("op_5837"), val = tensor([1, 1280, 1, -1])]; - tensor input_427_cast_fp16 = reshape(shape = var_5837, x = attn_57_cast_fp16)[name = tensor("input_427_cast_fp16")]; - tensor var_5844 = const()[name = tensor("op_5844"), val = tensor([1, 1])]; - tensor var_5846 = const()[name = tensor("op_5846"), val = tensor([1, 1])]; - tensor pretrained_out_287_pad_type_0 = const()[name = tensor("pretrained_out_287_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_287_pad_0 = const()[name = tensor("pretrained_out_287_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337367168))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338186432))), name = tensor("layers_14_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_14_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338186560)))]; - tensor pretrained_out_287_cast_fp16 = conv(bias = layers_14_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_5846, groups = var_5699, pad = pretrained_out_287_pad_0, pad_type = pretrained_out_287_pad_type_0, strides = var_5844, weight = layers_14_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_427_cast_fp16)[name = tensor("pretrained_out_287_cast_fp16")]; - tensor var_5850 = const()[name = tensor("op_5850"), val = tensor([1, 1])]; - tensor var_5852 = const()[name = tensor("op_5852"), val = tensor([1, 1])]; - tensor input_429_pad_type_0 = const()[name = tensor("input_429_pad_type_0"), val = tensor("custom")]; - tensor input_429_pad_0 = const()[name = tensor("input_429_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338189184)))]; - tensor input_429_cast_fp16 = conv(dilations = var_5852, groups = var_5699, pad = input_429_pad_0, pad_type = input_429_pad_type_0, strides = var_5850, weight = layers_14_self_attn_o_proj_loraA_weight_to_fp16, x = input_427_cast_fp16)[name = tensor("input_429_cast_fp16")]; - tensor var_5856 = const()[name = tensor("op_5856"), val = tensor([1, 1])]; - tensor var_5858 = const()[name = tensor("op_5858"), val = tensor([1, 1])]; - tensor lora_out_573_pad_type_0 = const()[name = tensor("lora_out_573_pad_type_0"), val = tensor("custom")]; - tensor lora_out_573_pad_0 = const()[name = tensor("lora_out_573_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_575_weight_0_to_fp16 = const()[name = tensor("lora_out_575_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338230208)))]; - tensor lora_out_575_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5858, groups = var_5699, pad = lora_out_573_pad_0, pad_type = lora_out_573_pad_type_0, strides = var_5856, weight = lora_out_575_weight_0_to_fp16, x = input_429_cast_fp16)[name = tensor("lora_out_575_cast_fp16")]; - tensor obj_175_cast_fp16 = add(x = pretrained_out_287_cast_fp16, y = lora_out_575_cast_fp16)[name = tensor("obj_175_cast_fp16")]; - tensor inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = obj_175_cast_fp16)[name = tensor("inputs_87_cast_fp16")]; - tensor var_5871 = const()[name = tensor("op_5871"), val = tensor([1])]; - tensor channels_mean_87_cast_fp16 = reduce_mean(axes = var_5871, keep_dims = var_5700, 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_5875 = const()[name = tensor("op_5875"), val = tensor([1])]; - tensor var_5876_cast_fp16 = reduce_mean(axes = var_5875, keep_dims = var_5700, x = zero_mean_sq_87_cast_fp16)[name = tensor("op_5876_cast_fp16")]; - tensor var_5877_to_fp16 = const()[name = tensor("op_5877_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_5878_cast_fp16 = add(x = var_5876_cast_fp16, y = var_5877_to_fp16)[name = tensor("op_5878_cast_fp16")]; - tensor denom_87_epsilon_0 = const()[name = tensor("denom_87_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_87_cast_fp16 = rsqrt(epsilon = denom_87_epsilon_0, x = var_5878_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_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(338271232)))]; - 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(338273856)))]; - 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_87_cast_fp16)[name = tensor("obj_177_cast_fp16")]; - tensor var_5896 = const()[name = tensor("op_5896"), val = tensor([1, 1])]; - tensor var_5898 = const()[name = tensor("op_5898"), val = tensor([1, 1])]; - tensor pretrained_out_289_pad_type_0 = const()[name = tensor("pretrained_out_289_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_289_pad_0 = const()[name = tensor("pretrained_out_289_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338276480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339095744))), name = tensor("layers_14_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_14_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339095872)))]; - tensor pretrained_out_289_cast_fp16 = conv(bias = layers_14_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_5898, groups = var_5699, pad = pretrained_out_289_pad_0, pad_type = pretrained_out_289_pad_type_0, strides = var_5896, weight = layers_14_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_177_cast_fp16)[name = tensor("pretrained_out_289_cast_fp16")]; - tensor var_5902 = const()[name = tensor("op_5902"), val = tensor([1, 1])]; - tensor var_5904 = const()[name = tensor("op_5904"), val = tensor([1, 1])]; - tensor input_431_pad_type_0 = const()[name = tensor("input_431_pad_type_0"), val = tensor("custom")]; - tensor input_431_pad_0 = const()[name = tensor("input_431_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339098496)))]; - tensor input_431_cast_fp16 = conv(dilations = var_5904, groups = var_5699, pad = input_431_pad_0, pad_type = input_431_pad_type_0, strides = var_5902, weight = layers_14_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_177_cast_fp16)[name = tensor("input_431_cast_fp16")]; - tensor var_5908 = const()[name = tensor("op_5908"), val = tensor([1, 1])]; - tensor var_5910 = const()[name = tensor("op_5910"), val = tensor([1, 1])]; - tensor lora_out_577_pad_type_0 = const()[name = tensor("lora_out_577_pad_type_0"), val = tensor("custom")]; - tensor lora_out_577_pad_0 = const()[name = tensor("lora_out_577_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_579_weight_0_to_fp16 = const()[name = tensor("lora_out_579_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339139520)))]; - tensor lora_out_579_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5910, groups = var_5699, pad = lora_out_577_pad_0, pad_type = lora_out_577_pad_type_0, strides = var_5908, weight = lora_out_579_weight_0_to_fp16, x = input_431_cast_fp16)[name = tensor("lora_out_579_cast_fp16")]; - tensor query_59_cast_fp16 = add(x = pretrained_out_289_cast_fp16, y = lora_out_579_cast_fp16)[name = tensor("query_59_cast_fp16")]; - tensor var_5920 = const()[name = tensor("op_5920"), val = tensor([1, 1])]; - tensor var_5922 = const()[name = tensor("op_5922"), val = tensor([1, 1])]; - tensor pretrained_out_291_pad_type_0 = const()[name = tensor("pretrained_out_291_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_291_pad_0 = const()[name = tensor("pretrained_out_291_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339180544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339999808))), name = tensor("layers_14_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_291_cast_fp16 = conv(dilations = var_5922, groups = var_5699, pad = pretrained_out_291_pad_0, pad_type = pretrained_out_291_pad_type_0, strides = var_5920, weight = layers_14_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_291_cast_fp16")]; - tensor var_5926 = const()[name = tensor("op_5926"), val = tensor([1, 1])]; - tensor var_5928 = const()[name = tensor("op_5928"), val = tensor([1, 1])]; - tensor input_433_pad_type_0 = const()[name = tensor("input_433_pad_type_0"), val = tensor("custom")]; - tensor input_433_pad_0 = const()[name = tensor("input_433_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339999936)))]; - tensor input_433_cast_fp16 = conv(dilations = var_5928, groups = var_5699, pad = input_433_pad_0, pad_type = input_433_pad_type_0, strides = var_5926, weight = layers_14_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_433_cast_fp16")]; - tensor var_5932 = const()[name = tensor("op_5932"), val = tensor([1, 1])]; - tensor var_5934 = const()[name = tensor("op_5934"), val = tensor([1, 1])]; - tensor lora_out_581_pad_type_0 = const()[name = tensor("lora_out_581_pad_type_0"), val = tensor("custom")]; - tensor lora_out_581_pad_0 = const()[name = tensor("lora_out_581_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_583_weight_0_to_fp16 = const()[name = tensor("lora_out_583_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(340040960)))]; - tensor lora_out_583_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5934, groups = var_5699, pad = lora_out_581_pad_0, pad_type = lora_out_581_pad_type_0, strides = var_5932, weight = lora_out_583_weight_0_to_fp16, x = input_433_cast_fp16)[name = tensor("lora_out_583_cast_fp16")]; - tensor key_59_cast_fp16 = add(x = pretrained_out_291_cast_fp16, y = lora_out_583_cast_fp16)[name = tensor("key_59_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 pretrained_out_293_pad_type_0 = const()[name = tensor("pretrained_out_293_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_293_pad_0 = const()[name = tensor("pretrained_out_293_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(340081984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(340901248))), name = tensor("layers_14_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_14_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(340901376)))]; - tensor pretrained_out_293_cast_fp16 = conv(bias = layers_14_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_5947, groups = var_5699, pad = pretrained_out_293_pad_0, pad_type = pretrained_out_293_pad_type_0, strides = var_5945, weight = layers_14_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_293_cast_fp16")]; - tensor var_5951 = const()[name = tensor("op_5951"), val = tensor([1, 1])]; - tensor var_5953 = const()[name = tensor("op_5953"), val = tensor([1, 1])]; - tensor input_435_pad_type_0 = const()[name = tensor("input_435_pad_type_0"), val = tensor("custom")]; - tensor input_435_pad_0 = const()[name = tensor("input_435_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(340904000)))]; - tensor input_435_cast_fp16 = conv(dilations = var_5953, groups = var_5699, pad = input_435_pad_0, pad_type = input_435_pad_type_0, strides = var_5951, weight = layers_14_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_435_cast_fp16")]; - tensor var_5957 = const()[name = tensor("op_5957"), val = tensor([1, 1])]; - tensor var_5959 = const()[name = tensor("op_5959"), val = tensor([1, 1])]; - tensor lora_out_585_pad_type_0 = const()[name = tensor("lora_out_585_pad_type_0"), val = tensor("custom")]; - tensor lora_out_585_pad_0 = const()[name = tensor("lora_out_585_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_587_weight_0_to_fp16 = const()[name = tensor("lora_out_587_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(340945024)))]; - tensor lora_out_587_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5959, groups = var_5699, pad = lora_out_585_pad_0, pad_type = lora_out_585_pad_type_0, strides = var_5957, weight = lora_out_587_weight_0_to_fp16, x = input_435_cast_fp16)[name = tensor("lora_out_587_cast_fp16")]; - tensor value_59_cast_fp16 = add(x = pretrained_out_293_cast_fp16, y = lora_out_587_cast_fp16)[name = tensor("value_59_cast_fp16")]; - tensor var_5966 = const()[name = tensor("op_5966"), val = tensor([1, 20, 64, -1])]; - tensor var_5967_cast_fp16 = reshape(shape = var_5966, x = query_59_cast_fp16)[name = tensor("op_5967_cast_fp16")]; - tensor var_5968_to_fp16 = const()[name = tensor("op_5968_to_fp16"), val = tensor(0x1p-3)]; - tensor var_5969_cast_fp16 = mul(x = var_5967_cast_fp16, y = var_5968_to_fp16)[name = tensor("op_5969_cast_fp16")]; - tensor var_5970 = const()[name = tensor("op_5970"), val = tensor([1, 20, 64, -1])]; - tensor var_5971_cast_fp16 = reshape(shape = var_5970, x = key_59_cast_fp16)[name = tensor("op_5971_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_5969_cast_fp16, y = var_5971_cast_fp16)[name = tensor("mh_w_89_cast_fp16")]; - tensor var_5974_cast_fp16 = softmax(axis = var_5692, x = mh_w_89_cast_fp16)[name = tensor("op_5974_cast_fp16")]; - tensor var_5975 = const()[name = tensor("op_5975"), val = tensor([1, 20, 64, -1])]; - tensor var_5976_cast_fp16 = reshape(shape = var_5975, x = value_59_cast_fp16)[name = tensor("op_5976_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_5976_cast_fp16, y = var_5974_cast_fp16)[name = tensor("attn_59_cast_fp16")]; - tensor var_5979 = const()[name = tensor("op_5979"), val = tensor([1, 1280, 1, -1])]; - tensor input_437_cast_fp16 = reshape(shape = var_5979, x = attn_59_cast_fp16)[name = tensor("input_437_cast_fp16")]; - tensor var_5986 = const()[name = tensor("op_5986"), val = tensor([1, 1])]; - tensor var_5988 = const()[name = tensor("op_5988"), val = tensor([1, 1])]; - tensor pretrained_out_295_pad_type_0 = const()[name = tensor("pretrained_out_295_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_295_pad_0 = const()[name = tensor("pretrained_out_295_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(340986048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341805312))), name = tensor("layers_14_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_14_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341805440)))]; - tensor pretrained_out_295_cast_fp16 = conv(bias = layers_14_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_5988, groups = var_5699, pad = pretrained_out_295_pad_0, pad_type = pretrained_out_295_pad_type_0, strides = var_5986, weight = layers_14_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_437_cast_fp16)[name = tensor("pretrained_out_295_cast_fp16")]; - tensor var_5992 = const()[name = tensor("op_5992"), val = tensor([1, 1])]; - tensor var_5994 = const()[name = tensor("op_5994"), val = tensor([1, 1])]; - tensor input_439_pad_type_0 = const()[name = tensor("input_439_pad_type_0"), val = tensor("custom")]; - tensor input_439_pad_0 = const()[name = tensor("input_439_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341808064)))]; - tensor input_439_cast_fp16 = conv(dilations = var_5994, groups = var_5699, pad = input_439_pad_0, pad_type = input_439_pad_type_0, strides = var_5992, weight = layers_14_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_437_cast_fp16)[name = tensor("input_439_cast_fp16")]; - tensor var_5998 = const()[name = tensor("op_5998"), val = tensor([1, 1])]; - tensor var_6000 = const()[name = tensor("op_6000"), val = tensor([1, 1])]; - tensor lora_out_589_pad_type_0 = const()[name = tensor("lora_out_589_pad_type_0"), val = tensor("custom")]; - tensor lora_out_589_pad_0 = const()[name = tensor("lora_out_589_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_591_weight_0_to_fp16 = const()[name = tensor("lora_out_591_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341849088)))]; - tensor lora_out_591_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6000, groups = var_5699, pad = lora_out_589_pad_0, pad_type = lora_out_589_pad_type_0, strides = var_5998, weight = lora_out_591_weight_0_to_fp16, x = input_439_cast_fp16)[name = tensor("lora_out_591_cast_fp16")]; - tensor obj_179_cast_fp16 = add(x = pretrained_out_295_cast_fp16, y = lora_out_591_cast_fp16)[name = tensor("obj_179_cast_fp16")]; - tensor inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = obj_179_cast_fp16)[name = tensor("inputs_89_cast_fp16")]; - tensor var_6009 = const()[name = tensor("op_6009"), val = tensor([1])]; - tensor channels_mean_89_cast_fp16 = reduce_mean(axes = var_6009, keep_dims = var_5700, 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_6013 = const()[name = tensor("op_6013"), val = tensor([1])]; - tensor var_6014_cast_fp16 = reduce_mean(axes = var_6013, keep_dims = var_5700, x = zero_mean_sq_89_cast_fp16)[name = tensor("op_6014_cast_fp16")]; - tensor var_6015_to_fp16 = const()[name = tensor("op_6015_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_6016_cast_fp16 = add(x = var_6014_cast_fp16, y = var_6015_to_fp16)[name = tensor("op_6016_cast_fp16")]; - tensor denom_89_epsilon_0 = const()[name = tensor("denom_89_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_89_cast_fp16 = rsqrt(epsilon = denom_89_epsilon_0, x = var_6016_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_441_gamma_0_to_fp16 = const()[name = tensor("input_441_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341890112)))]; - tensor input_441_beta_0_to_fp16 = const()[name = tensor("input_441_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341892736)))]; - tensor input_441_epsilon_0_to_fp16 = const()[name = tensor("input_441_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_441_cast_fp16 = batch_norm(beta = input_441_beta_0_to_fp16, epsilon = input_441_epsilon_0_to_fp16, gamma = input_441_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_441_cast_fp16")]; - tensor var_6030 = const()[name = tensor("op_6030"), val = tensor([1, 1])]; - tensor var_6032 = const()[name = tensor("op_6032"), val = tensor([1, 1])]; - tensor pretrained_out_297_pad_type_0 = const()[name = tensor("pretrained_out_297_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_297_pad_0 = const()[name = tensor("pretrained_out_297_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341895360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345172224))), name = tensor("layers_14_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_14_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_14_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345172352)))]; - tensor pretrained_out_297_cast_fp16 = conv(bias = layers_14_fc1_pretrained_bias_to_fp16, dilations = var_6032, groups = var_5699, pad = pretrained_out_297_pad_0, pad_type = pretrained_out_297_pad_type_0, strides = var_6030, weight = layers_14_fc1_pretrained_weight_to_fp16_palettized, x = input_441_cast_fp16)[name = tensor("pretrained_out_297_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 input_443_pad_type_0 = const()[name = tensor("input_443_pad_type_0"), val = tensor("custom")]; - tensor input_443_pad_0 = const()[name = tensor("input_443_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_14_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345182656)))]; - tensor input_443_cast_fp16 = conv(dilations = var_6038, groups = var_5699, pad = input_443_pad_0, pad_type = input_443_pad_type_0, strides = var_6036, weight = layers_14_fc1_loraA_weight_to_fp16, x = input_441_cast_fp16)[name = tensor("input_443_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 lora_out_593_pad_type_0 = const()[name = tensor("lora_out_593_pad_type_0"), val = tensor("custom")]; - tensor lora_out_593_pad_0 = const()[name = tensor("lora_out_593_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_595_weight_0_to_fp16 = const()[name = tensor("lora_out_595_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345223680)))]; - tensor lora_out_595_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_6044, groups = var_5699, pad = lora_out_593_pad_0, pad_type = lora_out_593_pad_type_0, strides = var_6042, weight = lora_out_595_weight_0_to_fp16, x = input_443_cast_fp16)[name = tensor("lora_out_595_cast_fp16")]; - tensor input_445_cast_fp16 = add(x = pretrained_out_297_cast_fp16, y = lora_out_595_cast_fp16)[name = tensor("input_445_cast_fp16")]; - tensor input_447_mode_0 = const()[name = tensor("input_447_mode_0"), val = tensor("EXACT")]; - tensor input_447_cast_fp16 = gelu(mode = input_447_mode_0, x = input_445_cast_fp16)[name = tensor("input_447_cast_fp16")]; - tensor var_6056 = const()[name = tensor("op_6056"), val = tensor([1, 1])]; - tensor var_6058 = const()[name = tensor("op_6058"), val = tensor([1, 1])]; - tensor pretrained_out_299_pad_type_0 = const()[name = tensor("pretrained_out_299_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_299_pad_0 = const()[name = tensor("pretrained_out_299_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345387584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348664448))), name = tensor("layers_14_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_14_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_14_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348664576)))]; - tensor pretrained_out_299_cast_fp16 = conv(bias = layers_14_fc2_pretrained_bias_to_fp16, dilations = var_6058, groups = var_5699, pad = pretrained_out_299_pad_0, pad_type = pretrained_out_299_pad_type_0, strides = var_6056, weight = layers_14_fc2_pretrained_weight_to_fp16_palettized, x = input_447_cast_fp16)[name = tensor("pretrained_out_299_cast_fp16")]; - tensor var_6062 = const()[name = tensor("op_6062"), val = tensor([1, 1])]; - tensor var_6064 = const()[name = tensor("op_6064"), val = tensor([1, 1])]; - tensor input_449_pad_type_0 = const()[name = tensor("input_449_pad_type_0"), val = tensor("custom")]; - tensor input_449_pad_0 = const()[name = tensor("input_449_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_14_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_14_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348667200)))]; - tensor input_449_cast_fp16 = conv(dilations = var_6064, groups = var_5699, pad = input_449_pad_0, pad_type = input_449_pad_type_0, strides = var_6062, weight = layers_14_fc2_loraA_weight_to_fp16, x = input_447_cast_fp16)[name = tensor("input_449_cast_fp16")]; - tensor var_6068 = const()[name = tensor("op_6068"), val = tensor([1, 1])]; - tensor var_6070 = const()[name = tensor("op_6070"), val = tensor([1, 1])]; - tensor lora_out_597_pad_type_0 = const()[name = tensor("lora_out_597_pad_type_0"), val = tensor("custom")]; - tensor lora_out_597_pad_0 = const()[name = tensor("lora_out_597_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_599_weight_0_to_fp16 = const()[name = tensor("lora_out_599_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348831104)))]; - tensor lora_out_599_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6070, groups = var_5699, pad = lora_out_597_pad_0, pad_type = lora_out_597_pad_type_0, strides = var_6068, weight = lora_out_599_weight_0_to_fp16, x = input_449_cast_fp16)[name = tensor("lora_out_599_cast_fp16")]; - tensor hidden_states_31_cast_fp16 = add(x = pretrained_out_299_cast_fp16, y = lora_out_599_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_6086 = const()[name = tensor("op_6086"), val = tensor(3)]; - tensor var_6093 = const()[name = tensor("op_6093"), val = tensor(1)]; - tensor var_6094 = const()[name = tensor("op_6094"), val = tensor(true)]; - tensor var_6106 = const()[name = tensor("op_6106"), val = tensor([1])]; - tensor channels_mean_91_cast_fp16 = reduce_mean(axes = var_6106, keep_dims = var_6094, 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_6110 = const()[name = tensor("op_6110"), val = tensor([1])]; - tensor var_6111_cast_fp16 = reduce_mean(axes = var_6110, keep_dims = var_6094, x = zero_mean_sq_91_cast_fp16)[name = tensor("op_6111_cast_fp16")]; - tensor var_6112_to_fp16 = const()[name = tensor("op_6112_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_6113_cast_fp16 = add(x = var_6111_cast_fp16, y = var_6112_to_fp16)[name = tensor("op_6113_cast_fp16")]; - tensor denom_91_epsilon_0 = const()[name = tensor("denom_91_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_91_cast_fp16 = rsqrt(epsilon = denom_91_epsilon_0, x = var_6113_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_181_gamma_0_to_fp16 = const()[name = tensor("obj_181_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348872128)))]; - tensor obj_181_beta_0_to_fp16 = const()[name = tensor("obj_181_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348874752)))]; - tensor obj_181_epsilon_0_to_fp16 = const()[name = tensor("obj_181_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_181_cast_fp16 = batch_norm(beta = obj_181_beta_0_to_fp16, epsilon = obj_181_epsilon_0_to_fp16, gamma = obj_181_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_181_cast_fp16")]; - tensor var_6131 = const()[name = tensor("op_6131"), val = tensor([1, 1])]; - tensor var_6133 = const()[name = tensor("op_6133"), val = tensor([1, 1])]; - tensor pretrained_out_301_pad_type_0 = const()[name = tensor("pretrained_out_301_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_301_pad_0 = const()[name = tensor("pretrained_out_301_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348877376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(349696640))), name = tensor("layers_15_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_15_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(349696768)))]; - tensor pretrained_out_301_cast_fp16 = conv(bias = layers_15_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_6133, groups = var_6093, pad = pretrained_out_301_pad_0, pad_type = pretrained_out_301_pad_type_0, strides = var_6131, weight = layers_15_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_181_cast_fp16)[name = tensor("pretrained_out_301_cast_fp16")]; - tensor var_6137 = const()[name = tensor("op_6137"), val = tensor([1, 1])]; - tensor var_6139 = const()[name = tensor("op_6139"), val = tensor([1, 1])]; - tensor input_451_pad_type_0 = const()[name = tensor("input_451_pad_type_0"), val = tensor("custom")]; - tensor input_451_pad_0 = const()[name = tensor("input_451_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(349699392)))]; - tensor input_451_cast_fp16 = conv(dilations = var_6139, groups = var_6093, pad = input_451_pad_0, pad_type = input_451_pad_type_0, strides = var_6137, weight = layers_15_self_attn_q_proj_loraA_weight_to_fp16, x = obj_181_cast_fp16)[name = tensor("input_451_cast_fp16")]; - tensor var_6143 = const()[name = tensor("op_6143"), val = tensor([1, 1])]; - tensor var_6145 = const()[name = tensor("op_6145"), val = tensor([1, 1])]; - tensor lora_out_601_pad_type_0 = const()[name = tensor("lora_out_601_pad_type_0"), val = tensor("custom")]; - tensor lora_out_601_pad_0 = const()[name = tensor("lora_out_601_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_603_weight_0_to_fp16 = const()[name = tensor("lora_out_603_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(349740416)))]; - tensor lora_out_603_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6145, groups = var_6093, pad = lora_out_601_pad_0, pad_type = lora_out_601_pad_type_0, strides = var_6143, weight = lora_out_603_weight_0_to_fp16, x = input_451_cast_fp16)[name = tensor("lora_out_603_cast_fp16")]; - tensor query_61_cast_fp16 = add(x = pretrained_out_301_cast_fp16, y = lora_out_603_cast_fp16)[name = tensor("query_61_cast_fp16")]; - tensor var_6155 = const()[name = tensor("op_6155"), val = tensor([1, 1])]; - tensor var_6157 = const()[name = tensor("op_6157"), val = tensor([1, 1])]; - tensor pretrained_out_303_pad_type_0 = const()[name = tensor("pretrained_out_303_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_303_pad_0 = const()[name = tensor("pretrained_out_303_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(349781440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350600704))), name = tensor("layers_15_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_303_cast_fp16 = conv(dilations = var_6157, groups = var_6093, pad = pretrained_out_303_pad_0, pad_type = pretrained_out_303_pad_type_0, strides = var_6155, weight = layers_15_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_181_cast_fp16)[name = tensor("pretrained_out_303_cast_fp16")]; - tensor var_6161 = const()[name = tensor("op_6161"), val = tensor([1, 1])]; - tensor var_6163 = const()[name = tensor("op_6163"), val = tensor([1, 1])]; - tensor input_453_pad_type_0 = const()[name = tensor("input_453_pad_type_0"), val = tensor("custom")]; - tensor input_453_pad_0 = const()[name = tensor("input_453_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350600832)))]; - tensor input_453_cast_fp16 = conv(dilations = var_6163, groups = var_6093, pad = input_453_pad_0, pad_type = input_453_pad_type_0, strides = var_6161, weight = layers_15_self_attn_k_proj_loraA_weight_to_fp16, x = obj_181_cast_fp16)[name = tensor("input_453_cast_fp16")]; - tensor var_6167 = const()[name = tensor("op_6167"), val = tensor([1, 1])]; - tensor var_6169 = const()[name = tensor("op_6169"), val = tensor([1, 1])]; - tensor lora_out_605_pad_type_0 = const()[name = tensor("lora_out_605_pad_type_0"), val = tensor("custom")]; - tensor lora_out_605_pad_0 = const()[name = tensor("lora_out_605_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_607_weight_0_to_fp16 = const()[name = tensor("lora_out_607_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350641856)))]; - tensor lora_out_607_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6169, groups = var_6093, pad = lora_out_605_pad_0, pad_type = lora_out_605_pad_type_0, strides = var_6167, weight = lora_out_607_weight_0_to_fp16, x = input_453_cast_fp16)[name = tensor("lora_out_607_cast_fp16")]; - tensor current_key_31_cast_fp16 = add(x = pretrained_out_303_cast_fp16, y = lora_out_607_cast_fp16)[name = tensor("current_key_31_cast_fp16")]; - tensor var_6180 = const()[name = tensor("op_6180"), val = tensor([1, 1])]; - tensor var_6182 = const()[name = tensor("op_6182"), val = tensor([1, 1])]; - tensor pretrained_out_305_pad_type_0 = const()[name = tensor("pretrained_out_305_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_305_pad_0 = const()[name = tensor("pretrained_out_305_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350682880))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351502144))), name = tensor("layers_15_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_15_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351502272)))]; - tensor pretrained_out_305_cast_fp16 = conv(bias = layers_15_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_6182, groups = var_6093, pad = pretrained_out_305_pad_0, pad_type = pretrained_out_305_pad_type_0, strides = var_6180, weight = layers_15_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_181_cast_fp16)[name = tensor("pretrained_out_305_cast_fp16")]; - tensor var_6186 = const()[name = tensor("op_6186"), val = tensor([1, 1])]; - tensor var_6188 = const()[name = tensor("op_6188"), val = tensor([1, 1])]; - tensor input_455_pad_type_0 = const()[name = tensor("input_455_pad_type_0"), val = tensor("custom")]; - tensor input_455_pad_0 = const()[name = tensor("input_455_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351504896)))]; - tensor input_455_cast_fp16 = conv(dilations = var_6188, groups = var_6093, pad = input_455_pad_0, pad_type = input_455_pad_type_0, strides = var_6186, weight = layers_15_self_attn_v_proj_loraA_weight_to_fp16, x = obj_181_cast_fp16)[name = tensor("input_455_cast_fp16")]; - tensor var_6192 = const()[name = tensor("op_6192"), val = tensor([1, 1])]; - tensor var_6194 = const()[name = tensor("op_6194"), val = tensor([1, 1])]; - tensor lora_out_609_pad_type_0 = const()[name = tensor("lora_out_609_pad_type_0"), val = tensor("custom")]; - tensor lora_out_609_pad_0 = const()[name = tensor("lora_out_609_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_611_weight_0_to_fp16 = const()[name = tensor("lora_out_611_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351545920)))]; - tensor lora_out_611_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6194, groups = var_6093, pad = lora_out_609_pad_0, pad_type = lora_out_609_pad_type_0, strides = var_6192, weight = lora_out_611_weight_0_to_fp16, x = input_455_cast_fp16)[name = tensor("lora_out_611_cast_fp16")]; - tensor current_value_31_cast_fp16 = add(x = pretrained_out_305_cast_fp16, y = lora_out_611_cast_fp16)[name = tensor("current_value_31_cast_fp16")]; - tensor var_6204_cast_fp16 = mul(x = current_key_31_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_6204_cast_fp16")]; - tensor var_6206_cast_fp16 = mul(x = var_103_cast_fp16_15, y = var_295_cast_fp16)[name = tensor("op_6206_cast_fp16")]; - tensor key_61_cast_fp16 = add(x = var_6204_cast_fp16, y = var_6206_cast_fp16)[name = tensor("key_61_cast_fp16")]; - tensor var_6208_cast_fp16 = mul(x = current_value_31_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_6208_cast_fp16")]; - tensor var_6210_cast_fp16 = mul(x = var_138_cast_fp16_15, y = var_295_cast_fp16)[name = tensor("op_6210_cast_fp16")]; - tensor value_61_cast_fp16 = add(x = var_6208_cast_fp16, y = var_6210_cast_fp16)[name = tensor("value_61_cast_fp16")]; - tensor var_6213 = const()[name = tensor("op_6213"), val = tensor([1, 20, 64, -1])]; - tensor var_6214_cast_fp16 = reshape(shape = var_6213, x = query_61_cast_fp16)[name = tensor("op_6214_cast_fp16")]; - tensor var_6215_to_fp16 = const()[name = tensor("op_6215_to_fp16"), val = tensor(0x1p-3)]; - tensor var_6216_cast_fp16 = mul(x = var_6214_cast_fp16, y = var_6215_to_fp16)[name = tensor("op_6216_cast_fp16")]; - tensor var_6217 = const()[name = tensor("op_6217"), val = tensor([1, 20, 64, -1])]; - tensor var_6218_cast_fp16 = reshape(shape = var_6217, x = key_61_cast_fp16)[name = tensor("op_6218_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_6216_cast_fp16, y = var_6218_cast_fp16)[name = tensor("mh_w_91_cast_fp16")]; - tensor mh_w_93_cast_fp16 = add(x = mh_w_91_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_93_cast_fp16")]; - tensor var_6226_cast_fp16 = softmax(axis = var_6086, x = mh_w_93_cast_fp16)[name = tensor("op_6226_cast_fp16")]; - tensor var_6227 = const()[name = tensor("op_6227"), val = tensor([1, 20, 64, -1])]; - tensor var_6228_cast_fp16 = reshape(shape = var_6227, x = value_61_cast_fp16)[name = tensor("op_6228_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_6228_cast_fp16, y = var_6226_cast_fp16)[name = tensor("attn_61_cast_fp16")]; - tensor var_6231 = const()[name = tensor("op_6231"), val = tensor([1, 1280, 1, -1])]; - tensor input_457_cast_fp16 = reshape(shape = var_6231, x = attn_61_cast_fp16)[name = tensor("input_457_cast_fp16")]; - tensor var_6238 = const()[name = tensor("op_6238"), val = tensor([1, 1])]; - tensor var_6240 = const()[name = tensor("op_6240"), val = tensor([1, 1])]; - tensor pretrained_out_307_pad_type_0 = const()[name = tensor("pretrained_out_307_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_307_pad_0 = const()[name = tensor("pretrained_out_307_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351586944))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352406208))), name = tensor("layers_15_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_15_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352406336)))]; - tensor pretrained_out_307_cast_fp16 = conv(bias = layers_15_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_6240, groups = var_6093, pad = pretrained_out_307_pad_0, pad_type = pretrained_out_307_pad_type_0, strides = var_6238, weight = layers_15_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_457_cast_fp16)[name = tensor("pretrained_out_307_cast_fp16")]; - tensor var_6244 = const()[name = tensor("op_6244"), val = tensor([1, 1])]; - tensor var_6246 = const()[name = tensor("op_6246"), val = tensor([1, 1])]; - tensor input_459_pad_type_0 = const()[name = tensor("input_459_pad_type_0"), val = tensor("custom")]; - tensor input_459_pad_0 = const()[name = tensor("input_459_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352408960)))]; - tensor input_459_cast_fp16 = conv(dilations = var_6246, groups = var_6093, pad = input_459_pad_0, pad_type = input_459_pad_type_0, strides = var_6244, weight = layers_15_self_attn_o_proj_loraA_weight_to_fp16, x = input_457_cast_fp16)[name = tensor("input_459_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 lora_out_613_pad_type_0 = const()[name = tensor("lora_out_613_pad_type_0"), val = tensor("custom")]; - tensor lora_out_613_pad_0 = const()[name = tensor("lora_out_613_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_615_weight_0_to_fp16 = const()[name = tensor("lora_out_615_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352449984)))]; - tensor lora_out_615_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6252, groups = var_6093, pad = lora_out_613_pad_0, pad_type = lora_out_613_pad_type_0, strides = var_6250, weight = lora_out_615_weight_0_to_fp16, x = input_459_cast_fp16)[name = tensor("lora_out_615_cast_fp16")]; - tensor obj_187_cast_fp16 = add(x = pretrained_out_307_cast_fp16, y = lora_out_615_cast_fp16)[name = tensor("obj_187_cast_fp16")]; - tensor inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = obj_187_cast_fp16)[name = tensor("inputs_93_cast_fp16")]; - tensor var_6265 = const()[name = tensor("op_6265"), val = tensor([1])]; - tensor channels_mean_93_cast_fp16 = reduce_mean(axes = var_6265, keep_dims = var_6094, 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_6269 = const()[name = tensor("op_6269"), val = tensor([1])]; - tensor var_6270_cast_fp16 = reduce_mean(axes = var_6269, keep_dims = var_6094, x = zero_mean_sq_93_cast_fp16)[name = tensor("op_6270_cast_fp16")]; - tensor var_6271_to_fp16 = const()[name = tensor("op_6271_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_6272_cast_fp16 = add(x = var_6270_cast_fp16, y = var_6271_to_fp16)[name = tensor("op_6272_cast_fp16")]; - tensor denom_93_epsilon_0 = const()[name = tensor("denom_93_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_93_cast_fp16 = rsqrt(epsilon = denom_93_epsilon_0, x = var_6272_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_189_gamma_0_to_fp16 = const()[name = tensor("obj_189_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352491008)))]; - tensor obj_189_beta_0_to_fp16 = const()[name = tensor("obj_189_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352493632)))]; - tensor obj_189_epsilon_0_to_fp16 = const()[name = tensor("obj_189_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_189_cast_fp16 = batch_norm(beta = obj_189_beta_0_to_fp16, epsilon = obj_189_epsilon_0_to_fp16, gamma = obj_189_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_189_cast_fp16")]; - tensor var_6290 = const()[name = tensor("op_6290"), val = tensor([1, 1])]; - tensor var_6292 = const()[name = tensor("op_6292"), val = tensor([1, 1])]; - tensor pretrained_out_309_pad_type_0 = const()[name = tensor("pretrained_out_309_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_309_pad_0 = const()[name = tensor("pretrained_out_309_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352496256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353315520))), name = tensor("layers_15_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_15_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353315648)))]; - tensor pretrained_out_309_cast_fp16 = conv(bias = layers_15_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_6292, groups = var_6093, pad = pretrained_out_309_pad_0, pad_type = pretrained_out_309_pad_type_0, strides = var_6290, weight = layers_15_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_189_cast_fp16)[name = tensor("pretrained_out_309_cast_fp16")]; - tensor var_6296 = const()[name = tensor("op_6296"), val = tensor([1, 1])]; - tensor var_6298 = const()[name = tensor("op_6298"), val = tensor([1, 1])]; - tensor input_461_pad_type_0 = const()[name = tensor("input_461_pad_type_0"), val = tensor("custom")]; - tensor input_461_pad_0 = const()[name = tensor("input_461_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353318272)))]; - tensor input_461_cast_fp16 = conv(dilations = var_6298, groups = var_6093, pad = input_461_pad_0, pad_type = input_461_pad_type_0, strides = var_6296, weight = layers_15_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_189_cast_fp16)[name = tensor("input_461_cast_fp16")]; - tensor var_6302 = const()[name = tensor("op_6302"), val = tensor([1, 1])]; - tensor var_6304 = const()[name = tensor("op_6304"), val = tensor([1, 1])]; - tensor lora_out_617_pad_type_0 = const()[name = tensor("lora_out_617_pad_type_0"), val = tensor("custom")]; - tensor lora_out_617_pad_0 = const()[name = tensor("lora_out_617_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_619_weight_0_to_fp16 = const()[name = tensor("lora_out_619_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353359296)))]; - tensor lora_out_619_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6304, groups = var_6093, pad = lora_out_617_pad_0, pad_type = lora_out_617_pad_type_0, strides = var_6302, weight = lora_out_619_weight_0_to_fp16, x = input_461_cast_fp16)[name = tensor("lora_out_619_cast_fp16")]; - tensor query_63_cast_fp16 = add(x = pretrained_out_309_cast_fp16, y = lora_out_619_cast_fp16)[name = tensor("query_63_cast_fp16")]; - tensor var_6314 = const()[name = tensor("op_6314"), val = tensor([1, 1])]; - tensor var_6316 = const()[name = tensor("op_6316"), val = tensor([1, 1])]; - tensor pretrained_out_311_pad_type_0 = const()[name = tensor("pretrained_out_311_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_311_pad_0 = const()[name = tensor("pretrained_out_311_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353400320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354219584))), name = tensor("layers_15_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_311_cast_fp16 = conv(dilations = var_6316, groups = var_6093, pad = pretrained_out_311_pad_0, pad_type = pretrained_out_311_pad_type_0, strides = var_6314, weight = layers_15_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_311_cast_fp16")]; - tensor var_6320 = const()[name = tensor("op_6320"), val = tensor([1, 1])]; - tensor var_6322 = const()[name = tensor("op_6322"), val = tensor([1, 1])]; - tensor input_463_pad_type_0 = const()[name = tensor("input_463_pad_type_0"), val = tensor("custom")]; - tensor input_463_pad_0 = const()[name = tensor("input_463_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354219712)))]; - tensor input_463_cast_fp16 = conv(dilations = var_6322, groups = var_6093, pad = input_463_pad_0, pad_type = input_463_pad_type_0, strides = var_6320, weight = layers_15_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_463_cast_fp16")]; - tensor var_6326 = const()[name = tensor("op_6326"), val = tensor([1, 1])]; - tensor var_6328 = const()[name = tensor("op_6328"), val = tensor([1, 1])]; - tensor lora_out_621_pad_type_0 = const()[name = tensor("lora_out_621_pad_type_0"), val = tensor("custom")]; - tensor lora_out_621_pad_0 = const()[name = tensor("lora_out_621_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_623_weight_0_to_fp16 = const()[name = tensor("lora_out_623_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354260736)))]; - tensor lora_out_623_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6328, groups = var_6093, pad = lora_out_621_pad_0, pad_type = lora_out_621_pad_type_0, strides = var_6326, weight = lora_out_623_weight_0_to_fp16, x = input_463_cast_fp16)[name = tensor("lora_out_623_cast_fp16")]; - tensor key_63_cast_fp16 = add(x = pretrained_out_311_cast_fp16, y = lora_out_623_cast_fp16)[name = tensor("key_63_cast_fp16")]; - tensor var_6339 = const()[name = tensor("op_6339"), val = tensor([1, 1])]; - tensor var_6341 = const()[name = tensor("op_6341"), val = tensor([1, 1])]; - tensor pretrained_out_313_pad_type_0 = const()[name = tensor("pretrained_out_313_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_313_pad_0 = const()[name = tensor("pretrained_out_313_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354301760))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355121024))), name = tensor("layers_15_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_15_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355121152)))]; - tensor pretrained_out_313_cast_fp16 = conv(bias = layers_15_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_6341, groups = var_6093, pad = pretrained_out_313_pad_0, pad_type = pretrained_out_313_pad_type_0, strides = var_6339, weight = layers_15_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_313_cast_fp16")]; - tensor var_6345 = const()[name = tensor("op_6345"), val = tensor([1, 1])]; - tensor var_6347 = const()[name = tensor("op_6347"), val = tensor([1, 1])]; - tensor input_465_pad_type_0 = const()[name = tensor("input_465_pad_type_0"), val = tensor("custom")]; - tensor input_465_pad_0 = const()[name = tensor("input_465_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355123776)))]; - tensor input_465_cast_fp16 = conv(dilations = var_6347, groups = var_6093, pad = input_465_pad_0, pad_type = input_465_pad_type_0, strides = var_6345, weight = layers_15_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_465_cast_fp16")]; - tensor var_6351 = const()[name = tensor("op_6351"), val = tensor([1, 1])]; - tensor var_6353 = const()[name = tensor("op_6353"), val = tensor([1, 1])]; - tensor lora_out_625_pad_type_0 = const()[name = tensor("lora_out_625_pad_type_0"), val = tensor("custom")]; - tensor lora_out_625_pad_0 = const()[name = tensor("lora_out_625_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_627_weight_0_to_fp16 = const()[name = tensor("lora_out_627_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355164800)))]; - tensor lora_out_627_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6353, groups = var_6093, pad = lora_out_625_pad_0, pad_type = lora_out_625_pad_type_0, strides = var_6351, weight = lora_out_627_weight_0_to_fp16, x = input_465_cast_fp16)[name = tensor("lora_out_627_cast_fp16")]; - tensor value_63_cast_fp16 = add(x = pretrained_out_313_cast_fp16, y = lora_out_627_cast_fp16)[name = tensor("value_63_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 = query_63_cast_fp16)[name = tensor("op_6361_cast_fp16")]; - tensor var_6362_to_fp16 = const()[name = tensor("op_6362_to_fp16"), val = tensor(0x1p-3)]; - tensor var_6363_cast_fp16 = mul(x = var_6361_cast_fp16, y = var_6362_to_fp16)[name = tensor("op_6363_cast_fp16")]; - tensor var_6364 = const()[name = tensor("op_6364"), val = tensor([1, 20, 64, -1])]; - tensor var_6365_cast_fp16 = reshape(shape = var_6364, x = key_63_cast_fp16)[name = tensor("op_6365_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_6363_cast_fp16, y = var_6365_cast_fp16)[name = tensor("mh_w_95_cast_fp16")]; - tensor var_6368_cast_fp16 = softmax(axis = var_6086, x = mh_w_95_cast_fp16)[name = tensor("op_6368_cast_fp16")]; - tensor var_6369 = const()[name = tensor("op_6369"), val = tensor([1, 20, 64, -1])]; - tensor var_6370_cast_fp16 = reshape(shape = var_6369, x = value_63_cast_fp16)[name = tensor("op_6370_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_6370_cast_fp16, y = var_6368_cast_fp16)[name = tensor("attn_63_cast_fp16")]; - tensor var_6373 = const()[name = tensor("op_6373"), val = tensor([1, 1280, 1, -1])]; - tensor input_467_cast_fp16 = reshape(shape = var_6373, x = attn_63_cast_fp16)[name = tensor("input_467_cast_fp16")]; - tensor var_6380 = const()[name = tensor("op_6380"), val = tensor([1, 1])]; - tensor var_6382 = const()[name = tensor("op_6382"), val = tensor([1, 1])]; - tensor pretrained_out_315_pad_type_0 = const()[name = tensor("pretrained_out_315_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_315_pad_0 = const()[name = tensor("pretrained_out_315_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355205824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356025088))), name = tensor("layers_15_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_15_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356025216)))]; - tensor pretrained_out_315_cast_fp16 = conv(bias = layers_15_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_6382, groups = var_6093, pad = pretrained_out_315_pad_0, pad_type = pretrained_out_315_pad_type_0, strides = var_6380, weight = layers_15_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_467_cast_fp16)[name = tensor("pretrained_out_315_cast_fp16")]; - tensor var_6386 = const()[name = tensor("op_6386"), val = tensor([1, 1])]; - tensor var_6388 = const()[name = tensor("op_6388"), val = tensor([1, 1])]; - tensor input_469_pad_type_0 = const()[name = tensor("input_469_pad_type_0"), val = tensor("custom")]; - tensor input_469_pad_0 = const()[name = tensor("input_469_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356027840)))]; - tensor input_469_cast_fp16 = conv(dilations = var_6388, groups = var_6093, pad = input_469_pad_0, pad_type = input_469_pad_type_0, strides = var_6386, weight = layers_15_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_467_cast_fp16)[name = tensor("input_469_cast_fp16")]; - tensor var_6392 = const()[name = tensor("op_6392"), val = tensor([1, 1])]; - tensor var_6394 = const()[name = tensor("op_6394"), val = tensor([1, 1])]; - tensor lora_out_629_pad_type_0 = const()[name = tensor("lora_out_629_pad_type_0"), val = tensor("custom")]; - tensor lora_out_629_pad_0 = const()[name = tensor("lora_out_629_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_631_weight_0_to_fp16 = const()[name = tensor("lora_out_631_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356068864)))]; - tensor lora_out_631_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6394, groups = var_6093, pad = lora_out_629_pad_0, pad_type = lora_out_629_pad_type_0, strides = var_6392, weight = lora_out_631_weight_0_to_fp16, x = input_469_cast_fp16)[name = tensor("lora_out_631_cast_fp16")]; - tensor obj_191_cast_fp16 = add(x = pretrained_out_315_cast_fp16, y = lora_out_631_cast_fp16)[name = tensor("obj_191_cast_fp16")]; - tensor inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = obj_191_cast_fp16)[name = tensor("inputs_95_cast_fp16")]; - tensor var_6403 = const()[name = tensor("op_6403"), val = tensor([1])]; - tensor channels_mean_95_cast_fp16 = reduce_mean(axes = var_6403, keep_dims = var_6094, 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_6407 = const()[name = tensor("op_6407"), val = tensor([1])]; - tensor var_6408_cast_fp16 = reduce_mean(axes = var_6407, keep_dims = var_6094, x = zero_mean_sq_95_cast_fp16)[name = tensor("op_6408_cast_fp16")]; - tensor var_6409_to_fp16 = const()[name = tensor("op_6409_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_6410_cast_fp16 = add(x = var_6408_cast_fp16, y = var_6409_to_fp16)[name = tensor("op_6410_cast_fp16")]; - tensor denom_95_epsilon_0 = const()[name = tensor("denom_95_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_95_cast_fp16 = rsqrt(epsilon = denom_95_epsilon_0, x = var_6410_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_471_gamma_0_to_fp16 = const()[name = tensor("input_471_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356109888)))]; - tensor input_471_beta_0_to_fp16 = const()[name = tensor("input_471_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356112512)))]; - tensor input_471_epsilon_0_to_fp16 = const()[name = tensor("input_471_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_471_cast_fp16 = batch_norm(beta = input_471_beta_0_to_fp16, epsilon = input_471_epsilon_0_to_fp16, gamma = input_471_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_471_cast_fp16")]; - tensor var_6424 = const()[name = tensor("op_6424"), val = tensor([1, 1])]; - tensor var_6426 = const()[name = tensor("op_6426"), val = tensor([1, 1])]; - tensor pretrained_out_317_pad_type_0 = const()[name = tensor("pretrained_out_317_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_317_pad_0 = const()[name = tensor("pretrained_out_317_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356115136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359392000))), name = tensor("layers_15_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_15_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_15_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359392128)))]; - tensor pretrained_out_317_cast_fp16 = conv(bias = layers_15_fc1_pretrained_bias_to_fp16, dilations = var_6426, groups = var_6093, pad = pretrained_out_317_pad_0, pad_type = pretrained_out_317_pad_type_0, strides = var_6424, weight = layers_15_fc1_pretrained_weight_to_fp16_palettized, x = input_471_cast_fp16)[name = tensor("pretrained_out_317_cast_fp16")]; - tensor var_6430 = const()[name = tensor("op_6430"), val = tensor([1, 1])]; - tensor var_6432 = const()[name = tensor("op_6432"), val = tensor([1, 1])]; - tensor input_473_pad_type_0 = const()[name = tensor("input_473_pad_type_0"), val = tensor("custom")]; - tensor input_473_pad_0 = const()[name = tensor("input_473_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_15_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359402432)))]; - tensor input_473_cast_fp16 = conv(dilations = var_6432, groups = var_6093, pad = input_473_pad_0, pad_type = input_473_pad_type_0, strides = var_6430, weight = layers_15_fc1_loraA_weight_to_fp16, x = input_471_cast_fp16)[name = tensor("input_473_cast_fp16")]; - tensor var_6436 = const()[name = tensor("op_6436"), val = tensor([1, 1])]; - tensor var_6438 = const()[name = tensor("op_6438"), val = tensor([1, 1])]; - tensor lora_out_633_pad_type_0 = const()[name = tensor("lora_out_633_pad_type_0"), val = tensor("custom")]; - tensor lora_out_633_pad_0 = const()[name = tensor("lora_out_633_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_635_weight_0_to_fp16 = const()[name = tensor("lora_out_635_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359443456)))]; - tensor lora_out_635_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_6438, groups = var_6093, pad = lora_out_633_pad_0, pad_type = lora_out_633_pad_type_0, strides = var_6436, weight = lora_out_635_weight_0_to_fp16, x = input_473_cast_fp16)[name = tensor("lora_out_635_cast_fp16")]; - tensor input_475_cast_fp16 = add(x = pretrained_out_317_cast_fp16, y = lora_out_635_cast_fp16)[name = tensor("input_475_cast_fp16")]; - tensor input_477_mode_0 = const()[name = tensor("input_477_mode_0"), val = tensor("EXACT")]; - tensor input_477_cast_fp16 = gelu(mode = input_477_mode_0, x = input_475_cast_fp16)[name = tensor("input_477_cast_fp16")]; - tensor var_6450 = const()[name = tensor("op_6450"), val = tensor([1, 1])]; - tensor var_6452 = const()[name = tensor("op_6452"), val = tensor([1, 1])]; - tensor pretrained_out_319_pad_type_0 = const()[name = tensor("pretrained_out_319_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_319_pad_0 = const()[name = tensor("pretrained_out_319_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359607360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362884224))), name = tensor("layers_15_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_15_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_15_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362884352)))]; - tensor pretrained_out_319_cast_fp16 = conv(bias = layers_15_fc2_pretrained_bias_to_fp16, dilations = var_6452, groups = var_6093, pad = pretrained_out_319_pad_0, pad_type = pretrained_out_319_pad_type_0, strides = var_6450, weight = layers_15_fc2_pretrained_weight_to_fp16_palettized, x = input_477_cast_fp16)[name = tensor("pretrained_out_319_cast_fp16")]; - tensor var_6456 = const()[name = tensor("op_6456"), val = tensor([1, 1])]; - tensor var_6458 = const()[name = tensor("op_6458"), val = tensor([1, 1])]; - tensor input_479_pad_type_0 = const()[name = tensor("input_479_pad_type_0"), val = tensor("custom")]; - tensor input_479_pad_0 = const()[name = tensor("input_479_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_15_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_15_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362886976)))]; - tensor input_479_cast_fp16 = conv(dilations = var_6458, groups = var_6093, pad = input_479_pad_0, pad_type = input_479_pad_type_0, strides = var_6456, weight = layers_15_fc2_loraA_weight_to_fp16, x = input_477_cast_fp16)[name = tensor("input_479_cast_fp16")]; - tensor var_6462 = const()[name = tensor("op_6462"), val = tensor([1, 1])]; - tensor var_6464 = const()[name = tensor("op_6464"), val = tensor([1, 1])]; - tensor lora_out_637_pad_type_0 = const()[name = tensor("lora_out_637_pad_type_0"), val = tensor("custom")]; - tensor lora_out_637_pad_0 = const()[name = tensor("lora_out_637_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_639_weight_0_to_fp16 = const()[name = tensor("lora_out_639_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363050880)))]; - tensor lora_out_639_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6464, groups = var_6093, pad = lora_out_637_pad_0, pad_type = lora_out_637_pad_type_0, strides = var_6462, weight = lora_out_639_weight_0_to_fp16, x = input_479_cast_fp16)[name = tensor("lora_out_639_cast_fp16")]; - tensor hidden_states_33_cast_fp16 = add(x = pretrained_out_319_cast_fp16, y = lora_out_639_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_6480 = const()[name = tensor("op_6480"), val = tensor(3)]; - tensor var_6487 = const()[name = tensor("op_6487"), val = tensor(1)]; - tensor var_6488 = const()[name = tensor("op_6488"), val = tensor(true)]; - tensor var_6500 = const()[name = tensor("op_6500"), val = tensor([1])]; - tensor channels_mean_97_cast_fp16 = reduce_mean(axes = var_6500, keep_dims = var_6488, 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_6504 = const()[name = tensor("op_6504"), val = tensor([1])]; - tensor var_6505_cast_fp16 = reduce_mean(axes = var_6504, keep_dims = var_6488, x = zero_mean_sq_97_cast_fp16)[name = tensor("op_6505_cast_fp16")]; - tensor var_6506_to_fp16 = const()[name = tensor("op_6506_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_6507_cast_fp16 = add(x = var_6505_cast_fp16, y = var_6506_to_fp16)[name = tensor("op_6507_cast_fp16")]; - tensor denom_97_epsilon_0 = const()[name = tensor("denom_97_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_97_cast_fp16 = rsqrt(epsilon = denom_97_epsilon_0, x = var_6507_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_193_gamma_0_to_fp16 = const()[name = tensor("obj_193_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363091904)))]; - tensor obj_193_beta_0_to_fp16 = const()[name = tensor("obj_193_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363094528)))]; - tensor obj_193_epsilon_0_to_fp16 = const()[name = tensor("obj_193_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_193_cast_fp16 = batch_norm(beta = obj_193_beta_0_to_fp16, epsilon = obj_193_epsilon_0_to_fp16, gamma = obj_193_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_193_cast_fp16")]; - tensor var_6525 = const()[name = tensor("op_6525"), val = tensor([1, 1])]; - tensor var_6527 = const()[name = tensor("op_6527"), val = tensor([1, 1])]; - tensor pretrained_out_321_pad_type_0 = const()[name = tensor("pretrained_out_321_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_321_pad_0 = const()[name = tensor("pretrained_out_321_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363097152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363916416))), name = tensor("layers_16_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_16_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363916544)))]; - tensor pretrained_out_321_cast_fp16 = conv(bias = layers_16_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_6527, groups = var_6487, pad = pretrained_out_321_pad_0, pad_type = pretrained_out_321_pad_type_0, strides = var_6525, weight = layers_16_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_193_cast_fp16)[name = tensor("pretrained_out_321_cast_fp16")]; - tensor var_6531 = const()[name = tensor("op_6531"), val = tensor([1, 1])]; - tensor var_6533 = const()[name = tensor("op_6533"), val = tensor([1, 1])]; - tensor input_481_pad_type_0 = const()[name = tensor("input_481_pad_type_0"), val = tensor("custom")]; - tensor input_481_pad_0 = const()[name = tensor("input_481_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363919168)))]; - tensor input_481_cast_fp16 = conv(dilations = var_6533, groups = var_6487, pad = input_481_pad_0, pad_type = input_481_pad_type_0, strides = var_6531, weight = layers_16_self_attn_q_proj_loraA_weight_to_fp16, x = obj_193_cast_fp16)[name = tensor("input_481_cast_fp16")]; - tensor var_6537 = const()[name = tensor("op_6537"), val = tensor([1, 1])]; - tensor var_6539 = const()[name = tensor("op_6539"), val = tensor([1, 1])]; - tensor lora_out_641_pad_type_0 = const()[name = tensor("lora_out_641_pad_type_0"), val = tensor("custom")]; - tensor lora_out_641_pad_0 = const()[name = tensor("lora_out_641_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_643_weight_0_to_fp16 = const()[name = tensor("lora_out_643_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363960192)))]; - tensor lora_out_643_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6539, groups = var_6487, pad = lora_out_641_pad_0, pad_type = lora_out_641_pad_type_0, strides = var_6537, weight = lora_out_643_weight_0_to_fp16, x = input_481_cast_fp16)[name = tensor("lora_out_643_cast_fp16")]; - tensor query_65_cast_fp16 = add(x = pretrained_out_321_cast_fp16, y = lora_out_643_cast_fp16)[name = tensor("query_65_cast_fp16")]; - tensor var_6549 = const()[name = tensor("op_6549"), val = tensor([1, 1])]; - tensor var_6551 = const()[name = tensor("op_6551"), val = tensor([1, 1])]; - tensor pretrained_out_323_pad_type_0 = const()[name = tensor("pretrained_out_323_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_323_pad_0 = const()[name = tensor("pretrained_out_323_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364001216))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364820480))), name = tensor("layers_16_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_323_cast_fp16 = conv(dilations = var_6551, groups = var_6487, pad = pretrained_out_323_pad_0, pad_type = pretrained_out_323_pad_type_0, strides = var_6549, weight = layers_16_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_193_cast_fp16)[name = tensor("pretrained_out_323_cast_fp16")]; - tensor var_6555 = const()[name = tensor("op_6555"), val = tensor([1, 1])]; - tensor var_6557 = const()[name = tensor("op_6557"), val = tensor([1, 1])]; - tensor input_483_pad_type_0 = const()[name = tensor("input_483_pad_type_0"), val = tensor("custom")]; - tensor input_483_pad_0 = const()[name = tensor("input_483_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364820608)))]; - tensor input_483_cast_fp16 = conv(dilations = var_6557, groups = var_6487, pad = input_483_pad_0, pad_type = input_483_pad_type_0, strides = var_6555, weight = layers_16_self_attn_k_proj_loraA_weight_to_fp16, x = obj_193_cast_fp16)[name = tensor("input_483_cast_fp16")]; - tensor var_6561 = const()[name = tensor("op_6561"), val = tensor([1, 1])]; - tensor var_6563 = const()[name = tensor("op_6563"), val = tensor([1, 1])]; - tensor lora_out_645_pad_type_0 = const()[name = tensor("lora_out_645_pad_type_0"), val = tensor("custom")]; - tensor lora_out_645_pad_0 = const()[name = tensor("lora_out_645_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_647_weight_0_to_fp16 = const()[name = tensor("lora_out_647_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364861632)))]; - tensor lora_out_647_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6563, groups = var_6487, pad = lora_out_645_pad_0, pad_type = lora_out_645_pad_type_0, strides = var_6561, weight = lora_out_647_weight_0_to_fp16, x = input_483_cast_fp16)[name = tensor("lora_out_647_cast_fp16")]; - tensor current_key_33_cast_fp16 = add(x = pretrained_out_323_cast_fp16, y = lora_out_647_cast_fp16)[name = tensor("current_key_33_cast_fp16")]; - tensor var_6574 = const()[name = tensor("op_6574"), val = tensor([1, 1])]; - tensor var_6576 = const()[name = tensor("op_6576"), val = tensor([1, 1])]; - tensor pretrained_out_325_pad_type_0 = const()[name = tensor("pretrained_out_325_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_325_pad_0 = const()[name = tensor("pretrained_out_325_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364902656))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(365721920))), name = tensor("layers_16_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_16_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(365722048)))]; - tensor pretrained_out_325_cast_fp16 = conv(bias = layers_16_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_6576, groups = var_6487, pad = pretrained_out_325_pad_0, pad_type = pretrained_out_325_pad_type_0, strides = var_6574, weight = layers_16_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_193_cast_fp16)[name = tensor("pretrained_out_325_cast_fp16")]; - tensor var_6580 = const()[name = tensor("op_6580"), val = tensor([1, 1])]; - tensor var_6582 = const()[name = tensor("op_6582"), val = tensor([1, 1])]; - tensor input_485_pad_type_0 = const()[name = tensor("input_485_pad_type_0"), val = tensor("custom")]; - tensor input_485_pad_0 = const()[name = tensor("input_485_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(365724672)))]; - tensor input_485_cast_fp16 = conv(dilations = var_6582, groups = var_6487, pad = input_485_pad_0, pad_type = input_485_pad_type_0, strides = var_6580, weight = layers_16_self_attn_v_proj_loraA_weight_to_fp16, x = obj_193_cast_fp16)[name = tensor("input_485_cast_fp16")]; - tensor var_6586 = const()[name = tensor("op_6586"), val = tensor([1, 1])]; - tensor var_6588 = const()[name = tensor("op_6588"), val = tensor([1, 1])]; - tensor lora_out_649_pad_type_0 = const()[name = tensor("lora_out_649_pad_type_0"), val = tensor("custom")]; - tensor lora_out_649_pad_0 = const()[name = tensor("lora_out_649_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_651_weight_0_to_fp16 = const()[name = tensor("lora_out_651_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(365765696)))]; - tensor lora_out_651_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6588, groups = var_6487, pad = lora_out_649_pad_0, pad_type = lora_out_649_pad_type_0, strides = var_6586, weight = lora_out_651_weight_0_to_fp16, x = input_485_cast_fp16)[name = tensor("lora_out_651_cast_fp16")]; - tensor current_value_33_cast_fp16 = add(x = pretrained_out_325_cast_fp16, y = lora_out_651_cast_fp16)[name = tensor("current_value_33_cast_fp16")]; - tensor var_6598_cast_fp16 = mul(x = current_key_33_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_6598_cast_fp16")]; - tensor var_6600_cast_fp16 = mul(x = var_103_cast_fp16_16, y = var_295_cast_fp16)[name = tensor("op_6600_cast_fp16")]; - tensor key_65_cast_fp16 = add(x = var_6598_cast_fp16, y = var_6600_cast_fp16)[name = tensor("key_65_cast_fp16")]; - tensor var_6602_cast_fp16 = mul(x = current_value_33_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_6602_cast_fp16")]; - tensor var_6604_cast_fp16 = mul(x = var_138_cast_fp16_16, y = var_295_cast_fp16)[name = tensor("op_6604_cast_fp16")]; - tensor value_65_cast_fp16 = add(x = var_6602_cast_fp16, y = var_6604_cast_fp16)[name = tensor("value_65_cast_fp16")]; - tensor var_6607 = const()[name = tensor("op_6607"), val = tensor([1, 20, 64, -1])]; - tensor var_6608_cast_fp16 = reshape(shape = var_6607, x = query_65_cast_fp16)[name = tensor("op_6608_cast_fp16")]; - tensor var_6609_to_fp16 = const()[name = tensor("op_6609_to_fp16"), val = tensor(0x1p-3)]; - tensor var_6610_cast_fp16 = mul(x = var_6608_cast_fp16, y = var_6609_to_fp16)[name = tensor("op_6610_cast_fp16")]; - tensor var_6611 = const()[name = tensor("op_6611"), val = tensor([1, 20, 64, -1])]; - tensor var_6612_cast_fp16 = reshape(shape = var_6611, x = key_65_cast_fp16)[name = tensor("op_6612_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_6610_cast_fp16, y = var_6612_cast_fp16)[name = tensor("mh_w_97_cast_fp16")]; - tensor mh_w_99_cast_fp16 = add(x = mh_w_97_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_99_cast_fp16")]; - tensor var_6620_cast_fp16 = softmax(axis = var_6480, x = mh_w_99_cast_fp16)[name = tensor("op_6620_cast_fp16")]; - tensor var_6621 = const()[name = tensor("op_6621"), val = tensor([1, 20, 64, -1])]; - tensor var_6622_cast_fp16 = reshape(shape = var_6621, x = value_65_cast_fp16)[name = tensor("op_6622_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_6622_cast_fp16, y = var_6620_cast_fp16)[name = tensor("attn_65_cast_fp16")]; - tensor var_6625 = const()[name = tensor("op_6625"), val = tensor([1, 1280, 1, -1])]; - tensor input_487_cast_fp16 = reshape(shape = var_6625, x = attn_65_cast_fp16)[name = tensor("input_487_cast_fp16")]; - tensor var_6632 = const()[name = tensor("op_6632"), val = tensor([1, 1])]; - tensor var_6634 = const()[name = tensor("op_6634"), val = tensor([1, 1])]; - tensor pretrained_out_327_pad_type_0 = const()[name = tensor("pretrained_out_327_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_327_pad_0 = const()[name = tensor("pretrained_out_327_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(365806720))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366625984))), name = tensor("layers_16_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_16_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366626112)))]; - tensor pretrained_out_327_cast_fp16 = conv(bias = layers_16_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_6634, groups = var_6487, pad = pretrained_out_327_pad_0, pad_type = pretrained_out_327_pad_type_0, strides = var_6632, weight = layers_16_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_487_cast_fp16)[name = tensor("pretrained_out_327_cast_fp16")]; - tensor var_6638 = const()[name = tensor("op_6638"), val = tensor([1, 1])]; - tensor var_6640 = const()[name = tensor("op_6640"), val = tensor([1, 1])]; - tensor input_489_pad_type_0 = const()[name = tensor("input_489_pad_type_0"), val = tensor("custom")]; - tensor input_489_pad_0 = const()[name = tensor("input_489_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366628736)))]; - tensor input_489_cast_fp16 = conv(dilations = var_6640, groups = var_6487, pad = input_489_pad_0, pad_type = input_489_pad_type_0, strides = var_6638, weight = layers_16_self_attn_o_proj_loraA_weight_to_fp16, x = input_487_cast_fp16)[name = tensor("input_489_cast_fp16")]; - tensor var_6644 = const()[name = tensor("op_6644"), val = tensor([1, 1])]; - tensor var_6646 = const()[name = tensor("op_6646"), val = tensor([1, 1])]; - tensor lora_out_653_pad_type_0 = const()[name = tensor("lora_out_653_pad_type_0"), val = tensor("custom")]; - tensor lora_out_653_pad_0 = const()[name = tensor("lora_out_653_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_655_weight_0_to_fp16 = const()[name = tensor("lora_out_655_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366669760)))]; - tensor lora_out_655_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6646, groups = var_6487, pad = lora_out_653_pad_0, pad_type = lora_out_653_pad_type_0, strides = var_6644, weight = lora_out_655_weight_0_to_fp16, x = input_489_cast_fp16)[name = tensor("lora_out_655_cast_fp16")]; - tensor obj_199_cast_fp16 = add(x = pretrained_out_327_cast_fp16, y = lora_out_655_cast_fp16)[name = tensor("obj_199_cast_fp16")]; - tensor inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = obj_199_cast_fp16)[name = tensor("inputs_99_cast_fp16")]; - tensor var_6659 = const()[name = tensor("op_6659"), val = tensor([1])]; - tensor channels_mean_99_cast_fp16 = reduce_mean(axes = var_6659, keep_dims = var_6488, 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_6663 = const()[name = tensor("op_6663"), val = tensor([1])]; - tensor var_6664_cast_fp16 = reduce_mean(axes = var_6663, keep_dims = var_6488, x = zero_mean_sq_99_cast_fp16)[name = tensor("op_6664_cast_fp16")]; - tensor var_6665_to_fp16 = const()[name = tensor("op_6665_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_6666_cast_fp16 = add(x = var_6664_cast_fp16, y = var_6665_to_fp16)[name = tensor("op_6666_cast_fp16")]; - tensor denom_99_epsilon_0 = const()[name = tensor("denom_99_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_99_cast_fp16 = rsqrt(epsilon = denom_99_epsilon_0, x = var_6666_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_201_gamma_0_to_fp16 = const()[name = tensor("obj_201_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366710784)))]; - tensor obj_201_beta_0_to_fp16 = const()[name = tensor("obj_201_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366713408)))]; - tensor obj_201_epsilon_0_to_fp16 = const()[name = tensor("obj_201_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_201_cast_fp16 = batch_norm(beta = obj_201_beta_0_to_fp16, epsilon = obj_201_epsilon_0_to_fp16, gamma = obj_201_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_201_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 pretrained_out_329_pad_type_0 = const()[name = tensor("pretrained_out_329_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_329_pad_0 = const()[name = tensor("pretrained_out_329_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366716032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(367535296))), name = tensor("layers_16_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_16_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(367535424)))]; - tensor pretrained_out_329_cast_fp16 = conv(bias = layers_16_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_6686, groups = var_6487, pad = pretrained_out_329_pad_0, pad_type = pretrained_out_329_pad_type_0, strides = var_6684, weight = layers_16_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_201_cast_fp16)[name = tensor("pretrained_out_329_cast_fp16")]; - tensor var_6690 = const()[name = tensor("op_6690"), val = tensor([1, 1])]; - tensor var_6692 = const()[name = tensor("op_6692"), val = tensor([1, 1])]; - tensor input_491_pad_type_0 = const()[name = tensor("input_491_pad_type_0"), val = tensor("custom")]; - tensor input_491_pad_0 = const()[name = tensor("input_491_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(367538048)))]; - tensor input_491_cast_fp16 = conv(dilations = var_6692, groups = var_6487, pad = input_491_pad_0, pad_type = input_491_pad_type_0, strides = var_6690, weight = layers_16_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_201_cast_fp16)[name = tensor("input_491_cast_fp16")]; - tensor var_6696 = const()[name = tensor("op_6696"), val = tensor([1, 1])]; - tensor var_6698 = const()[name = tensor("op_6698"), val = tensor([1, 1])]; - tensor lora_out_657_pad_type_0 = const()[name = tensor("lora_out_657_pad_type_0"), val = tensor("custom")]; - tensor lora_out_657_pad_0 = const()[name = tensor("lora_out_657_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_659_weight_0_to_fp16 = const()[name = tensor("lora_out_659_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(367579072)))]; - tensor lora_out_659_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6698, groups = var_6487, pad = lora_out_657_pad_0, pad_type = lora_out_657_pad_type_0, strides = var_6696, weight = lora_out_659_weight_0_to_fp16, x = input_491_cast_fp16)[name = tensor("lora_out_659_cast_fp16")]; - tensor query_67_cast_fp16 = add(x = pretrained_out_329_cast_fp16, y = lora_out_659_cast_fp16)[name = tensor("query_67_cast_fp16")]; - tensor var_6708 = const()[name = tensor("op_6708"), val = tensor([1, 1])]; - tensor var_6710 = const()[name = tensor("op_6710"), val = tensor([1, 1])]; - tensor pretrained_out_331_pad_type_0 = const()[name = tensor("pretrained_out_331_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_331_pad_0 = const()[name = tensor("pretrained_out_331_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(367620096))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368439360))), name = tensor("layers_16_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_331_cast_fp16 = conv(dilations = var_6710, groups = var_6487, pad = pretrained_out_331_pad_0, pad_type = pretrained_out_331_pad_type_0, strides = var_6708, weight = layers_16_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_331_cast_fp16")]; - tensor var_6714 = const()[name = tensor("op_6714"), val = tensor([1, 1])]; - tensor var_6716 = const()[name = tensor("op_6716"), val = tensor([1, 1])]; - tensor input_493_pad_type_0 = const()[name = tensor("input_493_pad_type_0"), val = tensor("custom")]; - tensor input_493_pad_0 = const()[name = tensor("input_493_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368439488)))]; - tensor input_493_cast_fp16 = conv(dilations = var_6716, groups = var_6487, pad = input_493_pad_0, pad_type = input_493_pad_type_0, strides = var_6714, weight = layers_16_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_493_cast_fp16")]; - tensor var_6720 = const()[name = tensor("op_6720"), val = tensor([1, 1])]; - tensor var_6722 = const()[name = tensor("op_6722"), val = tensor([1, 1])]; - tensor lora_out_661_pad_type_0 = const()[name = tensor("lora_out_661_pad_type_0"), val = tensor("custom")]; - tensor lora_out_661_pad_0 = const()[name = tensor("lora_out_661_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_663_weight_0_to_fp16 = const()[name = tensor("lora_out_663_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368480512)))]; - tensor lora_out_663_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6722, groups = var_6487, pad = lora_out_661_pad_0, pad_type = lora_out_661_pad_type_0, strides = var_6720, weight = lora_out_663_weight_0_to_fp16, x = input_493_cast_fp16)[name = tensor("lora_out_663_cast_fp16")]; - tensor key_67_cast_fp16 = add(x = pretrained_out_331_cast_fp16, y = lora_out_663_cast_fp16)[name = tensor("key_67_cast_fp16")]; - tensor var_6733 = const()[name = tensor("op_6733"), val = tensor([1, 1])]; - tensor var_6735 = const()[name = tensor("op_6735"), val = tensor([1, 1])]; - tensor pretrained_out_333_pad_type_0 = const()[name = tensor("pretrained_out_333_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_333_pad_0 = const()[name = tensor("pretrained_out_333_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368521536))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369340800))), name = tensor("layers_16_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_16_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369340928)))]; - tensor pretrained_out_333_cast_fp16 = conv(bias = layers_16_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_6735, groups = var_6487, pad = pretrained_out_333_pad_0, pad_type = pretrained_out_333_pad_type_0, strides = var_6733, weight = layers_16_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_333_cast_fp16")]; - tensor var_6739 = const()[name = tensor("op_6739"), val = tensor([1, 1])]; - tensor var_6741 = const()[name = tensor("op_6741"), val = tensor([1, 1])]; - tensor input_495_pad_type_0 = const()[name = tensor("input_495_pad_type_0"), val = tensor("custom")]; - tensor input_495_pad_0 = const()[name = tensor("input_495_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369343552)))]; - tensor input_495_cast_fp16 = conv(dilations = var_6741, groups = var_6487, pad = input_495_pad_0, pad_type = input_495_pad_type_0, strides = var_6739, weight = layers_16_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_495_cast_fp16")]; - tensor var_6745 = const()[name = tensor("op_6745"), val = tensor([1, 1])]; - tensor var_6747 = const()[name = tensor("op_6747"), val = tensor([1, 1])]; - tensor lora_out_665_pad_type_0 = const()[name = tensor("lora_out_665_pad_type_0"), val = tensor("custom")]; - tensor lora_out_665_pad_0 = const()[name = tensor("lora_out_665_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_667_weight_0_to_fp16 = const()[name = tensor("lora_out_667_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369384576)))]; - tensor lora_out_667_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6747, groups = var_6487, pad = lora_out_665_pad_0, pad_type = lora_out_665_pad_type_0, strides = var_6745, weight = lora_out_667_weight_0_to_fp16, x = input_495_cast_fp16)[name = tensor("lora_out_667_cast_fp16")]; - tensor value_67_cast_fp16 = add(x = pretrained_out_333_cast_fp16, y = lora_out_667_cast_fp16)[name = tensor("value_67_cast_fp16")]; - tensor var_6754 = const()[name = tensor("op_6754"), val = tensor([1, 20, 64, -1])]; - tensor var_6755_cast_fp16 = reshape(shape = var_6754, x = query_67_cast_fp16)[name = tensor("op_6755_cast_fp16")]; - tensor var_6756_to_fp16 = const()[name = tensor("op_6756_to_fp16"), val = tensor(0x1p-3)]; - tensor var_6757_cast_fp16 = mul(x = var_6755_cast_fp16, y = var_6756_to_fp16)[name = tensor("op_6757_cast_fp16")]; - tensor var_6758 = const()[name = tensor("op_6758"), val = tensor([1, 20, 64, -1])]; - tensor var_6759_cast_fp16 = reshape(shape = var_6758, x = key_67_cast_fp16)[name = tensor("op_6759_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_6757_cast_fp16, y = var_6759_cast_fp16)[name = tensor("mh_w_101_cast_fp16")]; - tensor var_6762_cast_fp16 = softmax(axis = var_6480, x = mh_w_101_cast_fp16)[name = tensor("op_6762_cast_fp16")]; - tensor var_6763 = const()[name = tensor("op_6763"), val = tensor([1, 20, 64, -1])]; - tensor var_6764_cast_fp16 = reshape(shape = var_6763, x = value_67_cast_fp16)[name = tensor("op_6764_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_6764_cast_fp16, y = var_6762_cast_fp16)[name = tensor("attn_67_cast_fp16")]; - tensor var_6767 = const()[name = tensor("op_6767"), val = tensor([1, 1280, 1, -1])]; - tensor input_497_cast_fp16 = reshape(shape = var_6767, x = attn_67_cast_fp16)[name = tensor("input_497_cast_fp16")]; - tensor var_6774 = const()[name = tensor("op_6774"), val = tensor([1, 1])]; - tensor var_6776 = const()[name = tensor("op_6776"), val = tensor([1, 1])]; - tensor pretrained_out_335_pad_type_0 = const()[name = tensor("pretrained_out_335_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_335_pad_0 = const()[name = tensor("pretrained_out_335_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369425600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370244864))), name = tensor("layers_16_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_16_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370244992)))]; - tensor pretrained_out_335_cast_fp16 = conv(bias = layers_16_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_6776, groups = var_6487, pad = pretrained_out_335_pad_0, pad_type = pretrained_out_335_pad_type_0, strides = var_6774, weight = layers_16_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_497_cast_fp16)[name = tensor("pretrained_out_335_cast_fp16")]; - tensor var_6780 = const()[name = tensor("op_6780"), val = tensor([1, 1])]; - tensor var_6782 = const()[name = tensor("op_6782"), val = tensor([1, 1])]; - tensor input_499_pad_type_0 = const()[name = tensor("input_499_pad_type_0"), val = tensor("custom")]; - tensor input_499_pad_0 = const()[name = tensor("input_499_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370247616)))]; - tensor input_499_cast_fp16 = conv(dilations = var_6782, groups = var_6487, pad = input_499_pad_0, pad_type = input_499_pad_type_0, strides = var_6780, weight = layers_16_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_497_cast_fp16)[name = tensor("input_499_cast_fp16")]; - tensor var_6786 = const()[name = tensor("op_6786"), val = tensor([1, 1])]; - tensor var_6788 = const()[name = tensor("op_6788"), val = tensor([1, 1])]; - tensor lora_out_669_pad_type_0 = const()[name = tensor("lora_out_669_pad_type_0"), val = tensor("custom")]; - tensor lora_out_669_pad_0 = const()[name = tensor("lora_out_669_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_671_weight_0_to_fp16 = const()[name = tensor("lora_out_671_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370288640)))]; - tensor lora_out_671_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6788, groups = var_6487, pad = lora_out_669_pad_0, pad_type = lora_out_669_pad_type_0, strides = var_6786, weight = lora_out_671_weight_0_to_fp16, x = input_499_cast_fp16)[name = tensor("lora_out_671_cast_fp16")]; - tensor obj_203_cast_fp16 = add(x = pretrained_out_335_cast_fp16, y = lora_out_671_cast_fp16)[name = tensor("obj_203_cast_fp16")]; - tensor inputs_101_cast_fp16 = add(x = inputs_99_cast_fp16, y = obj_203_cast_fp16)[name = tensor("inputs_101_cast_fp16")]; - tensor var_6797 = const()[name = tensor("op_6797"), val = tensor([1])]; - tensor channels_mean_101_cast_fp16 = reduce_mean(axes = var_6797, keep_dims = var_6488, 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_6801 = const()[name = tensor("op_6801"), val = tensor([1])]; - tensor var_6802_cast_fp16 = reduce_mean(axes = var_6801, keep_dims = var_6488, x = zero_mean_sq_101_cast_fp16)[name = tensor("op_6802_cast_fp16")]; - tensor var_6803_to_fp16 = const()[name = tensor("op_6803_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_6804_cast_fp16 = add(x = var_6802_cast_fp16, y = var_6803_to_fp16)[name = tensor("op_6804_cast_fp16")]; - tensor denom_101_epsilon_0 = const()[name = tensor("denom_101_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_101_cast_fp16 = rsqrt(epsilon = denom_101_epsilon_0, x = var_6804_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_501_gamma_0_to_fp16 = const()[name = tensor("input_501_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370329664)))]; - tensor input_501_beta_0_to_fp16 = const()[name = tensor("input_501_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370332288)))]; - tensor input_501_epsilon_0_to_fp16 = const()[name = tensor("input_501_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_501_cast_fp16 = batch_norm(beta = input_501_beta_0_to_fp16, epsilon = input_501_epsilon_0_to_fp16, gamma = input_501_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_501_cast_fp16")]; - tensor var_6818 = const()[name = tensor("op_6818"), val = tensor([1, 1])]; - tensor var_6820 = const()[name = tensor("op_6820"), val = tensor([1, 1])]; - tensor pretrained_out_337_pad_type_0 = const()[name = tensor("pretrained_out_337_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_337_pad_0 = const()[name = tensor("pretrained_out_337_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370334912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(373611776))), name = tensor("layers_16_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_16_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_16_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(373611904)))]; - tensor pretrained_out_337_cast_fp16 = conv(bias = layers_16_fc1_pretrained_bias_to_fp16, dilations = var_6820, groups = var_6487, pad = pretrained_out_337_pad_0, pad_type = pretrained_out_337_pad_type_0, strides = var_6818, weight = layers_16_fc1_pretrained_weight_to_fp16_palettized, x = input_501_cast_fp16)[name = tensor("pretrained_out_337_cast_fp16")]; - tensor var_6824 = const()[name = tensor("op_6824"), val = tensor([1, 1])]; - tensor var_6826 = const()[name = tensor("op_6826"), val = tensor([1, 1])]; - tensor input_503_pad_type_0 = const()[name = tensor("input_503_pad_type_0"), val = tensor("custom")]; - tensor input_503_pad_0 = const()[name = tensor("input_503_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_16_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(373622208)))]; - tensor input_503_cast_fp16 = conv(dilations = var_6826, groups = var_6487, pad = input_503_pad_0, pad_type = input_503_pad_type_0, strides = var_6824, weight = layers_16_fc1_loraA_weight_to_fp16, x = input_501_cast_fp16)[name = tensor("input_503_cast_fp16")]; - tensor var_6830 = const()[name = tensor("op_6830"), val = tensor([1, 1])]; - tensor var_6832 = const()[name = tensor("op_6832"), val = tensor([1, 1])]; - tensor lora_out_673_pad_type_0 = const()[name = tensor("lora_out_673_pad_type_0"), val = tensor("custom")]; - tensor lora_out_673_pad_0 = const()[name = tensor("lora_out_673_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_675_weight_0_to_fp16 = const()[name = tensor("lora_out_675_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(373663232)))]; - tensor lora_out_675_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_6832, groups = var_6487, pad = lora_out_673_pad_0, pad_type = lora_out_673_pad_type_0, strides = var_6830, weight = lora_out_675_weight_0_to_fp16, x = input_503_cast_fp16)[name = tensor("lora_out_675_cast_fp16")]; - tensor input_505_cast_fp16 = add(x = pretrained_out_337_cast_fp16, y = lora_out_675_cast_fp16)[name = tensor("input_505_cast_fp16")]; - tensor input_507_mode_0 = const()[name = tensor("input_507_mode_0"), val = tensor("EXACT")]; - tensor input_507_cast_fp16 = gelu(mode = input_507_mode_0, x = input_505_cast_fp16)[name = tensor("input_507_cast_fp16")]; - tensor var_6844 = const()[name = tensor("op_6844"), val = tensor([1, 1])]; - tensor var_6846 = const()[name = tensor("op_6846"), val = tensor([1, 1])]; - tensor pretrained_out_339_pad_type_0 = const()[name = tensor("pretrained_out_339_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_339_pad_0 = const()[name = tensor("pretrained_out_339_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(373827136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377104000))), name = tensor("layers_16_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_16_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_16_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377104128)))]; - tensor pretrained_out_339_cast_fp16 = conv(bias = layers_16_fc2_pretrained_bias_to_fp16, dilations = var_6846, groups = var_6487, pad = pretrained_out_339_pad_0, pad_type = pretrained_out_339_pad_type_0, strides = var_6844, weight = layers_16_fc2_pretrained_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = tensor("pretrained_out_339_cast_fp16")]; - tensor var_6850 = const()[name = tensor("op_6850"), val = tensor([1, 1])]; - tensor var_6852 = const()[name = tensor("op_6852"), val = tensor([1, 1])]; - tensor input_509_pad_type_0 = const()[name = tensor("input_509_pad_type_0"), val = tensor("custom")]; - tensor input_509_pad_0 = const()[name = tensor("input_509_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_16_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_16_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377106752)))]; - tensor input_509_cast_fp16 = conv(dilations = var_6852, groups = var_6487, pad = input_509_pad_0, pad_type = input_509_pad_type_0, strides = var_6850, weight = layers_16_fc2_loraA_weight_to_fp16, x = input_507_cast_fp16)[name = tensor("input_509_cast_fp16")]; - tensor var_6856 = const()[name = tensor("op_6856"), val = tensor([1, 1])]; - tensor var_6858 = const()[name = tensor("op_6858"), val = tensor([1, 1])]; - tensor lora_out_677_pad_type_0 = const()[name = tensor("lora_out_677_pad_type_0"), val = tensor("custom")]; - tensor lora_out_677_pad_0 = const()[name = tensor("lora_out_677_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_679_weight_0_to_fp16 = const()[name = tensor("lora_out_679_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377270656)))]; - tensor lora_out_679_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6858, groups = var_6487, pad = lora_out_677_pad_0, pad_type = lora_out_677_pad_type_0, strides = var_6856, weight = lora_out_679_weight_0_to_fp16, x = input_509_cast_fp16)[name = tensor("lora_out_679_cast_fp16")]; - tensor hidden_states_35_cast_fp16 = add(x = pretrained_out_339_cast_fp16, y = lora_out_679_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_6874 = const()[name = tensor("op_6874"), val = tensor(3)]; - tensor var_6881 = const()[name = tensor("op_6881"), val = tensor(1)]; - tensor var_6882 = const()[name = tensor("op_6882"), val = tensor(true)]; - tensor var_6894 = const()[name = tensor("op_6894"), val = tensor([1])]; - tensor channels_mean_103_cast_fp16 = reduce_mean(axes = var_6894, keep_dims = var_6882, 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_6898 = const()[name = tensor("op_6898"), val = tensor([1])]; - tensor var_6899_cast_fp16 = reduce_mean(axes = var_6898, keep_dims = var_6882, x = zero_mean_sq_103_cast_fp16)[name = tensor("op_6899_cast_fp16")]; - tensor var_6900_to_fp16 = const()[name = tensor("op_6900_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_6901_cast_fp16 = add(x = var_6899_cast_fp16, y = var_6900_to_fp16)[name = tensor("op_6901_cast_fp16")]; - tensor denom_103_epsilon_0 = const()[name = tensor("denom_103_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_103_cast_fp16 = rsqrt(epsilon = denom_103_epsilon_0, x = var_6901_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_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(377311680)))]; - 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(377314304)))]; - 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_103_cast_fp16)[name = tensor("obj_205_cast_fp16")]; - tensor var_6919 = const()[name = tensor("op_6919"), val = tensor([1, 1])]; - tensor var_6921 = const()[name = tensor("op_6921"), val = tensor([1, 1])]; - tensor pretrained_out_341_pad_type_0 = const()[name = tensor("pretrained_out_341_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_341_pad_0 = const()[name = tensor("pretrained_out_341_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377316928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378136192))), name = tensor("layers_17_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_17_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378136320)))]; - tensor pretrained_out_341_cast_fp16 = conv(bias = layers_17_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_6921, groups = var_6881, pad = pretrained_out_341_pad_0, pad_type = pretrained_out_341_pad_type_0, strides = var_6919, weight = layers_17_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_205_cast_fp16)[name = tensor("pretrained_out_341_cast_fp16")]; - tensor var_6925 = const()[name = tensor("op_6925"), val = tensor([1, 1])]; - tensor var_6927 = const()[name = tensor("op_6927"), val = tensor([1, 1])]; - tensor input_511_pad_type_0 = const()[name = tensor("input_511_pad_type_0"), val = tensor("custom")]; - tensor input_511_pad_0 = const()[name = tensor("input_511_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378138944)))]; - tensor input_511_cast_fp16 = conv(dilations = var_6927, groups = var_6881, pad = input_511_pad_0, pad_type = input_511_pad_type_0, strides = var_6925, weight = layers_17_self_attn_q_proj_loraA_weight_to_fp16, x = obj_205_cast_fp16)[name = tensor("input_511_cast_fp16")]; - tensor var_6931 = const()[name = tensor("op_6931"), val = tensor([1, 1])]; - tensor var_6933 = const()[name = tensor("op_6933"), val = tensor([1, 1])]; - tensor lora_out_681_pad_type_0 = const()[name = tensor("lora_out_681_pad_type_0"), val = tensor("custom")]; - tensor lora_out_681_pad_0 = const()[name = tensor("lora_out_681_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_683_weight_0_to_fp16 = const()[name = tensor("lora_out_683_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378179968)))]; - tensor lora_out_683_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6933, groups = var_6881, pad = lora_out_681_pad_0, pad_type = lora_out_681_pad_type_0, strides = var_6931, weight = lora_out_683_weight_0_to_fp16, x = input_511_cast_fp16)[name = tensor("lora_out_683_cast_fp16")]; - tensor query_69_cast_fp16 = add(x = pretrained_out_341_cast_fp16, y = lora_out_683_cast_fp16)[name = tensor("query_69_cast_fp16")]; - tensor var_6943 = const()[name = tensor("op_6943"), val = tensor([1, 1])]; - tensor var_6945 = const()[name = tensor("op_6945"), val = tensor([1, 1])]; - tensor pretrained_out_343_pad_type_0 = const()[name = tensor("pretrained_out_343_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_343_pad_0 = const()[name = tensor("pretrained_out_343_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378220992))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379040256))), name = tensor("layers_17_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_343_cast_fp16 = conv(dilations = var_6945, groups = var_6881, pad = pretrained_out_343_pad_0, pad_type = pretrained_out_343_pad_type_0, strides = var_6943, weight = layers_17_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_205_cast_fp16)[name = tensor("pretrained_out_343_cast_fp16")]; - tensor var_6949 = const()[name = tensor("op_6949"), val = tensor([1, 1])]; - tensor var_6951 = const()[name = tensor("op_6951"), val = tensor([1, 1])]; - tensor input_513_pad_type_0 = const()[name = tensor("input_513_pad_type_0"), val = tensor("custom")]; - tensor input_513_pad_0 = const()[name = tensor("input_513_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379040384)))]; - tensor input_513_cast_fp16 = conv(dilations = var_6951, groups = var_6881, pad = input_513_pad_0, pad_type = input_513_pad_type_0, strides = var_6949, weight = layers_17_self_attn_k_proj_loraA_weight_to_fp16, x = obj_205_cast_fp16)[name = tensor("input_513_cast_fp16")]; - tensor var_6955 = const()[name = tensor("op_6955"), val = tensor([1, 1])]; - tensor var_6957 = const()[name = tensor("op_6957"), val = tensor([1, 1])]; - tensor lora_out_685_pad_type_0 = const()[name = tensor("lora_out_685_pad_type_0"), val = tensor("custom")]; - tensor lora_out_685_pad_0 = const()[name = tensor("lora_out_685_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_687_weight_0_to_fp16 = const()[name = tensor("lora_out_687_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379081408)))]; - tensor lora_out_687_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6957, groups = var_6881, pad = lora_out_685_pad_0, pad_type = lora_out_685_pad_type_0, strides = var_6955, weight = lora_out_687_weight_0_to_fp16, x = input_513_cast_fp16)[name = tensor("lora_out_687_cast_fp16")]; - tensor current_key_35_cast_fp16 = add(x = pretrained_out_343_cast_fp16, y = lora_out_687_cast_fp16)[name = tensor("current_key_35_cast_fp16")]; - tensor var_6968 = const()[name = tensor("op_6968"), val = tensor([1, 1])]; - tensor var_6970 = const()[name = tensor("op_6970"), val = tensor([1, 1])]; - tensor pretrained_out_345_pad_type_0 = const()[name = tensor("pretrained_out_345_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_345_pad_0 = const()[name = tensor("pretrained_out_345_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379122432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379941696))), name = tensor("layers_17_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_17_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379941824)))]; - tensor pretrained_out_345_cast_fp16 = conv(bias = layers_17_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_6970, groups = var_6881, pad = pretrained_out_345_pad_0, pad_type = pretrained_out_345_pad_type_0, strides = var_6968, weight = layers_17_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_205_cast_fp16)[name = tensor("pretrained_out_345_cast_fp16")]; - tensor var_6974 = const()[name = tensor("op_6974"), val = tensor([1, 1])]; - tensor var_6976 = const()[name = tensor("op_6976"), val = tensor([1, 1])]; - tensor input_515_pad_type_0 = const()[name = tensor("input_515_pad_type_0"), val = tensor("custom")]; - tensor input_515_pad_0 = const()[name = tensor("input_515_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379944448)))]; - tensor input_515_cast_fp16 = conv(dilations = var_6976, groups = var_6881, pad = input_515_pad_0, pad_type = input_515_pad_type_0, strides = var_6974, weight = layers_17_self_attn_v_proj_loraA_weight_to_fp16, x = obj_205_cast_fp16)[name = tensor("input_515_cast_fp16")]; - tensor var_6980 = const()[name = tensor("op_6980"), val = tensor([1, 1])]; - tensor var_6982 = const()[name = tensor("op_6982"), val = tensor([1, 1])]; - tensor lora_out_689_pad_type_0 = const()[name = tensor("lora_out_689_pad_type_0"), val = tensor("custom")]; - tensor lora_out_689_pad_0 = const()[name = tensor("lora_out_689_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_691_weight_0_to_fp16 = const()[name = tensor("lora_out_691_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379985472)))]; - tensor lora_out_691_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6982, groups = var_6881, pad = lora_out_689_pad_0, pad_type = lora_out_689_pad_type_0, strides = var_6980, weight = lora_out_691_weight_0_to_fp16, x = input_515_cast_fp16)[name = tensor("lora_out_691_cast_fp16")]; - tensor current_value_35_cast_fp16 = add(x = pretrained_out_345_cast_fp16, y = lora_out_691_cast_fp16)[name = tensor("current_value_35_cast_fp16")]; - tensor var_6992_cast_fp16 = mul(x = current_key_35_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_6992_cast_fp16")]; - tensor var_6994_cast_fp16 = mul(x = var_103_cast_fp16_17, y = var_295_cast_fp16)[name = tensor("op_6994_cast_fp16")]; - tensor key_69_cast_fp16 = add(x = var_6992_cast_fp16, y = var_6994_cast_fp16)[name = tensor("key_69_cast_fp16")]; - tensor var_6996_cast_fp16 = mul(x = current_value_35_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_6996_cast_fp16")]; - tensor var_6998_cast_fp16 = mul(x = var_138_cast_fp16_17, y = var_295_cast_fp16)[name = tensor("op_6998_cast_fp16")]; - tensor value_69_cast_fp16 = add(x = var_6996_cast_fp16, y = var_6998_cast_fp16)[name = tensor("value_69_cast_fp16")]; - tensor var_7001 = const()[name = tensor("op_7001"), val = tensor([1, 20, 64, -1])]; - tensor var_7002_cast_fp16 = reshape(shape = var_7001, x = query_69_cast_fp16)[name = tensor("op_7002_cast_fp16")]; - tensor var_7003_to_fp16 = const()[name = tensor("op_7003_to_fp16"), val = tensor(0x1p-3)]; - tensor var_7004_cast_fp16 = mul(x = var_7002_cast_fp16, y = var_7003_to_fp16)[name = tensor("op_7004_cast_fp16")]; - tensor var_7005 = const()[name = tensor("op_7005"), val = tensor([1, 20, 64, -1])]; - tensor var_7006_cast_fp16 = reshape(shape = var_7005, x = key_69_cast_fp16)[name = tensor("op_7006_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_7004_cast_fp16, y = var_7006_cast_fp16)[name = tensor("mh_w_103_cast_fp16")]; - tensor mh_w_105_cast_fp16 = add(x = mh_w_103_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_105_cast_fp16")]; - tensor var_7014_cast_fp16 = softmax(axis = var_6874, x = mh_w_105_cast_fp16)[name = tensor("op_7014_cast_fp16")]; - tensor var_7015 = const()[name = tensor("op_7015"), val = tensor([1, 20, 64, -1])]; - tensor var_7016_cast_fp16 = reshape(shape = var_7015, x = value_69_cast_fp16)[name = tensor("op_7016_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_7016_cast_fp16, y = var_7014_cast_fp16)[name = tensor("attn_69_cast_fp16")]; - tensor var_7019 = const()[name = tensor("op_7019"), val = tensor([1, 1280, 1, -1])]; - tensor input_517_cast_fp16 = reshape(shape = var_7019, x = attn_69_cast_fp16)[name = tensor("input_517_cast_fp16")]; - tensor var_7026 = const()[name = tensor("op_7026"), val = tensor([1, 1])]; - tensor var_7028 = const()[name = tensor("op_7028"), val = tensor([1, 1])]; - tensor pretrained_out_347_pad_type_0 = const()[name = tensor("pretrained_out_347_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_347_pad_0 = const()[name = tensor("pretrained_out_347_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(380026496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(380845760))), name = tensor("layers_17_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_17_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(380845888)))]; - tensor pretrained_out_347_cast_fp16 = conv(bias = layers_17_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_7028, groups = var_6881, pad = pretrained_out_347_pad_0, pad_type = pretrained_out_347_pad_type_0, strides = var_7026, weight = layers_17_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_517_cast_fp16)[name = tensor("pretrained_out_347_cast_fp16")]; - tensor var_7032 = const()[name = tensor("op_7032"), val = tensor([1, 1])]; - tensor var_7034 = const()[name = tensor("op_7034"), val = tensor([1, 1])]; - tensor input_519_pad_type_0 = const()[name = tensor("input_519_pad_type_0"), val = tensor("custom")]; - tensor input_519_pad_0 = const()[name = tensor("input_519_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(380848512)))]; - tensor input_519_cast_fp16 = conv(dilations = var_7034, groups = var_6881, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = var_7032, weight = layers_17_self_attn_o_proj_loraA_weight_to_fp16, x = input_517_cast_fp16)[name = tensor("input_519_cast_fp16")]; - tensor var_7038 = const()[name = tensor("op_7038"), val = tensor([1, 1])]; - tensor var_7040 = const()[name = tensor("op_7040"), val = tensor([1, 1])]; - tensor lora_out_693_pad_type_0 = const()[name = tensor("lora_out_693_pad_type_0"), val = tensor("custom")]; - tensor lora_out_693_pad_0 = const()[name = tensor("lora_out_693_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_695_weight_0_to_fp16 = const()[name = tensor("lora_out_695_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(380889536)))]; - tensor lora_out_695_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7040, groups = var_6881, pad = lora_out_693_pad_0, pad_type = lora_out_693_pad_type_0, strides = var_7038, weight = lora_out_695_weight_0_to_fp16, x = input_519_cast_fp16)[name = tensor("lora_out_695_cast_fp16")]; - tensor obj_211_cast_fp16 = add(x = pretrained_out_347_cast_fp16, y = lora_out_695_cast_fp16)[name = tensor("obj_211_cast_fp16")]; - tensor inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = obj_211_cast_fp16)[name = tensor("inputs_105_cast_fp16")]; - tensor var_7053 = const()[name = tensor("op_7053"), val = tensor([1])]; - tensor channels_mean_105_cast_fp16 = reduce_mean(axes = var_7053, keep_dims = var_6882, 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_7057 = const()[name = tensor("op_7057"), val = tensor([1])]; - tensor var_7058_cast_fp16 = reduce_mean(axes = var_7057, keep_dims = var_6882, x = zero_mean_sq_105_cast_fp16)[name = tensor("op_7058_cast_fp16")]; - tensor var_7059_to_fp16 = const()[name = tensor("op_7059_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_7060_cast_fp16 = add(x = var_7058_cast_fp16, y = var_7059_to_fp16)[name = tensor("op_7060_cast_fp16")]; - tensor denom_105_epsilon_0 = const()[name = tensor("denom_105_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_105_cast_fp16 = rsqrt(epsilon = denom_105_epsilon_0, x = var_7060_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_213_gamma_0_to_fp16 = const()[name = tensor("obj_213_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(380930560)))]; - tensor obj_213_beta_0_to_fp16 = const()[name = tensor("obj_213_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(380933184)))]; - tensor obj_213_epsilon_0_to_fp16 = const()[name = tensor("obj_213_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_213_cast_fp16 = batch_norm(beta = obj_213_beta_0_to_fp16, epsilon = obj_213_epsilon_0_to_fp16, gamma = obj_213_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_213_cast_fp16")]; - tensor var_7078 = const()[name = tensor("op_7078"), val = tensor([1, 1])]; - tensor var_7080 = const()[name = tensor("op_7080"), val = tensor([1, 1])]; - tensor pretrained_out_349_pad_type_0 = const()[name = tensor("pretrained_out_349_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_349_pad_0 = const()[name = tensor("pretrained_out_349_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(380935808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381755072))), name = tensor("layers_17_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_17_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381755200)))]; - tensor pretrained_out_349_cast_fp16 = conv(bias = layers_17_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_7080, groups = var_6881, pad = pretrained_out_349_pad_0, pad_type = pretrained_out_349_pad_type_0, strides = var_7078, weight = layers_17_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_213_cast_fp16)[name = tensor("pretrained_out_349_cast_fp16")]; - tensor var_7084 = const()[name = tensor("op_7084"), val = tensor([1, 1])]; - tensor var_7086 = const()[name = tensor("op_7086"), val = tensor([1, 1])]; - tensor input_521_pad_type_0 = const()[name = tensor("input_521_pad_type_0"), val = tensor("custom")]; - tensor input_521_pad_0 = const()[name = tensor("input_521_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381757824)))]; - tensor input_521_cast_fp16 = conv(dilations = var_7086, groups = var_6881, pad = input_521_pad_0, pad_type = input_521_pad_type_0, strides = var_7084, weight = layers_17_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_213_cast_fp16)[name = tensor("input_521_cast_fp16")]; - tensor var_7090 = const()[name = tensor("op_7090"), val = tensor([1, 1])]; - tensor var_7092 = const()[name = tensor("op_7092"), val = tensor([1, 1])]; - tensor lora_out_697_pad_type_0 = const()[name = tensor("lora_out_697_pad_type_0"), val = tensor("custom")]; - tensor lora_out_697_pad_0 = const()[name = tensor("lora_out_697_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_699_weight_0_to_fp16 = const()[name = tensor("lora_out_699_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381798848)))]; - tensor lora_out_699_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7092, groups = var_6881, pad = lora_out_697_pad_0, pad_type = lora_out_697_pad_type_0, strides = var_7090, weight = lora_out_699_weight_0_to_fp16, x = input_521_cast_fp16)[name = tensor("lora_out_699_cast_fp16")]; - tensor query_71_cast_fp16 = add(x = pretrained_out_349_cast_fp16, y = lora_out_699_cast_fp16)[name = tensor("query_71_cast_fp16")]; - tensor var_7102 = const()[name = tensor("op_7102"), val = tensor([1, 1])]; - tensor var_7104 = const()[name = tensor("op_7104"), val = tensor([1, 1])]; - tensor pretrained_out_351_pad_type_0 = const()[name = tensor("pretrained_out_351_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_351_pad_0 = const()[name = tensor("pretrained_out_351_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381839872))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(382659136))), name = tensor("layers_17_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_351_cast_fp16 = conv(dilations = var_7104, groups = var_6881, pad = pretrained_out_351_pad_0, pad_type = pretrained_out_351_pad_type_0, strides = var_7102, weight = layers_17_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_351_cast_fp16")]; - tensor var_7108 = const()[name = tensor("op_7108"), val = tensor([1, 1])]; - tensor var_7110 = const()[name = tensor("op_7110"), val = tensor([1, 1])]; - tensor input_523_pad_type_0 = const()[name = tensor("input_523_pad_type_0"), val = tensor("custom")]; - tensor input_523_pad_0 = const()[name = tensor("input_523_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(382659264)))]; - tensor input_523_cast_fp16 = conv(dilations = var_7110, groups = var_6881, pad = input_523_pad_0, pad_type = input_523_pad_type_0, strides = var_7108, weight = layers_17_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_523_cast_fp16")]; - tensor var_7114 = const()[name = tensor("op_7114"), val = tensor([1, 1])]; - tensor var_7116 = const()[name = tensor("op_7116"), val = tensor([1, 1])]; - tensor lora_out_701_pad_type_0 = const()[name = tensor("lora_out_701_pad_type_0"), val = tensor("custom")]; - tensor lora_out_701_pad_0 = const()[name = tensor("lora_out_701_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_703_weight_0_to_fp16 = const()[name = tensor("lora_out_703_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(382700288)))]; - tensor lora_out_703_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7116, groups = var_6881, pad = lora_out_701_pad_0, pad_type = lora_out_701_pad_type_0, strides = var_7114, weight = lora_out_703_weight_0_to_fp16, x = input_523_cast_fp16)[name = tensor("lora_out_703_cast_fp16")]; - tensor key_71_cast_fp16 = add(x = pretrained_out_351_cast_fp16, y = lora_out_703_cast_fp16)[name = tensor("key_71_cast_fp16")]; - tensor var_7127 = const()[name = tensor("op_7127"), val = tensor([1, 1])]; - tensor var_7129 = const()[name = tensor("op_7129"), val = tensor([1, 1])]; - tensor pretrained_out_353_pad_type_0 = const()[name = tensor("pretrained_out_353_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_353_pad_0 = const()[name = tensor("pretrained_out_353_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(382741312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383560576))), name = tensor("layers_17_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_17_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383560704)))]; - tensor pretrained_out_353_cast_fp16 = conv(bias = layers_17_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_7129, groups = var_6881, pad = pretrained_out_353_pad_0, pad_type = pretrained_out_353_pad_type_0, strides = var_7127, weight = layers_17_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_353_cast_fp16")]; - tensor var_7133 = const()[name = tensor("op_7133"), val = tensor([1, 1])]; - tensor var_7135 = const()[name = tensor("op_7135"), val = tensor([1, 1])]; - tensor input_525_pad_type_0 = const()[name = tensor("input_525_pad_type_0"), val = tensor("custom")]; - tensor input_525_pad_0 = const()[name = tensor("input_525_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383563328)))]; - tensor input_525_cast_fp16 = conv(dilations = var_7135, groups = var_6881, pad = input_525_pad_0, pad_type = input_525_pad_type_0, strides = var_7133, weight = layers_17_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_525_cast_fp16")]; - tensor var_7139 = const()[name = tensor("op_7139"), val = tensor([1, 1])]; - tensor var_7141 = const()[name = tensor("op_7141"), val = tensor([1, 1])]; - tensor lora_out_705_pad_type_0 = const()[name = tensor("lora_out_705_pad_type_0"), val = tensor("custom")]; - tensor lora_out_705_pad_0 = const()[name = tensor("lora_out_705_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_707_weight_0_to_fp16 = const()[name = tensor("lora_out_707_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383604352)))]; - tensor lora_out_707_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7141, groups = var_6881, pad = lora_out_705_pad_0, pad_type = lora_out_705_pad_type_0, strides = var_7139, weight = lora_out_707_weight_0_to_fp16, x = input_525_cast_fp16)[name = tensor("lora_out_707_cast_fp16")]; - tensor value_71_cast_fp16 = add(x = pretrained_out_353_cast_fp16, y = lora_out_707_cast_fp16)[name = tensor("value_71_cast_fp16")]; - tensor var_7148 = const()[name = tensor("op_7148"), val = tensor([1, 20, 64, -1])]; - tensor var_7149_cast_fp16 = reshape(shape = var_7148, x = query_71_cast_fp16)[name = tensor("op_7149_cast_fp16")]; - tensor var_7150_to_fp16 = const()[name = tensor("op_7150_to_fp16"), val = tensor(0x1p-3)]; - tensor var_7151_cast_fp16 = mul(x = var_7149_cast_fp16, y = var_7150_to_fp16)[name = tensor("op_7151_cast_fp16")]; - tensor var_7152 = const()[name = tensor("op_7152"), val = tensor([1, 20, 64, -1])]; - tensor var_7153_cast_fp16 = reshape(shape = var_7152, x = key_71_cast_fp16)[name = tensor("op_7153_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_7151_cast_fp16, y = var_7153_cast_fp16)[name = tensor("mh_w_107_cast_fp16")]; - tensor var_7156_cast_fp16 = softmax(axis = var_6874, x = mh_w_107_cast_fp16)[name = tensor("op_7156_cast_fp16")]; - tensor var_7157 = const()[name = tensor("op_7157"), val = tensor([1, 20, 64, -1])]; - tensor var_7158_cast_fp16 = reshape(shape = var_7157, x = value_71_cast_fp16)[name = tensor("op_7158_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_7158_cast_fp16, y = var_7156_cast_fp16)[name = tensor("attn_71_cast_fp16")]; - tensor var_7161 = const()[name = tensor("op_7161"), val = tensor([1, 1280, 1, -1])]; - tensor input_527_cast_fp16 = reshape(shape = var_7161, x = attn_71_cast_fp16)[name = tensor("input_527_cast_fp16")]; - tensor var_7168 = const()[name = tensor("op_7168"), val = tensor([1, 1])]; - tensor var_7170 = const()[name = tensor("op_7170"), val = tensor([1, 1])]; - tensor pretrained_out_355_pad_type_0 = const()[name = tensor("pretrained_out_355_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_355_pad_0 = const()[name = tensor("pretrained_out_355_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383645376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384464640))), name = tensor("layers_17_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_17_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384464768)))]; - tensor pretrained_out_355_cast_fp16 = conv(bias = layers_17_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_7170, groups = var_6881, pad = pretrained_out_355_pad_0, pad_type = pretrained_out_355_pad_type_0, strides = var_7168, weight = layers_17_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_527_cast_fp16)[name = tensor("pretrained_out_355_cast_fp16")]; - tensor var_7174 = const()[name = tensor("op_7174"), val = tensor([1, 1])]; - tensor var_7176 = const()[name = tensor("op_7176"), val = tensor([1, 1])]; - tensor input_529_pad_type_0 = const()[name = tensor("input_529_pad_type_0"), val = tensor("custom")]; - tensor input_529_pad_0 = const()[name = tensor("input_529_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384467392)))]; - tensor input_529_cast_fp16 = conv(dilations = var_7176, groups = var_6881, pad = input_529_pad_0, pad_type = input_529_pad_type_0, strides = var_7174, weight = layers_17_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_527_cast_fp16)[name = tensor("input_529_cast_fp16")]; - tensor var_7180 = const()[name = tensor("op_7180"), val = tensor([1, 1])]; - tensor var_7182 = const()[name = tensor("op_7182"), val = tensor([1, 1])]; - tensor lora_out_709_pad_type_0 = const()[name = tensor("lora_out_709_pad_type_0"), val = tensor("custom")]; - tensor lora_out_709_pad_0 = const()[name = tensor("lora_out_709_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_711_weight_0_to_fp16 = const()[name = tensor("lora_out_711_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384508416)))]; - tensor lora_out_711_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7182, groups = var_6881, pad = lora_out_709_pad_0, pad_type = lora_out_709_pad_type_0, strides = var_7180, weight = lora_out_711_weight_0_to_fp16, x = input_529_cast_fp16)[name = tensor("lora_out_711_cast_fp16")]; - tensor obj_215_cast_fp16 = add(x = pretrained_out_355_cast_fp16, y = lora_out_711_cast_fp16)[name = tensor("obj_215_cast_fp16")]; - tensor inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = obj_215_cast_fp16)[name = tensor("inputs_107_cast_fp16")]; - tensor var_7191 = const()[name = tensor("op_7191"), val = tensor([1])]; - tensor channels_mean_107_cast_fp16 = reduce_mean(axes = var_7191, keep_dims = var_6882, 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_7195 = const()[name = tensor("op_7195"), val = tensor([1])]; - tensor var_7196_cast_fp16 = reduce_mean(axes = var_7195, keep_dims = var_6882, x = zero_mean_sq_107_cast_fp16)[name = tensor("op_7196_cast_fp16")]; - tensor var_7197_to_fp16 = const()[name = tensor("op_7197_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_7198_cast_fp16 = add(x = var_7196_cast_fp16, y = var_7197_to_fp16)[name = tensor("op_7198_cast_fp16")]; - tensor denom_107_epsilon_0 = const()[name = tensor("denom_107_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_107_cast_fp16 = rsqrt(epsilon = denom_107_epsilon_0, x = var_7198_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_531_gamma_0_to_fp16 = const()[name = tensor("input_531_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384549440)))]; - tensor input_531_beta_0_to_fp16 = const()[name = tensor("input_531_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384552064)))]; - tensor input_531_epsilon_0_to_fp16 = const()[name = tensor("input_531_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_531_cast_fp16 = batch_norm(beta = input_531_beta_0_to_fp16, epsilon = input_531_epsilon_0_to_fp16, gamma = input_531_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_531_cast_fp16")]; - tensor var_7212 = const()[name = tensor("op_7212"), val = tensor([1, 1])]; - tensor var_7214 = const()[name = tensor("op_7214"), val = tensor([1, 1])]; - tensor pretrained_out_357_pad_type_0 = const()[name = tensor("pretrained_out_357_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_357_pad_0 = const()[name = tensor("pretrained_out_357_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384554688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387831552))), name = tensor("layers_17_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_17_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_17_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387831680)))]; - tensor pretrained_out_357_cast_fp16 = conv(bias = layers_17_fc1_pretrained_bias_to_fp16, dilations = var_7214, groups = var_6881, pad = pretrained_out_357_pad_0, pad_type = pretrained_out_357_pad_type_0, strides = var_7212, weight = layers_17_fc1_pretrained_weight_to_fp16_palettized, x = input_531_cast_fp16)[name = tensor("pretrained_out_357_cast_fp16")]; - tensor var_7218 = const()[name = tensor("op_7218"), val = tensor([1, 1])]; - tensor var_7220 = const()[name = tensor("op_7220"), val = tensor([1, 1])]; - tensor input_533_pad_type_0 = const()[name = tensor("input_533_pad_type_0"), val = tensor("custom")]; - tensor input_533_pad_0 = const()[name = tensor("input_533_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_17_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387841984)))]; - tensor input_533_cast_fp16 = conv(dilations = var_7220, groups = var_6881, pad = input_533_pad_0, pad_type = input_533_pad_type_0, strides = var_7218, weight = layers_17_fc1_loraA_weight_to_fp16, x = input_531_cast_fp16)[name = tensor("input_533_cast_fp16")]; - tensor var_7224 = const()[name = tensor("op_7224"), val = tensor([1, 1])]; - tensor var_7226 = const()[name = tensor("op_7226"), val = tensor([1, 1])]; - tensor lora_out_713_pad_type_0 = const()[name = tensor("lora_out_713_pad_type_0"), val = tensor("custom")]; - tensor lora_out_713_pad_0 = const()[name = tensor("lora_out_713_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_715_weight_0_to_fp16 = const()[name = tensor("lora_out_715_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387883008)))]; - tensor lora_out_715_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_7226, groups = var_6881, pad = lora_out_713_pad_0, pad_type = lora_out_713_pad_type_0, strides = var_7224, weight = lora_out_715_weight_0_to_fp16, x = input_533_cast_fp16)[name = tensor("lora_out_715_cast_fp16")]; - tensor input_535_cast_fp16 = add(x = pretrained_out_357_cast_fp16, y = lora_out_715_cast_fp16)[name = tensor("input_535_cast_fp16")]; - tensor input_537_mode_0 = const()[name = tensor("input_537_mode_0"), val = tensor("EXACT")]; - tensor input_537_cast_fp16 = gelu(mode = input_537_mode_0, x = input_535_cast_fp16)[name = tensor("input_537_cast_fp16")]; - tensor var_7238 = const()[name = tensor("op_7238"), val = tensor([1, 1])]; - tensor var_7240 = const()[name = tensor("op_7240"), val = tensor([1, 1])]; - tensor pretrained_out_359_pad_type_0 = const()[name = tensor("pretrained_out_359_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_359_pad_0 = const()[name = tensor("pretrained_out_359_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(388046912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391323776))), name = tensor("layers_17_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_17_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_17_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391323904)))]; - tensor pretrained_out_359_cast_fp16 = conv(bias = layers_17_fc2_pretrained_bias_to_fp16, dilations = var_7240, groups = var_6881, pad = pretrained_out_359_pad_0, pad_type = pretrained_out_359_pad_type_0, strides = var_7238, weight = layers_17_fc2_pretrained_weight_to_fp16_palettized, x = input_537_cast_fp16)[name = tensor("pretrained_out_359_cast_fp16")]; - tensor var_7244 = const()[name = tensor("op_7244"), val = tensor([1, 1])]; - tensor var_7246 = const()[name = tensor("op_7246"), val = tensor([1, 1])]; - tensor input_539_pad_type_0 = const()[name = tensor("input_539_pad_type_0"), val = tensor("custom")]; - tensor input_539_pad_0 = const()[name = tensor("input_539_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_17_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_17_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391326528)))]; - tensor input_539_cast_fp16 = conv(dilations = var_7246, groups = var_6881, pad = input_539_pad_0, pad_type = input_539_pad_type_0, strides = var_7244, weight = layers_17_fc2_loraA_weight_to_fp16, x = input_537_cast_fp16)[name = tensor("input_539_cast_fp16")]; - tensor var_7250 = const()[name = tensor("op_7250"), val = tensor([1, 1])]; - tensor var_7252 = const()[name = tensor("op_7252"), val = tensor([1, 1])]; - tensor lora_out_717_pad_type_0 = const()[name = tensor("lora_out_717_pad_type_0"), val = tensor("custom")]; - tensor lora_out_717_pad_0 = const()[name = tensor("lora_out_717_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_719_weight_0_to_fp16 = const()[name = tensor("lora_out_719_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391490432)))]; - tensor lora_out_719_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7252, groups = var_6881, pad = lora_out_717_pad_0, pad_type = lora_out_717_pad_type_0, strides = var_7250, weight = lora_out_719_weight_0_to_fp16, x = input_539_cast_fp16)[name = tensor("lora_out_719_cast_fp16")]; - tensor hidden_states_37_cast_fp16 = add(x = pretrained_out_359_cast_fp16, y = lora_out_719_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_7268 = const()[name = tensor("op_7268"), val = tensor(3)]; - tensor var_7275 = const()[name = tensor("op_7275"), val = tensor(1)]; - tensor var_7276 = const()[name = tensor("op_7276"), val = tensor(true)]; - tensor var_7288 = const()[name = tensor("op_7288"), val = tensor([1])]; - tensor channels_mean_109_cast_fp16 = reduce_mean(axes = var_7288, keep_dims = var_7276, 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_7292 = const()[name = tensor("op_7292"), val = tensor([1])]; - tensor var_7293_cast_fp16 = reduce_mean(axes = var_7292, keep_dims = var_7276, x = zero_mean_sq_109_cast_fp16)[name = tensor("op_7293_cast_fp16")]; - tensor var_7294_to_fp16 = const()[name = tensor("op_7294_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_7295_cast_fp16 = add(x = var_7293_cast_fp16, y = var_7294_to_fp16)[name = tensor("op_7295_cast_fp16")]; - tensor denom_109_epsilon_0 = const()[name = tensor("denom_109_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_109_cast_fp16 = rsqrt(epsilon = denom_109_epsilon_0, x = var_7295_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_217_gamma_0_to_fp16 = const()[name = tensor("obj_217_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391531456)))]; - tensor obj_217_beta_0_to_fp16 = const()[name = tensor("obj_217_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391534080)))]; - tensor obj_217_epsilon_0_to_fp16 = const()[name = tensor("obj_217_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_217_cast_fp16 = batch_norm(beta = obj_217_beta_0_to_fp16, epsilon = obj_217_epsilon_0_to_fp16, gamma = obj_217_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_217_cast_fp16")]; - tensor var_7313 = const()[name = tensor("op_7313"), val = tensor([1, 1])]; - tensor var_7315 = const()[name = tensor("op_7315"), val = tensor([1, 1])]; - tensor pretrained_out_361_pad_type_0 = const()[name = tensor("pretrained_out_361_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_361_pad_0 = const()[name = tensor("pretrained_out_361_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391536704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(392355968))), name = tensor("layers_18_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_18_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(392356096)))]; - tensor pretrained_out_361_cast_fp16 = conv(bias = layers_18_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_7315, groups = var_7275, pad = pretrained_out_361_pad_0, pad_type = pretrained_out_361_pad_type_0, strides = var_7313, weight = layers_18_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_217_cast_fp16)[name = tensor("pretrained_out_361_cast_fp16")]; - tensor var_7319 = const()[name = tensor("op_7319"), val = tensor([1, 1])]; - tensor var_7321 = const()[name = tensor("op_7321"), val = tensor([1, 1])]; - tensor input_541_pad_type_0 = const()[name = tensor("input_541_pad_type_0"), val = tensor("custom")]; - tensor input_541_pad_0 = const()[name = tensor("input_541_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(392358720)))]; - tensor input_541_cast_fp16 = conv(dilations = var_7321, groups = var_7275, pad = input_541_pad_0, pad_type = input_541_pad_type_0, strides = var_7319, weight = layers_18_self_attn_q_proj_loraA_weight_to_fp16, x = obj_217_cast_fp16)[name = tensor("input_541_cast_fp16")]; - tensor var_7325 = const()[name = tensor("op_7325"), val = tensor([1, 1])]; - tensor var_7327 = const()[name = tensor("op_7327"), val = tensor([1, 1])]; - tensor lora_out_721_pad_type_0 = const()[name = tensor("lora_out_721_pad_type_0"), val = tensor("custom")]; - tensor lora_out_721_pad_0 = const()[name = tensor("lora_out_721_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_723_weight_0_to_fp16 = const()[name = tensor("lora_out_723_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(392399744)))]; - tensor lora_out_723_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7327, groups = var_7275, pad = lora_out_721_pad_0, pad_type = lora_out_721_pad_type_0, strides = var_7325, weight = lora_out_723_weight_0_to_fp16, x = input_541_cast_fp16)[name = tensor("lora_out_723_cast_fp16")]; - tensor query_73_cast_fp16 = add(x = pretrained_out_361_cast_fp16, y = lora_out_723_cast_fp16)[name = tensor("query_73_cast_fp16")]; - tensor var_7337 = const()[name = tensor("op_7337"), val = tensor([1, 1])]; - tensor var_7339 = const()[name = tensor("op_7339"), val = tensor([1, 1])]; - tensor pretrained_out_363_pad_type_0 = const()[name = tensor("pretrained_out_363_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_363_pad_0 = const()[name = tensor("pretrained_out_363_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(392440768))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(393260032))), name = tensor("layers_18_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_363_cast_fp16 = conv(dilations = var_7339, groups = var_7275, pad = pretrained_out_363_pad_0, pad_type = pretrained_out_363_pad_type_0, strides = var_7337, weight = layers_18_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_217_cast_fp16)[name = tensor("pretrained_out_363_cast_fp16")]; - tensor var_7343 = const()[name = tensor("op_7343"), val = tensor([1, 1])]; - tensor var_7345 = const()[name = tensor("op_7345"), val = tensor([1, 1])]; - tensor input_543_pad_type_0 = const()[name = tensor("input_543_pad_type_0"), val = tensor("custom")]; - tensor input_543_pad_0 = const()[name = tensor("input_543_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(393260160)))]; - tensor input_543_cast_fp16 = conv(dilations = var_7345, groups = var_7275, pad = input_543_pad_0, pad_type = input_543_pad_type_0, strides = var_7343, weight = layers_18_self_attn_k_proj_loraA_weight_to_fp16, x = obj_217_cast_fp16)[name = tensor("input_543_cast_fp16")]; - tensor var_7349 = const()[name = tensor("op_7349"), val = tensor([1, 1])]; - tensor var_7351 = const()[name = tensor("op_7351"), val = tensor([1, 1])]; - tensor lora_out_725_pad_type_0 = const()[name = tensor("lora_out_725_pad_type_0"), val = tensor("custom")]; - tensor lora_out_725_pad_0 = const()[name = tensor("lora_out_725_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_727_weight_0_to_fp16 = const()[name = tensor("lora_out_727_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(393301184)))]; - tensor lora_out_727_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7351, groups = var_7275, pad = lora_out_725_pad_0, pad_type = lora_out_725_pad_type_0, strides = var_7349, weight = lora_out_727_weight_0_to_fp16, x = input_543_cast_fp16)[name = tensor("lora_out_727_cast_fp16")]; - tensor current_key_37_cast_fp16 = add(x = pretrained_out_363_cast_fp16, y = lora_out_727_cast_fp16)[name = tensor("current_key_37_cast_fp16")]; - tensor var_7362 = const()[name = tensor("op_7362"), val = tensor([1, 1])]; - tensor var_7364 = const()[name = tensor("op_7364"), val = tensor([1, 1])]; - tensor pretrained_out_365_pad_type_0 = const()[name = tensor("pretrained_out_365_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_365_pad_0 = const()[name = tensor("pretrained_out_365_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(393342208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394161472))), name = tensor("layers_18_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_18_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394161600)))]; - tensor pretrained_out_365_cast_fp16 = conv(bias = layers_18_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_7364, groups = var_7275, pad = pretrained_out_365_pad_0, pad_type = pretrained_out_365_pad_type_0, strides = var_7362, weight = layers_18_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_217_cast_fp16)[name = tensor("pretrained_out_365_cast_fp16")]; - tensor var_7368 = const()[name = tensor("op_7368"), val = tensor([1, 1])]; - tensor var_7370 = const()[name = tensor("op_7370"), val = tensor([1, 1])]; - tensor input_545_pad_type_0 = const()[name = tensor("input_545_pad_type_0"), val = tensor("custom")]; - tensor input_545_pad_0 = const()[name = tensor("input_545_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394164224)))]; - tensor input_545_cast_fp16 = conv(dilations = var_7370, groups = var_7275, pad = input_545_pad_0, pad_type = input_545_pad_type_0, strides = var_7368, weight = layers_18_self_attn_v_proj_loraA_weight_to_fp16, x = obj_217_cast_fp16)[name = tensor("input_545_cast_fp16")]; - tensor var_7374 = const()[name = tensor("op_7374"), val = tensor([1, 1])]; - tensor var_7376 = const()[name = tensor("op_7376"), val = tensor([1, 1])]; - tensor lora_out_729_pad_type_0 = const()[name = tensor("lora_out_729_pad_type_0"), val = tensor("custom")]; - tensor lora_out_729_pad_0 = const()[name = tensor("lora_out_729_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_731_weight_0_to_fp16 = const()[name = tensor("lora_out_731_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394205248)))]; - tensor lora_out_731_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7376, groups = var_7275, pad = lora_out_729_pad_0, pad_type = lora_out_729_pad_type_0, strides = var_7374, weight = lora_out_731_weight_0_to_fp16, x = input_545_cast_fp16)[name = tensor("lora_out_731_cast_fp16")]; - tensor current_value_37_cast_fp16 = add(x = pretrained_out_365_cast_fp16, y = lora_out_731_cast_fp16)[name = tensor("current_value_37_cast_fp16")]; - tensor var_7386_cast_fp16 = mul(x = current_key_37_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_7386_cast_fp16")]; - tensor var_7388_cast_fp16 = mul(x = var_103_cast_fp16_18, y = var_295_cast_fp16)[name = tensor("op_7388_cast_fp16")]; - tensor key_73_cast_fp16 = add(x = var_7386_cast_fp16, y = var_7388_cast_fp16)[name = tensor("key_73_cast_fp16")]; - tensor var_7390_cast_fp16 = mul(x = current_value_37_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_7390_cast_fp16")]; - tensor var_7392_cast_fp16 = mul(x = var_138_cast_fp16_18, y = var_295_cast_fp16)[name = tensor("op_7392_cast_fp16")]; - tensor value_73_cast_fp16 = add(x = var_7390_cast_fp16, y = var_7392_cast_fp16)[name = tensor("value_73_cast_fp16")]; - tensor var_7395 = const()[name = tensor("op_7395"), val = tensor([1, 20, 64, -1])]; - tensor var_7396_cast_fp16 = reshape(shape = var_7395, x = query_73_cast_fp16)[name = tensor("op_7396_cast_fp16")]; - tensor var_7397_to_fp16 = const()[name = tensor("op_7397_to_fp16"), val = tensor(0x1p-3)]; - tensor var_7398_cast_fp16 = mul(x = var_7396_cast_fp16, y = var_7397_to_fp16)[name = tensor("op_7398_cast_fp16")]; - tensor var_7399 = const()[name = tensor("op_7399"), val = tensor([1, 20, 64, -1])]; - tensor var_7400_cast_fp16 = reshape(shape = var_7399, x = key_73_cast_fp16)[name = tensor("op_7400_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_7398_cast_fp16, y = var_7400_cast_fp16)[name = tensor("mh_w_109_cast_fp16")]; - tensor mh_w_111_cast_fp16 = add(x = mh_w_109_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_111_cast_fp16")]; - tensor var_7408_cast_fp16 = softmax(axis = var_7268, x = mh_w_111_cast_fp16)[name = tensor("op_7408_cast_fp16")]; - tensor var_7409 = const()[name = tensor("op_7409"), val = tensor([1, 20, 64, -1])]; - tensor var_7410_cast_fp16 = reshape(shape = var_7409, x = value_73_cast_fp16)[name = tensor("op_7410_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_7410_cast_fp16, y = var_7408_cast_fp16)[name = tensor("attn_73_cast_fp16")]; - tensor var_7413 = const()[name = tensor("op_7413"), val = tensor([1, 1280, 1, -1])]; - tensor input_547_cast_fp16 = reshape(shape = var_7413, x = attn_73_cast_fp16)[name = tensor("input_547_cast_fp16")]; - tensor var_7420 = const()[name = tensor("op_7420"), val = tensor([1, 1])]; - tensor var_7422 = const()[name = tensor("op_7422"), val = tensor([1, 1])]; - tensor pretrained_out_367_pad_type_0 = const()[name = tensor("pretrained_out_367_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_367_pad_0 = const()[name = tensor("pretrained_out_367_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394246272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395065536))), name = tensor("layers_18_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_18_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395065664)))]; - tensor pretrained_out_367_cast_fp16 = conv(bias = layers_18_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_7422, groups = var_7275, pad = pretrained_out_367_pad_0, pad_type = pretrained_out_367_pad_type_0, strides = var_7420, weight = layers_18_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_547_cast_fp16)[name = tensor("pretrained_out_367_cast_fp16")]; - tensor var_7426 = const()[name = tensor("op_7426"), val = tensor([1, 1])]; - tensor var_7428 = const()[name = tensor("op_7428"), val = tensor([1, 1])]; - tensor input_549_pad_type_0 = const()[name = tensor("input_549_pad_type_0"), val = tensor("custom")]; - tensor input_549_pad_0 = const()[name = tensor("input_549_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395068288)))]; - tensor input_549_cast_fp16 = conv(dilations = var_7428, groups = var_7275, pad = input_549_pad_0, pad_type = input_549_pad_type_0, strides = var_7426, weight = layers_18_self_attn_o_proj_loraA_weight_to_fp16, x = input_547_cast_fp16)[name = tensor("input_549_cast_fp16")]; - tensor var_7432 = const()[name = tensor("op_7432"), val = tensor([1, 1])]; - tensor var_7434 = const()[name = tensor("op_7434"), val = tensor([1, 1])]; - tensor lora_out_733_pad_type_0 = const()[name = tensor("lora_out_733_pad_type_0"), val = tensor("custom")]; - tensor lora_out_733_pad_0 = const()[name = tensor("lora_out_733_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_735_weight_0_to_fp16 = const()[name = tensor("lora_out_735_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395109312)))]; - tensor lora_out_735_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7434, groups = var_7275, pad = lora_out_733_pad_0, pad_type = lora_out_733_pad_type_0, strides = var_7432, weight = lora_out_735_weight_0_to_fp16, x = input_549_cast_fp16)[name = tensor("lora_out_735_cast_fp16")]; - tensor obj_223_cast_fp16 = add(x = pretrained_out_367_cast_fp16, y = lora_out_735_cast_fp16)[name = tensor("obj_223_cast_fp16")]; - tensor inputs_111_cast_fp16 = add(x = inputs_109_cast_fp16, y = obj_223_cast_fp16)[name = tensor("inputs_111_cast_fp16")]; - tensor var_7447 = const()[name = tensor("op_7447"), val = tensor([1])]; - tensor channels_mean_111_cast_fp16 = reduce_mean(axes = var_7447, keep_dims = var_7276, 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_7451 = const()[name = tensor("op_7451"), val = tensor([1])]; - tensor var_7452_cast_fp16 = reduce_mean(axes = var_7451, keep_dims = var_7276, x = zero_mean_sq_111_cast_fp16)[name = tensor("op_7452_cast_fp16")]; - tensor var_7453_to_fp16 = const()[name = tensor("op_7453_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_7454_cast_fp16 = add(x = var_7452_cast_fp16, y = var_7453_to_fp16)[name = tensor("op_7454_cast_fp16")]; - tensor denom_111_epsilon_0 = const()[name = tensor("denom_111_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_111_cast_fp16 = rsqrt(epsilon = denom_111_epsilon_0, x = var_7454_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_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(395150336)))]; - 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(395152960)))]; - 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_111_cast_fp16)[name = tensor("obj_225_cast_fp16")]; - tensor var_7472 = const()[name = tensor("op_7472"), val = tensor([1, 1])]; - tensor var_7474 = const()[name = tensor("op_7474"), val = tensor([1, 1])]; - tensor pretrained_out_369_pad_type_0 = const()[name = tensor("pretrained_out_369_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_369_pad_0 = const()[name = tensor("pretrained_out_369_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395155584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395974848))), name = tensor("layers_18_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_18_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395974976)))]; - tensor pretrained_out_369_cast_fp16 = conv(bias = layers_18_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_7474, groups = var_7275, pad = pretrained_out_369_pad_0, pad_type = pretrained_out_369_pad_type_0, strides = var_7472, weight = layers_18_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_225_cast_fp16)[name = tensor("pretrained_out_369_cast_fp16")]; - tensor var_7478 = const()[name = tensor("op_7478"), val = tensor([1, 1])]; - tensor var_7480 = const()[name = tensor("op_7480"), val = tensor([1, 1])]; - tensor input_551_pad_type_0 = const()[name = tensor("input_551_pad_type_0"), val = tensor("custom")]; - tensor input_551_pad_0 = const()[name = tensor("input_551_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395977600)))]; - tensor input_551_cast_fp16 = conv(dilations = var_7480, groups = var_7275, pad = input_551_pad_0, pad_type = input_551_pad_type_0, strides = var_7478, weight = layers_18_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("input_551_cast_fp16")]; - tensor var_7484 = const()[name = tensor("op_7484"), val = tensor([1, 1])]; - tensor var_7486 = const()[name = tensor("op_7486"), val = tensor([1, 1])]; - tensor lora_out_737_pad_type_0 = const()[name = tensor("lora_out_737_pad_type_0"), val = tensor("custom")]; - tensor lora_out_737_pad_0 = const()[name = tensor("lora_out_737_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_739_weight_0_to_fp16 = const()[name = tensor("lora_out_739_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396018624)))]; - tensor lora_out_739_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7486, groups = var_7275, pad = lora_out_737_pad_0, pad_type = lora_out_737_pad_type_0, strides = var_7484, weight = lora_out_739_weight_0_to_fp16, x = input_551_cast_fp16)[name = tensor("lora_out_739_cast_fp16")]; - tensor query_75_cast_fp16 = add(x = pretrained_out_369_cast_fp16, y = lora_out_739_cast_fp16)[name = tensor("query_75_cast_fp16")]; - tensor var_7496 = const()[name = tensor("op_7496"), val = tensor([1, 1])]; - tensor var_7498 = const()[name = tensor("op_7498"), val = tensor([1, 1])]; - tensor pretrained_out_371_pad_type_0 = const()[name = tensor("pretrained_out_371_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_371_pad_0 = const()[name = tensor("pretrained_out_371_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396059648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396878912))), name = tensor("layers_18_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_371_cast_fp16 = conv(dilations = var_7498, groups = var_7275, pad = pretrained_out_371_pad_0, pad_type = pretrained_out_371_pad_type_0, strides = var_7496, weight = layers_18_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_371_cast_fp16")]; - tensor var_7502 = const()[name = tensor("op_7502"), val = tensor([1, 1])]; - tensor var_7504 = const()[name = tensor("op_7504"), val = tensor([1, 1])]; - tensor input_553_pad_type_0 = const()[name = tensor("input_553_pad_type_0"), val = tensor("custom")]; - tensor input_553_pad_0 = const()[name = tensor("input_553_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396879040)))]; - tensor input_553_cast_fp16 = conv(dilations = var_7504, groups = var_7275, pad = input_553_pad_0, pad_type = input_553_pad_type_0, strides = var_7502, weight = layers_18_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_553_cast_fp16")]; - tensor var_7508 = const()[name = tensor("op_7508"), val = tensor([1, 1])]; - tensor var_7510 = const()[name = tensor("op_7510"), val = tensor([1, 1])]; - tensor lora_out_741_pad_type_0 = const()[name = tensor("lora_out_741_pad_type_0"), val = tensor("custom")]; - tensor lora_out_741_pad_0 = const()[name = tensor("lora_out_741_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_743_weight_0_to_fp16 = const()[name = tensor("lora_out_743_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396920064)))]; - tensor lora_out_743_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7510, groups = var_7275, pad = lora_out_741_pad_0, pad_type = lora_out_741_pad_type_0, strides = var_7508, weight = lora_out_743_weight_0_to_fp16, x = input_553_cast_fp16)[name = tensor("lora_out_743_cast_fp16")]; - tensor key_75_cast_fp16 = add(x = pretrained_out_371_cast_fp16, y = lora_out_743_cast_fp16)[name = tensor("key_75_cast_fp16")]; - tensor var_7521 = const()[name = tensor("op_7521"), val = tensor([1, 1])]; - tensor var_7523 = const()[name = tensor("op_7523"), val = tensor([1, 1])]; - tensor pretrained_out_373_pad_type_0 = const()[name = tensor("pretrained_out_373_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_373_pad_0 = const()[name = tensor("pretrained_out_373_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396961088))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(397780352))), name = tensor("layers_18_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_18_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(397780480)))]; - tensor pretrained_out_373_cast_fp16 = conv(bias = layers_18_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_7523, groups = var_7275, pad = pretrained_out_373_pad_0, pad_type = pretrained_out_373_pad_type_0, strides = var_7521, weight = layers_18_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_373_cast_fp16")]; - tensor var_7527 = const()[name = tensor("op_7527"), val = tensor([1, 1])]; - tensor var_7529 = const()[name = tensor("op_7529"), val = tensor([1, 1])]; - tensor input_555_pad_type_0 = const()[name = tensor("input_555_pad_type_0"), val = tensor("custom")]; - tensor input_555_pad_0 = const()[name = tensor("input_555_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(397783104)))]; - tensor input_555_cast_fp16 = conv(dilations = var_7529, groups = var_7275, pad = input_555_pad_0, pad_type = input_555_pad_type_0, strides = var_7527, weight = layers_18_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_555_cast_fp16")]; - tensor var_7533 = const()[name = tensor("op_7533"), val = tensor([1, 1])]; - tensor var_7535 = const()[name = tensor("op_7535"), val = tensor([1, 1])]; - tensor lora_out_745_pad_type_0 = const()[name = tensor("lora_out_745_pad_type_0"), val = tensor("custom")]; - tensor lora_out_745_pad_0 = const()[name = tensor("lora_out_745_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_747_weight_0_to_fp16 = const()[name = tensor("lora_out_747_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(397824128)))]; - tensor lora_out_747_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7535, groups = var_7275, pad = lora_out_745_pad_0, pad_type = lora_out_745_pad_type_0, strides = var_7533, weight = lora_out_747_weight_0_to_fp16, x = input_555_cast_fp16)[name = tensor("lora_out_747_cast_fp16")]; - tensor value_75_cast_fp16 = add(x = pretrained_out_373_cast_fp16, y = lora_out_747_cast_fp16)[name = tensor("value_75_cast_fp16")]; - tensor var_7542 = const()[name = tensor("op_7542"), val = tensor([1, 20, 64, -1])]; - tensor var_7543_cast_fp16 = reshape(shape = var_7542, x = query_75_cast_fp16)[name = tensor("op_7543_cast_fp16")]; - tensor var_7544_to_fp16 = const()[name = tensor("op_7544_to_fp16"), val = tensor(0x1p-3)]; - tensor var_7545_cast_fp16 = mul(x = var_7543_cast_fp16, y = var_7544_to_fp16)[name = tensor("op_7545_cast_fp16")]; - tensor var_7546 = const()[name = tensor("op_7546"), val = tensor([1, 20, 64, -1])]; - tensor var_7547_cast_fp16 = reshape(shape = var_7546, x = key_75_cast_fp16)[name = tensor("op_7547_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_7545_cast_fp16, y = var_7547_cast_fp16)[name = tensor("mh_w_113_cast_fp16")]; - tensor var_7550_cast_fp16 = softmax(axis = var_7268, x = mh_w_113_cast_fp16)[name = tensor("op_7550_cast_fp16")]; - tensor var_7551 = const()[name = tensor("op_7551"), val = tensor([1, 20, 64, -1])]; - tensor var_7552_cast_fp16 = reshape(shape = var_7551, x = value_75_cast_fp16)[name = tensor("op_7552_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_7552_cast_fp16, y = var_7550_cast_fp16)[name = tensor("attn_75_cast_fp16")]; - tensor var_7555 = const()[name = tensor("op_7555"), val = tensor([1, 1280, 1, -1])]; - tensor input_557_cast_fp16 = reshape(shape = var_7555, x = attn_75_cast_fp16)[name = tensor("input_557_cast_fp16")]; - tensor var_7562 = const()[name = tensor("op_7562"), val = tensor([1, 1])]; - tensor var_7564 = const()[name = tensor("op_7564"), val = tensor([1, 1])]; - tensor pretrained_out_375_pad_type_0 = const()[name = tensor("pretrained_out_375_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_375_pad_0 = const()[name = tensor("pretrained_out_375_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(397865152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398684416))), name = tensor("layers_18_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_18_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398684544)))]; - tensor pretrained_out_375_cast_fp16 = conv(bias = layers_18_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_7564, groups = var_7275, pad = pretrained_out_375_pad_0, pad_type = pretrained_out_375_pad_type_0, strides = var_7562, weight = layers_18_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_557_cast_fp16)[name = tensor("pretrained_out_375_cast_fp16")]; - tensor var_7568 = const()[name = tensor("op_7568"), val = tensor([1, 1])]; - tensor var_7570 = const()[name = tensor("op_7570"), val = tensor([1, 1])]; - tensor input_559_pad_type_0 = const()[name = tensor("input_559_pad_type_0"), val = tensor("custom")]; - tensor input_559_pad_0 = const()[name = tensor("input_559_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398687168)))]; - tensor input_559_cast_fp16 = conv(dilations = var_7570, groups = var_7275, pad = input_559_pad_0, pad_type = input_559_pad_type_0, strides = var_7568, weight = layers_18_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_557_cast_fp16)[name = tensor("input_559_cast_fp16")]; - tensor var_7574 = const()[name = tensor("op_7574"), val = tensor([1, 1])]; - tensor var_7576 = const()[name = tensor("op_7576"), val = tensor([1, 1])]; - tensor lora_out_749_pad_type_0 = const()[name = tensor("lora_out_749_pad_type_0"), val = tensor("custom")]; - tensor lora_out_749_pad_0 = const()[name = tensor("lora_out_749_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_751_weight_0_to_fp16 = const()[name = tensor("lora_out_751_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398728192)))]; - tensor lora_out_751_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7576, groups = var_7275, pad = lora_out_749_pad_0, pad_type = lora_out_749_pad_type_0, strides = var_7574, weight = lora_out_751_weight_0_to_fp16, x = input_559_cast_fp16)[name = tensor("lora_out_751_cast_fp16")]; - tensor obj_227_cast_fp16 = add(x = pretrained_out_375_cast_fp16, y = lora_out_751_cast_fp16)[name = tensor("obj_227_cast_fp16")]; - tensor inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = obj_227_cast_fp16)[name = tensor("inputs_113_cast_fp16")]; - tensor var_7585 = const()[name = tensor("op_7585"), val = tensor([1])]; - tensor channels_mean_113_cast_fp16 = reduce_mean(axes = var_7585, keep_dims = var_7276, 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_7589 = const()[name = tensor("op_7589"), val = tensor([1])]; - tensor var_7590_cast_fp16 = reduce_mean(axes = var_7589, keep_dims = var_7276, x = zero_mean_sq_113_cast_fp16)[name = tensor("op_7590_cast_fp16")]; - tensor var_7591_to_fp16 = const()[name = tensor("op_7591_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_7592_cast_fp16 = add(x = var_7590_cast_fp16, y = var_7591_to_fp16)[name = tensor("op_7592_cast_fp16")]; - tensor denom_113_epsilon_0 = const()[name = tensor("denom_113_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_113_cast_fp16 = rsqrt(epsilon = denom_113_epsilon_0, x = var_7592_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_561_gamma_0_to_fp16 = const()[name = tensor("input_561_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398769216)))]; - tensor input_561_beta_0_to_fp16 = const()[name = tensor("input_561_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398771840)))]; - tensor input_561_epsilon_0_to_fp16 = const()[name = tensor("input_561_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_561_cast_fp16 = batch_norm(beta = input_561_beta_0_to_fp16, epsilon = input_561_epsilon_0_to_fp16, gamma = input_561_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_561_cast_fp16")]; - tensor var_7606 = const()[name = tensor("op_7606"), val = tensor([1, 1])]; - tensor var_7608 = const()[name = tensor("op_7608"), val = tensor([1, 1])]; - tensor pretrained_out_377_pad_type_0 = const()[name = tensor("pretrained_out_377_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_377_pad_0 = const()[name = tensor("pretrained_out_377_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398774464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402051328))), name = tensor("layers_18_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_18_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_18_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402051456)))]; - tensor pretrained_out_377_cast_fp16 = conv(bias = layers_18_fc1_pretrained_bias_to_fp16, dilations = var_7608, groups = var_7275, pad = pretrained_out_377_pad_0, pad_type = pretrained_out_377_pad_type_0, strides = var_7606, weight = layers_18_fc1_pretrained_weight_to_fp16_palettized, x = input_561_cast_fp16)[name = tensor("pretrained_out_377_cast_fp16")]; - tensor var_7612 = const()[name = tensor("op_7612"), val = tensor([1, 1])]; - tensor var_7614 = const()[name = tensor("op_7614"), val = tensor([1, 1])]; - tensor input_563_pad_type_0 = const()[name = tensor("input_563_pad_type_0"), val = tensor("custom")]; - tensor input_563_pad_0 = const()[name = tensor("input_563_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_18_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402061760)))]; - tensor input_563_cast_fp16 = conv(dilations = var_7614, groups = var_7275, pad = input_563_pad_0, pad_type = input_563_pad_type_0, strides = var_7612, weight = layers_18_fc1_loraA_weight_to_fp16, x = input_561_cast_fp16)[name = tensor("input_563_cast_fp16")]; - tensor var_7618 = const()[name = tensor("op_7618"), val = tensor([1, 1])]; - tensor var_7620 = const()[name = tensor("op_7620"), val = tensor([1, 1])]; - tensor lora_out_753_pad_type_0 = const()[name = tensor("lora_out_753_pad_type_0"), val = tensor("custom")]; - tensor lora_out_753_pad_0 = const()[name = tensor("lora_out_753_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_755_weight_0_to_fp16 = const()[name = tensor("lora_out_755_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402102784)))]; - tensor lora_out_755_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_7620, groups = var_7275, pad = lora_out_753_pad_0, pad_type = lora_out_753_pad_type_0, strides = var_7618, weight = lora_out_755_weight_0_to_fp16, x = input_563_cast_fp16)[name = tensor("lora_out_755_cast_fp16")]; - tensor input_565_cast_fp16 = add(x = pretrained_out_377_cast_fp16, y = lora_out_755_cast_fp16)[name = tensor("input_565_cast_fp16")]; - tensor input_567_mode_0 = const()[name = tensor("input_567_mode_0"), val = tensor("EXACT")]; - tensor input_567_cast_fp16 = gelu(mode = input_567_mode_0, x = input_565_cast_fp16)[name = tensor("input_567_cast_fp16")]; - tensor var_7632 = const()[name = tensor("op_7632"), val = tensor([1, 1])]; - tensor var_7634 = const()[name = tensor("op_7634"), val = tensor([1, 1])]; - tensor pretrained_out_379_pad_type_0 = const()[name = tensor("pretrained_out_379_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_379_pad_0 = const()[name = tensor("pretrained_out_379_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402266688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(405543552))), name = tensor("layers_18_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_18_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_18_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(405543680)))]; - tensor pretrained_out_379_cast_fp16 = conv(bias = layers_18_fc2_pretrained_bias_to_fp16, dilations = var_7634, groups = var_7275, pad = pretrained_out_379_pad_0, pad_type = pretrained_out_379_pad_type_0, strides = var_7632, weight = layers_18_fc2_pretrained_weight_to_fp16_palettized, x = input_567_cast_fp16)[name = tensor("pretrained_out_379_cast_fp16")]; - tensor var_7638 = const()[name = tensor("op_7638"), val = tensor([1, 1])]; - tensor var_7640 = const()[name = tensor("op_7640"), val = tensor([1, 1])]; - tensor input_569_pad_type_0 = const()[name = tensor("input_569_pad_type_0"), val = tensor("custom")]; - tensor input_569_pad_0 = const()[name = tensor("input_569_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_18_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_18_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(405546304)))]; - tensor input_569_cast_fp16 = conv(dilations = var_7640, groups = var_7275, pad = input_569_pad_0, pad_type = input_569_pad_type_0, strides = var_7638, weight = layers_18_fc2_loraA_weight_to_fp16, x = input_567_cast_fp16)[name = tensor("input_569_cast_fp16")]; - tensor var_7644 = const()[name = tensor("op_7644"), val = tensor([1, 1])]; - tensor var_7646 = const()[name = tensor("op_7646"), val = tensor([1, 1])]; - tensor lora_out_757_pad_type_0 = const()[name = tensor("lora_out_757_pad_type_0"), val = tensor("custom")]; - tensor lora_out_757_pad_0 = const()[name = tensor("lora_out_757_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_759_weight_0_to_fp16 = const()[name = tensor("lora_out_759_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(405710208)))]; - tensor lora_out_759_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7646, groups = var_7275, pad = lora_out_757_pad_0, pad_type = lora_out_757_pad_type_0, strides = var_7644, weight = lora_out_759_weight_0_to_fp16, x = input_569_cast_fp16)[name = tensor("lora_out_759_cast_fp16")]; - tensor hidden_states_39_cast_fp16 = add(x = pretrained_out_379_cast_fp16, y = lora_out_759_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_7662 = const()[name = tensor("op_7662"), val = tensor(3)]; - tensor var_7669 = const()[name = tensor("op_7669"), val = tensor(1)]; - tensor var_7670 = const()[name = tensor("op_7670"), val = tensor(true)]; - tensor var_7682 = const()[name = tensor("op_7682"), val = tensor([1])]; - tensor channels_mean_115_cast_fp16 = reduce_mean(axes = var_7682, keep_dims = var_7670, 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_7686 = const()[name = tensor("op_7686"), val = tensor([1])]; - tensor var_7687_cast_fp16 = reduce_mean(axes = var_7686, keep_dims = var_7670, x = zero_mean_sq_115_cast_fp16)[name = tensor("op_7687_cast_fp16")]; - tensor var_7688_to_fp16 = const()[name = tensor("op_7688_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_7689_cast_fp16 = add(x = var_7687_cast_fp16, y = var_7688_to_fp16)[name = tensor("op_7689_cast_fp16")]; - tensor denom_115_epsilon_0 = const()[name = tensor("denom_115_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_115_cast_fp16 = rsqrt(epsilon = denom_115_epsilon_0, x = var_7689_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_229_gamma_0_to_fp16 = const()[name = tensor("obj_229_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(405751232)))]; - tensor obj_229_beta_0_to_fp16 = const()[name = tensor("obj_229_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(405753856)))]; - tensor obj_229_epsilon_0_to_fp16 = const()[name = tensor("obj_229_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_229_cast_fp16 = batch_norm(beta = obj_229_beta_0_to_fp16, epsilon = obj_229_epsilon_0_to_fp16, gamma = obj_229_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_229_cast_fp16")]; - tensor var_7707 = const()[name = tensor("op_7707"), val = tensor([1, 1])]; - tensor var_7709 = const()[name = tensor("op_7709"), val = tensor([1, 1])]; - tensor pretrained_out_381_pad_type_0 = const()[name = tensor("pretrained_out_381_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_381_pad_0 = const()[name = tensor("pretrained_out_381_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(405756480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(406575744))), name = tensor("layers_19_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_19_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(406575872)))]; - tensor pretrained_out_381_cast_fp16 = conv(bias = layers_19_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_7709, groups = var_7669, pad = pretrained_out_381_pad_0, pad_type = pretrained_out_381_pad_type_0, strides = var_7707, weight = layers_19_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_229_cast_fp16)[name = tensor("pretrained_out_381_cast_fp16")]; - tensor var_7713 = const()[name = tensor("op_7713"), val = tensor([1, 1])]; - tensor var_7715 = const()[name = tensor("op_7715"), val = tensor([1, 1])]; - tensor input_571_pad_type_0 = const()[name = tensor("input_571_pad_type_0"), val = tensor("custom")]; - tensor input_571_pad_0 = const()[name = tensor("input_571_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(406578496)))]; - tensor input_571_cast_fp16 = conv(dilations = var_7715, groups = var_7669, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = var_7713, weight = layers_19_self_attn_q_proj_loraA_weight_to_fp16, x = obj_229_cast_fp16)[name = tensor("input_571_cast_fp16")]; - tensor var_7719 = const()[name = tensor("op_7719"), val = tensor([1, 1])]; - tensor var_7721 = const()[name = tensor("op_7721"), val = tensor([1, 1])]; - tensor lora_out_761_pad_type_0 = const()[name = tensor("lora_out_761_pad_type_0"), val = tensor("custom")]; - tensor lora_out_761_pad_0 = const()[name = tensor("lora_out_761_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_763_weight_0_to_fp16 = const()[name = tensor("lora_out_763_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(406619520)))]; - tensor lora_out_763_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7721, groups = var_7669, pad = lora_out_761_pad_0, pad_type = lora_out_761_pad_type_0, strides = var_7719, weight = lora_out_763_weight_0_to_fp16, x = input_571_cast_fp16)[name = tensor("lora_out_763_cast_fp16")]; - tensor query_77_cast_fp16 = add(x = pretrained_out_381_cast_fp16, y = lora_out_763_cast_fp16)[name = tensor("query_77_cast_fp16")]; - tensor var_7731 = const()[name = tensor("op_7731"), val = tensor([1, 1])]; - tensor var_7733 = const()[name = tensor("op_7733"), val = tensor([1, 1])]; - tensor pretrained_out_383_pad_type_0 = const()[name = tensor("pretrained_out_383_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_383_pad_0 = const()[name = tensor("pretrained_out_383_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(406660544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407479808))), name = tensor("layers_19_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_383_cast_fp16 = conv(dilations = var_7733, groups = var_7669, pad = pretrained_out_383_pad_0, pad_type = pretrained_out_383_pad_type_0, strides = var_7731, weight = layers_19_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_229_cast_fp16)[name = tensor("pretrained_out_383_cast_fp16")]; - tensor var_7737 = const()[name = tensor("op_7737"), val = tensor([1, 1])]; - tensor var_7739 = const()[name = tensor("op_7739"), val = tensor([1, 1])]; - tensor input_573_pad_type_0 = const()[name = tensor("input_573_pad_type_0"), val = tensor("custom")]; - tensor input_573_pad_0 = const()[name = tensor("input_573_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407479936)))]; - tensor input_573_cast_fp16 = conv(dilations = var_7739, groups = var_7669, pad = input_573_pad_0, pad_type = input_573_pad_type_0, strides = var_7737, weight = layers_19_self_attn_k_proj_loraA_weight_to_fp16, x = obj_229_cast_fp16)[name = tensor("input_573_cast_fp16")]; - tensor var_7743 = const()[name = tensor("op_7743"), val = tensor([1, 1])]; - tensor var_7745 = const()[name = tensor("op_7745"), val = tensor([1, 1])]; - tensor lora_out_765_pad_type_0 = const()[name = tensor("lora_out_765_pad_type_0"), val = tensor("custom")]; - tensor lora_out_765_pad_0 = const()[name = tensor("lora_out_765_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_767_weight_0_to_fp16 = const()[name = tensor("lora_out_767_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407520960)))]; - tensor lora_out_767_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7745, groups = var_7669, pad = lora_out_765_pad_0, pad_type = lora_out_765_pad_type_0, strides = var_7743, weight = lora_out_767_weight_0_to_fp16, x = input_573_cast_fp16)[name = tensor("lora_out_767_cast_fp16")]; - tensor current_key_39_cast_fp16 = add(x = pretrained_out_383_cast_fp16, y = lora_out_767_cast_fp16)[name = tensor("current_key_39_cast_fp16")]; - tensor var_7756 = const()[name = tensor("op_7756"), val = tensor([1, 1])]; - tensor var_7758 = const()[name = tensor("op_7758"), val = tensor([1, 1])]; - tensor pretrained_out_385_pad_type_0 = const()[name = tensor("pretrained_out_385_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_385_pad_0 = const()[name = tensor("pretrained_out_385_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407561984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408381248))), name = tensor("layers_19_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_19_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408381376)))]; - tensor pretrained_out_385_cast_fp16 = conv(bias = layers_19_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_7758, groups = var_7669, pad = pretrained_out_385_pad_0, pad_type = pretrained_out_385_pad_type_0, strides = var_7756, weight = layers_19_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_229_cast_fp16)[name = tensor("pretrained_out_385_cast_fp16")]; - tensor var_7762 = const()[name = tensor("op_7762"), val = tensor([1, 1])]; - tensor var_7764 = const()[name = tensor("op_7764"), val = tensor([1, 1])]; - tensor input_575_pad_type_0 = const()[name = tensor("input_575_pad_type_0"), val = tensor("custom")]; - tensor input_575_pad_0 = const()[name = tensor("input_575_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408384000)))]; - tensor input_575_cast_fp16 = conv(dilations = var_7764, groups = var_7669, pad = input_575_pad_0, pad_type = input_575_pad_type_0, strides = var_7762, weight = layers_19_self_attn_v_proj_loraA_weight_to_fp16, x = obj_229_cast_fp16)[name = tensor("input_575_cast_fp16")]; - tensor var_7768 = const()[name = tensor("op_7768"), val = tensor([1, 1])]; - tensor var_7770 = const()[name = tensor("op_7770"), val = tensor([1, 1])]; - tensor lora_out_769_pad_type_0 = const()[name = tensor("lora_out_769_pad_type_0"), val = tensor("custom")]; - tensor lora_out_769_pad_0 = const()[name = tensor("lora_out_769_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_771_weight_0_to_fp16 = const()[name = tensor("lora_out_771_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408425024)))]; - tensor lora_out_771_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7770, groups = var_7669, pad = lora_out_769_pad_0, pad_type = lora_out_769_pad_type_0, strides = var_7768, weight = lora_out_771_weight_0_to_fp16, x = input_575_cast_fp16)[name = tensor("lora_out_771_cast_fp16")]; - tensor current_value_39_cast_fp16 = add(x = pretrained_out_385_cast_fp16, y = lora_out_771_cast_fp16)[name = tensor("current_value_39_cast_fp16")]; - tensor var_7780_cast_fp16 = mul(x = current_key_39_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_7780_cast_fp16")]; - tensor var_7782_cast_fp16 = mul(x = var_103_cast_fp16_19, y = var_295_cast_fp16)[name = tensor("op_7782_cast_fp16")]; - tensor key_77_cast_fp16 = add(x = var_7780_cast_fp16, y = var_7782_cast_fp16)[name = tensor("key_77_cast_fp16")]; - tensor var_7784_cast_fp16 = mul(x = current_value_39_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_7784_cast_fp16")]; - tensor var_7786_cast_fp16 = mul(x = var_138_cast_fp16_19, y = var_295_cast_fp16)[name = tensor("op_7786_cast_fp16")]; - tensor value_77_cast_fp16 = add(x = var_7784_cast_fp16, y = var_7786_cast_fp16)[name = tensor("value_77_cast_fp16")]; - tensor var_7789 = const()[name = tensor("op_7789"), val = tensor([1, 20, 64, -1])]; - tensor var_7790_cast_fp16 = reshape(shape = var_7789, x = query_77_cast_fp16)[name = tensor("op_7790_cast_fp16")]; - tensor var_7791_to_fp16 = const()[name = tensor("op_7791_to_fp16"), val = tensor(0x1p-3)]; - tensor var_7792_cast_fp16 = mul(x = var_7790_cast_fp16, y = var_7791_to_fp16)[name = tensor("op_7792_cast_fp16")]; - tensor var_7793 = const()[name = tensor("op_7793"), val = tensor([1, 20, 64, -1])]; - tensor var_7794_cast_fp16 = reshape(shape = var_7793, x = key_77_cast_fp16)[name = tensor("op_7794_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_7792_cast_fp16, y = var_7794_cast_fp16)[name = tensor("mh_w_115_cast_fp16")]; - tensor mh_w_117_cast_fp16 = add(x = mh_w_115_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_117_cast_fp16")]; - tensor var_7802_cast_fp16 = softmax(axis = var_7662, x = mh_w_117_cast_fp16)[name = tensor("op_7802_cast_fp16")]; - tensor var_7803 = const()[name = tensor("op_7803"), val = tensor([1, 20, 64, -1])]; - tensor var_7804_cast_fp16 = reshape(shape = var_7803, x = value_77_cast_fp16)[name = tensor("op_7804_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_7804_cast_fp16, y = var_7802_cast_fp16)[name = tensor("attn_77_cast_fp16")]; - tensor var_7807 = const()[name = tensor("op_7807"), val = tensor([1, 1280, 1, -1])]; - tensor input_577_cast_fp16 = reshape(shape = var_7807, x = attn_77_cast_fp16)[name = tensor("input_577_cast_fp16")]; - tensor var_7814 = const()[name = tensor("op_7814"), val = tensor([1, 1])]; - tensor var_7816 = const()[name = tensor("op_7816"), val = tensor([1, 1])]; - tensor pretrained_out_387_pad_type_0 = const()[name = tensor("pretrained_out_387_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_387_pad_0 = const()[name = tensor("pretrained_out_387_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408466048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409285312))), name = tensor("layers_19_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_19_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409285440)))]; - tensor pretrained_out_387_cast_fp16 = conv(bias = layers_19_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_7816, groups = var_7669, pad = pretrained_out_387_pad_0, pad_type = pretrained_out_387_pad_type_0, strides = var_7814, weight = layers_19_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_577_cast_fp16)[name = tensor("pretrained_out_387_cast_fp16")]; - tensor var_7820 = const()[name = tensor("op_7820"), val = tensor([1, 1])]; - tensor var_7822 = const()[name = tensor("op_7822"), val = tensor([1, 1])]; - tensor input_579_pad_type_0 = const()[name = tensor("input_579_pad_type_0"), val = tensor("custom")]; - tensor input_579_pad_0 = const()[name = tensor("input_579_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409288064)))]; - tensor input_579_cast_fp16 = conv(dilations = var_7822, groups = var_7669, pad = input_579_pad_0, pad_type = input_579_pad_type_0, strides = var_7820, weight = layers_19_self_attn_o_proj_loraA_weight_to_fp16, x = input_577_cast_fp16)[name = tensor("input_579_cast_fp16")]; - tensor var_7826 = const()[name = tensor("op_7826"), val = tensor([1, 1])]; - tensor var_7828 = const()[name = tensor("op_7828"), val = tensor([1, 1])]; - tensor lora_out_773_pad_type_0 = const()[name = tensor("lora_out_773_pad_type_0"), val = tensor("custom")]; - tensor lora_out_773_pad_0 = const()[name = tensor("lora_out_773_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_775_weight_0_to_fp16 = const()[name = tensor("lora_out_775_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409329088)))]; - tensor lora_out_775_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7828, groups = var_7669, pad = lora_out_773_pad_0, pad_type = lora_out_773_pad_type_0, strides = var_7826, weight = lora_out_775_weight_0_to_fp16, x = input_579_cast_fp16)[name = tensor("lora_out_775_cast_fp16")]; - tensor obj_235_cast_fp16 = add(x = pretrained_out_387_cast_fp16, y = lora_out_775_cast_fp16)[name = tensor("obj_235_cast_fp16")]; - tensor inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = obj_235_cast_fp16)[name = tensor("inputs_117_cast_fp16")]; - tensor var_7841 = const()[name = tensor("op_7841"), val = tensor([1])]; - tensor channels_mean_117_cast_fp16 = reduce_mean(axes = var_7841, keep_dims = var_7670, 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_7845 = const()[name = tensor("op_7845"), val = tensor([1])]; - tensor var_7846_cast_fp16 = reduce_mean(axes = var_7845, keep_dims = var_7670, x = zero_mean_sq_117_cast_fp16)[name = tensor("op_7846_cast_fp16")]; - tensor var_7847_to_fp16 = const()[name = tensor("op_7847_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_7848_cast_fp16 = add(x = var_7846_cast_fp16, y = var_7847_to_fp16)[name = tensor("op_7848_cast_fp16")]; - tensor denom_117_epsilon_0 = const()[name = tensor("denom_117_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_117_cast_fp16 = rsqrt(epsilon = denom_117_epsilon_0, x = var_7848_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_237_gamma_0_to_fp16 = const()[name = tensor("obj_237_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409370112)))]; - tensor obj_237_beta_0_to_fp16 = const()[name = tensor("obj_237_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409372736)))]; - tensor obj_237_epsilon_0_to_fp16 = const()[name = tensor("obj_237_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_237_cast_fp16 = batch_norm(beta = obj_237_beta_0_to_fp16, epsilon = obj_237_epsilon_0_to_fp16, gamma = obj_237_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_237_cast_fp16")]; - tensor var_7866 = const()[name = tensor("op_7866"), val = tensor([1, 1])]; - tensor var_7868 = const()[name = tensor("op_7868"), val = tensor([1, 1])]; - tensor pretrained_out_389_pad_type_0 = const()[name = tensor("pretrained_out_389_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_389_pad_0 = const()[name = tensor("pretrained_out_389_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409375360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410194624))), name = tensor("layers_19_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_19_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410194752)))]; - tensor pretrained_out_389_cast_fp16 = conv(bias = layers_19_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_7868, groups = var_7669, pad = pretrained_out_389_pad_0, pad_type = pretrained_out_389_pad_type_0, strides = var_7866, weight = layers_19_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_237_cast_fp16)[name = tensor("pretrained_out_389_cast_fp16")]; - tensor var_7872 = const()[name = tensor("op_7872"), val = tensor([1, 1])]; - tensor var_7874 = const()[name = tensor("op_7874"), val = tensor([1, 1])]; - tensor input_581_pad_type_0 = const()[name = tensor("input_581_pad_type_0"), val = tensor("custom")]; - tensor input_581_pad_0 = const()[name = tensor("input_581_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410197376)))]; - tensor input_581_cast_fp16 = conv(dilations = var_7874, groups = var_7669, pad = input_581_pad_0, pad_type = input_581_pad_type_0, strides = var_7872, weight = layers_19_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_237_cast_fp16)[name = tensor("input_581_cast_fp16")]; - tensor var_7878 = const()[name = tensor("op_7878"), val = tensor([1, 1])]; - tensor var_7880 = const()[name = tensor("op_7880"), val = tensor([1, 1])]; - tensor lora_out_777_pad_type_0 = const()[name = tensor("lora_out_777_pad_type_0"), val = tensor("custom")]; - tensor lora_out_777_pad_0 = const()[name = tensor("lora_out_777_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_779_weight_0_to_fp16 = const()[name = tensor("lora_out_779_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410238400)))]; - tensor lora_out_779_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7880, groups = var_7669, pad = lora_out_777_pad_0, pad_type = lora_out_777_pad_type_0, strides = var_7878, weight = lora_out_779_weight_0_to_fp16, x = input_581_cast_fp16)[name = tensor("lora_out_779_cast_fp16")]; - tensor query_79_cast_fp16 = add(x = pretrained_out_389_cast_fp16, y = lora_out_779_cast_fp16)[name = tensor("query_79_cast_fp16")]; - tensor var_7890 = const()[name = tensor("op_7890"), val = tensor([1, 1])]; - tensor var_7892 = const()[name = tensor("op_7892"), val = tensor([1, 1])]; - tensor pretrained_out_391_pad_type_0 = const()[name = tensor("pretrained_out_391_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_391_pad_0 = const()[name = tensor("pretrained_out_391_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410279424))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411098688))), name = tensor("layers_19_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_391_cast_fp16 = conv(dilations = var_7892, groups = var_7669, pad = pretrained_out_391_pad_0, pad_type = pretrained_out_391_pad_type_0, strides = var_7890, weight = layers_19_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_391_cast_fp16")]; - tensor var_7896 = const()[name = tensor("op_7896"), val = tensor([1, 1])]; - tensor var_7898 = const()[name = tensor("op_7898"), val = tensor([1, 1])]; - tensor input_583_pad_type_0 = const()[name = tensor("input_583_pad_type_0"), val = tensor("custom")]; - tensor input_583_pad_0 = const()[name = tensor("input_583_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411098816)))]; - tensor input_583_cast_fp16 = conv(dilations = var_7898, groups = var_7669, pad = input_583_pad_0, pad_type = input_583_pad_type_0, strides = var_7896, weight = layers_19_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_583_cast_fp16")]; - tensor var_7902 = const()[name = tensor("op_7902"), val = tensor([1, 1])]; - tensor var_7904 = const()[name = tensor("op_7904"), val = tensor([1, 1])]; - tensor lora_out_781_pad_type_0 = const()[name = tensor("lora_out_781_pad_type_0"), val = tensor("custom")]; - tensor lora_out_781_pad_0 = const()[name = tensor("lora_out_781_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_783_weight_0_to_fp16 = const()[name = tensor("lora_out_783_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411139840)))]; - tensor lora_out_783_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7904, groups = var_7669, pad = lora_out_781_pad_0, pad_type = lora_out_781_pad_type_0, strides = var_7902, weight = lora_out_783_weight_0_to_fp16, x = input_583_cast_fp16)[name = tensor("lora_out_783_cast_fp16")]; - tensor key_79_cast_fp16 = add(x = pretrained_out_391_cast_fp16, y = lora_out_783_cast_fp16)[name = tensor("key_79_cast_fp16")]; - tensor var_7915 = const()[name = tensor("op_7915"), val = tensor([1, 1])]; - tensor var_7917 = const()[name = tensor("op_7917"), val = tensor([1, 1])]; - tensor pretrained_out_393_pad_type_0 = const()[name = tensor("pretrained_out_393_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_393_pad_0 = const()[name = tensor("pretrained_out_393_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411180864))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412000128))), name = tensor("layers_19_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_19_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412000256)))]; - tensor pretrained_out_393_cast_fp16 = conv(bias = layers_19_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_7917, groups = var_7669, pad = pretrained_out_393_pad_0, pad_type = pretrained_out_393_pad_type_0, strides = var_7915, weight = layers_19_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_393_cast_fp16")]; - tensor var_7921 = const()[name = tensor("op_7921"), val = tensor([1, 1])]; - tensor var_7923 = const()[name = tensor("op_7923"), val = tensor([1, 1])]; - tensor input_585_pad_type_0 = const()[name = tensor("input_585_pad_type_0"), val = tensor("custom")]; - tensor input_585_pad_0 = const()[name = tensor("input_585_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412002880)))]; - tensor input_585_cast_fp16 = conv(dilations = var_7923, groups = var_7669, pad = input_585_pad_0, pad_type = input_585_pad_type_0, strides = var_7921, weight = layers_19_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_585_cast_fp16")]; - tensor var_7927 = const()[name = tensor("op_7927"), val = tensor([1, 1])]; - tensor var_7929 = const()[name = tensor("op_7929"), val = tensor([1, 1])]; - tensor lora_out_785_pad_type_0 = const()[name = tensor("lora_out_785_pad_type_0"), val = tensor("custom")]; - tensor lora_out_785_pad_0 = const()[name = tensor("lora_out_785_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_787_weight_0_to_fp16 = const()[name = tensor("lora_out_787_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412043904)))]; - tensor lora_out_787_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7929, groups = var_7669, pad = lora_out_785_pad_0, pad_type = lora_out_785_pad_type_0, strides = var_7927, weight = lora_out_787_weight_0_to_fp16, x = input_585_cast_fp16)[name = tensor("lora_out_787_cast_fp16")]; - tensor value_79_cast_fp16 = add(x = pretrained_out_393_cast_fp16, y = lora_out_787_cast_fp16)[name = tensor("value_79_cast_fp16")]; - tensor var_7936 = const()[name = tensor("op_7936"), val = tensor([1, 20, 64, -1])]; - tensor var_7937_cast_fp16 = reshape(shape = var_7936, x = query_79_cast_fp16)[name = tensor("op_7937_cast_fp16")]; - tensor var_7938_to_fp16 = const()[name = tensor("op_7938_to_fp16"), val = tensor(0x1p-3)]; - tensor var_7939_cast_fp16 = mul(x = var_7937_cast_fp16, y = var_7938_to_fp16)[name = tensor("op_7939_cast_fp16")]; - tensor var_7940 = const()[name = tensor("op_7940"), val = tensor([1, 20, 64, -1])]; - tensor var_7941_cast_fp16 = reshape(shape = var_7940, x = key_79_cast_fp16)[name = tensor("op_7941_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_7939_cast_fp16, y = var_7941_cast_fp16)[name = tensor("mh_w_119_cast_fp16")]; - tensor var_7944_cast_fp16 = softmax(axis = var_7662, x = mh_w_119_cast_fp16)[name = tensor("op_7944_cast_fp16")]; - tensor var_7945 = const()[name = tensor("op_7945"), val = tensor([1, 20, 64, -1])]; - tensor var_7946_cast_fp16 = reshape(shape = var_7945, x = value_79_cast_fp16)[name = tensor("op_7946_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_7946_cast_fp16, y = var_7944_cast_fp16)[name = tensor("attn_79_cast_fp16")]; - tensor var_7949 = const()[name = tensor("op_7949"), val = tensor([1, 1280, 1, -1])]; - tensor input_587_cast_fp16 = reshape(shape = var_7949, x = attn_79_cast_fp16)[name = tensor("input_587_cast_fp16")]; - tensor var_7956 = const()[name = tensor("op_7956"), val = tensor([1, 1])]; - tensor var_7958 = const()[name = tensor("op_7958"), val = tensor([1, 1])]; - tensor pretrained_out_395_pad_type_0 = const()[name = tensor("pretrained_out_395_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_395_pad_0 = const()[name = tensor("pretrained_out_395_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412084928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412904192))), name = tensor("layers_19_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_19_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412904320)))]; - tensor pretrained_out_395_cast_fp16 = conv(bias = layers_19_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_7958, groups = var_7669, pad = pretrained_out_395_pad_0, pad_type = pretrained_out_395_pad_type_0, strides = var_7956, weight = layers_19_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_587_cast_fp16)[name = tensor("pretrained_out_395_cast_fp16")]; - tensor var_7962 = const()[name = tensor("op_7962"), val = tensor([1, 1])]; - tensor var_7964 = const()[name = tensor("op_7964"), val = tensor([1, 1])]; - tensor input_589_pad_type_0 = const()[name = tensor("input_589_pad_type_0"), val = tensor("custom")]; - tensor input_589_pad_0 = const()[name = tensor("input_589_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412906944)))]; - tensor input_589_cast_fp16 = conv(dilations = var_7964, groups = var_7669, pad = input_589_pad_0, pad_type = input_589_pad_type_0, strides = var_7962, weight = layers_19_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_587_cast_fp16)[name = tensor("input_589_cast_fp16")]; - tensor var_7968 = const()[name = tensor("op_7968"), val = tensor([1, 1])]; - tensor var_7970 = const()[name = tensor("op_7970"), val = tensor([1, 1])]; - tensor lora_out_789_pad_type_0 = const()[name = tensor("lora_out_789_pad_type_0"), val = tensor("custom")]; - tensor lora_out_789_pad_0 = const()[name = tensor("lora_out_789_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_791_weight_0_to_fp16 = const()[name = tensor("lora_out_791_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412947968)))]; - tensor lora_out_791_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7970, groups = var_7669, pad = lora_out_789_pad_0, pad_type = lora_out_789_pad_type_0, strides = var_7968, weight = lora_out_791_weight_0_to_fp16, x = input_589_cast_fp16)[name = tensor("lora_out_791_cast_fp16")]; - tensor obj_239_cast_fp16 = add(x = pretrained_out_395_cast_fp16, y = lora_out_791_cast_fp16)[name = tensor("obj_239_cast_fp16")]; - tensor inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = obj_239_cast_fp16)[name = tensor("inputs_119_cast_fp16")]; - tensor var_7979 = const()[name = tensor("op_7979"), val = tensor([1])]; - tensor channels_mean_119_cast_fp16 = reduce_mean(axes = var_7979, keep_dims = var_7670, 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_7983 = const()[name = tensor("op_7983"), val = tensor([1])]; - tensor var_7984_cast_fp16 = reduce_mean(axes = var_7983, keep_dims = var_7670, x = zero_mean_sq_119_cast_fp16)[name = tensor("op_7984_cast_fp16")]; - tensor var_7985_to_fp16 = const()[name = tensor("op_7985_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_7986_cast_fp16 = add(x = var_7984_cast_fp16, y = var_7985_to_fp16)[name = tensor("op_7986_cast_fp16")]; - tensor denom_119_epsilon_0 = const()[name = tensor("denom_119_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_119_cast_fp16 = rsqrt(epsilon = denom_119_epsilon_0, x = var_7986_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_591_gamma_0_to_fp16 = const()[name = tensor("input_591_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412988992)))]; - tensor input_591_beta_0_to_fp16 = const()[name = tensor("input_591_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412991616)))]; - tensor input_591_epsilon_0_to_fp16 = const()[name = tensor("input_591_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_591_cast_fp16 = batch_norm(beta = input_591_beta_0_to_fp16, epsilon = input_591_epsilon_0_to_fp16, gamma = input_591_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_591_cast_fp16")]; - tensor var_8000 = const()[name = tensor("op_8000"), val = tensor([1, 1])]; - tensor var_8002 = const()[name = tensor("op_8002"), val = tensor([1, 1])]; - tensor pretrained_out_397_pad_type_0 = const()[name = tensor("pretrained_out_397_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_397_pad_0 = const()[name = tensor("pretrained_out_397_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412994240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(416271104))), name = tensor("layers_19_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_19_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_19_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(416271232)))]; - tensor pretrained_out_397_cast_fp16 = conv(bias = layers_19_fc1_pretrained_bias_to_fp16, dilations = var_8002, groups = var_7669, pad = pretrained_out_397_pad_0, pad_type = pretrained_out_397_pad_type_0, strides = var_8000, weight = layers_19_fc1_pretrained_weight_to_fp16_palettized, x = input_591_cast_fp16)[name = tensor("pretrained_out_397_cast_fp16")]; - tensor var_8006 = const()[name = tensor("op_8006"), val = tensor([1, 1])]; - tensor var_8008 = const()[name = tensor("op_8008"), val = tensor([1, 1])]; - tensor input_593_pad_type_0 = const()[name = tensor("input_593_pad_type_0"), val = tensor("custom")]; - tensor input_593_pad_0 = const()[name = tensor("input_593_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_19_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(416281536)))]; - tensor input_593_cast_fp16 = conv(dilations = var_8008, groups = var_7669, pad = input_593_pad_0, pad_type = input_593_pad_type_0, strides = var_8006, weight = layers_19_fc1_loraA_weight_to_fp16, x = input_591_cast_fp16)[name = tensor("input_593_cast_fp16")]; - tensor var_8012 = const()[name = tensor("op_8012"), val = tensor([1, 1])]; - tensor var_8014 = const()[name = tensor("op_8014"), val = tensor([1, 1])]; - tensor lora_out_793_pad_type_0 = const()[name = tensor("lora_out_793_pad_type_0"), val = tensor("custom")]; - tensor lora_out_793_pad_0 = const()[name = tensor("lora_out_793_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_795_weight_0_to_fp16 = const()[name = tensor("lora_out_795_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(416322560)))]; - tensor lora_out_795_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_8014, groups = var_7669, pad = lora_out_793_pad_0, pad_type = lora_out_793_pad_type_0, strides = var_8012, weight = lora_out_795_weight_0_to_fp16, x = input_593_cast_fp16)[name = tensor("lora_out_795_cast_fp16")]; - tensor input_595_cast_fp16 = add(x = pretrained_out_397_cast_fp16, y = lora_out_795_cast_fp16)[name = tensor("input_595_cast_fp16")]; - tensor input_597_mode_0 = const()[name = tensor("input_597_mode_0"), val = tensor("EXACT")]; - tensor input_597_cast_fp16 = gelu(mode = input_597_mode_0, x = input_595_cast_fp16)[name = tensor("input_597_cast_fp16")]; - tensor var_8026 = const()[name = tensor("op_8026"), val = tensor([1, 1])]; - tensor var_8028 = const()[name = tensor("op_8028"), val = tensor([1, 1])]; - tensor pretrained_out_399_pad_type_0 = const()[name = tensor("pretrained_out_399_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_399_pad_0 = const()[name = tensor("pretrained_out_399_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(416486464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419763328))), name = tensor("layers_19_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_19_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_19_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419763456)))]; - tensor pretrained_out_399_cast_fp16 = conv(bias = layers_19_fc2_pretrained_bias_to_fp16, dilations = var_8028, groups = var_7669, pad = pretrained_out_399_pad_0, pad_type = pretrained_out_399_pad_type_0, strides = var_8026, weight = layers_19_fc2_pretrained_weight_to_fp16_palettized, x = input_597_cast_fp16)[name = tensor("pretrained_out_399_cast_fp16")]; - tensor var_8032 = const()[name = tensor("op_8032"), val = tensor([1, 1])]; - tensor var_8034 = const()[name = tensor("op_8034"), val = tensor([1, 1])]; - tensor input_599_pad_type_0 = const()[name = tensor("input_599_pad_type_0"), val = tensor("custom")]; - tensor input_599_pad_0 = const()[name = tensor("input_599_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_19_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_19_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419766080)))]; - tensor input_599_cast_fp16 = conv(dilations = var_8034, groups = var_7669, pad = input_599_pad_0, pad_type = input_599_pad_type_0, strides = var_8032, weight = layers_19_fc2_loraA_weight_to_fp16, x = input_597_cast_fp16)[name = tensor("input_599_cast_fp16")]; - tensor var_8038 = const()[name = tensor("op_8038"), val = tensor([1, 1])]; - tensor var_8040 = const()[name = tensor("op_8040"), val = tensor([1, 1])]; - tensor lora_out_797_pad_type_0 = const()[name = tensor("lora_out_797_pad_type_0"), val = tensor("custom")]; - tensor lora_out_797_pad_0 = const()[name = tensor("lora_out_797_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_799_weight_0_to_fp16 = const()[name = tensor("lora_out_799_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419929984)))]; - tensor lora_out_799_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8040, groups = var_7669, pad = lora_out_797_pad_0, pad_type = lora_out_797_pad_type_0, strides = var_8038, weight = lora_out_799_weight_0_to_fp16, x = input_599_cast_fp16)[name = tensor("lora_out_799_cast_fp16")]; - tensor hidden_states_41_cast_fp16 = add(x = pretrained_out_399_cast_fp16, y = lora_out_799_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_8056 = const()[name = tensor("op_8056"), val = tensor(3)]; - tensor var_8063 = const()[name = tensor("op_8063"), val = tensor(1)]; - tensor var_8064 = const()[name = tensor("op_8064"), val = tensor(true)]; - tensor var_8076 = const()[name = tensor("op_8076"), val = tensor([1])]; - tensor channels_mean_121_cast_fp16 = reduce_mean(axes = var_8076, keep_dims = var_8064, 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_8080 = const()[name = tensor("op_8080"), val = tensor([1])]; - tensor var_8081_cast_fp16 = reduce_mean(axes = var_8080, keep_dims = var_8064, x = zero_mean_sq_121_cast_fp16)[name = tensor("op_8081_cast_fp16")]; - tensor var_8082_to_fp16 = const()[name = tensor("op_8082_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_8083_cast_fp16 = add(x = var_8081_cast_fp16, y = var_8082_to_fp16)[name = tensor("op_8083_cast_fp16")]; - tensor denom_121_epsilon_0 = const()[name = tensor("denom_121_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_121_cast_fp16 = rsqrt(epsilon = denom_121_epsilon_0, x = var_8083_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_241_gamma_0_to_fp16 = const()[name = tensor("obj_241_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419971008)))]; - tensor obj_241_beta_0_to_fp16 = const()[name = tensor("obj_241_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419973632)))]; - tensor obj_241_epsilon_0_to_fp16 = const()[name = tensor("obj_241_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_241_cast_fp16 = batch_norm(beta = obj_241_beta_0_to_fp16, epsilon = obj_241_epsilon_0_to_fp16, gamma = obj_241_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_241_cast_fp16")]; - tensor var_8101 = const()[name = tensor("op_8101"), val = tensor([1, 1])]; - tensor var_8103 = const()[name = tensor("op_8103"), val = tensor([1, 1])]; - tensor pretrained_out_401_pad_type_0 = const()[name = tensor("pretrained_out_401_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_401_pad_0 = const()[name = tensor("pretrained_out_401_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419976256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420795520))), name = tensor("layers_20_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_20_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420795648)))]; - tensor pretrained_out_401_cast_fp16 = conv(bias = layers_20_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_8103, groups = var_8063, pad = pretrained_out_401_pad_0, pad_type = pretrained_out_401_pad_type_0, strides = var_8101, weight = layers_20_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_241_cast_fp16)[name = tensor("pretrained_out_401_cast_fp16")]; - tensor var_8107 = const()[name = tensor("op_8107"), val = tensor([1, 1])]; - tensor var_8109 = const()[name = tensor("op_8109"), val = tensor([1, 1])]; - tensor input_601_pad_type_0 = const()[name = tensor("input_601_pad_type_0"), val = tensor("custom")]; - tensor input_601_pad_0 = const()[name = tensor("input_601_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420798272)))]; - tensor input_601_cast_fp16 = conv(dilations = var_8109, groups = var_8063, pad = input_601_pad_0, pad_type = input_601_pad_type_0, strides = var_8107, weight = layers_20_self_attn_q_proj_loraA_weight_to_fp16, x = obj_241_cast_fp16)[name = tensor("input_601_cast_fp16")]; - tensor var_8113 = const()[name = tensor("op_8113"), val = tensor([1, 1])]; - tensor var_8115 = const()[name = tensor("op_8115"), val = tensor([1, 1])]; - tensor lora_out_801_pad_type_0 = const()[name = tensor("lora_out_801_pad_type_0"), val = tensor("custom")]; - tensor lora_out_801_pad_0 = const()[name = tensor("lora_out_801_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_803_weight_0_to_fp16 = const()[name = tensor("lora_out_803_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420839296)))]; - tensor lora_out_803_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8115, groups = var_8063, pad = lora_out_801_pad_0, pad_type = lora_out_801_pad_type_0, strides = var_8113, weight = lora_out_803_weight_0_to_fp16, x = input_601_cast_fp16)[name = tensor("lora_out_803_cast_fp16")]; - tensor query_81_cast_fp16 = add(x = pretrained_out_401_cast_fp16, y = lora_out_803_cast_fp16)[name = tensor("query_81_cast_fp16")]; - tensor var_8125 = const()[name = tensor("op_8125"), val = tensor([1, 1])]; - tensor var_8127 = const()[name = tensor("op_8127"), val = tensor([1, 1])]; - tensor pretrained_out_403_pad_type_0 = const()[name = tensor("pretrained_out_403_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_403_pad_0 = const()[name = tensor("pretrained_out_403_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420880320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421699584))), name = tensor("layers_20_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_403_cast_fp16 = conv(dilations = var_8127, groups = var_8063, pad = pretrained_out_403_pad_0, pad_type = pretrained_out_403_pad_type_0, strides = var_8125, weight = layers_20_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_241_cast_fp16)[name = tensor("pretrained_out_403_cast_fp16")]; - tensor var_8131 = const()[name = tensor("op_8131"), val = tensor([1, 1])]; - tensor var_8133 = const()[name = tensor("op_8133"), val = tensor([1, 1])]; - tensor input_603_pad_type_0 = const()[name = tensor("input_603_pad_type_0"), val = tensor("custom")]; - tensor input_603_pad_0 = const()[name = tensor("input_603_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421699712)))]; - tensor input_603_cast_fp16 = conv(dilations = var_8133, groups = var_8063, pad = input_603_pad_0, pad_type = input_603_pad_type_0, strides = var_8131, weight = layers_20_self_attn_k_proj_loraA_weight_to_fp16, x = obj_241_cast_fp16)[name = tensor("input_603_cast_fp16")]; - tensor var_8137 = const()[name = tensor("op_8137"), val = tensor([1, 1])]; - tensor var_8139 = const()[name = tensor("op_8139"), val = tensor([1, 1])]; - tensor lora_out_805_pad_type_0 = const()[name = tensor("lora_out_805_pad_type_0"), val = tensor("custom")]; - tensor lora_out_805_pad_0 = const()[name = tensor("lora_out_805_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_807_weight_0_to_fp16 = const()[name = tensor("lora_out_807_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421740736)))]; - tensor lora_out_807_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8139, groups = var_8063, pad = lora_out_805_pad_0, pad_type = lora_out_805_pad_type_0, strides = var_8137, weight = lora_out_807_weight_0_to_fp16, x = input_603_cast_fp16)[name = tensor("lora_out_807_cast_fp16")]; - tensor current_key_41_cast_fp16 = add(x = pretrained_out_403_cast_fp16, y = lora_out_807_cast_fp16)[name = tensor("current_key_41_cast_fp16")]; - tensor var_8150 = const()[name = tensor("op_8150"), val = tensor([1, 1])]; - tensor var_8152 = const()[name = tensor("op_8152"), val = tensor([1, 1])]; - tensor pretrained_out_405_pad_type_0 = const()[name = tensor("pretrained_out_405_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_405_pad_0 = const()[name = tensor("pretrained_out_405_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421781760))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422601024))), name = tensor("layers_20_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_20_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422601152)))]; - tensor pretrained_out_405_cast_fp16 = conv(bias = layers_20_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_8152, groups = var_8063, pad = pretrained_out_405_pad_0, pad_type = pretrained_out_405_pad_type_0, strides = var_8150, weight = layers_20_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_241_cast_fp16)[name = tensor("pretrained_out_405_cast_fp16")]; - tensor var_8156 = const()[name = tensor("op_8156"), val = tensor([1, 1])]; - tensor var_8158 = const()[name = tensor("op_8158"), val = tensor([1, 1])]; - tensor input_605_pad_type_0 = const()[name = tensor("input_605_pad_type_0"), val = tensor("custom")]; - tensor input_605_pad_0 = const()[name = tensor("input_605_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422603776)))]; - tensor input_605_cast_fp16 = conv(dilations = var_8158, groups = var_8063, pad = input_605_pad_0, pad_type = input_605_pad_type_0, strides = var_8156, weight = layers_20_self_attn_v_proj_loraA_weight_to_fp16, x = obj_241_cast_fp16)[name = tensor("input_605_cast_fp16")]; - tensor var_8162 = const()[name = tensor("op_8162"), val = tensor([1, 1])]; - tensor var_8164 = const()[name = tensor("op_8164"), val = tensor([1, 1])]; - tensor lora_out_809_pad_type_0 = const()[name = tensor("lora_out_809_pad_type_0"), val = tensor("custom")]; - tensor lora_out_809_pad_0 = const()[name = tensor("lora_out_809_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_811_weight_0_to_fp16 = const()[name = tensor("lora_out_811_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422644800)))]; - tensor lora_out_811_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8164, groups = var_8063, pad = lora_out_809_pad_0, pad_type = lora_out_809_pad_type_0, strides = var_8162, weight = lora_out_811_weight_0_to_fp16, x = input_605_cast_fp16)[name = tensor("lora_out_811_cast_fp16")]; - tensor current_value_41_cast_fp16 = add(x = pretrained_out_405_cast_fp16, y = lora_out_811_cast_fp16)[name = tensor("current_value_41_cast_fp16")]; - tensor var_8174_cast_fp16 = mul(x = current_key_41_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_8174_cast_fp16")]; - tensor var_8176_cast_fp16 = mul(x = var_103_cast_fp16_20, y = var_295_cast_fp16)[name = tensor("op_8176_cast_fp16")]; - tensor key_81_cast_fp16 = add(x = var_8174_cast_fp16, y = var_8176_cast_fp16)[name = tensor("key_81_cast_fp16")]; - tensor var_8178_cast_fp16 = mul(x = current_value_41_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_8178_cast_fp16")]; - tensor var_8180_cast_fp16 = mul(x = var_138_cast_fp16_20, y = var_295_cast_fp16)[name = tensor("op_8180_cast_fp16")]; - tensor value_81_cast_fp16 = add(x = var_8178_cast_fp16, y = var_8180_cast_fp16)[name = tensor("value_81_cast_fp16")]; - tensor var_8183 = const()[name = tensor("op_8183"), val = tensor([1, 20, 64, -1])]; - tensor var_8184_cast_fp16 = reshape(shape = var_8183, x = query_81_cast_fp16)[name = tensor("op_8184_cast_fp16")]; - tensor var_8185_to_fp16 = const()[name = tensor("op_8185_to_fp16"), val = tensor(0x1p-3)]; - tensor var_8186_cast_fp16 = mul(x = var_8184_cast_fp16, y = var_8185_to_fp16)[name = tensor("op_8186_cast_fp16")]; - tensor var_8187 = const()[name = tensor("op_8187"), val = tensor([1, 20, 64, -1])]; - tensor var_8188_cast_fp16 = reshape(shape = var_8187, x = key_81_cast_fp16)[name = tensor("op_8188_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_8186_cast_fp16, y = var_8188_cast_fp16)[name = tensor("mh_w_121_cast_fp16")]; - tensor mh_w_123_cast_fp16 = add(x = mh_w_121_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_123_cast_fp16")]; - tensor var_8196_cast_fp16 = softmax(axis = var_8056, x = mh_w_123_cast_fp16)[name = tensor("op_8196_cast_fp16")]; - tensor var_8197 = const()[name = tensor("op_8197"), val = tensor([1, 20, 64, -1])]; - tensor var_8198_cast_fp16 = reshape(shape = var_8197, x = value_81_cast_fp16)[name = tensor("op_8198_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_8198_cast_fp16, y = var_8196_cast_fp16)[name = tensor("attn_81_cast_fp16")]; - tensor var_8201 = const()[name = tensor("op_8201"), val = tensor([1, 1280, 1, -1])]; - tensor input_607_cast_fp16 = reshape(shape = var_8201, x = attn_81_cast_fp16)[name = tensor("input_607_cast_fp16")]; - tensor var_8208 = const()[name = tensor("op_8208"), val = tensor([1, 1])]; - tensor var_8210 = const()[name = tensor("op_8210"), val = tensor([1, 1])]; - tensor pretrained_out_407_pad_type_0 = const()[name = tensor("pretrained_out_407_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_407_pad_0 = const()[name = tensor("pretrained_out_407_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422685824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(423505088))), name = tensor("layers_20_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_20_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(423505216)))]; - tensor pretrained_out_407_cast_fp16 = conv(bias = layers_20_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_8210, groups = var_8063, pad = pretrained_out_407_pad_0, pad_type = pretrained_out_407_pad_type_0, strides = var_8208, weight = layers_20_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_607_cast_fp16)[name = tensor("pretrained_out_407_cast_fp16")]; - tensor var_8214 = const()[name = tensor("op_8214"), val = tensor([1, 1])]; - tensor var_8216 = const()[name = tensor("op_8216"), val = tensor([1, 1])]; - tensor input_609_pad_type_0 = const()[name = tensor("input_609_pad_type_0"), val = tensor("custom")]; - tensor input_609_pad_0 = const()[name = tensor("input_609_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(423507840)))]; - tensor input_609_cast_fp16 = conv(dilations = var_8216, groups = var_8063, pad = input_609_pad_0, pad_type = input_609_pad_type_0, strides = var_8214, weight = layers_20_self_attn_o_proj_loraA_weight_to_fp16, x = input_607_cast_fp16)[name = tensor("input_609_cast_fp16")]; - tensor var_8220 = const()[name = tensor("op_8220"), val = tensor([1, 1])]; - tensor var_8222 = const()[name = tensor("op_8222"), val = tensor([1, 1])]; - tensor lora_out_813_pad_type_0 = const()[name = tensor("lora_out_813_pad_type_0"), val = tensor("custom")]; - tensor lora_out_813_pad_0 = const()[name = tensor("lora_out_813_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_815_weight_0_to_fp16 = const()[name = tensor("lora_out_815_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(423548864)))]; - tensor lora_out_815_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8222, groups = var_8063, pad = lora_out_813_pad_0, pad_type = lora_out_813_pad_type_0, strides = var_8220, weight = lora_out_815_weight_0_to_fp16, x = input_609_cast_fp16)[name = tensor("lora_out_815_cast_fp16")]; - tensor obj_247_cast_fp16 = add(x = pretrained_out_407_cast_fp16, y = lora_out_815_cast_fp16)[name = tensor("obj_247_cast_fp16")]; - tensor inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = obj_247_cast_fp16)[name = tensor("inputs_123_cast_fp16")]; - tensor var_8235 = const()[name = tensor("op_8235"), val = tensor([1])]; - tensor channels_mean_123_cast_fp16 = reduce_mean(axes = var_8235, keep_dims = var_8064, 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_8239 = const()[name = tensor("op_8239"), val = tensor([1])]; - tensor var_8240_cast_fp16 = reduce_mean(axes = var_8239, keep_dims = var_8064, x = zero_mean_sq_123_cast_fp16)[name = tensor("op_8240_cast_fp16")]; - tensor var_8241_to_fp16 = const()[name = tensor("op_8241_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_8242_cast_fp16 = add(x = var_8240_cast_fp16, y = var_8241_to_fp16)[name = tensor("op_8242_cast_fp16")]; - tensor denom_123_epsilon_0 = const()[name = tensor("denom_123_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_123_cast_fp16 = rsqrt(epsilon = denom_123_epsilon_0, x = var_8242_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_249_gamma_0_to_fp16 = const()[name = tensor("obj_249_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(423589888)))]; - tensor obj_249_beta_0_to_fp16 = const()[name = tensor("obj_249_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(423592512)))]; - tensor obj_249_epsilon_0_to_fp16 = const()[name = tensor("obj_249_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_249_cast_fp16 = batch_norm(beta = obj_249_beta_0_to_fp16, epsilon = obj_249_epsilon_0_to_fp16, gamma = obj_249_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_249_cast_fp16")]; - tensor var_8260 = const()[name = tensor("op_8260"), val = tensor([1, 1])]; - tensor var_8262 = const()[name = tensor("op_8262"), val = tensor([1, 1])]; - tensor pretrained_out_409_pad_type_0 = const()[name = tensor("pretrained_out_409_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_409_pad_0 = const()[name = tensor("pretrained_out_409_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(423595136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424414400))), name = tensor("layers_20_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_20_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424414528)))]; - tensor pretrained_out_409_cast_fp16 = conv(bias = layers_20_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_8262, groups = var_8063, pad = pretrained_out_409_pad_0, pad_type = pretrained_out_409_pad_type_0, strides = var_8260, weight = layers_20_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_249_cast_fp16)[name = tensor("pretrained_out_409_cast_fp16")]; - tensor var_8266 = const()[name = tensor("op_8266"), val = tensor([1, 1])]; - tensor var_8268 = const()[name = tensor("op_8268"), val = tensor([1, 1])]; - tensor input_611_pad_type_0 = const()[name = tensor("input_611_pad_type_0"), val = tensor("custom")]; - tensor input_611_pad_0 = const()[name = tensor("input_611_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424417152)))]; - tensor input_611_cast_fp16 = conv(dilations = var_8268, groups = var_8063, pad = input_611_pad_0, pad_type = input_611_pad_type_0, strides = var_8266, weight = layers_20_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_249_cast_fp16)[name = tensor("input_611_cast_fp16")]; - tensor var_8272 = const()[name = tensor("op_8272"), val = tensor([1, 1])]; - tensor var_8274 = const()[name = tensor("op_8274"), val = tensor([1, 1])]; - tensor lora_out_817_pad_type_0 = const()[name = tensor("lora_out_817_pad_type_0"), val = tensor("custom")]; - tensor lora_out_817_pad_0 = const()[name = tensor("lora_out_817_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_819_weight_0_to_fp16 = const()[name = tensor("lora_out_819_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424458176)))]; - tensor lora_out_819_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8274, groups = var_8063, pad = lora_out_817_pad_0, pad_type = lora_out_817_pad_type_0, strides = var_8272, weight = lora_out_819_weight_0_to_fp16, x = input_611_cast_fp16)[name = tensor("lora_out_819_cast_fp16")]; - tensor query_83_cast_fp16 = add(x = pretrained_out_409_cast_fp16, y = lora_out_819_cast_fp16)[name = tensor("query_83_cast_fp16")]; - tensor var_8284 = const()[name = tensor("op_8284"), val = tensor([1, 1])]; - tensor var_8286 = const()[name = tensor("op_8286"), val = tensor([1, 1])]; - tensor pretrained_out_411_pad_type_0 = const()[name = tensor("pretrained_out_411_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_411_pad_0 = const()[name = tensor("pretrained_out_411_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424499200))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(425318464))), name = tensor("layers_20_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_411_cast_fp16 = conv(dilations = var_8286, groups = var_8063, pad = pretrained_out_411_pad_0, pad_type = pretrained_out_411_pad_type_0, strides = var_8284, weight = layers_20_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_411_cast_fp16")]; - tensor var_8290 = const()[name = tensor("op_8290"), val = tensor([1, 1])]; - tensor var_8292 = const()[name = tensor("op_8292"), val = tensor([1, 1])]; - tensor input_613_pad_type_0 = const()[name = tensor("input_613_pad_type_0"), val = tensor("custom")]; - tensor input_613_pad_0 = const()[name = tensor("input_613_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(425318592)))]; - tensor input_613_cast_fp16 = conv(dilations = var_8292, groups = var_8063, pad = input_613_pad_0, pad_type = input_613_pad_type_0, strides = var_8290, weight = layers_20_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_613_cast_fp16")]; - tensor var_8296 = const()[name = tensor("op_8296"), val = tensor([1, 1])]; - tensor var_8298 = const()[name = tensor("op_8298"), val = tensor([1, 1])]; - tensor lora_out_821_pad_type_0 = const()[name = tensor("lora_out_821_pad_type_0"), val = tensor("custom")]; - tensor lora_out_821_pad_0 = const()[name = tensor("lora_out_821_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_823_weight_0_to_fp16 = const()[name = tensor("lora_out_823_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(425359616)))]; - tensor lora_out_823_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8298, groups = var_8063, pad = lora_out_821_pad_0, pad_type = lora_out_821_pad_type_0, strides = var_8296, weight = lora_out_823_weight_0_to_fp16, x = input_613_cast_fp16)[name = tensor("lora_out_823_cast_fp16")]; - tensor key_83_cast_fp16 = add(x = pretrained_out_411_cast_fp16, y = lora_out_823_cast_fp16)[name = tensor("key_83_cast_fp16")]; - tensor var_8309 = const()[name = tensor("op_8309"), val = tensor([1, 1])]; - tensor var_8311 = const()[name = tensor("op_8311"), val = tensor([1, 1])]; - tensor pretrained_out_413_pad_type_0 = const()[name = tensor("pretrained_out_413_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_413_pad_0 = const()[name = tensor("pretrained_out_413_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(425400640))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426219904))), name = tensor("layers_20_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_20_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426220032)))]; - tensor pretrained_out_413_cast_fp16 = conv(bias = layers_20_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_8311, groups = var_8063, pad = pretrained_out_413_pad_0, pad_type = pretrained_out_413_pad_type_0, strides = var_8309, weight = layers_20_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_413_cast_fp16")]; - tensor var_8315 = const()[name = tensor("op_8315"), val = tensor([1, 1])]; - tensor var_8317 = const()[name = tensor("op_8317"), val = tensor([1, 1])]; - tensor input_615_pad_type_0 = const()[name = tensor("input_615_pad_type_0"), val = tensor("custom")]; - tensor input_615_pad_0 = const()[name = tensor("input_615_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426222656)))]; - tensor input_615_cast_fp16 = conv(dilations = var_8317, groups = var_8063, pad = input_615_pad_0, pad_type = input_615_pad_type_0, strides = var_8315, weight = layers_20_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_615_cast_fp16")]; - tensor var_8321 = const()[name = tensor("op_8321"), val = tensor([1, 1])]; - tensor var_8323 = const()[name = tensor("op_8323"), val = tensor([1, 1])]; - tensor lora_out_825_pad_type_0 = const()[name = tensor("lora_out_825_pad_type_0"), val = tensor("custom")]; - tensor lora_out_825_pad_0 = const()[name = tensor("lora_out_825_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_827_weight_0_to_fp16 = const()[name = tensor("lora_out_827_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426263680)))]; - tensor lora_out_827_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8323, groups = var_8063, pad = lora_out_825_pad_0, pad_type = lora_out_825_pad_type_0, strides = var_8321, weight = lora_out_827_weight_0_to_fp16, x = input_615_cast_fp16)[name = tensor("lora_out_827_cast_fp16")]; - tensor value_83_cast_fp16 = add(x = pretrained_out_413_cast_fp16, y = lora_out_827_cast_fp16)[name = tensor("value_83_cast_fp16")]; - tensor var_8330 = const()[name = tensor("op_8330"), val = tensor([1, 20, 64, -1])]; - tensor var_8331_cast_fp16 = reshape(shape = var_8330, x = query_83_cast_fp16)[name = tensor("op_8331_cast_fp16")]; - tensor var_8332_to_fp16 = const()[name = tensor("op_8332_to_fp16"), val = tensor(0x1p-3)]; - tensor var_8333_cast_fp16 = mul(x = var_8331_cast_fp16, y = var_8332_to_fp16)[name = tensor("op_8333_cast_fp16")]; - tensor var_8334 = const()[name = tensor("op_8334"), val = tensor([1, 20, 64, -1])]; - tensor var_8335_cast_fp16 = reshape(shape = var_8334, x = key_83_cast_fp16)[name = tensor("op_8335_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_8333_cast_fp16, y = var_8335_cast_fp16)[name = tensor("mh_w_125_cast_fp16")]; - tensor var_8338_cast_fp16 = softmax(axis = var_8056, x = mh_w_125_cast_fp16)[name = tensor("op_8338_cast_fp16")]; - tensor var_8339 = const()[name = tensor("op_8339"), val = tensor([1, 20, 64, -1])]; - tensor var_8340_cast_fp16 = reshape(shape = var_8339, x = value_83_cast_fp16)[name = tensor("op_8340_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_8340_cast_fp16, y = var_8338_cast_fp16)[name = tensor("attn_83_cast_fp16")]; - tensor var_8343 = const()[name = tensor("op_8343"), val = tensor([1, 1280, 1, -1])]; - tensor input_617_cast_fp16 = reshape(shape = var_8343, x = attn_83_cast_fp16)[name = tensor("input_617_cast_fp16")]; - tensor var_8350 = const()[name = tensor("op_8350"), val = tensor([1, 1])]; - tensor var_8352 = const()[name = tensor("op_8352"), val = tensor([1, 1])]; - tensor pretrained_out_415_pad_type_0 = const()[name = tensor("pretrained_out_415_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_415_pad_0 = const()[name = tensor("pretrained_out_415_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426304704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427123968))), name = tensor("layers_20_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_20_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427124096)))]; - tensor pretrained_out_415_cast_fp16 = conv(bias = layers_20_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_8352, groups = var_8063, pad = pretrained_out_415_pad_0, pad_type = pretrained_out_415_pad_type_0, strides = var_8350, weight = layers_20_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_617_cast_fp16)[name = tensor("pretrained_out_415_cast_fp16")]; - tensor var_8356 = const()[name = tensor("op_8356"), val = tensor([1, 1])]; - tensor var_8358 = const()[name = tensor("op_8358"), val = tensor([1, 1])]; - tensor input_619_pad_type_0 = const()[name = tensor("input_619_pad_type_0"), val = tensor("custom")]; - tensor input_619_pad_0 = const()[name = tensor("input_619_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427126720)))]; - tensor input_619_cast_fp16 = conv(dilations = var_8358, groups = var_8063, pad = input_619_pad_0, pad_type = input_619_pad_type_0, strides = var_8356, weight = layers_20_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_617_cast_fp16)[name = tensor("input_619_cast_fp16")]; - tensor var_8362 = const()[name = tensor("op_8362"), val = tensor([1, 1])]; - tensor var_8364 = const()[name = tensor("op_8364"), val = tensor([1, 1])]; - tensor lora_out_829_pad_type_0 = const()[name = tensor("lora_out_829_pad_type_0"), val = tensor("custom")]; - tensor lora_out_829_pad_0 = const()[name = tensor("lora_out_829_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_831_weight_0_to_fp16 = const()[name = tensor("lora_out_831_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427167744)))]; - tensor lora_out_831_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8364, groups = var_8063, pad = lora_out_829_pad_0, pad_type = lora_out_829_pad_type_0, strides = var_8362, weight = lora_out_831_weight_0_to_fp16, x = input_619_cast_fp16)[name = tensor("lora_out_831_cast_fp16")]; - tensor obj_251_cast_fp16 = add(x = pretrained_out_415_cast_fp16, y = lora_out_831_cast_fp16)[name = tensor("obj_251_cast_fp16")]; - tensor inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = obj_251_cast_fp16)[name = tensor("inputs_125_cast_fp16")]; - tensor var_8373 = const()[name = tensor("op_8373"), val = tensor([1])]; - tensor channels_mean_125_cast_fp16 = reduce_mean(axes = var_8373, keep_dims = var_8064, 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_8377 = const()[name = tensor("op_8377"), val = tensor([1])]; - tensor var_8378_cast_fp16 = reduce_mean(axes = var_8377, keep_dims = var_8064, x = zero_mean_sq_125_cast_fp16)[name = tensor("op_8378_cast_fp16")]; - tensor var_8379_to_fp16 = const()[name = tensor("op_8379_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_8380_cast_fp16 = add(x = var_8378_cast_fp16, y = var_8379_to_fp16)[name = tensor("op_8380_cast_fp16")]; - tensor denom_125_epsilon_0 = const()[name = tensor("denom_125_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_125_cast_fp16 = rsqrt(epsilon = denom_125_epsilon_0, x = var_8380_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_621_gamma_0_to_fp16 = const()[name = tensor("input_621_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427208768)))]; - tensor input_621_beta_0_to_fp16 = const()[name = tensor("input_621_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427211392)))]; - tensor input_621_epsilon_0_to_fp16 = const()[name = tensor("input_621_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_621_cast_fp16 = batch_norm(beta = input_621_beta_0_to_fp16, epsilon = input_621_epsilon_0_to_fp16, gamma = input_621_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_621_cast_fp16")]; - tensor var_8394 = const()[name = tensor("op_8394"), val = tensor([1, 1])]; - tensor var_8396 = const()[name = tensor("op_8396"), val = tensor([1, 1])]; - tensor pretrained_out_417_pad_type_0 = const()[name = tensor("pretrained_out_417_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_417_pad_0 = const()[name = tensor("pretrained_out_417_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427214016))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(430490880))), name = tensor("layers_20_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_20_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_20_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(430491008)))]; - tensor pretrained_out_417_cast_fp16 = conv(bias = layers_20_fc1_pretrained_bias_to_fp16, dilations = var_8396, groups = var_8063, pad = pretrained_out_417_pad_0, pad_type = pretrained_out_417_pad_type_0, strides = var_8394, weight = layers_20_fc1_pretrained_weight_to_fp16_palettized, x = input_621_cast_fp16)[name = tensor("pretrained_out_417_cast_fp16")]; - tensor var_8400 = const()[name = tensor("op_8400"), val = tensor([1, 1])]; - tensor var_8402 = const()[name = tensor("op_8402"), val = tensor([1, 1])]; - tensor input_623_pad_type_0 = const()[name = tensor("input_623_pad_type_0"), val = tensor("custom")]; - tensor input_623_pad_0 = const()[name = tensor("input_623_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_20_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(430501312)))]; - tensor input_623_cast_fp16 = conv(dilations = var_8402, groups = var_8063, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = var_8400, weight = layers_20_fc1_loraA_weight_to_fp16, x = input_621_cast_fp16)[name = tensor("input_623_cast_fp16")]; - tensor var_8406 = const()[name = tensor("op_8406"), val = tensor([1, 1])]; - tensor var_8408 = const()[name = tensor("op_8408"), val = tensor([1, 1])]; - tensor lora_out_833_pad_type_0 = const()[name = tensor("lora_out_833_pad_type_0"), val = tensor("custom")]; - tensor lora_out_833_pad_0 = const()[name = tensor("lora_out_833_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_835_weight_0_to_fp16 = const()[name = tensor("lora_out_835_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(430542336)))]; - tensor lora_out_835_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_8408, groups = var_8063, pad = lora_out_833_pad_0, pad_type = lora_out_833_pad_type_0, strides = var_8406, weight = lora_out_835_weight_0_to_fp16, x = input_623_cast_fp16)[name = tensor("lora_out_835_cast_fp16")]; - tensor input_625_cast_fp16 = add(x = pretrained_out_417_cast_fp16, y = lora_out_835_cast_fp16)[name = tensor("input_625_cast_fp16")]; - tensor input_627_mode_0 = const()[name = tensor("input_627_mode_0"), val = tensor("EXACT")]; - tensor input_627_cast_fp16 = gelu(mode = input_627_mode_0, x = input_625_cast_fp16)[name = tensor("input_627_cast_fp16")]; - tensor var_8420 = const()[name = tensor("op_8420"), val = tensor([1, 1])]; - tensor var_8422 = const()[name = tensor("op_8422"), val = tensor([1, 1])]; - tensor pretrained_out_419_pad_type_0 = const()[name = tensor("pretrained_out_419_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_419_pad_0 = const()[name = tensor("pretrained_out_419_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(430706240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(433983104))), name = tensor("layers_20_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_20_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_20_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(433983232)))]; - tensor pretrained_out_419_cast_fp16 = conv(bias = layers_20_fc2_pretrained_bias_to_fp16, dilations = var_8422, groups = var_8063, pad = pretrained_out_419_pad_0, pad_type = pretrained_out_419_pad_type_0, strides = var_8420, weight = layers_20_fc2_pretrained_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = tensor("pretrained_out_419_cast_fp16")]; - tensor var_8426 = const()[name = tensor("op_8426"), val = tensor([1, 1])]; - tensor var_8428 = const()[name = tensor("op_8428"), val = tensor([1, 1])]; - tensor input_629_pad_type_0 = const()[name = tensor("input_629_pad_type_0"), val = tensor("custom")]; - tensor input_629_pad_0 = const()[name = tensor("input_629_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_20_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_20_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(433985856)))]; - tensor input_629_cast_fp16 = conv(dilations = var_8428, groups = var_8063, pad = input_629_pad_0, pad_type = input_629_pad_type_0, strides = var_8426, weight = layers_20_fc2_loraA_weight_to_fp16, x = input_627_cast_fp16)[name = tensor("input_629_cast_fp16")]; - tensor var_8432 = const()[name = tensor("op_8432"), val = tensor([1, 1])]; - tensor var_8434 = const()[name = tensor("op_8434"), val = tensor([1, 1])]; - tensor lora_out_837_pad_type_0 = const()[name = tensor("lora_out_837_pad_type_0"), val = tensor("custom")]; - tensor lora_out_837_pad_0 = const()[name = tensor("lora_out_837_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_839_weight_0_to_fp16 = const()[name = tensor("lora_out_839_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434149760)))]; - tensor lora_out_839_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8434, groups = var_8063, pad = lora_out_837_pad_0, pad_type = lora_out_837_pad_type_0, strides = var_8432, weight = lora_out_839_weight_0_to_fp16, x = input_629_cast_fp16)[name = tensor("lora_out_839_cast_fp16")]; - tensor hidden_states_43_cast_fp16 = add(x = pretrained_out_419_cast_fp16, y = lora_out_839_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_8450 = const()[name = tensor("op_8450"), val = tensor(3)]; - tensor var_8457 = const()[name = tensor("op_8457"), val = tensor(1)]; - tensor var_8458 = const()[name = tensor("op_8458"), val = tensor(true)]; - tensor var_8470 = const()[name = tensor("op_8470"), val = tensor([1])]; - tensor channels_mean_127_cast_fp16 = reduce_mean(axes = var_8470, keep_dims = var_8458, 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_8474 = const()[name = tensor("op_8474"), val = tensor([1])]; - tensor var_8475_cast_fp16 = reduce_mean(axes = var_8474, keep_dims = var_8458, x = zero_mean_sq_127_cast_fp16)[name = tensor("op_8475_cast_fp16")]; - tensor var_8476_to_fp16 = const()[name = tensor("op_8476_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_8477_cast_fp16 = add(x = var_8475_cast_fp16, y = var_8476_to_fp16)[name = tensor("op_8477_cast_fp16")]; - tensor denom_127_epsilon_0 = const()[name = tensor("denom_127_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_127_cast_fp16 = rsqrt(epsilon = denom_127_epsilon_0, x = var_8477_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_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(434190784)))]; - 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(434193408)))]; - 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_127_cast_fp16)[name = tensor("obj_253_cast_fp16")]; - tensor var_8495 = const()[name = tensor("op_8495"), val = tensor([1, 1])]; - tensor var_8497 = const()[name = tensor("op_8497"), val = tensor([1, 1])]; - tensor pretrained_out_421_pad_type_0 = const()[name = tensor("pretrained_out_421_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_421_pad_0 = const()[name = tensor("pretrained_out_421_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434196032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435015296))), name = tensor("layers_21_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_21_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435015424)))]; - tensor pretrained_out_421_cast_fp16 = conv(bias = layers_21_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_8497, groups = var_8457, pad = pretrained_out_421_pad_0, pad_type = pretrained_out_421_pad_type_0, strides = var_8495, weight = layers_21_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_253_cast_fp16)[name = tensor("pretrained_out_421_cast_fp16")]; - tensor var_8501 = const()[name = tensor("op_8501"), val = tensor([1, 1])]; - tensor var_8503 = const()[name = tensor("op_8503"), val = tensor([1, 1])]; - tensor input_631_pad_type_0 = const()[name = tensor("input_631_pad_type_0"), val = tensor("custom")]; - tensor input_631_pad_0 = const()[name = tensor("input_631_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435018048)))]; - tensor input_631_cast_fp16 = conv(dilations = var_8503, groups = var_8457, pad = input_631_pad_0, pad_type = input_631_pad_type_0, strides = var_8501, weight = layers_21_self_attn_q_proj_loraA_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("input_631_cast_fp16")]; - tensor var_8507 = const()[name = tensor("op_8507"), val = tensor([1, 1])]; - tensor var_8509 = const()[name = tensor("op_8509"), val = tensor([1, 1])]; - tensor lora_out_841_pad_type_0 = const()[name = tensor("lora_out_841_pad_type_0"), val = tensor("custom")]; - tensor lora_out_841_pad_0 = const()[name = tensor("lora_out_841_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_843_weight_0_to_fp16 = const()[name = tensor("lora_out_843_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435059072)))]; - tensor lora_out_843_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8509, groups = var_8457, pad = lora_out_841_pad_0, pad_type = lora_out_841_pad_type_0, strides = var_8507, weight = lora_out_843_weight_0_to_fp16, x = input_631_cast_fp16)[name = tensor("lora_out_843_cast_fp16")]; - tensor query_85_cast_fp16 = add(x = pretrained_out_421_cast_fp16, y = lora_out_843_cast_fp16)[name = tensor("query_85_cast_fp16")]; - tensor var_8519 = const()[name = tensor("op_8519"), val = tensor([1, 1])]; - tensor var_8521 = const()[name = tensor("op_8521"), val = tensor([1, 1])]; - tensor pretrained_out_423_pad_type_0 = const()[name = tensor("pretrained_out_423_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_423_pad_0 = const()[name = tensor("pretrained_out_423_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435100096))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435919360))), name = tensor("layers_21_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_423_cast_fp16 = conv(dilations = var_8521, groups = var_8457, pad = pretrained_out_423_pad_0, pad_type = pretrained_out_423_pad_type_0, strides = var_8519, weight = layers_21_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_253_cast_fp16)[name = tensor("pretrained_out_423_cast_fp16")]; - tensor var_8525 = const()[name = tensor("op_8525"), val = tensor([1, 1])]; - tensor var_8527 = const()[name = tensor("op_8527"), val = tensor([1, 1])]; - tensor input_633_pad_type_0 = const()[name = tensor("input_633_pad_type_0"), val = tensor("custom")]; - tensor input_633_pad_0 = const()[name = tensor("input_633_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435919488)))]; - tensor input_633_cast_fp16 = conv(dilations = var_8527, groups = var_8457, pad = input_633_pad_0, pad_type = input_633_pad_type_0, strides = var_8525, weight = layers_21_self_attn_k_proj_loraA_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("input_633_cast_fp16")]; - tensor var_8531 = const()[name = tensor("op_8531"), val = tensor([1, 1])]; - tensor var_8533 = const()[name = tensor("op_8533"), val = tensor([1, 1])]; - tensor lora_out_845_pad_type_0 = const()[name = tensor("lora_out_845_pad_type_0"), val = tensor("custom")]; - tensor lora_out_845_pad_0 = const()[name = tensor("lora_out_845_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_847_weight_0_to_fp16 = const()[name = tensor("lora_out_847_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435960512)))]; - tensor lora_out_847_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8533, groups = var_8457, pad = lora_out_845_pad_0, pad_type = lora_out_845_pad_type_0, strides = var_8531, weight = lora_out_847_weight_0_to_fp16, x = input_633_cast_fp16)[name = tensor("lora_out_847_cast_fp16")]; - tensor current_key_43_cast_fp16 = add(x = pretrained_out_423_cast_fp16, y = lora_out_847_cast_fp16)[name = tensor("current_key_43_cast_fp16")]; - tensor var_8544 = const()[name = tensor("op_8544"), val = tensor([1, 1])]; - tensor var_8546 = const()[name = tensor("op_8546"), val = tensor([1, 1])]; - tensor pretrained_out_425_pad_type_0 = const()[name = tensor("pretrained_out_425_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_425_pad_0 = const()[name = tensor("pretrained_out_425_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436001536))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436820800))), name = tensor("layers_21_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_21_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436820928)))]; - tensor pretrained_out_425_cast_fp16 = conv(bias = layers_21_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_8546, groups = var_8457, pad = pretrained_out_425_pad_0, pad_type = pretrained_out_425_pad_type_0, strides = var_8544, weight = layers_21_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_253_cast_fp16)[name = tensor("pretrained_out_425_cast_fp16")]; - tensor var_8550 = const()[name = tensor("op_8550"), val = tensor([1, 1])]; - tensor var_8552 = const()[name = tensor("op_8552"), val = tensor([1, 1])]; - tensor input_635_pad_type_0 = const()[name = tensor("input_635_pad_type_0"), val = tensor("custom")]; - tensor input_635_pad_0 = const()[name = tensor("input_635_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436823552)))]; - tensor input_635_cast_fp16 = conv(dilations = var_8552, groups = var_8457, pad = input_635_pad_0, pad_type = input_635_pad_type_0, strides = var_8550, weight = layers_21_self_attn_v_proj_loraA_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("input_635_cast_fp16")]; - tensor var_8556 = const()[name = tensor("op_8556"), val = tensor([1, 1])]; - tensor var_8558 = const()[name = tensor("op_8558"), val = tensor([1, 1])]; - tensor lora_out_849_pad_type_0 = const()[name = tensor("lora_out_849_pad_type_0"), val = tensor("custom")]; - tensor lora_out_849_pad_0 = const()[name = tensor("lora_out_849_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_851_weight_0_to_fp16 = const()[name = tensor("lora_out_851_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436864576)))]; - tensor lora_out_851_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8558, groups = var_8457, pad = lora_out_849_pad_0, pad_type = lora_out_849_pad_type_0, strides = var_8556, weight = lora_out_851_weight_0_to_fp16, x = input_635_cast_fp16)[name = tensor("lora_out_851_cast_fp16")]; - tensor current_value_43_cast_fp16 = add(x = pretrained_out_425_cast_fp16, y = lora_out_851_cast_fp16)[name = tensor("current_value_43_cast_fp16")]; - tensor var_8568_cast_fp16 = mul(x = current_key_43_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_8568_cast_fp16")]; - tensor var_8570_cast_fp16 = mul(x = var_103_cast_fp16_21, y = var_295_cast_fp16)[name = tensor("op_8570_cast_fp16")]; - tensor key_85_cast_fp16 = add(x = var_8568_cast_fp16, y = var_8570_cast_fp16)[name = tensor("key_85_cast_fp16")]; - tensor var_8572_cast_fp16 = mul(x = current_value_43_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_8572_cast_fp16")]; - tensor var_8574_cast_fp16 = mul(x = var_138_cast_fp16_21, y = var_295_cast_fp16)[name = tensor("op_8574_cast_fp16")]; - tensor value_85_cast_fp16 = add(x = var_8572_cast_fp16, y = var_8574_cast_fp16)[name = tensor("value_85_cast_fp16")]; - tensor var_8577 = const()[name = tensor("op_8577"), val = tensor([1, 20, 64, -1])]; - tensor var_8578_cast_fp16 = reshape(shape = var_8577, x = query_85_cast_fp16)[name = tensor("op_8578_cast_fp16")]; - tensor var_8579_to_fp16 = const()[name = tensor("op_8579_to_fp16"), val = tensor(0x1p-3)]; - tensor var_8580_cast_fp16 = mul(x = var_8578_cast_fp16, y = var_8579_to_fp16)[name = tensor("op_8580_cast_fp16")]; - tensor var_8581 = const()[name = tensor("op_8581"), val = tensor([1, 20, 64, -1])]; - tensor var_8582_cast_fp16 = reshape(shape = var_8581, x = key_85_cast_fp16)[name = tensor("op_8582_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_8580_cast_fp16, y = var_8582_cast_fp16)[name = tensor("mh_w_127_cast_fp16")]; - tensor mh_w_129_cast_fp16 = add(x = mh_w_127_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_129_cast_fp16")]; - tensor var_8590_cast_fp16 = softmax(axis = var_8450, x = mh_w_129_cast_fp16)[name = tensor("op_8590_cast_fp16")]; - tensor var_8591 = const()[name = tensor("op_8591"), val = tensor([1, 20, 64, -1])]; - tensor var_8592_cast_fp16 = reshape(shape = var_8591, x = value_85_cast_fp16)[name = tensor("op_8592_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_8592_cast_fp16, y = var_8590_cast_fp16)[name = tensor("attn_85_cast_fp16")]; - tensor var_8595 = const()[name = tensor("op_8595"), val = tensor([1, 1280, 1, -1])]; - tensor input_637_cast_fp16 = reshape(shape = var_8595, x = attn_85_cast_fp16)[name = tensor("input_637_cast_fp16")]; - tensor var_8602 = const()[name = tensor("op_8602"), val = tensor([1, 1])]; - tensor var_8604 = const()[name = tensor("op_8604"), val = tensor([1, 1])]; - tensor pretrained_out_427_pad_type_0 = const()[name = tensor("pretrained_out_427_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_427_pad_0 = const()[name = tensor("pretrained_out_427_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436905600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437724864))), name = tensor("layers_21_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_21_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437724992)))]; - tensor pretrained_out_427_cast_fp16 = conv(bias = layers_21_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_8604, groups = var_8457, pad = pretrained_out_427_pad_0, pad_type = pretrained_out_427_pad_type_0, strides = var_8602, weight = layers_21_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_637_cast_fp16)[name = tensor("pretrained_out_427_cast_fp16")]; - tensor var_8608 = const()[name = tensor("op_8608"), val = tensor([1, 1])]; - tensor var_8610 = const()[name = tensor("op_8610"), val = tensor([1, 1])]; - tensor input_639_pad_type_0 = const()[name = tensor("input_639_pad_type_0"), val = tensor("custom")]; - tensor input_639_pad_0 = const()[name = tensor("input_639_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437727616)))]; - tensor input_639_cast_fp16 = conv(dilations = var_8610, groups = var_8457, pad = input_639_pad_0, pad_type = input_639_pad_type_0, strides = var_8608, weight = layers_21_self_attn_o_proj_loraA_weight_to_fp16, x = input_637_cast_fp16)[name = tensor("input_639_cast_fp16")]; - tensor var_8614 = const()[name = tensor("op_8614"), val = tensor([1, 1])]; - tensor var_8616 = const()[name = tensor("op_8616"), val = tensor([1, 1])]; - tensor lora_out_853_pad_type_0 = const()[name = tensor("lora_out_853_pad_type_0"), val = tensor("custom")]; - tensor lora_out_853_pad_0 = const()[name = tensor("lora_out_853_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_855_weight_0_to_fp16 = const()[name = tensor("lora_out_855_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437768640)))]; - tensor lora_out_855_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8616, groups = var_8457, pad = lora_out_853_pad_0, pad_type = lora_out_853_pad_type_0, strides = var_8614, weight = lora_out_855_weight_0_to_fp16, x = input_639_cast_fp16)[name = tensor("lora_out_855_cast_fp16")]; - tensor obj_259_cast_fp16 = add(x = pretrained_out_427_cast_fp16, y = lora_out_855_cast_fp16)[name = tensor("obj_259_cast_fp16")]; - tensor inputs_129_cast_fp16 = add(x = inputs_127_cast_fp16, y = obj_259_cast_fp16)[name = tensor("inputs_129_cast_fp16")]; - tensor var_8629 = const()[name = tensor("op_8629"), val = tensor([1])]; - tensor channels_mean_129_cast_fp16 = reduce_mean(axes = var_8629, keep_dims = var_8458, 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_8633 = const()[name = tensor("op_8633"), val = tensor([1])]; - tensor var_8634_cast_fp16 = reduce_mean(axes = var_8633, keep_dims = var_8458, x = zero_mean_sq_129_cast_fp16)[name = tensor("op_8634_cast_fp16")]; - tensor var_8635_to_fp16 = const()[name = tensor("op_8635_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_8636_cast_fp16 = add(x = var_8634_cast_fp16, y = var_8635_to_fp16)[name = tensor("op_8636_cast_fp16")]; - tensor denom_129_epsilon_0 = const()[name = tensor("denom_129_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_129_cast_fp16 = rsqrt(epsilon = denom_129_epsilon_0, x = var_8636_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_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(437809664)))]; - 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(437812288)))]; - 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_129_cast_fp16)[name = tensor("obj_261_cast_fp16")]; - tensor var_8654 = const()[name = tensor("op_8654"), val = tensor([1, 1])]; - tensor var_8656 = const()[name = tensor("op_8656"), val = tensor([1, 1])]; - tensor pretrained_out_429_pad_type_0 = const()[name = tensor("pretrained_out_429_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_429_pad_0 = const()[name = tensor("pretrained_out_429_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437814912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438634176))), name = tensor("layers_21_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_21_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438634304)))]; - tensor pretrained_out_429_cast_fp16 = conv(bias = layers_21_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_8656, groups = var_8457, pad = pretrained_out_429_pad_0, pad_type = pretrained_out_429_pad_type_0, strides = var_8654, weight = layers_21_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_261_cast_fp16)[name = tensor("pretrained_out_429_cast_fp16")]; - tensor var_8660 = const()[name = tensor("op_8660"), val = tensor([1, 1])]; - tensor var_8662 = const()[name = tensor("op_8662"), val = tensor([1, 1])]; - tensor input_641_pad_type_0 = const()[name = tensor("input_641_pad_type_0"), val = tensor("custom")]; - tensor input_641_pad_0 = const()[name = tensor("input_641_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438636928)))]; - tensor input_641_cast_fp16 = conv(dilations = var_8662, groups = var_8457, pad = input_641_pad_0, pad_type = input_641_pad_type_0, strides = var_8660, weight = layers_21_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_261_cast_fp16)[name = tensor("input_641_cast_fp16")]; - tensor var_8666 = const()[name = tensor("op_8666"), val = tensor([1, 1])]; - tensor var_8668 = const()[name = tensor("op_8668"), val = tensor([1, 1])]; - tensor lora_out_857_pad_type_0 = const()[name = tensor("lora_out_857_pad_type_0"), val = tensor("custom")]; - tensor lora_out_857_pad_0 = const()[name = tensor("lora_out_857_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_859_weight_0_to_fp16 = const()[name = tensor("lora_out_859_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438677952)))]; - tensor lora_out_859_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8668, groups = var_8457, pad = lora_out_857_pad_0, pad_type = lora_out_857_pad_type_0, strides = var_8666, weight = lora_out_859_weight_0_to_fp16, x = input_641_cast_fp16)[name = tensor("lora_out_859_cast_fp16")]; - tensor query_87_cast_fp16 = add(x = pretrained_out_429_cast_fp16, y = lora_out_859_cast_fp16)[name = tensor("query_87_cast_fp16")]; - tensor var_8678 = const()[name = tensor("op_8678"), val = tensor([1, 1])]; - tensor var_8680 = const()[name = tensor("op_8680"), val = tensor([1, 1])]; - tensor pretrained_out_431_pad_type_0 = const()[name = tensor("pretrained_out_431_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_431_pad_0 = const()[name = tensor("pretrained_out_431_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438718976))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439538240))), name = tensor("layers_21_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_431_cast_fp16 = conv(dilations = var_8680, groups = var_8457, pad = pretrained_out_431_pad_0, pad_type = pretrained_out_431_pad_type_0, strides = var_8678, weight = layers_21_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_431_cast_fp16")]; - tensor var_8684 = const()[name = tensor("op_8684"), val = tensor([1, 1])]; - tensor var_8686 = const()[name = tensor("op_8686"), val = tensor([1, 1])]; - tensor input_643_pad_type_0 = const()[name = tensor("input_643_pad_type_0"), val = tensor("custom")]; - tensor input_643_pad_0 = const()[name = tensor("input_643_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439538368)))]; - tensor input_643_cast_fp16 = conv(dilations = var_8686, groups = var_8457, pad = input_643_pad_0, pad_type = input_643_pad_type_0, strides = var_8684, weight = layers_21_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_643_cast_fp16")]; - tensor var_8690 = const()[name = tensor("op_8690"), val = tensor([1, 1])]; - tensor var_8692 = const()[name = tensor("op_8692"), val = tensor([1, 1])]; - tensor lora_out_861_pad_type_0 = const()[name = tensor("lora_out_861_pad_type_0"), val = tensor("custom")]; - tensor lora_out_861_pad_0 = const()[name = tensor("lora_out_861_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_863_weight_0_to_fp16 = const()[name = tensor("lora_out_863_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439579392)))]; - tensor lora_out_863_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8692, groups = var_8457, pad = lora_out_861_pad_0, pad_type = lora_out_861_pad_type_0, strides = var_8690, weight = lora_out_863_weight_0_to_fp16, x = input_643_cast_fp16)[name = tensor("lora_out_863_cast_fp16")]; - tensor key_87_cast_fp16 = add(x = pretrained_out_431_cast_fp16, y = lora_out_863_cast_fp16)[name = tensor("key_87_cast_fp16")]; - tensor var_8703 = const()[name = tensor("op_8703"), val = tensor([1, 1])]; - tensor var_8705 = const()[name = tensor("op_8705"), val = tensor([1, 1])]; - tensor pretrained_out_433_pad_type_0 = const()[name = tensor("pretrained_out_433_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_433_pad_0 = const()[name = tensor("pretrained_out_433_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439620416))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440439680))), name = tensor("layers_21_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_21_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440439808)))]; - tensor pretrained_out_433_cast_fp16 = conv(bias = layers_21_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_8705, groups = var_8457, pad = pretrained_out_433_pad_0, pad_type = pretrained_out_433_pad_type_0, strides = var_8703, weight = layers_21_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_433_cast_fp16")]; - tensor var_8709 = const()[name = tensor("op_8709"), val = tensor([1, 1])]; - tensor var_8711 = const()[name = tensor("op_8711"), val = tensor([1, 1])]; - tensor input_645_pad_type_0 = const()[name = tensor("input_645_pad_type_0"), val = tensor("custom")]; - tensor input_645_pad_0 = const()[name = tensor("input_645_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440442432)))]; - tensor input_645_cast_fp16 = conv(dilations = var_8711, groups = var_8457, pad = input_645_pad_0, pad_type = input_645_pad_type_0, strides = var_8709, weight = layers_21_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_645_cast_fp16")]; - tensor var_8715 = const()[name = tensor("op_8715"), val = tensor([1, 1])]; - tensor var_8717 = const()[name = tensor("op_8717"), val = tensor([1, 1])]; - tensor lora_out_865_pad_type_0 = const()[name = tensor("lora_out_865_pad_type_0"), val = tensor("custom")]; - tensor lora_out_865_pad_0 = const()[name = tensor("lora_out_865_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_867_weight_0_to_fp16 = const()[name = tensor("lora_out_867_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440483456)))]; - tensor lora_out_867_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8717, groups = var_8457, pad = lora_out_865_pad_0, pad_type = lora_out_865_pad_type_0, strides = var_8715, weight = lora_out_867_weight_0_to_fp16, x = input_645_cast_fp16)[name = tensor("lora_out_867_cast_fp16")]; - tensor value_87_cast_fp16 = add(x = pretrained_out_433_cast_fp16, y = lora_out_867_cast_fp16)[name = tensor("value_87_cast_fp16")]; - tensor var_8724 = const()[name = tensor("op_8724"), val = tensor([1, 20, 64, -1])]; - tensor var_8725_cast_fp16 = reshape(shape = var_8724, x = query_87_cast_fp16)[name = tensor("op_8725_cast_fp16")]; - tensor var_8726_to_fp16 = const()[name = tensor("op_8726_to_fp16"), val = tensor(0x1p-3)]; - tensor var_8727_cast_fp16 = mul(x = var_8725_cast_fp16, y = var_8726_to_fp16)[name = tensor("op_8727_cast_fp16")]; - tensor var_8728 = const()[name = tensor("op_8728"), val = tensor([1, 20, 64, -1])]; - tensor var_8729_cast_fp16 = reshape(shape = var_8728, x = key_87_cast_fp16)[name = tensor("op_8729_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_8727_cast_fp16, y = var_8729_cast_fp16)[name = tensor("mh_w_131_cast_fp16")]; - tensor var_8732_cast_fp16 = softmax(axis = var_8450, x = mh_w_131_cast_fp16)[name = tensor("op_8732_cast_fp16")]; - tensor var_8733 = const()[name = tensor("op_8733"), val = tensor([1, 20, 64, -1])]; - tensor var_8734_cast_fp16 = reshape(shape = var_8733, x = value_87_cast_fp16)[name = tensor("op_8734_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_8734_cast_fp16, y = var_8732_cast_fp16)[name = tensor("attn_87_cast_fp16")]; - tensor var_8737 = const()[name = tensor("op_8737"), val = tensor([1, 1280, 1, -1])]; - tensor input_647_cast_fp16 = reshape(shape = var_8737, x = attn_87_cast_fp16)[name = tensor("input_647_cast_fp16")]; - tensor var_8744 = const()[name = tensor("op_8744"), val = tensor([1, 1])]; - tensor var_8746 = const()[name = tensor("op_8746"), val = tensor([1, 1])]; - tensor pretrained_out_435_pad_type_0 = const()[name = tensor("pretrained_out_435_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_435_pad_0 = const()[name = tensor("pretrained_out_435_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440524480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441343744))), name = tensor("layers_21_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_21_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441343872)))]; - tensor pretrained_out_435_cast_fp16 = conv(bias = layers_21_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_8746, groups = var_8457, pad = pretrained_out_435_pad_0, pad_type = pretrained_out_435_pad_type_0, strides = var_8744, weight = layers_21_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_647_cast_fp16)[name = tensor("pretrained_out_435_cast_fp16")]; - tensor var_8750 = const()[name = tensor("op_8750"), val = tensor([1, 1])]; - tensor var_8752 = const()[name = tensor("op_8752"), val = tensor([1, 1])]; - tensor input_649_pad_type_0 = const()[name = tensor("input_649_pad_type_0"), val = tensor("custom")]; - tensor input_649_pad_0 = const()[name = tensor("input_649_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441346496)))]; - tensor input_649_cast_fp16 = conv(dilations = var_8752, groups = var_8457, pad = input_649_pad_0, pad_type = input_649_pad_type_0, strides = var_8750, weight = layers_21_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_647_cast_fp16)[name = tensor("input_649_cast_fp16")]; - tensor var_8756 = const()[name = tensor("op_8756"), val = tensor([1, 1])]; - tensor var_8758 = const()[name = tensor("op_8758"), val = tensor([1, 1])]; - tensor lora_out_869_pad_type_0 = const()[name = tensor("lora_out_869_pad_type_0"), val = tensor("custom")]; - tensor lora_out_869_pad_0 = const()[name = tensor("lora_out_869_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_871_weight_0_to_fp16 = const()[name = tensor("lora_out_871_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441387520)))]; - tensor lora_out_871_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8758, groups = var_8457, pad = lora_out_869_pad_0, pad_type = lora_out_869_pad_type_0, strides = var_8756, weight = lora_out_871_weight_0_to_fp16, x = input_649_cast_fp16)[name = tensor("lora_out_871_cast_fp16")]; - tensor obj_263_cast_fp16 = add(x = pretrained_out_435_cast_fp16, y = lora_out_871_cast_fp16)[name = tensor("obj_263_cast_fp16")]; - tensor inputs_131_cast_fp16 = add(x = inputs_129_cast_fp16, y = obj_263_cast_fp16)[name = tensor("inputs_131_cast_fp16")]; - tensor var_8767 = const()[name = tensor("op_8767"), val = tensor([1])]; - tensor channels_mean_131_cast_fp16 = reduce_mean(axes = var_8767, keep_dims = var_8458, 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_8771 = const()[name = tensor("op_8771"), val = tensor([1])]; - tensor var_8772_cast_fp16 = reduce_mean(axes = var_8771, keep_dims = var_8458, x = zero_mean_sq_131_cast_fp16)[name = tensor("op_8772_cast_fp16")]; - tensor var_8773_to_fp16 = const()[name = tensor("op_8773_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_8774_cast_fp16 = add(x = var_8772_cast_fp16, y = var_8773_to_fp16)[name = tensor("op_8774_cast_fp16")]; - tensor denom_131_epsilon_0 = const()[name = tensor("denom_131_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_131_cast_fp16 = rsqrt(epsilon = denom_131_epsilon_0, x = var_8774_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_651_gamma_0_to_fp16 = const()[name = tensor("input_651_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441428544)))]; - tensor input_651_beta_0_to_fp16 = const()[name = tensor("input_651_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441431168)))]; - tensor input_651_epsilon_0_to_fp16 = const()[name = tensor("input_651_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_651_cast_fp16 = batch_norm(beta = input_651_beta_0_to_fp16, epsilon = input_651_epsilon_0_to_fp16, gamma = input_651_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_651_cast_fp16")]; - tensor var_8788 = const()[name = tensor("op_8788"), val = tensor([1, 1])]; - tensor var_8790 = const()[name = tensor("op_8790"), val = tensor([1, 1])]; - tensor pretrained_out_437_pad_type_0 = const()[name = tensor("pretrained_out_437_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_437_pad_0 = const()[name = tensor("pretrained_out_437_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441433792))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444710656))), name = tensor("layers_21_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_21_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_21_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444710784)))]; - tensor pretrained_out_437_cast_fp16 = conv(bias = layers_21_fc1_pretrained_bias_to_fp16, dilations = var_8790, groups = var_8457, pad = pretrained_out_437_pad_0, pad_type = pretrained_out_437_pad_type_0, strides = var_8788, weight = layers_21_fc1_pretrained_weight_to_fp16_palettized, x = input_651_cast_fp16)[name = tensor("pretrained_out_437_cast_fp16")]; - tensor var_8794 = const()[name = tensor("op_8794"), val = tensor([1, 1])]; - tensor var_8796 = const()[name = tensor("op_8796"), val = tensor([1, 1])]; - tensor input_653_pad_type_0 = const()[name = tensor("input_653_pad_type_0"), val = tensor("custom")]; - tensor input_653_pad_0 = const()[name = tensor("input_653_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_21_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444721088)))]; - tensor input_653_cast_fp16 = conv(dilations = var_8796, groups = var_8457, pad = input_653_pad_0, pad_type = input_653_pad_type_0, strides = var_8794, weight = layers_21_fc1_loraA_weight_to_fp16, x = input_651_cast_fp16)[name = tensor("input_653_cast_fp16")]; - tensor var_8800 = const()[name = tensor("op_8800"), val = tensor([1, 1])]; - tensor var_8802 = const()[name = tensor("op_8802"), val = tensor([1, 1])]; - tensor lora_out_873_pad_type_0 = const()[name = tensor("lora_out_873_pad_type_0"), val = tensor("custom")]; - tensor lora_out_873_pad_0 = const()[name = tensor("lora_out_873_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_875_weight_0_to_fp16 = const()[name = tensor("lora_out_875_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444762112)))]; - tensor lora_out_875_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_8802, groups = var_8457, pad = lora_out_873_pad_0, pad_type = lora_out_873_pad_type_0, strides = var_8800, weight = lora_out_875_weight_0_to_fp16, x = input_653_cast_fp16)[name = tensor("lora_out_875_cast_fp16")]; - tensor input_655_cast_fp16 = add(x = pretrained_out_437_cast_fp16, y = lora_out_875_cast_fp16)[name = tensor("input_655_cast_fp16")]; - tensor input_657_mode_0 = const()[name = tensor("input_657_mode_0"), val = tensor("EXACT")]; - tensor input_657_cast_fp16 = gelu(mode = input_657_mode_0, x = input_655_cast_fp16)[name = tensor("input_657_cast_fp16")]; - tensor var_8814 = const()[name = tensor("op_8814"), val = tensor([1, 1])]; - tensor var_8816 = const()[name = tensor("op_8816"), val = tensor([1, 1])]; - tensor pretrained_out_439_pad_type_0 = const()[name = tensor("pretrained_out_439_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_439_pad_0 = const()[name = tensor("pretrained_out_439_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444926016))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448202880))), name = tensor("layers_21_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_21_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_21_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448203008)))]; - tensor pretrained_out_439_cast_fp16 = conv(bias = layers_21_fc2_pretrained_bias_to_fp16, dilations = var_8816, groups = var_8457, pad = pretrained_out_439_pad_0, pad_type = pretrained_out_439_pad_type_0, strides = var_8814, weight = layers_21_fc2_pretrained_weight_to_fp16_palettized, x = input_657_cast_fp16)[name = tensor("pretrained_out_439_cast_fp16")]; - tensor var_8820 = const()[name = tensor("op_8820"), val = tensor([1, 1])]; - tensor var_8822 = const()[name = tensor("op_8822"), val = tensor([1, 1])]; - tensor input_659_pad_type_0 = const()[name = tensor("input_659_pad_type_0"), val = tensor("custom")]; - tensor input_659_pad_0 = const()[name = tensor("input_659_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_21_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_21_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448205632)))]; - tensor input_659_cast_fp16 = conv(dilations = var_8822, groups = var_8457, pad = input_659_pad_0, pad_type = input_659_pad_type_0, strides = var_8820, weight = layers_21_fc2_loraA_weight_to_fp16, x = input_657_cast_fp16)[name = tensor("input_659_cast_fp16")]; - tensor var_8826 = const()[name = tensor("op_8826"), val = tensor([1, 1])]; - tensor var_8828 = const()[name = tensor("op_8828"), val = tensor([1, 1])]; - tensor lora_out_877_pad_type_0 = const()[name = tensor("lora_out_877_pad_type_0"), val = tensor("custom")]; - tensor lora_out_877_pad_0 = const()[name = tensor("lora_out_877_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_879_weight_0_to_fp16 = const()[name = tensor("lora_out_879_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448369536)))]; - tensor lora_out_879_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8828, groups = var_8457, pad = lora_out_877_pad_0, pad_type = lora_out_877_pad_type_0, strides = var_8826, weight = lora_out_879_weight_0_to_fp16, x = input_659_cast_fp16)[name = tensor("lora_out_879_cast_fp16")]; - tensor hidden_states_45_cast_fp16 = add(x = pretrained_out_439_cast_fp16, y = lora_out_879_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_8844 = const()[name = tensor("op_8844"), val = tensor(3)]; - tensor var_8851 = const()[name = tensor("op_8851"), val = tensor(1)]; - tensor var_8852 = const()[name = tensor("op_8852"), val = tensor(true)]; - tensor var_8864 = const()[name = tensor("op_8864"), val = tensor([1])]; - tensor channels_mean_133_cast_fp16 = reduce_mean(axes = var_8864, keep_dims = var_8852, 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_8868 = const()[name = tensor("op_8868"), val = tensor([1])]; - tensor var_8869_cast_fp16 = reduce_mean(axes = var_8868, keep_dims = var_8852, x = zero_mean_sq_133_cast_fp16)[name = tensor("op_8869_cast_fp16")]; - tensor var_8870_to_fp16 = const()[name = tensor("op_8870_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_8871_cast_fp16 = add(x = var_8869_cast_fp16, y = var_8870_to_fp16)[name = tensor("op_8871_cast_fp16")]; - tensor denom_133_epsilon_0 = const()[name = tensor("denom_133_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_133_cast_fp16 = rsqrt(epsilon = denom_133_epsilon_0, x = var_8871_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_265_gamma_0_to_fp16 = const()[name = tensor("obj_265_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448410560)))]; - tensor obj_265_beta_0_to_fp16 = const()[name = tensor("obj_265_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448413184)))]; - tensor obj_265_epsilon_0_to_fp16 = const()[name = tensor("obj_265_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_265_cast_fp16 = batch_norm(beta = obj_265_beta_0_to_fp16, epsilon = obj_265_epsilon_0_to_fp16, gamma = obj_265_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_265_cast_fp16")]; - tensor var_8889 = const()[name = tensor("op_8889"), val = tensor([1, 1])]; - tensor var_8891 = const()[name = tensor("op_8891"), val = tensor([1, 1])]; - tensor pretrained_out_441_pad_type_0 = const()[name = tensor("pretrained_out_441_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_441_pad_0 = const()[name = tensor("pretrained_out_441_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448415808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(449235072))), name = tensor("layers_22_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_22_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(449235200)))]; - tensor pretrained_out_441_cast_fp16 = conv(bias = layers_22_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_8891, groups = var_8851, pad = pretrained_out_441_pad_0, pad_type = pretrained_out_441_pad_type_0, strides = var_8889, weight = layers_22_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_265_cast_fp16)[name = tensor("pretrained_out_441_cast_fp16")]; - tensor var_8895 = const()[name = tensor("op_8895"), val = tensor([1, 1])]; - tensor var_8897 = const()[name = tensor("op_8897"), val = tensor([1, 1])]; - tensor input_661_pad_type_0 = const()[name = tensor("input_661_pad_type_0"), val = tensor("custom")]; - tensor input_661_pad_0 = const()[name = tensor("input_661_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(449237824)))]; - tensor input_661_cast_fp16 = conv(dilations = var_8897, groups = var_8851, pad = input_661_pad_0, pad_type = input_661_pad_type_0, strides = var_8895, weight = layers_22_self_attn_q_proj_loraA_weight_to_fp16, x = obj_265_cast_fp16)[name = tensor("input_661_cast_fp16")]; - tensor var_8901 = const()[name = tensor("op_8901"), val = tensor([1, 1])]; - tensor var_8903 = const()[name = tensor("op_8903"), val = tensor([1, 1])]; - tensor lora_out_881_pad_type_0 = const()[name = tensor("lora_out_881_pad_type_0"), val = tensor("custom")]; - tensor lora_out_881_pad_0 = const()[name = tensor("lora_out_881_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_883_weight_0_to_fp16 = const()[name = tensor("lora_out_883_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(449278848)))]; - tensor lora_out_883_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8903, groups = var_8851, pad = lora_out_881_pad_0, pad_type = lora_out_881_pad_type_0, strides = var_8901, weight = lora_out_883_weight_0_to_fp16, x = input_661_cast_fp16)[name = tensor("lora_out_883_cast_fp16")]; - tensor query_89_cast_fp16 = add(x = pretrained_out_441_cast_fp16, y = lora_out_883_cast_fp16)[name = tensor("query_89_cast_fp16")]; - tensor var_8913 = const()[name = tensor("op_8913"), val = tensor([1, 1])]; - tensor var_8915 = const()[name = tensor("op_8915"), val = tensor([1, 1])]; - tensor pretrained_out_443_pad_type_0 = const()[name = tensor("pretrained_out_443_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_443_pad_0 = const()[name = tensor("pretrained_out_443_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(449319872))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450139136))), name = tensor("layers_22_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_443_cast_fp16 = conv(dilations = var_8915, groups = var_8851, pad = pretrained_out_443_pad_0, pad_type = pretrained_out_443_pad_type_0, strides = var_8913, weight = layers_22_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_265_cast_fp16)[name = tensor("pretrained_out_443_cast_fp16")]; - tensor var_8919 = const()[name = tensor("op_8919"), val = tensor([1, 1])]; - tensor var_8921 = const()[name = tensor("op_8921"), val = tensor([1, 1])]; - tensor input_663_pad_type_0 = const()[name = tensor("input_663_pad_type_0"), val = tensor("custom")]; - tensor input_663_pad_0 = const()[name = tensor("input_663_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450139264)))]; - tensor input_663_cast_fp16 = conv(dilations = var_8921, groups = var_8851, pad = input_663_pad_0, pad_type = input_663_pad_type_0, strides = var_8919, weight = layers_22_self_attn_k_proj_loraA_weight_to_fp16, x = obj_265_cast_fp16)[name = tensor("input_663_cast_fp16")]; - tensor var_8925 = const()[name = tensor("op_8925"), val = tensor([1, 1])]; - tensor var_8927 = const()[name = tensor("op_8927"), val = tensor([1, 1])]; - tensor lora_out_885_pad_type_0 = const()[name = tensor("lora_out_885_pad_type_0"), val = tensor("custom")]; - tensor lora_out_885_pad_0 = const()[name = tensor("lora_out_885_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_887_weight_0_to_fp16 = const()[name = tensor("lora_out_887_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450180288)))]; - tensor lora_out_887_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8927, groups = var_8851, pad = lora_out_885_pad_0, pad_type = lora_out_885_pad_type_0, strides = var_8925, weight = lora_out_887_weight_0_to_fp16, x = input_663_cast_fp16)[name = tensor("lora_out_887_cast_fp16")]; - tensor current_key_45_cast_fp16 = add(x = pretrained_out_443_cast_fp16, y = lora_out_887_cast_fp16)[name = tensor("current_key_45_cast_fp16")]; - tensor var_8938 = const()[name = tensor("op_8938"), val = tensor([1, 1])]; - tensor var_8940 = const()[name = tensor("op_8940"), val = tensor([1, 1])]; - tensor pretrained_out_445_pad_type_0 = const()[name = tensor("pretrained_out_445_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_445_pad_0 = const()[name = tensor("pretrained_out_445_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450221312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451040576))), name = tensor("layers_22_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_22_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451040704)))]; - tensor pretrained_out_445_cast_fp16 = conv(bias = layers_22_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_8940, groups = var_8851, pad = pretrained_out_445_pad_0, pad_type = pretrained_out_445_pad_type_0, strides = var_8938, weight = layers_22_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_265_cast_fp16)[name = tensor("pretrained_out_445_cast_fp16")]; - tensor var_8944 = const()[name = tensor("op_8944"), val = tensor([1, 1])]; - tensor var_8946 = const()[name = tensor("op_8946"), val = tensor([1, 1])]; - tensor input_665_pad_type_0 = const()[name = tensor("input_665_pad_type_0"), val = tensor("custom")]; - tensor input_665_pad_0 = const()[name = tensor("input_665_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451043328)))]; - tensor input_665_cast_fp16 = conv(dilations = var_8946, groups = var_8851, pad = input_665_pad_0, pad_type = input_665_pad_type_0, strides = var_8944, weight = layers_22_self_attn_v_proj_loraA_weight_to_fp16, x = obj_265_cast_fp16)[name = tensor("input_665_cast_fp16")]; - tensor var_8950 = const()[name = tensor("op_8950"), val = tensor([1, 1])]; - tensor var_8952 = const()[name = tensor("op_8952"), val = tensor([1, 1])]; - tensor lora_out_889_pad_type_0 = const()[name = tensor("lora_out_889_pad_type_0"), val = tensor("custom")]; - tensor lora_out_889_pad_0 = const()[name = tensor("lora_out_889_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_891_weight_0_to_fp16 = const()[name = tensor("lora_out_891_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451084352)))]; - tensor lora_out_891_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8952, groups = var_8851, pad = lora_out_889_pad_0, pad_type = lora_out_889_pad_type_0, strides = var_8950, weight = lora_out_891_weight_0_to_fp16, x = input_665_cast_fp16)[name = tensor("lora_out_891_cast_fp16")]; - tensor current_value_45_cast_fp16 = add(x = pretrained_out_445_cast_fp16, y = lora_out_891_cast_fp16)[name = tensor("current_value_45_cast_fp16")]; - tensor var_8962_cast_fp16 = mul(x = current_key_45_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_8962_cast_fp16")]; - tensor var_8964_cast_fp16 = mul(x = var_103_cast_fp16_22, y = var_295_cast_fp16)[name = tensor("op_8964_cast_fp16")]; - tensor key_89_cast_fp16 = add(x = var_8962_cast_fp16, y = var_8964_cast_fp16)[name = tensor("key_89_cast_fp16")]; - tensor var_8966_cast_fp16 = mul(x = current_value_45_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_8966_cast_fp16")]; - tensor var_8968_cast_fp16 = mul(x = var_138_cast_fp16_22, y = var_295_cast_fp16)[name = tensor("op_8968_cast_fp16")]; - tensor value_89_cast_fp16 = add(x = var_8966_cast_fp16, y = var_8968_cast_fp16)[name = tensor("value_89_cast_fp16")]; - tensor var_8971 = const()[name = tensor("op_8971"), val = tensor([1, 20, 64, -1])]; - tensor var_8972_cast_fp16 = reshape(shape = var_8971, x = query_89_cast_fp16)[name = tensor("op_8972_cast_fp16")]; - tensor var_8973_to_fp16 = const()[name = tensor("op_8973_to_fp16"), val = tensor(0x1p-3)]; - tensor var_8974_cast_fp16 = mul(x = var_8972_cast_fp16, y = var_8973_to_fp16)[name = tensor("op_8974_cast_fp16")]; - tensor var_8975 = const()[name = tensor("op_8975"), val = tensor([1, 20, 64, -1])]; - tensor var_8976_cast_fp16 = reshape(shape = var_8975, x = key_89_cast_fp16)[name = tensor("op_8976_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_8974_cast_fp16, y = var_8976_cast_fp16)[name = tensor("mh_w_133_cast_fp16")]; - tensor mh_w_135_cast_fp16 = add(x = mh_w_133_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_135_cast_fp16")]; - tensor var_8984_cast_fp16 = softmax(axis = var_8844, x = mh_w_135_cast_fp16)[name = tensor("op_8984_cast_fp16")]; - tensor var_8985 = const()[name = tensor("op_8985"), val = tensor([1, 20, 64, -1])]; - tensor var_8986_cast_fp16 = reshape(shape = var_8985, x = value_89_cast_fp16)[name = tensor("op_8986_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_8986_cast_fp16, y = var_8984_cast_fp16)[name = tensor("attn_89_cast_fp16")]; - tensor var_8989 = const()[name = tensor("op_8989"), val = tensor([1, 1280, 1, -1])]; - tensor input_667_cast_fp16 = reshape(shape = var_8989, x = attn_89_cast_fp16)[name = tensor("input_667_cast_fp16")]; - tensor var_8996 = const()[name = tensor("op_8996"), val = tensor([1, 1])]; - tensor var_8998 = const()[name = tensor("op_8998"), val = tensor([1, 1])]; - tensor pretrained_out_447_pad_type_0 = const()[name = tensor("pretrained_out_447_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_447_pad_0 = const()[name = tensor("pretrained_out_447_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451125376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451944640))), name = tensor("layers_22_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_22_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451944768)))]; - tensor pretrained_out_447_cast_fp16 = conv(bias = layers_22_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_8998, groups = var_8851, pad = pretrained_out_447_pad_0, pad_type = pretrained_out_447_pad_type_0, strides = var_8996, weight = layers_22_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_667_cast_fp16)[name = tensor("pretrained_out_447_cast_fp16")]; - tensor var_9002 = const()[name = tensor("op_9002"), val = tensor([1, 1])]; - tensor var_9004 = const()[name = tensor("op_9004"), val = tensor([1, 1])]; - tensor input_669_pad_type_0 = const()[name = tensor("input_669_pad_type_0"), val = tensor("custom")]; - tensor input_669_pad_0 = const()[name = tensor("input_669_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451947392)))]; - tensor input_669_cast_fp16 = conv(dilations = var_9004, groups = var_8851, pad = input_669_pad_0, pad_type = input_669_pad_type_0, strides = var_9002, weight = layers_22_self_attn_o_proj_loraA_weight_to_fp16, x = input_667_cast_fp16)[name = tensor("input_669_cast_fp16")]; - tensor var_9008 = const()[name = tensor("op_9008"), val = tensor([1, 1])]; - tensor var_9010 = const()[name = tensor("op_9010"), val = tensor([1, 1])]; - tensor lora_out_893_pad_type_0 = const()[name = tensor("lora_out_893_pad_type_0"), val = tensor("custom")]; - tensor lora_out_893_pad_0 = const()[name = tensor("lora_out_893_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_895_weight_0_to_fp16 = const()[name = tensor("lora_out_895_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451988416)))]; - tensor lora_out_895_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9010, groups = var_8851, pad = lora_out_893_pad_0, pad_type = lora_out_893_pad_type_0, strides = var_9008, weight = lora_out_895_weight_0_to_fp16, x = input_669_cast_fp16)[name = tensor("lora_out_895_cast_fp16")]; - tensor obj_271_cast_fp16 = add(x = pretrained_out_447_cast_fp16, y = lora_out_895_cast_fp16)[name = tensor("obj_271_cast_fp16")]; - tensor inputs_135_cast_fp16 = add(x = inputs_133_cast_fp16, y = obj_271_cast_fp16)[name = tensor("inputs_135_cast_fp16")]; - tensor var_9023 = const()[name = tensor("op_9023"), val = tensor([1])]; - tensor channels_mean_135_cast_fp16 = reduce_mean(axes = var_9023, keep_dims = var_8852, 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_9027 = const()[name = tensor("op_9027"), val = tensor([1])]; - tensor var_9028_cast_fp16 = reduce_mean(axes = var_9027, keep_dims = var_8852, x = zero_mean_sq_135_cast_fp16)[name = tensor("op_9028_cast_fp16")]; - tensor var_9029_to_fp16 = const()[name = tensor("op_9029_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_9030_cast_fp16 = add(x = var_9028_cast_fp16, y = var_9029_to_fp16)[name = tensor("op_9030_cast_fp16")]; - tensor denom_135_epsilon_0 = const()[name = tensor("denom_135_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_135_cast_fp16 = rsqrt(epsilon = denom_135_epsilon_0, x = var_9030_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_273_gamma_0_to_fp16 = const()[name = tensor("obj_273_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452029440)))]; - tensor obj_273_beta_0_to_fp16 = const()[name = tensor("obj_273_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452032064)))]; - tensor obj_273_epsilon_0_to_fp16 = const()[name = tensor("obj_273_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_273_cast_fp16 = batch_norm(beta = obj_273_beta_0_to_fp16, epsilon = obj_273_epsilon_0_to_fp16, gamma = obj_273_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_273_cast_fp16")]; - tensor var_9048 = const()[name = tensor("op_9048"), val = tensor([1, 1])]; - tensor var_9050 = const()[name = tensor("op_9050"), val = tensor([1, 1])]; - tensor pretrained_out_449_pad_type_0 = const()[name = tensor("pretrained_out_449_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_449_pad_0 = const()[name = tensor("pretrained_out_449_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452034688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452853952))), name = tensor("layers_22_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_22_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452854080)))]; - tensor pretrained_out_449_cast_fp16 = conv(bias = layers_22_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_9050, groups = var_8851, pad = pretrained_out_449_pad_0, pad_type = pretrained_out_449_pad_type_0, strides = var_9048, weight = layers_22_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_273_cast_fp16)[name = tensor("pretrained_out_449_cast_fp16")]; - tensor var_9054 = const()[name = tensor("op_9054"), val = tensor([1, 1])]; - tensor var_9056 = const()[name = tensor("op_9056"), val = tensor([1, 1])]; - tensor input_671_pad_type_0 = const()[name = tensor("input_671_pad_type_0"), val = tensor("custom")]; - tensor input_671_pad_0 = const()[name = tensor("input_671_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452856704)))]; - tensor input_671_cast_fp16 = conv(dilations = var_9056, groups = var_8851, pad = input_671_pad_0, pad_type = input_671_pad_type_0, strides = var_9054, weight = layers_22_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_273_cast_fp16)[name = tensor("input_671_cast_fp16")]; - tensor var_9060 = const()[name = tensor("op_9060"), val = tensor([1, 1])]; - tensor var_9062 = const()[name = tensor("op_9062"), val = tensor([1, 1])]; - tensor lora_out_897_pad_type_0 = const()[name = tensor("lora_out_897_pad_type_0"), val = tensor("custom")]; - tensor lora_out_897_pad_0 = const()[name = tensor("lora_out_897_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_899_weight_0_to_fp16 = const()[name = tensor("lora_out_899_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452897728)))]; - tensor lora_out_899_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9062, groups = var_8851, pad = lora_out_897_pad_0, pad_type = lora_out_897_pad_type_0, strides = var_9060, weight = lora_out_899_weight_0_to_fp16, x = input_671_cast_fp16)[name = tensor("lora_out_899_cast_fp16")]; - tensor query_91_cast_fp16 = add(x = pretrained_out_449_cast_fp16, y = lora_out_899_cast_fp16)[name = tensor("query_91_cast_fp16")]; - tensor var_9072 = const()[name = tensor("op_9072"), val = tensor([1, 1])]; - tensor var_9074 = const()[name = tensor("op_9074"), val = tensor([1, 1])]; - tensor pretrained_out_451_pad_type_0 = const()[name = tensor("pretrained_out_451_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_451_pad_0 = const()[name = tensor("pretrained_out_451_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452938752))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453758016))), name = tensor("layers_22_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_451_cast_fp16 = conv(dilations = var_9074, groups = var_8851, pad = pretrained_out_451_pad_0, pad_type = pretrained_out_451_pad_type_0, strides = var_9072, weight = layers_22_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_451_cast_fp16")]; - tensor var_9078 = const()[name = tensor("op_9078"), val = tensor([1, 1])]; - tensor var_9080 = const()[name = tensor("op_9080"), val = tensor([1, 1])]; - tensor input_673_pad_type_0 = const()[name = tensor("input_673_pad_type_0"), val = tensor("custom")]; - tensor input_673_pad_0 = const()[name = tensor("input_673_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453758144)))]; - tensor input_673_cast_fp16 = conv(dilations = var_9080, groups = var_8851, pad = input_673_pad_0, pad_type = input_673_pad_type_0, strides = var_9078, weight = layers_22_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_673_cast_fp16")]; - tensor var_9084 = const()[name = tensor("op_9084"), val = tensor([1, 1])]; - tensor var_9086 = const()[name = tensor("op_9086"), val = tensor([1, 1])]; - tensor lora_out_901_pad_type_0 = const()[name = tensor("lora_out_901_pad_type_0"), val = tensor("custom")]; - tensor lora_out_901_pad_0 = const()[name = tensor("lora_out_901_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_903_weight_0_to_fp16 = const()[name = tensor("lora_out_903_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453799168)))]; - tensor lora_out_903_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9086, groups = var_8851, pad = lora_out_901_pad_0, pad_type = lora_out_901_pad_type_0, strides = var_9084, weight = lora_out_903_weight_0_to_fp16, x = input_673_cast_fp16)[name = tensor("lora_out_903_cast_fp16")]; - tensor key_91_cast_fp16 = add(x = pretrained_out_451_cast_fp16, y = lora_out_903_cast_fp16)[name = tensor("key_91_cast_fp16")]; - tensor var_9097 = const()[name = tensor("op_9097"), val = tensor([1, 1])]; - tensor var_9099 = const()[name = tensor("op_9099"), val = tensor([1, 1])]; - tensor pretrained_out_453_pad_type_0 = const()[name = tensor("pretrained_out_453_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_453_pad_0 = const()[name = tensor("pretrained_out_453_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453840192))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454659456))), name = tensor("layers_22_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_22_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454659584)))]; - tensor pretrained_out_453_cast_fp16 = conv(bias = layers_22_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_9099, groups = var_8851, pad = pretrained_out_453_pad_0, pad_type = pretrained_out_453_pad_type_0, strides = var_9097, weight = layers_22_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_453_cast_fp16")]; - tensor var_9103 = const()[name = tensor("op_9103"), val = tensor([1, 1])]; - tensor var_9105 = const()[name = tensor("op_9105"), val = tensor([1, 1])]; - tensor input_675_pad_type_0 = const()[name = tensor("input_675_pad_type_0"), val = tensor("custom")]; - tensor input_675_pad_0 = const()[name = tensor("input_675_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454662208)))]; - tensor input_675_cast_fp16 = conv(dilations = var_9105, groups = var_8851, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = var_9103, weight = layers_22_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_675_cast_fp16")]; - tensor var_9109 = const()[name = tensor("op_9109"), val = tensor([1, 1])]; - tensor var_9111 = const()[name = tensor("op_9111"), val = tensor([1, 1])]; - tensor lora_out_905_pad_type_0 = const()[name = tensor("lora_out_905_pad_type_0"), val = tensor("custom")]; - tensor lora_out_905_pad_0 = const()[name = tensor("lora_out_905_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_907_weight_0_to_fp16 = const()[name = tensor("lora_out_907_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454703232)))]; - tensor lora_out_907_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9111, groups = var_8851, pad = lora_out_905_pad_0, pad_type = lora_out_905_pad_type_0, strides = var_9109, weight = lora_out_907_weight_0_to_fp16, x = input_675_cast_fp16)[name = tensor("lora_out_907_cast_fp16")]; - tensor value_91_cast_fp16 = add(x = pretrained_out_453_cast_fp16, y = lora_out_907_cast_fp16)[name = tensor("value_91_cast_fp16")]; - tensor var_9118 = const()[name = tensor("op_9118"), val = tensor([1, 20, 64, -1])]; - tensor var_9119_cast_fp16 = reshape(shape = var_9118, x = query_91_cast_fp16)[name = tensor("op_9119_cast_fp16")]; - tensor var_9120_to_fp16 = const()[name = tensor("op_9120_to_fp16"), val = tensor(0x1p-3)]; - tensor var_9121_cast_fp16 = mul(x = var_9119_cast_fp16, y = var_9120_to_fp16)[name = tensor("op_9121_cast_fp16")]; - tensor var_9122 = const()[name = tensor("op_9122"), val = tensor([1, 20, 64, -1])]; - tensor var_9123_cast_fp16 = reshape(shape = var_9122, x = key_91_cast_fp16)[name = tensor("op_9123_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_9121_cast_fp16, y = var_9123_cast_fp16)[name = tensor("mh_w_137_cast_fp16")]; - tensor var_9126_cast_fp16 = softmax(axis = var_8844, x = mh_w_137_cast_fp16)[name = tensor("op_9126_cast_fp16")]; - tensor var_9127 = const()[name = tensor("op_9127"), val = tensor([1, 20, 64, -1])]; - tensor var_9128_cast_fp16 = reshape(shape = var_9127, x = value_91_cast_fp16)[name = tensor("op_9128_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_9128_cast_fp16, y = var_9126_cast_fp16)[name = tensor("attn_91_cast_fp16")]; - tensor var_9131 = const()[name = tensor("op_9131"), val = tensor([1, 1280, 1, -1])]; - tensor input_677_cast_fp16 = reshape(shape = var_9131, x = attn_91_cast_fp16)[name = tensor("input_677_cast_fp16")]; - tensor var_9138 = const()[name = tensor("op_9138"), val = tensor([1, 1])]; - tensor var_9140 = const()[name = tensor("op_9140"), val = tensor([1, 1])]; - tensor pretrained_out_455_pad_type_0 = const()[name = tensor("pretrained_out_455_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_455_pad_0 = const()[name = tensor("pretrained_out_455_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454744256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455563520))), name = tensor("layers_22_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_22_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455563648)))]; - tensor pretrained_out_455_cast_fp16 = conv(bias = layers_22_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_9140, groups = var_8851, pad = pretrained_out_455_pad_0, pad_type = pretrained_out_455_pad_type_0, strides = var_9138, weight = layers_22_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_677_cast_fp16)[name = tensor("pretrained_out_455_cast_fp16")]; - tensor var_9144 = const()[name = tensor("op_9144"), val = tensor([1, 1])]; - tensor var_9146 = const()[name = tensor("op_9146"), val = tensor([1, 1])]; - tensor input_679_pad_type_0 = const()[name = tensor("input_679_pad_type_0"), val = tensor("custom")]; - tensor input_679_pad_0 = const()[name = tensor("input_679_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455566272)))]; - tensor input_679_cast_fp16 = conv(dilations = var_9146, groups = var_8851, pad = input_679_pad_0, pad_type = input_679_pad_type_0, strides = var_9144, weight = layers_22_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_677_cast_fp16)[name = tensor("input_679_cast_fp16")]; - tensor var_9150 = const()[name = tensor("op_9150"), val = tensor([1, 1])]; - tensor var_9152 = const()[name = tensor("op_9152"), val = tensor([1, 1])]; - tensor lora_out_909_pad_type_0 = const()[name = tensor("lora_out_909_pad_type_0"), val = tensor("custom")]; - tensor lora_out_909_pad_0 = const()[name = tensor("lora_out_909_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_911_weight_0_to_fp16 = const()[name = tensor("lora_out_911_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455607296)))]; - tensor lora_out_911_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9152, groups = var_8851, pad = lora_out_909_pad_0, pad_type = lora_out_909_pad_type_0, strides = var_9150, weight = lora_out_911_weight_0_to_fp16, x = input_679_cast_fp16)[name = tensor("lora_out_911_cast_fp16")]; - tensor obj_275_cast_fp16 = add(x = pretrained_out_455_cast_fp16, y = lora_out_911_cast_fp16)[name = tensor("obj_275_cast_fp16")]; - tensor inputs_137_cast_fp16 = add(x = inputs_135_cast_fp16, y = obj_275_cast_fp16)[name = tensor("inputs_137_cast_fp16")]; - tensor var_9161 = const()[name = tensor("op_9161"), val = tensor([1])]; - tensor channels_mean_137_cast_fp16 = reduce_mean(axes = var_9161, keep_dims = var_8852, 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_9165 = const()[name = tensor("op_9165"), val = tensor([1])]; - tensor var_9166_cast_fp16 = reduce_mean(axes = var_9165, keep_dims = var_8852, x = zero_mean_sq_137_cast_fp16)[name = tensor("op_9166_cast_fp16")]; - tensor var_9167_to_fp16 = const()[name = tensor("op_9167_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_9168_cast_fp16 = add(x = var_9166_cast_fp16, y = var_9167_to_fp16)[name = tensor("op_9168_cast_fp16")]; - tensor denom_137_epsilon_0 = const()[name = tensor("denom_137_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_137_cast_fp16 = rsqrt(epsilon = denom_137_epsilon_0, x = var_9168_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_681_gamma_0_to_fp16 = const()[name = tensor("input_681_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455648320)))]; - tensor input_681_beta_0_to_fp16 = const()[name = tensor("input_681_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455650944)))]; - tensor input_681_epsilon_0_to_fp16 = const()[name = tensor("input_681_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_681_cast_fp16 = batch_norm(beta = input_681_beta_0_to_fp16, epsilon = input_681_epsilon_0_to_fp16, gamma = input_681_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_681_cast_fp16")]; - tensor var_9182 = const()[name = tensor("op_9182"), val = tensor([1, 1])]; - tensor var_9184 = const()[name = tensor("op_9184"), val = tensor([1, 1])]; - tensor pretrained_out_457_pad_type_0 = const()[name = tensor("pretrained_out_457_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_457_pad_0 = const()[name = tensor("pretrained_out_457_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455653568))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(458930432))), name = tensor("layers_22_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_22_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_22_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(458930560)))]; - tensor pretrained_out_457_cast_fp16 = conv(bias = layers_22_fc1_pretrained_bias_to_fp16, dilations = var_9184, groups = var_8851, pad = pretrained_out_457_pad_0, pad_type = pretrained_out_457_pad_type_0, strides = var_9182, weight = layers_22_fc1_pretrained_weight_to_fp16_palettized, x = input_681_cast_fp16)[name = tensor("pretrained_out_457_cast_fp16")]; - tensor var_9188 = const()[name = tensor("op_9188"), val = tensor([1, 1])]; - tensor var_9190 = const()[name = tensor("op_9190"), val = tensor([1, 1])]; - tensor input_683_pad_type_0 = const()[name = tensor("input_683_pad_type_0"), val = tensor("custom")]; - tensor input_683_pad_0 = const()[name = tensor("input_683_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_22_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(458940864)))]; - tensor input_683_cast_fp16 = conv(dilations = var_9190, groups = var_8851, pad = input_683_pad_0, pad_type = input_683_pad_type_0, strides = var_9188, weight = layers_22_fc1_loraA_weight_to_fp16, x = input_681_cast_fp16)[name = tensor("input_683_cast_fp16")]; - tensor var_9194 = const()[name = tensor("op_9194"), val = tensor([1, 1])]; - tensor var_9196 = const()[name = tensor("op_9196"), val = tensor([1, 1])]; - tensor lora_out_913_pad_type_0 = const()[name = tensor("lora_out_913_pad_type_0"), val = tensor("custom")]; - tensor lora_out_913_pad_0 = const()[name = tensor("lora_out_913_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_915_weight_0_to_fp16 = const()[name = tensor("lora_out_915_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(458981888)))]; - tensor lora_out_915_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_9196, groups = var_8851, pad = lora_out_913_pad_0, pad_type = lora_out_913_pad_type_0, strides = var_9194, weight = lora_out_915_weight_0_to_fp16, x = input_683_cast_fp16)[name = tensor("lora_out_915_cast_fp16")]; - tensor input_685_cast_fp16 = add(x = pretrained_out_457_cast_fp16, y = lora_out_915_cast_fp16)[name = tensor("input_685_cast_fp16")]; - tensor input_687_mode_0 = const()[name = tensor("input_687_mode_0"), val = tensor("EXACT")]; - tensor input_687_cast_fp16 = gelu(mode = input_687_mode_0, x = input_685_cast_fp16)[name = tensor("input_687_cast_fp16")]; - tensor var_9208 = const()[name = tensor("op_9208"), val = tensor([1, 1])]; - tensor var_9210 = const()[name = tensor("op_9210"), val = tensor([1, 1])]; - tensor pretrained_out_459_pad_type_0 = const()[name = tensor("pretrained_out_459_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_459_pad_0 = const()[name = tensor("pretrained_out_459_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459145792))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462422656))), name = tensor("layers_22_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_22_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_22_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462422784)))]; - tensor pretrained_out_459_cast_fp16 = conv(bias = layers_22_fc2_pretrained_bias_to_fp16, dilations = var_9210, groups = var_8851, pad = pretrained_out_459_pad_0, pad_type = pretrained_out_459_pad_type_0, strides = var_9208, weight = layers_22_fc2_pretrained_weight_to_fp16_palettized, x = input_687_cast_fp16)[name = tensor("pretrained_out_459_cast_fp16")]; - tensor var_9214 = const()[name = tensor("op_9214"), val = tensor([1, 1])]; - tensor var_9216 = const()[name = tensor("op_9216"), val = tensor([1, 1])]; - tensor input_689_pad_type_0 = const()[name = tensor("input_689_pad_type_0"), val = tensor("custom")]; - tensor input_689_pad_0 = const()[name = tensor("input_689_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_22_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_22_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462425408)))]; - tensor input_689_cast_fp16 = conv(dilations = var_9216, groups = var_8851, pad = input_689_pad_0, pad_type = input_689_pad_type_0, strides = var_9214, weight = layers_22_fc2_loraA_weight_to_fp16, x = input_687_cast_fp16)[name = tensor("input_689_cast_fp16")]; - tensor var_9220 = const()[name = tensor("op_9220"), val = tensor([1, 1])]; - tensor var_9222 = const()[name = tensor("op_9222"), val = tensor([1, 1])]; - tensor lora_out_917_pad_type_0 = const()[name = tensor("lora_out_917_pad_type_0"), val = tensor("custom")]; - tensor lora_out_917_pad_0 = const()[name = tensor("lora_out_917_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_919_weight_0_to_fp16 = const()[name = tensor("lora_out_919_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462589312)))]; - tensor lora_out_919_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9222, groups = var_8851, pad = lora_out_917_pad_0, pad_type = lora_out_917_pad_type_0, strides = var_9220, weight = lora_out_919_weight_0_to_fp16, x = input_689_cast_fp16)[name = tensor("lora_out_919_cast_fp16")]; - tensor hidden_states_47_cast_fp16 = add(x = pretrained_out_459_cast_fp16, y = lora_out_919_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_9238 = const()[name = tensor("op_9238"), val = tensor(3)]; - tensor var_9245 = const()[name = tensor("op_9245"), val = tensor(1)]; - tensor var_9246 = const()[name = tensor("op_9246"), val = tensor(true)]; - tensor var_9258 = const()[name = tensor("op_9258"), val = tensor([1])]; - tensor channels_mean_139_cast_fp16 = reduce_mean(axes = var_9258, keep_dims = var_9246, 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_9262 = const()[name = tensor("op_9262"), val = tensor([1])]; - tensor var_9263_cast_fp16 = reduce_mean(axes = var_9262, keep_dims = var_9246, x = zero_mean_sq_139_cast_fp16)[name = tensor("op_9263_cast_fp16")]; - tensor var_9264_to_fp16 = const()[name = tensor("op_9264_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_9265_cast_fp16 = add(x = var_9263_cast_fp16, y = var_9264_to_fp16)[name = tensor("op_9265_cast_fp16")]; - tensor denom_139_epsilon_0 = const()[name = tensor("denom_139_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_139_cast_fp16 = rsqrt(epsilon = denom_139_epsilon_0, x = var_9265_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_277_gamma_0_to_fp16 = const()[name = tensor("obj_277_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462630336)))]; - tensor obj_277_beta_0_to_fp16 = const()[name = tensor("obj_277_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462632960)))]; - tensor obj_277_epsilon_0_to_fp16 = const()[name = tensor("obj_277_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_277_cast_fp16 = batch_norm(beta = obj_277_beta_0_to_fp16, epsilon = obj_277_epsilon_0_to_fp16, gamma = obj_277_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_277_cast_fp16")]; - tensor var_9283 = const()[name = tensor("op_9283"), val = tensor([1, 1])]; - tensor var_9285 = const()[name = tensor("op_9285"), val = tensor([1, 1])]; - tensor pretrained_out_461_pad_type_0 = const()[name = tensor("pretrained_out_461_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_461_pad_0 = const()[name = tensor("pretrained_out_461_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462635584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463454848))), name = tensor("layers_23_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_23_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463454976)))]; - tensor pretrained_out_461_cast_fp16 = conv(bias = layers_23_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_9285, groups = var_9245, pad = pretrained_out_461_pad_0, pad_type = pretrained_out_461_pad_type_0, strides = var_9283, weight = layers_23_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_277_cast_fp16)[name = tensor("pretrained_out_461_cast_fp16")]; - tensor var_9289 = const()[name = tensor("op_9289"), val = tensor([1, 1])]; - tensor var_9291 = const()[name = tensor("op_9291"), val = tensor([1, 1])]; - tensor input_691_pad_type_0 = const()[name = tensor("input_691_pad_type_0"), val = tensor("custom")]; - tensor input_691_pad_0 = const()[name = tensor("input_691_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463457600)))]; - tensor input_691_cast_fp16 = conv(dilations = var_9291, groups = var_9245, pad = input_691_pad_0, pad_type = input_691_pad_type_0, strides = var_9289, weight = layers_23_self_attn_q_proj_loraA_weight_to_fp16, x = obj_277_cast_fp16)[name = tensor("input_691_cast_fp16")]; - tensor var_9295 = const()[name = tensor("op_9295"), val = tensor([1, 1])]; - tensor var_9297 = const()[name = tensor("op_9297"), val = tensor([1, 1])]; - tensor lora_out_921_pad_type_0 = const()[name = tensor("lora_out_921_pad_type_0"), val = tensor("custom")]; - tensor lora_out_921_pad_0 = const()[name = tensor("lora_out_921_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_923_weight_0_to_fp16 = const()[name = tensor("lora_out_923_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463498624)))]; - tensor lora_out_923_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9297, groups = var_9245, pad = lora_out_921_pad_0, pad_type = lora_out_921_pad_type_0, strides = var_9295, weight = lora_out_923_weight_0_to_fp16, x = input_691_cast_fp16)[name = tensor("lora_out_923_cast_fp16")]; - tensor query_93_cast_fp16 = add(x = pretrained_out_461_cast_fp16, y = lora_out_923_cast_fp16)[name = tensor("query_93_cast_fp16")]; - tensor var_9307 = const()[name = tensor("op_9307"), val = tensor([1, 1])]; - tensor var_9309 = const()[name = tensor("op_9309"), val = tensor([1, 1])]; - tensor pretrained_out_463_pad_type_0 = const()[name = tensor("pretrained_out_463_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_463_pad_0 = const()[name = tensor("pretrained_out_463_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463539648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(464358912))), name = tensor("layers_23_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_463_cast_fp16 = conv(dilations = var_9309, groups = var_9245, pad = pretrained_out_463_pad_0, pad_type = pretrained_out_463_pad_type_0, strides = var_9307, weight = layers_23_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_277_cast_fp16)[name = tensor("pretrained_out_463_cast_fp16")]; - tensor var_9313 = const()[name = tensor("op_9313"), val = tensor([1, 1])]; - tensor var_9315 = const()[name = tensor("op_9315"), val = tensor([1, 1])]; - tensor input_693_pad_type_0 = const()[name = tensor("input_693_pad_type_0"), val = tensor("custom")]; - tensor input_693_pad_0 = const()[name = tensor("input_693_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(464359040)))]; - tensor input_693_cast_fp16 = conv(dilations = var_9315, groups = var_9245, pad = input_693_pad_0, pad_type = input_693_pad_type_0, strides = var_9313, weight = layers_23_self_attn_k_proj_loraA_weight_to_fp16, x = obj_277_cast_fp16)[name = tensor("input_693_cast_fp16")]; - tensor var_9319 = const()[name = tensor("op_9319"), val = tensor([1, 1])]; - tensor var_9321 = const()[name = tensor("op_9321"), val = tensor([1, 1])]; - tensor lora_out_925_pad_type_0 = const()[name = tensor("lora_out_925_pad_type_0"), val = tensor("custom")]; - tensor lora_out_925_pad_0 = const()[name = tensor("lora_out_925_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_927_weight_0_to_fp16 = const()[name = tensor("lora_out_927_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(464400064)))]; - tensor lora_out_927_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9321, groups = var_9245, pad = lora_out_925_pad_0, pad_type = lora_out_925_pad_type_0, strides = var_9319, weight = lora_out_927_weight_0_to_fp16, x = input_693_cast_fp16)[name = tensor("lora_out_927_cast_fp16")]; - tensor current_key_47_cast_fp16 = add(x = pretrained_out_463_cast_fp16, y = lora_out_927_cast_fp16)[name = tensor("current_key_47_cast_fp16")]; - tensor var_9332 = const()[name = tensor("op_9332"), val = tensor([1, 1])]; - tensor var_9334 = const()[name = tensor("op_9334"), val = tensor([1, 1])]; - tensor pretrained_out_465_pad_type_0 = const()[name = tensor("pretrained_out_465_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_465_pad_0 = const()[name = tensor("pretrained_out_465_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(464441088))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465260352))), name = tensor("layers_23_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_23_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465260480)))]; - tensor pretrained_out_465_cast_fp16 = conv(bias = layers_23_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_9334, groups = var_9245, pad = pretrained_out_465_pad_0, pad_type = pretrained_out_465_pad_type_0, strides = var_9332, weight = layers_23_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_277_cast_fp16)[name = tensor("pretrained_out_465_cast_fp16")]; - tensor var_9338 = const()[name = tensor("op_9338"), val = tensor([1, 1])]; - tensor var_9340 = const()[name = tensor("op_9340"), val = tensor([1, 1])]; - tensor input_695_pad_type_0 = const()[name = tensor("input_695_pad_type_0"), val = tensor("custom")]; - tensor input_695_pad_0 = const()[name = tensor("input_695_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465263104)))]; - tensor input_695_cast_fp16 = conv(dilations = var_9340, groups = var_9245, pad = input_695_pad_0, pad_type = input_695_pad_type_0, strides = var_9338, weight = layers_23_self_attn_v_proj_loraA_weight_to_fp16, x = obj_277_cast_fp16)[name = tensor("input_695_cast_fp16")]; - tensor var_9344 = const()[name = tensor("op_9344"), val = tensor([1, 1])]; - tensor var_9346 = const()[name = tensor("op_9346"), val = tensor([1, 1])]; - tensor lora_out_929_pad_type_0 = const()[name = tensor("lora_out_929_pad_type_0"), val = tensor("custom")]; - tensor lora_out_929_pad_0 = const()[name = tensor("lora_out_929_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_931_weight_0_to_fp16 = const()[name = tensor("lora_out_931_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465304128)))]; - tensor lora_out_931_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9346, groups = var_9245, pad = lora_out_929_pad_0, pad_type = lora_out_929_pad_type_0, strides = var_9344, weight = lora_out_931_weight_0_to_fp16, x = input_695_cast_fp16)[name = tensor("lora_out_931_cast_fp16")]; - tensor current_value_47_cast_fp16 = add(x = pretrained_out_465_cast_fp16, y = lora_out_931_cast_fp16)[name = tensor("current_value_47_cast_fp16")]; - tensor var_9356_cast_fp16 = mul(x = current_key_47_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_9356_cast_fp16")]; - tensor var_9358_cast_fp16 = mul(x = var_103_cast_fp16_23, y = var_295_cast_fp16)[name = tensor("op_9358_cast_fp16")]; - tensor key_93_cast_fp16 = add(x = var_9356_cast_fp16, y = var_9358_cast_fp16)[name = tensor("key_93_cast_fp16")]; - tensor var_9360_cast_fp16 = mul(x = current_value_47_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_9360_cast_fp16")]; - tensor var_9362_cast_fp16 = mul(x = var_138_cast_fp16_23, y = var_295_cast_fp16)[name = tensor("op_9362_cast_fp16")]; - tensor value_93_cast_fp16 = add(x = var_9360_cast_fp16, y = var_9362_cast_fp16)[name = tensor("value_93_cast_fp16")]; - tensor var_9365 = const()[name = tensor("op_9365"), val = tensor([1, 20, 64, -1])]; - tensor var_9366_cast_fp16 = reshape(shape = var_9365, x = query_93_cast_fp16)[name = tensor("op_9366_cast_fp16")]; - tensor var_9367_to_fp16 = const()[name = tensor("op_9367_to_fp16"), val = tensor(0x1p-3)]; - tensor var_9368_cast_fp16 = mul(x = var_9366_cast_fp16, y = var_9367_to_fp16)[name = tensor("op_9368_cast_fp16")]; - tensor var_9369 = const()[name = tensor("op_9369"), val = tensor([1, 20, 64, -1])]; - tensor var_9370_cast_fp16 = reshape(shape = var_9369, x = key_93_cast_fp16)[name = tensor("op_9370_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_9368_cast_fp16, y = var_9370_cast_fp16)[name = tensor("mh_w_139_cast_fp16")]; - tensor mh_w_141_cast_fp16 = add(x = mh_w_139_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_141_cast_fp16")]; - tensor var_9378_cast_fp16 = softmax(axis = var_9238, x = mh_w_141_cast_fp16)[name = tensor("op_9378_cast_fp16")]; - tensor var_9379 = const()[name = tensor("op_9379"), val = tensor([1, 20, 64, -1])]; - tensor var_9380_cast_fp16 = reshape(shape = var_9379, x = value_93_cast_fp16)[name = tensor("op_9380_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_9380_cast_fp16, y = var_9378_cast_fp16)[name = tensor("attn_93_cast_fp16")]; - tensor var_9383 = const()[name = tensor("op_9383"), val = tensor([1, 1280, 1, -1])]; - tensor input_697_cast_fp16 = reshape(shape = var_9383, x = attn_93_cast_fp16)[name = tensor("input_697_cast_fp16")]; - tensor var_9390 = const()[name = tensor("op_9390"), val = tensor([1, 1])]; - tensor var_9392 = const()[name = tensor("op_9392"), val = tensor([1, 1])]; - tensor pretrained_out_467_pad_type_0 = const()[name = tensor("pretrained_out_467_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_467_pad_0 = const()[name = tensor("pretrained_out_467_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465345152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466164416))), name = tensor("layers_23_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_23_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466164544)))]; - tensor pretrained_out_467_cast_fp16 = conv(bias = layers_23_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_9392, groups = var_9245, pad = pretrained_out_467_pad_0, pad_type = pretrained_out_467_pad_type_0, strides = var_9390, weight = layers_23_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_697_cast_fp16)[name = tensor("pretrained_out_467_cast_fp16")]; - tensor var_9396 = const()[name = tensor("op_9396"), val = tensor([1, 1])]; - tensor var_9398 = const()[name = tensor("op_9398"), val = tensor([1, 1])]; - tensor input_699_pad_type_0 = const()[name = tensor("input_699_pad_type_0"), val = tensor("custom")]; - tensor input_699_pad_0 = const()[name = tensor("input_699_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466167168)))]; - tensor input_699_cast_fp16 = conv(dilations = var_9398, groups = var_9245, pad = input_699_pad_0, pad_type = input_699_pad_type_0, strides = var_9396, weight = layers_23_self_attn_o_proj_loraA_weight_to_fp16, x = input_697_cast_fp16)[name = tensor("input_699_cast_fp16")]; - tensor var_9402 = const()[name = tensor("op_9402"), val = tensor([1, 1])]; - tensor var_9404 = const()[name = tensor("op_9404"), val = tensor([1, 1])]; - tensor lora_out_933_pad_type_0 = const()[name = tensor("lora_out_933_pad_type_0"), val = tensor("custom")]; - tensor lora_out_933_pad_0 = const()[name = tensor("lora_out_933_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_935_weight_0_to_fp16 = const()[name = tensor("lora_out_935_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466208192)))]; - tensor lora_out_935_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9404, groups = var_9245, pad = lora_out_933_pad_0, pad_type = lora_out_933_pad_type_0, strides = var_9402, weight = lora_out_935_weight_0_to_fp16, x = input_699_cast_fp16)[name = tensor("lora_out_935_cast_fp16")]; - tensor obj_283_cast_fp16 = add(x = pretrained_out_467_cast_fp16, y = lora_out_935_cast_fp16)[name = tensor("obj_283_cast_fp16")]; - tensor inputs_141_cast_fp16 = add(x = inputs_139_cast_fp16, y = obj_283_cast_fp16)[name = tensor("inputs_141_cast_fp16")]; - tensor var_9417 = const()[name = tensor("op_9417"), val = tensor([1])]; - tensor channels_mean_141_cast_fp16 = reduce_mean(axes = var_9417, keep_dims = var_9246, 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_9421 = const()[name = tensor("op_9421"), val = tensor([1])]; - tensor var_9422_cast_fp16 = reduce_mean(axes = var_9421, keep_dims = var_9246, x = zero_mean_sq_141_cast_fp16)[name = tensor("op_9422_cast_fp16")]; - tensor var_9423_to_fp16 = const()[name = tensor("op_9423_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_9424_cast_fp16 = add(x = var_9422_cast_fp16, y = var_9423_to_fp16)[name = tensor("op_9424_cast_fp16")]; - tensor denom_141_epsilon_0 = const()[name = tensor("denom_141_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_141_cast_fp16 = rsqrt(epsilon = denom_141_epsilon_0, x = var_9424_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_285_gamma_0_to_fp16 = const()[name = tensor("obj_285_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466249216)))]; - tensor obj_285_beta_0_to_fp16 = const()[name = tensor("obj_285_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466251840)))]; - tensor obj_285_epsilon_0_to_fp16 = const()[name = tensor("obj_285_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_285_cast_fp16 = batch_norm(beta = obj_285_beta_0_to_fp16, epsilon = obj_285_epsilon_0_to_fp16, gamma = obj_285_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_285_cast_fp16")]; - tensor var_9442 = const()[name = tensor("op_9442"), val = tensor([1, 1])]; - tensor var_9444 = const()[name = tensor("op_9444"), val = tensor([1, 1])]; - tensor pretrained_out_469_pad_type_0 = const()[name = tensor("pretrained_out_469_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_469_pad_0 = const()[name = tensor("pretrained_out_469_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466254464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467073728))), name = tensor("layers_23_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_23_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467073856)))]; - tensor pretrained_out_469_cast_fp16 = conv(bias = layers_23_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_9444, groups = var_9245, pad = pretrained_out_469_pad_0, pad_type = pretrained_out_469_pad_type_0, strides = var_9442, weight = layers_23_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_285_cast_fp16)[name = tensor("pretrained_out_469_cast_fp16")]; - tensor var_9448 = const()[name = tensor("op_9448"), val = tensor([1, 1])]; - tensor var_9450 = const()[name = tensor("op_9450"), val = tensor([1, 1])]; - tensor input_701_pad_type_0 = const()[name = tensor("input_701_pad_type_0"), val = tensor("custom")]; - tensor input_701_pad_0 = const()[name = tensor("input_701_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467076480)))]; - tensor input_701_cast_fp16 = conv(dilations = var_9450, groups = var_9245, pad = input_701_pad_0, pad_type = input_701_pad_type_0, strides = var_9448, weight = layers_23_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_285_cast_fp16)[name = tensor("input_701_cast_fp16")]; - tensor var_9454 = const()[name = tensor("op_9454"), val = tensor([1, 1])]; - tensor var_9456 = const()[name = tensor("op_9456"), val = tensor([1, 1])]; - tensor lora_out_937_pad_type_0 = const()[name = tensor("lora_out_937_pad_type_0"), val = tensor("custom")]; - tensor lora_out_937_pad_0 = const()[name = tensor("lora_out_937_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_939_weight_0_to_fp16 = const()[name = tensor("lora_out_939_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467117504)))]; - tensor lora_out_939_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9456, groups = var_9245, pad = lora_out_937_pad_0, pad_type = lora_out_937_pad_type_0, strides = var_9454, weight = lora_out_939_weight_0_to_fp16, x = input_701_cast_fp16)[name = tensor("lora_out_939_cast_fp16")]; - tensor query_95_cast_fp16 = add(x = pretrained_out_469_cast_fp16, y = lora_out_939_cast_fp16)[name = tensor("query_95_cast_fp16")]; - tensor var_9466 = const()[name = tensor("op_9466"), val = tensor([1, 1])]; - tensor var_9468 = const()[name = tensor("op_9468"), val = tensor([1, 1])]; - tensor pretrained_out_471_pad_type_0 = const()[name = tensor("pretrained_out_471_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_471_pad_0 = const()[name = tensor("pretrained_out_471_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467158528))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467977792))), name = tensor("layers_23_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_471_cast_fp16 = conv(dilations = var_9468, groups = var_9245, pad = pretrained_out_471_pad_0, pad_type = pretrained_out_471_pad_type_0, strides = var_9466, weight = layers_23_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_471_cast_fp16")]; - tensor var_9472 = const()[name = tensor("op_9472"), val = tensor([1, 1])]; - tensor var_9474 = const()[name = tensor("op_9474"), val = tensor([1, 1])]; - tensor input_703_pad_type_0 = const()[name = tensor("input_703_pad_type_0"), val = tensor("custom")]; - tensor input_703_pad_0 = const()[name = tensor("input_703_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467977920)))]; - tensor input_703_cast_fp16 = conv(dilations = var_9474, groups = var_9245, pad = input_703_pad_0, pad_type = input_703_pad_type_0, strides = var_9472, weight = layers_23_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_703_cast_fp16")]; - tensor var_9478 = const()[name = tensor("op_9478"), val = tensor([1, 1])]; - tensor var_9480 = const()[name = tensor("op_9480"), val = tensor([1, 1])]; - tensor lora_out_941_pad_type_0 = const()[name = tensor("lora_out_941_pad_type_0"), val = tensor("custom")]; - tensor lora_out_941_pad_0 = const()[name = tensor("lora_out_941_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_943_weight_0_to_fp16 = const()[name = tensor("lora_out_943_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468018944)))]; - tensor lora_out_943_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9480, groups = var_9245, pad = lora_out_941_pad_0, pad_type = lora_out_941_pad_type_0, strides = var_9478, weight = lora_out_943_weight_0_to_fp16, x = input_703_cast_fp16)[name = tensor("lora_out_943_cast_fp16")]; - tensor key_95_cast_fp16 = add(x = pretrained_out_471_cast_fp16, y = lora_out_943_cast_fp16)[name = tensor("key_95_cast_fp16")]; - tensor var_9491 = const()[name = tensor("op_9491"), val = tensor([1, 1])]; - tensor var_9493 = const()[name = tensor("op_9493"), val = tensor([1, 1])]; - tensor pretrained_out_473_pad_type_0 = const()[name = tensor("pretrained_out_473_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_473_pad_0 = const()[name = tensor("pretrained_out_473_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468059968))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468879232))), name = tensor("layers_23_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_23_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468879360)))]; - tensor pretrained_out_473_cast_fp16 = conv(bias = layers_23_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_9493, groups = var_9245, pad = pretrained_out_473_pad_0, pad_type = pretrained_out_473_pad_type_0, strides = var_9491, weight = layers_23_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_473_cast_fp16")]; - tensor var_9497 = const()[name = tensor("op_9497"), val = tensor([1, 1])]; - tensor var_9499 = const()[name = tensor("op_9499"), val = tensor([1, 1])]; - tensor input_705_pad_type_0 = const()[name = tensor("input_705_pad_type_0"), val = tensor("custom")]; - tensor input_705_pad_0 = const()[name = tensor("input_705_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468881984)))]; - tensor input_705_cast_fp16 = conv(dilations = var_9499, groups = var_9245, pad = input_705_pad_0, pad_type = input_705_pad_type_0, strides = var_9497, weight = layers_23_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_705_cast_fp16")]; - tensor var_9503 = const()[name = tensor("op_9503"), val = tensor([1, 1])]; - tensor var_9505 = const()[name = tensor("op_9505"), val = tensor([1, 1])]; - tensor lora_out_945_pad_type_0 = const()[name = tensor("lora_out_945_pad_type_0"), val = tensor("custom")]; - tensor lora_out_945_pad_0 = const()[name = tensor("lora_out_945_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_947_weight_0_to_fp16 = const()[name = tensor("lora_out_947_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468923008)))]; - tensor lora_out_947_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9505, groups = var_9245, pad = lora_out_945_pad_0, pad_type = lora_out_945_pad_type_0, strides = var_9503, weight = lora_out_947_weight_0_to_fp16, x = input_705_cast_fp16)[name = tensor("lora_out_947_cast_fp16")]; - tensor value_95_cast_fp16 = add(x = pretrained_out_473_cast_fp16, y = lora_out_947_cast_fp16)[name = tensor("value_95_cast_fp16")]; - tensor var_9512 = const()[name = tensor("op_9512"), val = tensor([1, 20, 64, -1])]; - tensor var_9513_cast_fp16 = reshape(shape = var_9512, x = query_95_cast_fp16)[name = tensor("op_9513_cast_fp16")]; - tensor var_9514_to_fp16 = const()[name = tensor("op_9514_to_fp16"), val = tensor(0x1p-3)]; - tensor var_9515_cast_fp16 = mul(x = var_9513_cast_fp16, y = var_9514_to_fp16)[name = tensor("op_9515_cast_fp16")]; - tensor var_9516 = const()[name = tensor("op_9516"), val = tensor([1, 20, 64, -1])]; - tensor var_9517_cast_fp16 = reshape(shape = var_9516, x = key_95_cast_fp16)[name = tensor("op_9517_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_9515_cast_fp16, y = var_9517_cast_fp16)[name = tensor("mh_w_143_cast_fp16")]; - tensor var_9520_cast_fp16 = softmax(axis = var_9238, x = mh_w_143_cast_fp16)[name = tensor("op_9520_cast_fp16")]; - tensor var_9521 = const()[name = tensor("op_9521"), val = tensor([1, 20, 64, -1])]; - tensor var_9522_cast_fp16 = reshape(shape = var_9521, x = value_95_cast_fp16)[name = tensor("op_9522_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_9522_cast_fp16, y = var_9520_cast_fp16)[name = tensor("attn_95_cast_fp16")]; - tensor var_9525 = const()[name = tensor("op_9525"), val = tensor([1, 1280, 1, -1])]; - tensor input_707_cast_fp16 = reshape(shape = var_9525, x = attn_95_cast_fp16)[name = tensor("input_707_cast_fp16")]; - tensor var_9532 = const()[name = tensor("op_9532"), val = tensor([1, 1])]; - tensor var_9534 = const()[name = tensor("op_9534"), val = tensor([1, 1])]; - tensor pretrained_out_475_pad_type_0 = const()[name = tensor("pretrained_out_475_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_475_pad_0 = const()[name = tensor("pretrained_out_475_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468964032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469783296))), name = tensor("layers_23_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_23_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469783424)))]; - tensor pretrained_out_475_cast_fp16 = conv(bias = layers_23_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_9534, groups = var_9245, pad = pretrained_out_475_pad_0, pad_type = pretrained_out_475_pad_type_0, strides = var_9532, weight = layers_23_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_707_cast_fp16)[name = tensor("pretrained_out_475_cast_fp16")]; - tensor var_9538 = const()[name = tensor("op_9538"), val = tensor([1, 1])]; - tensor var_9540 = const()[name = tensor("op_9540"), val = tensor([1, 1])]; - tensor input_709_pad_type_0 = const()[name = tensor("input_709_pad_type_0"), val = tensor("custom")]; - tensor input_709_pad_0 = const()[name = tensor("input_709_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469786048)))]; - tensor input_709_cast_fp16 = conv(dilations = var_9540, groups = var_9245, pad = input_709_pad_0, pad_type = input_709_pad_type_0, strides = var_9538, weight = layers_23_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_707_cast_fp16)[name = tensor("input_709_cast_fp16")]; - tensor var_9544 = const()[name = tensor("op_9544"), val = tensor([1, 1])]; - tensor var_9546 = const()[name = tensor("op_9546"), val = tensor([1, 1])]; - tensor lora_out_949_pad_type_0 = const()[name = tensor("lora_out_949_pad_type_0"), val = tensor("custom")]; - tensor lora_out_949_pad_0 = const()[name = tensor("lora_out_949_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_951_weight_0_to_fp16 = const()[name = tensor("lora_out_951_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469827072)))]; - tensor lora_out_951_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9546, groups = var_9245, pad = lora_out_949_pad_0, pad_type = lora_out_949_pad_type_0, strides = var_9544, weight = lora_out_951_weight_0_to_fp16, x = input_709_cast_fp16)[name = tensor("lora_out_951_cast_fp16")]; - tensor obj_287_cast_fp16 = add(x = pretrained_out_475_cast_fp16, y = lora_out_951_cast_fp16)[name = tensor("obj_287_cast_fp16")]; - tensor inputs_143_cast_fp16 = add(x = inputs_141_cast_fp16, y = obj_287_cast_fp16)[name = tensor("inputs_143_cast_fp16")]; - tensor var_9555 = const()[name = tensor("op_9555"), val = tensor([1])]; - tensor channels_mean_143_cast_fp16 = reduce_mean(axes = var_9555, keep_dims = var_9246, 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_9559 = const()[name = tensor("op_9559"), val = tensor([1])]; - tensor var_9560_cast_fp16 = reduce_mean(axes = var_9559, keep_dims = var_9246, x = zero_mean_sq_143_cast_fp16)[name = tensor("op_9560_cast_fp16")]; - tensor var_9561_to_fp16 = const()[name = tensor("op_9561_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_9562_cast_fp16 = add(x = var_9560_cast_fp16, y = var_9561_to_fp16)[name = tensor("op_9562_cast_fp16")]; - tensor denom_143_epsilon_0 = const()[name = tensor("denom_143_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_143_cast_fp16 = rsqrt(epsilon = denom_143_epsilon_0, x = var_9562_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_711_gamma_0_to_fp16 = const()[name = tensor("input_711_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469868096)))]; - tensor input_711_beta_0_to_fp16 = const()[name = tensor("input_711_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469870720)))]; - tensor input_711_epsilon_0_to_fp16 = const()[name = tensor("input_711_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_711_cast_fp16 = batch_norm(beta = input_711_beta_0_to_fp16, epsilon = input_711_epsilon_0_to_fp16, gamma = input_711_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_711_cast_fp16")]; - tensor var_9576 = const()[name = tensor("op_9576"), val = tensor([1, 1])]; - tensor var_9578 = const()[name = tensor("op_9578"), val = tensor([1, 1])]; - tensor pretrained_out_477_pad_type_0 = const()[name = tensor("pretrained_out_477_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_477_pad_0 = const()[name = tensor("pretrained_out_477_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469873344))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473150208))), name = tensor("layers_23_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_23_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_23_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473150336)))]; - tensor pretrained_out_477_cast_fp16 = conv(bias = layers_23_fc1_pretrained_bias_to_fp16, dilations = var_9578, groups = var_9245, pad = pretrained_out_477_pad_0, pad_type = pretrained_out_477_pad_type_0, strides = var_9576, weight = layers_23_fc1_pretrained_weight_to_fp16_palettized, x = input_711_cast_fp16)[name = tensor("pretrained_out_477_cast_fp16")]; - tensor var_9582 = const()[name = tensor("op_9582"), val = tensor([1, 1])]; - tensor var_9584 = const()[name = tensor("op_9584"), val = tensor([1, 1])]; - tensor input_713_pad_type_0 = const()[name = tensor("input_713_pad_type_0"), val = tensor("custom")]; - tensor input_713_pad_0 = const()[name = tensor("input_713_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_23_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473160640)))]; - tensor input_713_cast_fp16 = conv(dilations = var_9584, groups = var_9245, pad = input_713_pad_0, pad_type = input_713_pad_type_0, strides = var_9582, weight = layers_23_fc1_loraA_weight_to_fp16, x = input_711_cast_fp16)[name = tensor("input_713_cast_fp16")]; - tensor var_9588 = const()[name = tensor("op_9588"), val = tensor([1, 1])]; - tensor var_9590 = const()[name = tensor("op_9590"), val = tensor([1, 1])]; - tensor lora_out_953_pad_type_0 = const()[name = tensor("lora_out_953_pad_type_0"), val = tensor("custom")]; - tensor lora_out_953_pad_0 = const()[name = tensor("lora_out_953_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_955_weight_0_to_fp16 = const()[name = tensor("lora_out_955_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473201664)))]; - tensor lora_out_955_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_9590, groups = var_9245, pad = lora_out_953_pad_0, pad_type = lora_out_953_pad_type_0, strides = var_9588, weight = lora_out_955_weight_0_to_fp16, x = input_713_cast_fp16)[name = tensor("lora_out_955_cast_fp16")]; - tensor input_715_cast_fp16 = add(x = pretrained_out_477_cast_fp16, y = lora_out_955_cast_fp16)[name = tensor("input_715_cast_fp16")]; - tensor input_717_mode_0 = const()[name = tensor("input_717_mode_0"), val = tensor("EXACT")]; - tensor input_717_cast_fp16 = gelu(mode = input_717_mode_0, x = input_715_cast_fp16)[name = tensor("input_717_cast_fp16")]; - tensor var_9602 = const()[name = tensor("op_9602"), val = tensor([1, 1])]; - tensor var_9604 = const()[name = tensor("op_9604"), val = tensor([1, 1])]; - tensor pretrained_out_479_pad_type_0 = const()[name = tensor("pretrained_out_479_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_479_pad_0 = const()[name = tensor("pretrained_out_479_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473365568))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(476642432))), name = tensor("layers_23_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_23_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_23_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(476642560)))]; - tensor pretrained_out_479_cast_fp16 = conv(bias = layers_23_fc2_pretrained_bias_to_fp16, dilations = var_9604, groups = var_9245, pad = pretrained_out_479_pad_0, pad_type = pretrained_out_479_pad_type_0, strides = var_9602, weight = layers_23_fc2_pretrained_weight_to_fp16_palettized, x = input_717_cast_fp16)[name = tensor("pretrained_out_479_cast_fp16")]; - tensor var_9608 = const()[name = tensor("op_9608"), val = tensor([1, 1])]; - tensor var_9610 = const()[name = tensor("op_9610"), val = tensor([1, 1])]; - tensor input_719_pad_type_0 = const()[name = tensor("input_719_pad_type_0"), val = tensor("custom")]; - tensor input_719_pad_0 = const()[name = tensor("input_719_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_23_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_23_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(476645184)))]; - tensor input_719_cast_fp16 = conv(dilations = var_9610, groups = var_9245, pad = input_719_pad_0, pad_type = input_719_pad_type_0, strides = var_9608, weight = layers_23_fc2_loraA_weight_to_fp16, x = input_717_cast_fp16)[name = tensor("input_719_cast_fp16")]; - tensor var_9614 = const()[name = tensor("op_9614"), val = tensor([1, 1])]; - tensor var_9616 = const()[name = tensor("op_9616"), val = tensor([1, 1])]; - tensor lora_out_957_pad_type_0 = const()[name = tensor("lora_out_957_pad_type_0"), val = tensor("custom")]; - tensor lora_out_957_pad_0 = const()[name = tensor("lora_out_957_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_959_weight_0_to_fp16 = const()[name = tensor("lora_out_959_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(476809088)))]; - tensor lora_out_959_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9616, groups = var_9245, pad = lora_out_957_pad_0, pad_type = lora_out_957_pad_type_0, strides = var_9614, weight = lora_out_959_weight_0_to_fp16, x = input_719_cast_fp16)[name = tensor("lora_out_959_cast_fp16")]; - tensor hidden_states_49_cast_fp16 = add(x = pretrained_out_479_cast_fp16, y = lora_out_959_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_9632 = const()[name = tensor("op_9632"), val = tensor(3)]; - tensor var_9639 = const()[name = tensor("op_9639"), val = tensor(1)]; - tensor var_9640 = const()[name = tensor("op_9640"), val = tensor(true)]; - tensor var_9652 = const()[name = tensor("op_9652"), val = tensor([1])]; - tensor channels_mean_145_cast_fp16 = reduce_mean(axes = var_9652, keep_dims = var_9640, 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_9656 = const()[name = tensor("op_9656"), val = tensor([1])]; - tensor var_9657_cast_fp16 = reduce_mean(axes = var_9656, keep_dims = var_9640, x = zero_mean_sq_145_cast_fp16)[name = tensor("op_9657_cast_fp16")]; - tensor var_9658_to_fp16 = const()[name = tensor("op_9658_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_9659_cast_fp16 = add(x = var_9657_cast_fp16, y = var_9658_to_fp16)[name = tensor("op_9659_cast_fp16")]; - tensor denom_145_epsilon_0 = const()[name = tensor("denom_145_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_145_cast_fp16 = rsqrt(epsilon = denom_145_epsilon_0, x = var_9659_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_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(476850112)))]; - 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(476852736)))]; - 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_145_cast_fp16)[name = tensor("obj_289_cast_fp16")]; - tensor var_9677 = const()[name = tensor("op_9677"), val = tensor([1, 1])]; - tensor var_9679 = const()[name = tensor("op_9679"), val = tensor([1, 1])]; - tensor pretrained_out_481_pad_type_0 = const()[name = tensor("pretrained_out_481_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_481_pad_0 = const()[name = tensor("pretrained_out_481_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(476855360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477674624))), name = tensor("layers_24_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_24_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477674752)))]; - tensor pretrained_out_481_cast_fp16 = conv(bias = layers_24_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_9679, groups = var_9639, pad = pretrained_out_481_pad_0, pad_type = pretrained_out_481_pad_type_0, strides = var_9677, weight = layers_24_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_289_cast_fp16)[name = tensor("pretrained_out_481_cast_fp16")]; - tensor var_9683 = const()[name = tensor("op_9683"), val = tensor([1, 1])]; - tensor var_9685 = const()[name = tensor("op_9685"), val = tensor([1, 1])]; - tensor input_721_pad_type_0 = const()[name = tensor("input_721_pad_type_0"), val = tensor("custom")]; - tensor input_721_pad_0 = const()[name = tensor("input_721_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477677376)))]; - tensor input_721_cast_fp16 = conv(dilations = var_9685, groups = var_9639, pad = input_721_pad_0, pad_type = input_721_pad_type_0, strides = var_9683, weight = layers_24_self_attn_q_proj_loraA_weight_to_fp16, x = obj_289_cast_fp16)[name = tensor("input_721_cast_fp16")]; - tensor var_9689 = const()[name = tensor("op_9689"), val = tensor([1, 1])]; - tensor var_9691 = const()[name = tensor("op_9691"), val = tensor([1, 1])]; - tensor lora_out_961_pad_type_0 = const()[name = tensor("lora_out_961_pad_type_0"), val = tensor("custom")]; - tensor lora_out_961_pad_0 = const()[name = tensor("lora_out_961_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_963_weight_0_to_fp16 = const()[name = tensor("lora_out_963_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477718400)))]; - tensor lora_out_963_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9691, groups = var_9639, pad = lora_out_961_pad_0, pad_type = lora_out_961_pad_type_0, strides = var_9689, weight = lora_out_963_weight_0_to_fp16, x = input_721_cast_fp16)[name = tensor("lora_out_963_cast_fp16")]; - tensor query_97_cast_fp16 = add(x = pretrained_out_481_cast_fp16, y = lora_out_963_cast_fp16)[name = tensor("query_97_cast_fp16")]; - tensor var_9701 = const()[name = tensor("op_9701"), val = tensor([1, 1])]; - tensor var_9703 = const()[name = tensor("op_9703"), val = tensor([1, 1])]; - tensor pretrained_out_483_pad_type_0 = const()[name = tensor("pretrained_out_483_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_483_pad_0 = const()[name = tensor("pretrained_out_483_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477759424))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478578688))), name = tensor("layers_24_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_483_cast_fp16 = conv(dilations = var_9703, groups = var_9639, pad = pretrained_out_483_pad_0, pad_type = pretrained_out_483_pad_type_0, strides = var_9701, weight = layers_24_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_289_cast_fp16)[name = tensor("pretrained_out_483_cast_fp16")]; - tensor var_9707 = const()[name = tensor("op_9707"), val = tensor([1, 1])]; - tensor var_9709 = const()[name = tensor("op_9709"), val = tensor([1, 1])]; - tensor input_723_pad_type_0 = const()[name = tensor("input_723_pad_type_0"), val = tensor("custom")]; - tensor input_723_pad_0 = const()[name = tensor("input_723_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478578816)))]; - tensor input_723_cast_fp16 = conv(dilations = var_9709, groups = var_9639, pad = input_723_pad_0, pad_type = input_723_pad_type_0, strides = var_9707, weight = layers_24_self_attn_k_proj_loraA_weight_to_fp16, x = obj_289_cast_fp16)[name = tensor("input_723_cast_fp16")]; - tensor var_9713 = const()[name = tensor("op_9713"), val = tensor([1, 1])]; - tensor var_9715 = const()[name = tensor("op_9715"), val = tensor([1, 1])]; - tensor lora_out_965_pad_type_0 = const()[name = tensor("lora_out_965_pad_type_0"), val = tensor("custom")]; - tensor lora_out_965_pad_0 = const()[name = tensor("lora_out_965_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_967_weight_0_to_fp16 = const()[name = tensor("lora_out_967_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478619840)))]; - tensor lora_out_967_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9715, groups = var_9639, pad = lora_out_965_pad_0, pad_type = lora_out_965_pad_type_0, strides = var_9713, weight = lora_out_967_weight_0_to_fp16, x = input_723_cast_fp16)[name = tensor("lora_out_967_cast_fp16")]; - tensor current_key_49_cast_fp16 = add(x = pretrained_out_483_cast_fp16, y = lora_out_967_cast_fp16)[name = tensor("current_key_49_cast_fp16")]; - tensor var_9726 = const()[name = tensor("op_9726"), val = tensor([1, 1])]; - tensor var_9728 = const()[name = tensor("op_9728"), val = tensor([1, 1])]; - tensor pretrained_out_485_pad_type_0 = const()[name = tensor("pretrained_out_485_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_485_pad_0 = const()[name = tensor("pretrained_out_485_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478660864))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479480128))), name = tensor("layers_24_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_24_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479480256)))]; - tensor pretrained_out_485_cast_fp16 = conv(bias = layers_24_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_9728, groups = var_9639, pad = pretrained_out_485_pad_0, pad_type = pretrained_out_485_pad_type_0, strides = var_9726, weight = layers_24_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_289_cast_fp16)[name = tensor("pretrained_out_485_cast_fp16")]; - tensor var_9732 = const()[name = tensor("op_9732"), val = tensor([1, 1])]; - tensor var_9734 = const()[name = tensor("op_9734"), val = tensor([1, 1])]; - tensor input_725_pad_type_0 = const()[name = tensor("input_725_pad_type_0"), val = tensor("custom")]; - tensor input_725_pad_0 = const()[name = tensor("input_725_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479482880)))]; - tensor input_725_cast_fp16 = conv(dilations = var_9734, groups = var_9639, pad = input_725_pad_0, pad_type = input_725_pad_type_0, strides = var_9732, weight = layers_24_self_attn_v_proj_loraA_weight_to_fp16, x = obj_289_cast_fp16)[name = tensor("input_725_cast_fp16")]; - tensor var_9738 = const()[name = tensor("op_9738"), val = tensor([1, 1])]; - tensor var_9740 = const()[name = tensor("op_9740"), val = tensor([1, 1])]; - tensor lora_out_969_pad_type_0 = const()[name = tensor("lora_out_969_pad_type_0"), val = tensor("custom")]; - tensor lora_out_969_pad_0 = const()[name = tensor("lora_out_969_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_971_weight_0_to_fp16 = const()[name = tensor("lora_out_971_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479523904)))]; - tensor lora_out_971_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9740, groups = var_9639, pad = lora_out_969_pad_0, pad_type = lora_out_969_pad_type_0, strides = var_9738, weight = lora_out_971_weight_0_to_fp16, x = input_725_cast_fp16)[name = tensor("lora_out_971_cast_fp16")]; - tensor current_value_49_cast_fp16 = add(x = pretrained_out_485_cast_fp16, y = lora_out_971_cast_fp16)[name = tensor("current_value_49_cast_fp16")]; - tensor var_9750_cast_fp16 = mul(x = current_key_49_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_9750_cast_fp16")]; - tensor var_9752_cast_fp16 = mul(x = var_103_cast_fp16_24, y = var_295_cast_fp16)[name = tensor("op_9752_cast_fp16")]; - tensor key_97_cast_fp16 = add(x = var_9750_cast_fp16, y = var_9752_cast_fp16)[name = tensor("key_97_cast_fp16")]; - tensor var_9754_cast_fp16 = mul(x = current_value_49_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_9754_cast_fp16")]; - tensor var_9756_cast_fp16 = mul(x = var_138_cast_fp16_24, y = var_295_cast_fp16)[name = tensor("op_9756_cast_fp16")]; - tensor value_97_cast_fp16 = add(x = var_9754_cast_fp16, y = var_9756_cast_fp16)[name = tensor("value_97_cast_fp16")]; - tensor var_9759 = const()[name = tensor("op_9759"), val = tensor([1, 20, 64, -1])]; - tensor var_9760_cast_fp16 = reshape(shape = var_9759, x = query_97_cast_fp16)[name = tensor("op_9760_cast_fp16")]; - tensor var_9761_to_fp16 = const()[name = tensor("op_9761_to_fp16"), val = tensor(0x1p-3)]; - tensor var_9762_cast_fp16 = mul(x = var_9760_cast_fp16, y = var_9761_to_fp16)[name = tensor("op_9762_cast_fp16")]; - tensor var_9763 = const()[name = tensor("op_9763"), val = tensor([1, 20, 64, -1])]; - tensor var_9764_cast_fp16 = reshape(shape = var_9763, x = key_97_cast_fp16)[name = tensor("op_9764_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_9762_cast_fp16, y = var_9764_cast_fp16)[name = tensor("mh_w_145_cast_fp16")]; - tensor mh_w_147_cast_fp16 = add(x = mh_w_145_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_147_cast_fp16")]; - tensor var_9772_cast_fp16 = softmax(axis = var_9632, x = mh_w_147_cast_fp16)[name = tensor("op_9772_cast_fp16")]; - tensor var_9773 = const()[name = tensor("op_9773"), val = tensor([1, 20, 64, -1])]; - tensor var_9774_cast_fp16 = reshape(shape = var_9773, x = value_97_cast_fp16)[name = tensor("op_9774_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_9774_cast_fp16, y = var_9772_cast_fp16)[name = tensor("attn_97_cast_fp16")]; - tensor var_9777 = const()[name = tensor("op_9777"), val = tensor([1, 1280, 1, -1])]; - tensor input_727_cast_fp16 = reshape(shape = var_9777, x = attn_97_cast_fp16)[name = tensor("input_727_cast_fp16")]; - tensor var_9784 = const()[name = tensor("op_9784"), val = tensor([1, 1])]; - tensor var_9786 = const()[name = tensor("op_9786"), val = tensor([1, 1])]; - tensor pretrained_out_487_pad_type_0 = const()[name = tensor("pretrained_out_487_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_487_pad_0 = const()[name = tensor("pretrained_out_487_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479564928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480384192))), name = tensor("layers_24_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_24_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480384320)))]; - tensor pretrained_out_487_cast_fp16 = conv(bias = layers_24_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_9786, groups = var_9639, pad = pretrained_out_487_pad_0, pad_type = pretrained_out_487_pad_type_0, strides = var_9784, weight = layers_24_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_727_cast_fp16)[name = tensor("pretrained_out_487_cast_fp16")]; - tensor var_9790 = const()[name = tensor("op_9790"), val = tensor([1, 1])]; - tensor var_9792 = const()[name = tensor("op_9792"), val = tensor([1, 1])]; - tensor input_729_pad_type_0 = const()[name = tensor("input_729_pad_type_0"), val = tensor("custom")]; - tensor input_729_pad_0 = const()[name = tensor("input_729_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480386944)))]; - tensor input_729_cast_fp16 = conv(dilations = var_9792, groups = var_9639, pad = input_729_pad_0, pad_type = input_729_pad_type_0, strides = var_9790, weight = layers_24_self_attn_o_proj_loraA_weight_to_fp16, x = input_727_cast_fp16)[name = tensor("input_729_cast_fp16")]; - tensor var_9796 = const()[name = tensor("op_9796"), val = tensor([1, 1])]; - tensor var_9798 = const()[name = tensor("op_9798"), val = tensor([1, 1])]; - tensor lora_out_973_pad_type_0 = const()[name = tensor("lora_out_973_pad_type_0"), val = tensor("custom")]; - tensor lora_out_973_pad_0 = const()[name = tensor("lora_out_973_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_975_weight_0_to_fp16 = const()[name = tensor("lora_out_975_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480427968)))]; - tensor lora_out_975_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9798, groups = var_9639, pad = lora_out_973_pad_0, pad_type = lora_out_973_pad_type_0, strides = var_9796, weight = lora_out_975_weight_0_to_fp16, x = input_729_cast_fp16)[name = tensor("lora_out_975_cast_fp16")]; - tensor obj_295_cast_fp16 = add(x = pretrained_out_487_cast_fp16, y = lora_out_975_cast_fp16)[name = tensor("obj_295_cast_fp16")]; - tensor inputs_147_cast_fp16 = add(x = inputs_145_cast_fp16, y = obj_295_cast_fp16)[name = tensor("inputs_147_cast_fp16")]; - tensor var_9811 = const()[name = tensor("op_9811"), val = tensor([1])]; - tensor channels_mean_147_cast_fp16 = reduce_mean(axes = var_9811, keep_dims = var_9640, 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_9815 = const()[name = tensor("op_9815"), val = tensor([1])]; - tensor var_9816_cast_fp16 = reduce_mean(axes = var_9815, keep_dims = var_9640, x = zero_mean_sq_147_cast_fp16)[name = tensor("op_9816_cast_fp16")]; - tensor var_9817_to_fp16 = const()[name = tensor("op_9817_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_9818_cast_fp16 = add(x = var_9816_cast_fp16, y = var_9817_to_fp16)[name = tensor("op_9818_cast_fp16")]; - tensor denom_147_epsilon_0 = const()[name = tensor("denom_147_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_147_cast_fp16 = rsqrt(epsilon = denom_147_epsilon_0, x = var_9818_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_297_gamma_0_to_fp16 = const()[name = tensor("obj_297_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480468992)))]; - tensor obj_297_beta_0_to_fp16 = const()[name = tensor("obj_297_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480471616)))]; - tensor obj_297_epsilon_0_to_fp16 = const()[name = tensor("obj_297_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_297_cast_fp16 = batch_norm(beta = obj_297_beta_0_to_fp16, epsilon = obj_297_epsilon_0_to_fp16, gamma = obj_297_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_297_cast_fp16")]; - tensor var_9836 = const()[name = tensor("op_9836"), val = tensor([1, 1])]; - tensor var_9838 = const()[name = tensor("op_9838"), val = tensor([1, 1])]; - tensor pretrained_out_489_pad_type_0 = const()[name = tensor("pretrained_out_489_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_489_pad_0 = const()[name = tensor("pretrained_out_489_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480474240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(481293504))), name = tensor("layers_24_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_24_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_24_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(481293632)))]; - tensor pretrained_out_489_cast_fp16 = conv(bias = layers_24_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_9838, groups = var_9639, pad = pretrained_out_489_pad_0, pad_type = pretrained_out_489_pad_type_0, strides = var_9836, weight = layers_24_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_297_cast_fp16)[name = tensor("pretrained_out_489_cast_fp16")]; - tensor var_9842 = const()[name = tensor("op_9842"), val = tensor([1, 1])]; - tensor var_9844 = const()[name = tensor("op_9844"), val = tensor([1, 1])]; - tensor input_731_pad_type_0 = const()[name = tensor("input_731_pad_type_0"), val = tensor("custom")]; - tensor input_731_pad_0 = const()[name = tensor("input_731_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_24_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(481296256)))]; - tensor input_731_cast_fp16 = conv(dilations = var_9844, groups = var_9639, pad = input_731_pad_0, pad_type = input_731_pad_type_0, strides = var_9842, weight = layers_24_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_297_cast_fp16)[name = tensor("input_731_cast_fp16")]; - tensor var_9848 = const()[name = tensor("op_9848"), val = tensor([1, 1])]; - tensor var_9850 = const()[name = tensor("op_9850"), val = tensor([1, 1])]; - tensor lora_out_977_pad_type_0 = const()[name = tensor("lora_out_977_pad_type_0"), val = tensor("custom")]; - tensor lora_out_977_pad_0 = const()[name = tensor("lora_out_977_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_979_weight_0_to_fp16 = const()[name = tensor("lora_out_979_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(481337280)))]; - tensor lora_out_979_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9850, groups = var_9639, pad = lora_out_977_pad_0, pad_type = lora_out_977_pad_type_0, strides = var_9848, weight = lora_out_979_weight_0_to_fp16, x = input_731_cast_fp16)[name = tensor("lora_out_979_cast_fp16")]; - tensor query_99_cast_fp16 = add(x = pretrained_out_489_cast_fp16, y = lora_out_979_cast_fp16)[name = tensor("query_99_cast_fp16")]; - tensor var_9860 = const()[name = tensor("op_9860"), val = tensor([1, 1])]; - tensor var_9862 = const()[name = tensor("op_9862"), val = tensor([1, 1])]; - tensor pretrained_out_491_pad_type_0 = const()[name = tensor("pretrained_out_491_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_491_pad_0 = const()[name = tensor("pretrained_out_491_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(481378304))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(482197568))), name = tensor("layers_24_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_491_cast_fp16 = conv(dilations = var_9862, groups = var_9639, pad = pretrained_out_491_pad_0, pad_type = pretrained_out_491_pad_type_0, strides = var_9860, weight = layers_24_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_491_cast_fp16")]; - tensor var_9866 = const()[name = tensor("op_9866"), val = tensor([1, 1])]; - tensor var_9868 = const()[name = tensor("op_9868"), val = tensor([1, 1])]; - tensor input_733_pad_type_0 = const()[name = tensor("input_733_pad_type_0"), val = tensor("custom")]; - tensor input_733_pad_0 = const()[name = tensor("input_733_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_24_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(482197696)))]; - tensor input_733_cast_fp16 = conv(dilations = var_9868, groups = var_9639, pad = input_733_pad_0, pad_type = input_733_pad_type_0, strides = var_9866, weight = layers_24_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_733_cast_fp16")]; - tensor var_9872 = const()[name = tensor("op_9872"), val = tensor([1, 1])]; - tensor var_9874 = const()[name = tensor("op_9874"), val = tensor([1, 1])]; - tensor lora_out_981_pad_type_0 = const()[name = tensor("lora_out_981_pad_type_0"), val = tensor("custom")]; - tensor lora_out_981_pad_0 = const()[name = tensor("lora_out_981_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_983_weight_0_to_fp16 = const()[name = tensor("lora_out_983_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(482238720)))]; - tensor lora_out_983_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9874, groups = var_9639, pad = lora_out_981_pad_0, pad_type = lora_out_981_pad_type_0, strides = var_9872, weight = lora_out_983_weight_0_to_fp16, x = input_733_cast_fp16)[name = tensor("lora_out_983_cast_fp16")]; - tensor key_99_cast_fp16 = add(x = pretrained_out_491_cast_fp16, y = lora_out_983_cast_fp16)[name = tensor("key_99_cast_fp16")]; - tensor var_9885 = const()[name = tensor("op_9885"), val = tensor([1, 1])]; - tensor var_9887 = const()[name = tensor("op_9887"), val = tensor([1, 1])]; - tensor pretrained_out_493_pad_type_0 = const()[name = tensor("pretrained_out_493_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_493_pad_0 = const()[name = tensor("pretrained_out_493_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(482279744))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(483099008))), name = tensor("layers_24_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_24_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_24_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(483099136)))]; - tensor pretrained_out_493_cast_fp16 = conv(bias = layers_24_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_9887, groups = var_9639, pad = pretrained_out_493_pad_0, pad_type = pretrained_out_493_pad_type_0, strides = var_9885, weight = layers_24_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_493_cast_fp16")]; - tensor var_9891 = const()[name = tensor("op_9891"), val = tensor([1, 1])]; - tensor var_9893 = const()[name = tensor("op_9893"), val = tensor([1, 1])]; - tensor input_735_pad_type_0 = const()[name = tensor("input_735_pad_type_0"), val = tensor("custom")]; - tensor input_735_pad_0 = const()[name = tensor("input_735_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_24_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(483101760)))]; - tensor input_735_cast_fp16 = conv(dilations = var_9893, groups = var_9639, pad = input_735_pad_0, pad_type = input_735_pad_type_0, strides = var_9891, weight = layers_24_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_735_cast_fp16")]; - tensor var_9897 = const()[name = tensor("op_9897"), val = tensor([1, 1])]; - tensor var_9899 = const()[name = tensor("op_9899"), val = tensor([1, 1])]; - tensor lora_out_985_pad_type_0 = const()[name = tensor("lora_out_985_pad_type_0"), val = tensor("custom")]; - tensor lora_out_985_pad_0 = const()[name = tensor("lora_out_985_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_987_weight_0_to_fp16 = const()[name = tensor("lora_out_987_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(483142784)))]; - tensor lora_out_987_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9899, groups = var_9639, pad = lora_out_985_pad_0, pad_type = lora_out_985_pad_type_0, strides = var_9897, weight = lora_out_987_weight_0_to_fp16, x = input_735_cast_fp16)[name = tensor("lora_out_987_cast_fp16")]; - tensor value_99_cast_fp16 = add(x = pretrained_out_493_cast_fp16, y = lora_out_987_cast_fp16)[name = tensor("value_99_cast_fp16")]; - tensor var_9906 = const()[name = tensor("op_9906"), val = tensor([1, 20, 64, -1])]; - tensor var_9907_cast_fp16 = reshape(shape = var_9906, x = query_99_cast_fp16)[name = tensor("op_9907_cast_fp16")]; - tensor var_9908_to_fp16 = const()[name = tensor("op_9908_to_fp16"), val = tensor(0x1p-3)]; - tensor var_9909_cast_fp16 = mul(x = var_9907_cast_fp16, y = var_9908_to_fp16)[name = tensor("op_9909_cast_fp16")]; - tensor var_9910 = const()[name = tensor("op_9910"), val = tensor([1, 20, 64, -1])]; - tensor var_9911_cast_fp16 = reshape(shape = var_9910, x = key_99_cast_fp16)[name = tensor("op_9911_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_9909_cast_fp16, y = var_9911_cast_fp16)[name = tensor("mh_w_149_cast_fp16")]; - tensor var_9914_cast_fp16 = softmax(axis = var_9632, x = mh_w_149_cast_fp16)[name = tensor("op_9914_cast_fp16")]; - tensor var_9915 = const()[name = tensor("op_9915"), val = tensor([1, 20, 64, -1])]; - tensor var_9916_cast_fp16 = reshape(shape = var_9915, x = value_99_cast_fp16)[name = tensor("op_9916_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_9916_cast_fp16, y = var_9914_cast_fp16)[name = tensor("attn_99_cast_fp16")]; - tensor var_9919 = const()[name = tensor("op_9919"), val = tensor([1, 1280, 1, -1])]; - tensor input_737_cast_fp16 = reshape(shape = var_9919, x = attn_99_cast_fp16)[name = tensor("input_737_cast_fp16")]; - tensor var_9926 = const()[name = tensor("op_9926"), val = tensor([1, 1])]; - tensor var_9928 = const()[name = tensor("op_9928"), val = tensor([1, 1])]; - tensor pretrained_out_495_pad_type_0 = const()[name = tensor("pretrained_out_495_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_495_pad_0 = const()[name = tensor("pretrained_out_495_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(483183808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484003072))), name = tensor("layers_24_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_24_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_24_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484003200)))]; - tensor pretrained_out_495_cast_fp16 = conv(bias = layers_24_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_9928, groups = var_9639, pad = pretrained_out_495_pad_0, pad_type = pretrained_out_495_pad_type_0, strides = var_9926, weight = layers_24_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_737_cast_fp16)[name = tensor("pretrained_out_495_cast_fp16")]; - tensor var_9932 = const()[name = tensor("op_9932"), val = tensor([1, 1])]; - tensor var_9934 = const()[name = tensor("op_9934"), val = tensor([1, 1])]; - tensor input_739_pad_type_0 = const()[name = tensor("input_739_pad_type_0"), val = tensor("custom")]; - tensor input_739_pad_0 = const()[name = tensor("input_739_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_24_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484005824)))]; - tensor input_739_cast_fp16 = conv(dilations = var_9934, groups = var_9639, pad = input_739_pad_0, pad_type = input_739_pad_type_0, strides = var_9932, weight = layers_24_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_737_cast_fp16)[name = tensor("input_739_cast_fp16")]; - tensor var_9938 = const()[name = tensor("op_9938"), val = tensor([1, 1])]; - tensor var_9940 = const()[name = tensor("op_9940"), val = tensor([1, 1])]; - tensor lora_out_989_pad_type_0 = const()[name = tensor("lora_out_989_pad_type_0"), val = tensor("custom")]; - tensor lora_out_989_pad_0 = const()[name = tensor("lora_out_989_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_991_weight_0_to_fp16 = const()[name = tensor("lora_out_991_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484046848)))]; - tensor lora_out_991_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9940, groups = var_9639, pad = lora_out_989_pad_0, pad_type = lora_out_989_pad_type_0, strides = var_9938, weight = lora_out_991_weight_0_to_fp16, x = input_739_cast_fp16)[name = tensor("lora_out_991_cast_fp16")]; - tensor obj_299_cast_fp16 = add(x = pretrained_out_495_cast_fp16, y = lora_out_991_cast_fp16)[name = tensor("obj_299_cast_fp16")]; - tensor inputs_149_cast_fp16 = add(x = inputs_147_cast_fp16, y = obj_299_cast_fp16)[name = tensor("inputs_149_cast_fp16")]; - tensor var_9949 = const()[name = tensor("op_9949"), val = tensor([1])]; - tensor channels_mean_149_cast_fp16 = reduce_mean(axes = var_9949, keep_dims = var_9640, 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_9953 = const()[name = tensor("op_9953"), val = tensor([1])]; - tensor var_9954_cast_fp16 = reduce_mean(axes = var_9953, keep_dims = var_9640, x = zero_mean_sq_149_cast_fp16)[name = tensor("op_9954_cast_fp16")]; - tensor var_9955_to_fp16 = const()[name = tensor("op_9955_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_9956_cast_fp16 = add(x = var_9954_cast_fp16, y = var_9955_to_fp16)[name = tensor("op_9956_cast_fp16")]; - tensor denom_149_epsilon_0 = const()[name = tensor("denom_149_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_149_cast_fp16 = rsqrt(epsilon = denom_149_epsilon_0, x = var_9956_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_741_gamma_0_to_fp16 = const()[name = tensor("input_741_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484087872)))]; - tensor input_741_beta_0_to_fp16 = const()[name = tensor("input_741_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484090496)))]; - tensor input_741_epsilon_0_to_fp16 = const()[name = tensor("input_741_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_741_cast_fp16 = batch_norm(beta = input_741_beta_0_to_fp16, epsilon = input_741_epsilon_0_to_fp16, gamma = input_741_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_741_cast_fp16")]; - tensor var_9970 = const()[name = tensor("op_9970"), val = tensor([1, 1])]; - tensor var_9972 = const()[name = tensor("op_9972"), val = tensor([1, 1])]; - tensor pretrained_out_497_pad_type_0 = const()[name = tensor("pretrained_out_497_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_497_pad_0 = const()[name = tensor("pretrained_out_497_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484093120))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487369984))), name = tensor("layers_24_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_24_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_24_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487370112)))]; - tensor pretrained_out_497_cast_fp16 = conv(bias = layers_24_fc1_pretrained_bias_to_fp16, dilations = var_9972, groups = var_9639, pad = pretrained_out_497_pad_0, pad_type = pretrained_out_497_pad_type_0, strides = var_9970, weight = layers_24_fc1_pretrained_weight_to_fp16_palettized, x = input_741_cast_fp16)[name = tensor("pretrained_out_497_cast_fp16")]; - tensor var_9976 = const()[name = tensor("op_9976"), val = tensor([1, 1])]; - tensor var_9978 = const()[name = tensor("op_9978"), val = tensor([1, 1])]; - tensor input_743_pad_type_0 = const()[name = tensor("input_743_pad_type_0"), val = tensor("custom")]; - tensor input_743_pad_0 = const()[name = tensor("input_743_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_24_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487380416)))]; - tensor input_743_cast_fp16 = conv(dilations = var_9978, groups = var_9639, pad = input_743_pad_0, pad_type = input_743_pad_type_0, strides = var_9976, weight = layers_24_fc1_loraA_weight_to_fp16, x = input_741_cast_fp16)[name = tensor("input_743_cast_fp16")]; - tensor var_9982 = const()[name = tensor("op_9982"), val = tensor([1, 1])]; - tensor var_9984 = const()[name = tensor("op_9984"), val = tensor([1, 1])]; - tensor lora_out_993_pad_type_0 = const()[name = tensor("lora_out_993_pad_type_0"), val = tensor("custom")]; - tensor lora_out_993_pad_0 = const()[name = tensor("lora_out_993_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_995_weight_0_to_fp16 = const()[name = tensor("lora_out_995_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487421440)))]; - tensor lora_out_995_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_9984, groups = var_9639, pad = lora_out_993_pad_0, pad_type = lora_out_993_pad_type_0, strides = var_9982, weight = lora_out_995_weight_0_to_fp16, x = input_743_cast_fp16)[name = tensor("lora_out_995_cast_fp16")]; - tensor input_745_cast_fp16 = add(x = pretrained_out_497_cast_fp16, y = lora_out_995_cast_fp16)[name = tensor("input_745_cast_fp16")]; - tensor input_747_mode_0 = const()[name = tensor("input_747_mode_0"), val = tensor("EXACT")]; - tensor input_747_cast_fp16 = gelu(mode = input_747_mode_0, x = input_745_cast_fp16)[name = tensor("input_747_cast_fp16")]; - tensor var_9996 = const()[name = tensor("op_9996"), val = tensor([1, 1])]; - tensor var_9998 = const()[name = tensor("op_9998"), val = tensor([1, 1])]; - tensor pretrained_out_499_pad_type_0 = const()[name = tensor("pretrained_out_499_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_499_pad_0 = const()[name = tensor("pretrained_out_499_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487585344))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492500608))), name = tensor("layers_24_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_24_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_24_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492500800)))]; - tensor pretrained_out_499_cast_fp16 = conv(bias = layers_24_fc2_pretrained_bias_to_fp16, dilations = var_9998, groups = var_9639, pad = pretrained_out_499_pad_0, pad_type = pretrained_out_499_pad_type_0, strides = var_9996, weight = layers_24_fc2_pretrained_weight_to_fp16_palettized, x = input_747_cast_fp16)[name = tensor("pretrained_out_499_cast_fp16")]; - tensor var_10002 = const()[name = tensor("op_10002"), val = tensor([1, 1])]; - tensor var_10004 = const()[name = tensor("op_10004"), val = tensor([1, 1])]; - tensor input_749_pad_type_0 = const()[name = tensor("input_749_pad_type_0"), val = tensor("custom")]; - tensor input_749_pad_0 = const()[name = tensor("input_749_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_24_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_24_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492503424)))]; - tensor input_749_cast_fp16 = conv(dilations = var_10004, groups = var_9639, pad = input_749_pad_0, pad_type = input_749_pad_type_0, strides = var_10002, weight = layers_24_fc2_loraA_weight_to_fp16, x = input_747_cast_fp16)[name = tensor("input_749_cast_fp16")]; - tensor var_10008 = const()[name = tensor("op_10008"), val = tensor([1, 1])]; - tensor var_10010 = const()[name = tensor("op_10010"), val = tensor([1, 1])]; - tensor lora_out_997_pad_type_0 = const()[name = tensor("lora_out_997_pad_type_0"), val = tensor("custom")]; - tensor lora_out_997_pad_0 = const()[name = tensor("lora_out_997_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_999_weight_0_to_fp16 = const()[name = tensor("lora_out_999_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492667328)))]; - tensor lora_out_999_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10010, groups = var_9639, pad = lora_out_997_pad_0, pad_type = lora_out_997_pad_type_0, strides = var_10008, weight = lora_out_999_weight_0_to_fp16, x = input_749_cast_fp16)[name = tensor("lora_out_999_cast_fp16")]; - tensor hidden_states_51_cast_fp16 = add(x = pretrained_out_499_cast_fp16, y = lora_out_999_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_10026 = const()[name = tensor("op_10026"), val = tensor(3)]; - tensor var_10033 = const()[name = tensor("op_10033"), val = tensor(1)]; - tensor var_10034 = const()[name = tensor("op_10034"), val = tensor(true)]; - tensor var_10046 = const()[name = tensor("op_10046"), val = tensor([1])]; - tensor channels_mean_151_cast_fp16 = reduce_mean(axes = var_10046, keep_dims = var_10034, 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_10050 = const()[name = tensor("op_10050"), val = tensor([1])]; - tensor var_10051_cast_fp16 = reduce_mean(axes = var_10050, keep_dims = var_10034, x = zero_mean_sq_151_cast_fp16)[name = tensor("op_10051_cast_fp16")]; - tensor var_10052_to_fp16 = const()[name = tensor("op_10052_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_10053_cast_fp16 = add(x = var_10051_cast_fp16, y = var_10052_to_fp16)[name = tensor("op_10053_cast_fp16")]; - tensor denom_151_epsilon_0 = const()[name = tensor("denom_151_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_151_cast_fp16 = rsqrt(epsilon = denom_151_epsilon_0, x = var_10053_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_301_gamma_0_to_fp16 = const()[name = tensor("obj_301_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492708352)))]; - tensor obj_301_beta_0_to_fp16 = const()[name = tensor("obj_301_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492710976)))]; - tensor obj_301_epsilon_0_to_fp16 = const()[name = tensor("obj_301_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_301_cast_fp16 = batch_norm(beta = obj_301_beta_0_to_fp16, epsilon = obj_301_epsilon_0_to_fp16, gamma = obj_301_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_301_cast_fp16")]; - tensor var_10071 = const()[name = tensor("op_10071"), val = tensor([1, 1])]; - tensor var_10073 = const()[name = tensor("op_10073"), val = tensor([1, 1])]; - tensor pretrained_out_501_pad_type_0 = const()[name = tensor("pretrained_out_501_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_501_pad_0 = const()[name = tensor("pretrained_out_501_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492713600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493532864))), name = tensor("layers_25_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_25_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493532992)))]; - tensor pretrained_out_501_cast_fp16 = conv(bias = layers_25_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_10073, groups = var_10033, pad = pretrained_out_501_pad_0, pad_type = pretrained_out_501_pad_type_0, strides = var_10071, weight = layers_25_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_301_cast_fp16)[name = tensor("pretrained_out_501_cast_fp16")]; - tensor var_10077 = const()[name = tensor("op_10077"), val = tensor([1, 1])]; - tensor var_10079 = const()[name = tensor("op_10079"), val = tensor([1, 1])]; - tensor input_751_pad_type_0 = const()[name = tensor("input_751_pad_type_0"), val = tensor("custom")]; - tensor input_751_pad_0 = const()[name = tensor("input_751_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493535616)))]; - tensor input_751_cast_fp16 = conv(dilations = var_10079, groups = var_10033, pad = input_751_pad_0, pad_type = input_751_pad_type_0, strides = var_10077, weight = layers_25_self_attn_q_proj_loraA_weight_to_fp16, x = obj_301_cast_fp16)[name = tensor("input_751_cast_fp16")]; - tensor var_10083 = const()[name = tensor("op_10083"), val = tensor([1, 1])]; - tensor var_10085 = const()[name = tensor("op_10085"), val = tensor([1, 1])]; - tensor lora_out_1001_pad_type_0 = const()[name = tensor("lora_out_1001_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1001_pad_0 = const()[name = tensor("lora_out_1001_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1003_weight_0_to_fp16 = const()[name = tensor("lora_out_1003_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493576640)))]; - tensor lora_out_1003_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10085, groups = var_10033, pad = lora_out_1001_pad_0, pad_type = lora_out_1001_pad_type_0, strides = var_10083, weight = lora_out_1003_weight_0_to_fp16, x = input_751_cast_fp16)[name = tensor("lora_out_1003_cast_fp16")]; - tensor query_101_cast_fp16 = add(x = pretrained_out_501_cast_fp16, y = lora_out_1003_cast_fp16)[name = tensor("query_101_cast_fp16")]; - tensor var_10095 = const()[name = tensor("op_10095"), val = tensor([1, 1])]; - tensor var_10097 = const()[name = tensor("op_10097"), val = tensor([1, 1])]; - tensor pretrained_out_503_pad_type_0 = const()[name = tensor("pretrained_out_503_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_503_pad_0 = const()[name = tensor("pretrained_out_503_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493617664))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494436928))), name = tensor("layers_25_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_503_cast_fp16 = conv(dilations = var_10097, groups = var_10033, pad = pretrained_out_503_pad_0, pad_type = pretrained_out_503_pad_type_0, strides = var_10095, weight = layers_25_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_301_cast_fp16)[name = tensor("pretrained_out_503_cast_fp16")]; - tensor var_10101 = const()[name = tensor("op_10101"), val = tensor([1, 1])]; - tensor var_10103 = const()[name = tensor("op_10103"), val = tensor([1, 1])]; - tensor input_753_pad_type_0 = const()[name = tensor("input_753_pad_type_0"), val = tensor("custom")]; - tensor input_753_pad_0 = const()[name = tensor("input_753_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494437056)))]; - tensor input_753_cast_fp16 = conv(dilations = var_10103, groups = var_10033, pad = input_753_pad_0, pad_type = input_753_pad_type_0, strides = var_10101, weight = layers_25_self_attn_k_proj_loraA_weight_to_fp16, x = obj_301_cast_fp16)[name = tensor("input_753_cast_fp16")]; - tensor var_10107 = const()[name = tensor("op_10107"), val = tensor([1, 1])]; - tensor var_10109 = const()[name = tensor("op_10109"), val = tensor([1, 1])]; - tensor lora_out_1005_pad_type_0 = const()[name = tensor("lora_out_1005_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1005_pad_0 = const()[name = tensor("lora_out_1005_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1007_weight_0_to_fp16 = const()[name = tensor("lora_out_1007_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494478080)))]; - tensor lora_out_1007_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10109, groups = var_10033, pad = lora_out_1005_pad_0, pad_type = lora_out_1005_pad_type_0, strides = var_10107, weight = lora_out_1007_weight_0_to_fp16, x = input_753_cast_fp16)[name = tensor("lora_out_1007_cast_fp16")]; - tensor current_key_51_cast_fp16 = add(x = pretrained_out_503_cast_fp16, y = lora_out_1007_cast_fp16)[name = tensor("current_key_51_cast_fp16")]; - tensor var_10120 = const()[name = tensor("op_10120"), val = tensor([1, 1])]; - tensor var_10122 = const()[name = tensor("op_10122"), val = tensor([1, 1])]; - tensor pretrained_out_505_pad_type_0 = const()[name = tensor("pretrained_out_505_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_505_pad_0 = const()[name = tensor("pretrained_out_505_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494519104))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495338368))), name = tensor("layers_25_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_25_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495338496)))]; - tensor pretrained_out_505_cast_fp16 = conv(bias = layers_25_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_10122, groups = var_10033, pad = pretrained_out_505_pad_0, pad_type = pretrained_out_505_pad_type_0, strides = var_10120, weight = layers_25_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_301_cast_fp16)[name = tensor("pretrained_out_505_cast_fp16")]; - tensor var_10126 = const()[name = tensor("op_10126"), val = tensor([1, 1])]; - tensor var_10128 = const()[name = tensor("op_10128"), val = tensor([1, 1])]; - tensor input_755_pad_type_0 = const()[name = tensor("input_755_pad_type_0"), val = tensor("custom")]; - tensor input_755_pad_0 = const()[name = tensor("input_755_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495341120)))]; - tensor input_755_cast_fp16 = conv(dilations = var_10128, groups = var_10033, pad = input_755_pad_0, pad_type = input_755_pad_type_0, strides = var_10126, weight = layers_25_self_attn_v_proj_loraA_weight_to_fp16, x = obj_301_cast_fp16)[name = tensor("input_755_cast_fp16")]; - tensor var_10132 = const()[name = tensor("op_10132"), val = tensor([1, 1])]; - tensor var_10134 = const()[name = tensor("op_10134"), val = tensor([1, 1])]; - tensor lora_out_1009_pad_type_0 = const()[name = tensor("lora_out_1009_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1009_pad_0 = const()[name = tensor("lora_out_1009_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1011_weight_0_to_fp16 = const()[name = tensor("lora_out_1011_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495382144)))]; - tensor lora_out_1011_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10134, groups = var_10033, pad = lora_out_1009_pad_0, pad_type = lora_out_1009_pad_type_0, strides = var_10132, weight = lora_out_1011_weight_0_to_fp16, x = input_755_cast_fp16)[name = tensor("lora_out_1011_cast_fp16")]; - tensor current_value_51_cast_fp16 = add(x = pretrained_out_505_cast_fp16, y = lora_out_1011_cast_fp16)[name = tensor("current_value_51_cast_fp16")]; - tensor var_10144_cast_fp16 = mul(x = current_key_51_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_10144_cast_fp16")]; - tensor var_10146_cast_fp16 = mul(x = var_103_cast_fp16_25, y = var_295_cast_fp16)[name = tensor("op_10146_cast_fp16")]; - tensor key_101_cast_fp16 = add(x = var_10144_cast_fp16, y = var_10146_cast_fp16)[name = tensor("key_101_cast_fp16")]; - tensor var_10148_cast_fp16 = mul(x = current_value_51_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_10148_cast_fp16")]; - tensor var_10150_cast_fp16 = mul(x = var_138_cast_fp16_25, y = var_295_cast_fp16)[name = tensor("op_10150_cast_fp16")]; - tensor value_101_cast_fp16 = add(x = var_10148_cast_fp16, y = var_10150_cast_fp16)[name = tensor("value_101_cast_fp16")]; - tensor var_10153 = const()[name = tensor("op_10153"), val = tensor([1, 20, 64, -1])]; - tensor var_10154_cast_fp16 = reshape(shape = var_10153, x = query_101_cast_fp16)[name = tensor("op_10154_cast_fp16")]; - tensor var_10155_to_fp16 = const()[name = tensor("op_10155_to_fp16"), val = tensor(0x1p-3)]; - tensor var_10156_cast_fp16 = mul(x = var_10154_cast_fp16, y = var_10155_to_fp16)[name = tensor("op_10156_cast_fp16")]; - tensor var_10157 = const()[name = tensor("op_10157"), val = tensor([1, 20, 64, -1])]; - tensor var_10158_cast_fp16 = reshape(shape = var_10157, x = key_101_cast_fp16)[name = tensor("op_10158_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_10156_cast_fp16, y = var_10158_cast_fp16)[name = tensor("mh_w_151_cast_fp16")]; - tensor mh_w_153_cast_fp16 = add(x = mh_w_151_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_153_cast_fp16")]; - tensor var_10166_cast_fp16 = softmax(axis = var_10026, x = mh_w_153_cast_fp16)[name = tensor("op_10166_cast_fp16")]; - tensor var_10167 = const()[name = tensor("op_10167"), val = tensor([1, 20, 64, -1])]; - tensor var_10168_cast_fp16 = reshape(shape = var_10167, x = value_101_cast_fp16)[name = tensor("op_10168_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_10168_cast_fp16, y = var_10166_cast_fp16)[name = tensor("attn_101_cast_fp16")]; - tensor var_10171 = const()[name = tensor("op_10171"), val = tensor([1, 1280, 1, -1])]; - tensor input_757_cast_fp16 = reshape(shape = var_10171, x = attn_101_cast_fp16)[name = tensor("input_757_cast_fp16")]; - tensor var_10178 = const()[name = tensor("op_10178"), val = tensor([1, 1])]; - tensor var_10180 = const()[name = tensor("op_10180"), val = tensor([1, 1])]; - tensor pretrained_out_507_pad_type_0 = const()[name = tensor("pretrained_out_507_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_507_pad_0 = const()[name = tensor("pretrained_out_507_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495423168))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496242432))), name = tensor("layers_25_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_25_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496242560)))]; - tensor pretrained_out_507_cast_fp16 = conv(bias = layers_25_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_10180, groups = var_10033, pad = pretrained_out_507_pad_0, pad_type = pretrained_out_507_pad_type_0, strides = var_10178, weight = layers_25_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_757_cast_fp16)[name = tensor("pretrained_out_507_cast_fp16")]; - tensor var_10184 = const()[name = tensor("op_10184"), val = tensor([1, 1])]; - tensor var_10186 = const()[name = tensor("op_10186"), val = tensor([1, 1])]; - tensor input_759_pad_type_0 = const()[name = tensor("input_759_pad_type_0"), val = tensor("custom")]; - tensor input_759_pad_0 = const()[name = tensor("input_759_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496245184)))]; - tensor input_759_cast_fp16 = conv(dilations = var_10186, groups = var_10033, pad = input_759_pad_0, pad_type = input_759_pad_type_0, strides = var_10184, weight = layers_25_self_attn_o_proj_loraA_weight_to_fp16, x = input_757_cast_fp16)[name = tensor("input_759_cast_fp16")]; - tensor var_10190 = const()[name = tensor("op_10190"), val = tensor([1, 1])]; - tensor var_10192 = const()[name = tensor("op_10192"), val = tensor([1, 1])]; - tensor lora_out_1013_pad_type_0 = const()[name = tensor("lora_out_1013_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1013_pad_0 = const()[name = tensor("lora_out_1013_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1015_weight_0_to_fp16 = const()[name = tensor("lora_out_1015_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496286208)))]; - tensor lora_out_1015_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10192, groups = var_10033, pad = lora_out_1013_pad_0, pad_type = lora_out_1013_pad_type_0, strides = var_10190, weight = lora_out_1015_weight_0_to_fp16, x = input_759_cast_fp16)[name = tensor("lora_out_1015_cast_fp16")]; - tensor obj_307_cast_fp16 = add(x = pretrained_out_507_cast_fp16, y = lora_out_1015_cast_fp16)[name = tensor("obj_307_cast_fp16")]; - tensor inputs_153_cast_fp16 = add(x = inputs_151_cast_fp16, y = obj_307_cast_fp16)[name = tensor("inputs_153_cast_fp16")]; - tensor var_10205 = const()[name = tensor("op_10205"), val = tensor([1])]; - tensor channels_mean_153_cast_fp16 = reduce_mean(axes = var_10205, keep_dims = var_10034, 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_10209 = const()[name = tensor("op_10209"), val = tensor([1])]; - tensor var_10210_cast_fp16 = reduce_mean(axes = var_10209, keep_dims = var_10034, x = zero_mean_sq_153_cast_fp16)[name = tensor("op_10210_cast_fp16")]; - tensor var_10211_to_fp16 = const()[name = tensor("op_10211_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_10212_cast_fp16 = add(x = var_10210_cast_fp16, y = var_10211_to_fp16)[name = tensor("op_10212_cast_fp16")]; - tensor denom_153_epsilon_0 = const()[name = tensor("denom_153_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_153_cast_fp16 = rsqrt(epsilon = denom_153_epsilon_0, x = var_10212_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_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(496327232)))]; - 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(496329856)))]; - 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_153_cast_fp16)[name = tensor("obj_309_cast_fp16")]; - tensor var_10230 = const()[name = tensor("op_10230"), val = tensor([1, 1])]; - tensor var_10232 = const()[name = tensor("op_10232"), val = tensor([1, 1])]; - tensor pretrained_out_509_pad_type_0 = const()[name = tensor("pretrained_out_509_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_509_pad_0 = const()[name = tensor("pretrained_out_509_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496332480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(497151744))), name = tensor("layers_25_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_25_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_25_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(497151872)))]; - tensor pretrained_out_509_cast_fp16 = conv(bias = layers_25_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_10232, groups = var_10033, pad = pretrained_out_509_pad_0, pad_type = pretrained_out_509_pad_type_0, strides = var_10230, weight = layers_25_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_309_cast_fp16)[name = tensor("pretrained_out_509_cast_fp16")]; - tensor var_10236 = const()[name = tensor("op_10236"), val = tensor([1, 1])]; - tensor var_10238 = const()[name = tensor("op_10238"), val = tensor([1, 1])]; - tensor input_761_pad_type_0 = const()[name = tensor("input_761_pad_type_0"), val = tensor("custom")]; - tensor input_761_pad_0 = const()[name = tensor("input_761_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_25_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(497154496)))]; - tensor input_761_cast_fp16 = conv(dilations = var_10238, groups = var_10033, pad = input_761_pad_0, pad_type = input_761_pad_type_0, strides = var_10236, weight = layers_25_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("input_761_cast_fp16")]; - tensor var_10242 = const()[name = tensor("op_10242"), val = tensor([1, 1])]; - tensor var_10244 = const()[name = tensor("op_10244"), val = tensor([1, 1])]; - tensor lora_out_1017_pad_type_0 = const()[name = tensor("lora_out_1017_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1017_pad_0 = const()[name = tensor("lora_out_1017_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1019_weight_0_to_fp16 = const()[name = tensor("lora_out_1019_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(497195520)))]; - tensor lora_out_1019_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10244, groups = var_10033, pad = lora_out_1017_pad_0, pad_type = lora_out_1017_pad_type_0, strides = var_10242, weight = lora_out_1019_weight_0_to_fp16, x = input_761_cast_fp16)[name = tensor("lora_out_1019_cast_fp16")]; - tensor query_103_cast_fp16 = add(x = pretrained_out_509_cast_fp16, y = lora_out_1019_cast_fp16)[name = tensor("query_103_cast_fp16")]; - tensor var_10254 = const()[name = tensor("op_10254"), val = tensor([1, 1])]; - tensor var_10256 = const()[name = tensor("op_10256"), val = tensor([1, 1])]; - tensor pretrained_out_511_pad_type_0 = const()[name = tensor("pretrained_out_511_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_511_pad_0 = const()[name = tensor("pretrained_out_511_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(497236544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(498055808))), name = tensor("layers_25_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_511_cast_fp16 = conv(dilations = var_10256, groups = var_10033, pad = pretrained_out_511_pad_0, pad_type = pretrained_out_511_pad_type_0, strides = var_10254, weight = layers_25_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_511_cast_fp16")]; - tensor var_10260 = const()[name = tensor("op_10260"), val = tensor([1, 1])]; - tensor var_10262 = const()[name = tensor("op_10262"), val = tensor([1, 1])]; - tensor input_763_pad_type_0 = const()[name = tensor("input_763_pad_type_0"), val = tensor("custom")]; - tensor input_763_pad_0 = const()[name = tensor("input_763_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_25_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(498055936)))]; - tensor input_763_cast_fp16 = conv(dilations = var_10262, groups = var_10033, pad = input_763_pad_0, pad_type = input_763_pad_type_0, strides = var_10260, weight = layers_25_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_763_cast_fp16")]; - tensor var_10266 = const()[name = tensor("op_10266"), val = tensor([1, 1])]; - tensor var_10268 = const()[name = tensor("op_10268"), val = tensor([1, 1])]; - tensor lora_out_1021_pad_type_0 = const()[name = tensor("lora_out_1021_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1021_pad_0 = const()[name = tensor("lora_out_1021_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1023_weight_0_to_fp16 = const()[name = tensor("lora_out_1023_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(498096960)))]; - tensor lora_out_1023_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10268, groups = var_10033, pad = lora_out_1021_pad_0, pad_type = lora_out_1021_pad_type_0, strides = var_10266, weight = lora_out_1023_weight_0_to_fp16, x = input_763_cast_fp16)[name = tensor("lora_out_1023_cast_fp16")]; - tensor key_103_cast_fp16 = add(x = pretrained_out_511_cast_fp16, y = lora_out_1023_cast_fp16)[name = tensor("key_103_cast_fp16")]; - tensor var_10279 = const()[name = tensor("op_10279"), val = tensor([1, 1])]; - tensor var_10281 = const()[name = tensor("op_10281"), val = tensor([1, 1])]; - tensor pretrained_out_513_pad_type_0 = const()[name = tensor("pretrained_out_513_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_513_pad_0 = const()[name = tensor("pretrained_out_513_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(498137984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(498957248))), name = tensor("layers_25_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_25_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_25_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(498957376)))]; - tensor pretrained_out_513_cast_fp16 = conv(bias = layers_25_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_10281, groups = var_10033, pad = pretrained_out_513_pad_0, pad_type = pretrained_out_513_pad_type_0, strides = var_10279, weight = layers_25_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_513_cast_fp16")]; - tensor var_10285 = const()[name = tensor("op_10285"), val = tensor([1, 1])]; - tensor var_10287 = const()[name = tensor("op_10287"), val = tensor([1, 1])]; - tensor input_765_pad_type_0 = const()[name = tensor("input_765_pad_type_0"), val = tensor("custom")]; - tensor input_765_pad_0 = const()[name = tensor("input_765_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_25_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(498960000)))]; - tensor input_765_cast_fp16 = conv(dilations = var_10287, groups = var_10033, pad = input_765_pad_0, pad_type = input_765_pad_type_0, strides = var_10285, weight = layers_25_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_765_cast_fp16")]; - tensor var_10291 = const()[name = tensor("op_10291"), val = tensor([1, 1])]; - tensor var_10293 = const()[name = tensor("op_10293"), val = tensor([1, 1])]; - tensor lora_out_1025_pad_type_0 = const()[name = tensor("lora_out_1025_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1025_pad_0 = const()[name = tensor("lora_out_1025_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1027_weight_0_to_fp16 = const()[name = tensor("lora_out_1027_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499001024)))]; - tensor lora_out_1027_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10293, groups = var_10033, pad = lora_out_1025_pad_0, pad_type = lora_out_1025_pad_type_0, strides = var_10291, weight = lora_out_1027_weight_0_to_fp16, x = input_765_cast_fp16)[name = tensor("lora_out_1027_cast_fp16")]; - tensor value_103_cast_fp16 = add(x = pretrained_out_513_cast_fp16, y = lora_out_1027_cast_fp16)[name = tensor("value_103_cast_fp16")]; - tensor var_10300 = const()[name = tensor("op_10300"), val = tensor([1, 20, 64, -1])]; - tensor var_10301_cast_fp16 = reshape(shape = var_10300, x = query_103_cast_fp16)[name = tensor("op_10301_cast_fp16")]; - tensor var_10302_to_fp16 = const()[name = tensor("op_10302_to_fp16"), val = tensor(0x1p-3)]; - tensor var_10303_cast_fp16 = mul(x = var_10301_cast_fp16, y = var_10302_to_fp16)[name = tensor("op_10303_cast_fp16")]; - tensor var_10304 = const()[name = tensor("op_10304"), val = tensor([1, 20, 64, -1])]; - tensor var_10305_cast_fp16 = reshape(shape = var_10304, x = key_103_cast_fp16)[name = tensor("op_10305_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_10303_cast_fp16, y = var_10305_cast_fp16)[name = tensor("mh_w_155_cast_fp16")]; - tensor var_10308_cast_fp16 = softmax(axis = var_10026, x = mh_w_155_cast_fp16)[name = tensor("op_10308_cast_fp16")]; - tensor var_10309 = const()[name = tensor("op_10309"), val = tensor([1, 20, 64, -1])]; - tensor var_10310_cast_fp16 = reshape(shape = var_10309, x = value_103_cast_fp16)[name = tensor("op_10310_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_10310_cast_fp16, y = var_10308_cast_fp16)[name = tensor("attn_103_cast_fp16")]; - tensor var_10313 = const()[name = tensor("op_10313"), val = tensor([1, 1280, 1, -1])]; - tensor input_767_cast_fp16 = reshape(shape = var_10313, x = attn_103_cast_fp16)[name = tensor("input_767_cast_fp16")]; - tensor var_10320 = const()[name = tensor("op_10320"), val = tensor([1, 1])]; - tensor var_10322 = const()[name = tensor("op_10322"), val = tensor([1, 1])]; - tensor pretrained_out_515_pad_type_0 = const()[name = tensor("pretrained_out_515_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_515_pad_0 = const()[name = tensor("pretrained_out_515_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499042048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499861312))), name = tensor("layers_25_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_25_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_25_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499861440)))]; - tensor pretrained_out_515_cast_fp16 = conv(bias = layers_25_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_10322, groups = var_10033, pad = pretrained_out_515_pad_0, pad_type = pretrained_out_515_pad_type_0, strides = var_10320, weight = layers_25_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = tensor("pretrained_out_515_cast_fp16")]; - tensor var_10326 = const()[name = tensor("op_10326"), val = tensor([1, 1])]; - tensor var_10328 = const()[name = tensor("op_10328"), val = tensor([1, 1])]; - tensor input_769_pad_type_0 = const()[name = tensor("input_769_pad_type_0"), val = tensor("custom")]; - tensor input_769_pad_0 = const()[name = tensor("input_769_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_25_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499864064)))]; - tensor input_769_cast_fp16 = conv(dilations = var_10328, groups = var_10033, pad = input_769_pad_0, pad_type = input_769_pad_type_0, strides = var_10326, weight = layers_25_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_767_cast_fp16)[name = tensor("input_769_cast_fp16")]; - tensor var_10332 = const()[name = tensor("op_10332"), val = tensor([1, 1])]; - tensor var_10334 = const()[name = tensor("op_10334"), val = tensor([1, 1])]; - tensor lora_out_1029_pad_type_0 = const()[name = tensor("lora_out_1029_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1029_pad_0 = const()[name = tensor("lora_out_1029_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1031_weight_0_to_fp16 = const()[name = tensor("lora_out_1031_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499905088)))]; - tensor lora_out_1031_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10334, groups = var_10033, pad = lora_out_1029_pad_0, pad_type = lora_out_1029_pad_type_0, strides = var_10332, weight = lora_out_1031_weight_0_to_fp16, x = input_769_cast_fp16)[name = tensor("lora_out_1031_cast_fp16")]; - tensor obj_311_cast_fp16 = add(x = pretrained_out_515_cast_fp16, y = lora_out_1031_cast_fp16)[name = tensor("obj_311_cast_fp16")]; - tensor inputs_155_cast_fp16 = add(x = inputs_153_cast_fp16, y = obj_311_cast_fp16)[name = tensor("inputs_155_cast_fp16")]; - tensor var_10343 = const()[name = tensor("op_10343"), val = tensor([1])]; - tensor channels_mean_155_cast_fp16 = reduce_mean(axes = var_10343, keep_dims = var_10034, 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_10347 = const()[name = tensor("op_10347"), val = tensor([1])]; - tensor var_10348_cast_fp16 = reduce_mean(axes = var_10347, keep_dims = var_10034, x = zero_mean_sq_155_cast_fp16)[name = tensor("op_10348_cast_fp16")]; - tensor var_10349_to_fp16 = const()[name = tensor("op_10349_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_10350_cast_fp16 = add(x = var_10348_cast_fp16, y = var_10349_to_fp16)[name = tensor("op_10350_cast_fp16")]; - tensor denom_155_epsilon_0 = const()[name = tensor("denom_155_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_155_cast_fp16 = rsqrt(epsilon = denom_155_epsilon_0, x = var_10350_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_771_gamma_0_to_fp16 = const()[name = tensor("input_771_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499946112)))]; - tensor input_771_beta_0_to_fp16 = const()[name = tensor("input_771_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499948736)))]; - tensor input_771_epsilon_0_to_fp16 = const()[name = tensor("input_771_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_771_cast_fp16 = batch_norm(beta = input_771_beta_0_to_fp16, epsilon = input_771_epsilon_0_to_fp16, gamma = input_771_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_771_cast_fp16")]; - tensor var_10364 = const()[name = tensor("op_10364"), val = tensor([1, 1])]; - tensor var_10366 = const()[name = tensor("op_10366"), val = tensor([1, 1])]; - tensor pretrained_out_517_pad_type_0 = const()[name = tensor("pretrained_out_517_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_517_pad_0 = const()[name = tensor("pretrained_out_517_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499951360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503228224))), name = tensor("layers_25_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_25_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_25_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503228352)))]; - tensor pretrained_out_517_cast_fp16 = conv(bias = layers_25_fc1_pretrained_bias_to_fp16, dilations = var_10366, groups = var_10033, pad = pretrained_out_517_pad_0, pad_type = pretrained_out_517_pad_type_0, strides = var_10364, weight = layers_25_fc1_pretrained_weight_to_fp16_palettized, x = input_771_cast_fp16)[name = tensor("pretrained_out_517_cast_fp16")]; - tensor var_10370 = const()[name = tensor("op_10370"), val = tensor([1, 1])]; - tensor var_10372 = const()[name = tensor("op_10372"), val = tensor([1, 1])]; - tensor input_773_pad_type_0 = const()[name = tensor("input_773_pad_type_0"), val = tensor("custom")]; - tensor input_773_pad_0 = const()[name = tensor("input_773_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_25_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503238656)))]; - tensor input_773_cast_fp16 = conv(dilations = var_10372, groups = var_10033, pad = input_773_pad_0, pad_type = input_773_pad_type_0, strides = var_10370, weight = layers_25_fc1_loraA_weight_to_fp16, x = input_771_cast_fp16)[name = tensor("input_773_cast_fp16")]; - tensor var_10376 = const()[name = tensor("op_10376"), val = tensor([1, 1])]; - tensor var_10378 = const()[name = tensor("op_10378"), val = tensor([1, 1])]; - tensor lora_out_1033_pad_type_0 = const()[name = tensor("lora_out_1033_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1033_pad_0 = const()[name = tensor("lora_out_1033_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1035_weight_0_to_fp16 = const()[name = tensor("lora_out_1035_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503279680)))]; - tensor lora_out_1035_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_10378, groups = var_10033, pad = lora_out_1033_pad_0, pad_type = lora_out_1033_pad_type_0, strides = var_10376, weight = lora_out_1035_weight_0_to_fp16, x = input_773_cast_fp16)[name = tensor("lora_out_1035_cast_fp16")]; - tensor input_775_cast_fp16 = add(x = pretrained_out_517_cast_fp16, y = lora_out_1035_cast_fp16)[name = tensor("input_775_cast_fp16")]; - tensor input_777_mode_0 = const()[name = tensor("input_777_mode_0"), val = tensor("EXACT")]; - tensor input_777_cast_fp16 = gelu(mode = input_777_mode_0, x = input_775_cast_fp16)[name = tensor("input_777_cast_fp16")]; - tensor var_10390 = const()[name = tensor("op_10390"), val = tensor([1, 1])]; - tensor var_10392 = const()[name = tensor("op_10392"), val = tensor([1, 1])]; - tensor pretrained_out_519_pad_type_0 = const()[name = tensor("pretrained_out_519_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_519_pad_0 = const()[name = tensor("pretrained_out_519_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503443584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506720448))), name = tensor("layers_25_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_25_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_25_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506720576)))]; - tensor pretrained_out_519_cast_fp16 = conv(bias = layers_25_fc2_pretrained_bias_to_fp16, dilations = var_10392, groups = var_10033, pad = pretrained_out_519_pad_0, pad_type = pretrained_out_519_pad_type_0, strides = var_10390, weight = layers_25_fc2_pretrained_weight_to_fp16_palettized, x = input_777_cast_fp16)[name = tensor("pretrained_out_519_cast_fp16")]; - tensor var_10396 = const()[name = tensor("op_10396"), val = tensor([1, 1])]; - tensor var_10398 = const()[name = tensor("op_10398"), val = tensor([1, 1])]; - tensor input_779_pad_type_0 = const()[name = tensor("input_779_pad_type_0"), val = tensor("custom")]; - tensor input_779_pad_0 = const()[name = tensor("input_779_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_25_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_25_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506723200)))]; - tensor input_779_cast_fp16 = conv(dilations = var_10398, groups = var_10033, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = var_10396, weight = layers_25_fc2_loraA_weight_to_fp16, x = input_777_cast_fp16)[name = tensor("input_779_cast_fp16")]; - tensor var_10402 = const()[name = tensor("op_10402"), val = tensor([1, 1])]; - tensor var_10404 = const()[name = tensor("op_10404"), val = tensor([1, 1])]; - tensor lora_out_1037_pad_type_0 = const()[name = tensor("lora_out_1037_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1037_pad_0 = const()[name = tensor("lora_out_1037_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1039_weight_0_to_fp16 = const()[name = tensor("lora_out_1039_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506887104)))]; - tensor lora_out_1039_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10404, groups = var_10033, pad = lora_out_1037_pad_0, pad_type = lora_out_1037_pad_type_0, strides = var_10402, weight = lora_out_1039_weight_0_to_fp16, x = input_779_cast_fp16)[name = tensor("lora_out_1039_cast_fp16")]; - tensor hidden_states_53_cast_fp16 = add(x = pretrained_out_519_cast_fp16, y = lora_out_1039_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_10420 = const()[name = tensor("op_10420"), val = tensor(3)]; - tensor var_10427 = const()[name = tensor("op_10427"), val = tensor(1)]; - tensor var_10428 = const()[name = tensor("op_10428"), val = tensor(true)]; - tensor var_10440 = const()[name = tensor("op_10440"), val = tensor([1])]; - tensor channels_mean_157_cast_fp16 = reduce_mean(axes = var_10440, keep_dims = var_10428, 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_10444 = const()[name = tensor("op_10444"), val = tensor([1])]; - tensor var_10445_cast_fp16 = reduce_mean(axes = var_10444, keep_dims = var_10428, x = zero_mean_sq_157_cast_fp16)[name = tensor("op_10445_cast_fp16")]; - tensor var_10446_to_fp16 = const()[name = tensor("op_10446_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_10447_cast_fp16 = add(x = var_10445_cast_fp16, y = var_10446_to_fp16)[name = tensor("op_10447_cast_fp16")]; - tensor denom_157_epsilon_0 = const()[name = tensor("denom_157_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_157_cast_fp16 = rsqrt(epsilon = denom_157_epsilon_0, x = var_10447_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_313_gamma_0_to_fp16 = const()[name = tensor("obj_313_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506928128)))]; - tensor obj_313_beta_0_to_fp16 = const()[name = tensor("obj_313_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506930752)))]; - tensor obj_313_epsilon_0_to_fp16 = const()[name = tensor("obj_313_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_313_cast_fp16 = batch_norm(beta = obj_313_beta_0_to_fp16, epsilon = obj_313_epsilon_0_to_fp16, gamma = obj_313_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_313_cast_fp16")]; - tensor var_10465 = const()[name = tensor("op_10465"), val = tensor([1, 1])]; - tensor var_10467 = const()[name = tensor("op_10467"), val = tensor([1, 1])]; - tensor pretrained_out_521_pad_type_0 = const()[name = tensor("pretrained_out_521_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_521_pad_0 = const()[name = tensor("pretrained_out_521_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506933376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507752640))), name = tensor("layers_26_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_26_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507752768)))]; - tensor pretrained_out_521_cast_fp16 = conv(bias = layers_26_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_10467, groups = var_10427, pad = pretrained_out_521_pad_0, pad_type = pretrained_out_521_pad_type_0, strides = var_10465, weight = layers_26_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_313_cast_fp16)[name = tensor("pretrained_out_521_cast_fp16")]; - tensor var_10471 = const()[name = tensor("op_10471"), val = tensor([1, 1])]; - tensor var_10473 = const()[name = tensor("op_10473"), val = tensor([1, 1])]; - tensor input_781_pad_type_0 = const()[name = tensor("input_781_pad_type_0"), val = tensor("custom")]; - tensor input_781_pad_0 = const()[name = tensor("input_781_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507755392)))]; - tensor input_781_cast_fp16 = conv(dilations = var_10473, groups = var_10427, pad = input_781_pad_0, pad_type = input_781_pad_type_0, strides = var_10471, weight = layers_26_self_attn_q_proj_loraA_weight_to_fp16, x = obj_313_cast_fp16)[name = tensor("input_781_cast_fp16")]; - tensor var_10477 = const()[name = tensor("op_10477"), val = tensor([1, 1])]; - tensor var_10479 = const()[name = tensor("op_10479"), val = tensor([1, 1])]; - tensor lora_out_1041_pad_type_0 = const()[name = tensor("lora_out_1041_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1041_pad_0 = const()[name = tensor("lora_out_1041_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1043_weight_0_to_fp16 = const()[name = tensor("lora_out_1043_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507796416)))]; - tensor lora_out_1043_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10479, groups = var_10427, pad = lora_out_1041_pad_0, pad_type = lora_out_1041_pad_type_0, strides = var_10477, weight = lora_out_1043_weight_0_to_fp16, x = input_781_cast_fp16)[name = tensor("lora_out_1043_cast_fp16")]; - tensor query_105_cast_fp16 = add(x = pretrained_out_521_cast_fp16, y = lora_out_1043_cast_fp16)[name = tensor("query_105_cast_fp16")]; - tensor var_10489 = const()[name = tensor("op_10489"), val = tensor([1, 1])]; - tensor var_10491 = const()[name = tensor("op_10491"), val = tensor([1, 1])]; - tensor pretrained_out_523_pad_type_0 = const()[name = tensor("pretrained_out_523_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_523_pad_0 = const()[name = tensor("pretrained_out_523_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507837440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(508656704))), name = tensor("layers_26_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_523_cast_fp16 = conv(dilations = var_10491, groups = var_10427, pad = pretrained_out_523_pad_0, pad_type = pretrained_out_523_pad_type_0, strides = var_10489, weight = layers_26_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_313_cast_fp16)[name = tensor("pretrained_out_523_cast_fp16")]; - tensor var_10495 = const()[name = tensor("op_10495"), val = tensor([1, 1])]; - tensor var_10497 = const()[name = tensor("op_10497"), val = tensor([1, 1])]; - tensor input_783_pad_type_0 = const()[name = tensor("input_783_pad_type_0"), val = tensor("custom")]; - tensor input_783_pad_0 = const()[name = tensor("input_783_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(508656832)))]; - tensor input_783_cast_fp16 = conv(dilations = var_10497, groups = var_10427, pad = input_783_pad_0, pad_type = input_783_pad_type_0, strides = var_10495, weight = layers_26_self_attn_k_proj_loraA_weight_to_fp16, x = obj_313_cast_fp16)[name = tensor("input_783_cast_fp16")]; - tensor var_10501 = const()[name = tensor("op_10501"), val = tensor([1, 1])]; - tensor var_10503 = const()[name = tensor("op_10503"), val = tensor([1, 1])]; - tensor lora_out_1045_pad_type_0 = const()[name = tensor("lora_out_1045_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1045_pad_0 = const()[name = tensor("lora_out_1045_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1047_weight_0_to_fp16 = const()[name = tensor("lora_out_1047_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(508697856)))]; - tensor lora_out_1047_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10503, groups = var_10427, pad = lora_out_1045_pad_0, pad_type = lora_out_1045_pad_type_0, strides = var_10501, weight = lora_out_1047_weight_0_to_fp16, x = input_783_cast_fp16)[name = tensor("lora_out_1047_cast_fp16")]; - tensor current_key_53_cast_fp16 = add(x = pretrained_out_523_cast_fp16, y = lora_out_1047_cast_fp16)[name = tensor("current_key_53_cast_fp16")]; - tensor var_10514 = const()[name = tensor("op_10514"), val = tensor([1, 1])]; - tensor var_10516 = const()[name = tensor("op_10516"), val = tensor([1, 1])]; - tensor pretrained_out_525_pad_type_0 = const()[name = tensor("pretrained_out_525_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_525_pad_0 = const()[name = tensor("pretrained_out_525_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(508738880))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509558144))), name = tensor("layers_26_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_26_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509558272)))]; - tensor pretrained_out_525_cast_fp16 = conv(bias = layers_26_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_10516, groups = var_10427, pad = pretrained_out_525_pad_0, pad_type = pretrained_out_525_pad_type_0, strides = var_10514, weight = layers_26_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_313_cast_fp16)[name = tensor("pretrained_out_525_cast_fp16")]; - tensor var_10520 = const()[name = tensor("op_10520"), val = tensor([1, 1])]; - tensor var_10522 = const()[name = tensor("op_10522"), val = tensor([1, 1])]; - tensor input_785_pad_type_0 = const()[name = tensor("input_785_pad_type_0"), val = tensor("custom")]; - tensor input_785_pad_0 = const()[name = tensor("input_785_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509560896)))]; - tensor input_785_cast_fp16 = conv(dilations = var_10522, groups = var_10427, pad = input_785_pad_0, pad_type = input_785_pad_type_0, strides = var_10520, weight = layers_26_self_attn_v_proj_loraA_weight_to_fp16, x = obj_313_cast_fp16)[name = tensor("input_785_cast_fp16")]; - tensor var_10526 = const()[name = tensor("op_10526"), val = tensor([1, 1])]; - tensor var_10528 = const()[name = tensor("op_10528"), val = tensor([1, 1])]; - tensor lora_out_1049_pad_type_0 = const()[name = tensor("lora_out_1049_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1049_pad_0 = const()[name = tensor("lora_out_1049_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1051_weight_0_to_fp16 = const()[name = tensor("lora_out_1051_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509601920)))]; - tensor lora_out_1051_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10528, groups = var_10427, pad = lora_out_1049_pad_0, pad_type = lora_out_1049_pad_type_0, strides = var_10526, weight = lora_out_1051_weight_0_to_fp16, x = input_785_cast_fp16)[name = tensor("lora_out_1051_cast_fp16")]; - tensor current_value_53_cast_fp16 = add(x = pretrained_out_525_cast_fp16, y = lora_out_1051_cast_fp16)[name = tensor("current_value_53_cast_fp16")]; - tensor var_10538_cast_fp16 = mul(x = current_key_53_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_10538_cast_fp16")]; - tensor var_10540_cast_fp16 = mul(x = var_103_cast_fp16_26, y = var_295_cast_fp16)[name = tensor("op_10540_cast_fp16")]; - tensor key_105_cast_fp16 = add(x = var_10538_cast_fp16, y = var_10540_cast_fp16)[name = tensor("key_105_cast_fp16")]; - tensor var_10542_cast_fp16 = mul(x = current_value_53_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_10542_cast_fp16")]; - tensor var_10544_cast_fp16 = mul(x = var_138_cast_fp16_26, y = var_295_cast_fp16)[name = tensor("op_10544_cast_fp16")]; - tensor value_105_cast_fp16 = add(x = var_10542_cast_fp16, y = var_10544_cast_fp16)[name = tensor("value_105_cast_fp16")]; - tensor var_10547 = const()[name = tensor("op_10547"), val = tensor([1, 20, 64, -1])]; - tensor var_10548_cast_fp16 = reshape(shape = var_10547, x = query_105_cast_fp16)[name = tensor("op_10548_cast_fp16")]; - tensor var_10549_to_fp16 = const()[name = tensor("op_10549_to_fp16"), val = tensor(0x1p-3)]; - tensor var_10550_cast_fp16 = mul(x = var_10548_cast_fp16, y = var_10549_to_fp16)[name = tensor("op_10550_cast_fp16")]; - tensor var_10551 = const()[name = tensor("op_10551"), val = tensor([1, 20, 64, -1])]; - tensor var_10552_cast_fp16 = reshape(shape = var_10551, x = key_105_cast_fp16)[name = tensor("op_10552_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_10550_cast_fp16, y = var_10552_cast_fp16)[name = tensor("mh_w_157_cast_fp16")]; - tensor mh_w_159_cast_fp16 = add(x = mh_w_157_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_159_cast_fp16")]; - tensor var_10560_cast_fp16 = softmax(axis = var_10420, x = mh_w_159_cast_fp16)[name = tensor("op_10560_cast_fp16")]; - tensor var_10561 = const()[name = tensor("op_10561"), val = tensor([1, 20, 64, -1])]; - tensor var_10562_cast_fp16 = reshape(shape = var_10561, x = value_105_cast_fp16)[name = tensor("op_10562_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_10562_cast_fp16, y = var_10560_cast_fp16)[name = tensor("attn_105_cast_fp16")]; - tensor var_10565 = const()[name = tensor("op_10565"), val = tensor([1, 1280, 1, -1])]; - tensor input_787_cast_fp16 = reshape(shape = var_10565, x = attn_105_cast_fp16)[name = tensor("input_787_cast_fp16")]; - tensor var_10572 = const()[name = tensor("op_10572"), val = tensor([1, 1])]; - tensor var_10574 = const()[name = tensor("op_10574"), val = tensor([1, 1])]; - tensor pretrained_out_527_pad_type_0 = const()[name = tensor("pretrained_out_527_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_527_pad_0 = const()[name = tensor("pretrained_out_527_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509642944))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510462208))), name = tensor("layers_26_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_26_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510462336)))]; - tensor pretrained_out_527_cast_fp16 = conv(bias = layers_26_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_10574, groups = var_10427, pad = pretrained_out_527_pad_0, pad_type = pretrained_out_527_pad_type_0, strides = var_10572, weight = layers_26_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_787_cast_fp16)[name = tensor("pretrained_out_527_cast_fp16")]; - tensor var_10578 = const()[name = tensor("op_10578"), val = tensor([1, 1])]; - tensor var_10580 = const()[name = tensor("op_10580"), val = tensor([1, 1])]; - tensor input_789_pad_type_0 = const()[name = tensor("input_789_pad_type_0"), val = tensor("custom")]; - tensor input_789_pad_0 = const()[name = tensor("input_789_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510464960)))]; - tensor input_789_cast_fp16 = conv(dilations = var_10580, groups = var_10427, pad = input_789_pad_0, pad_type = input_789_pad_type_0, strides = var_10578, weight = layers_26_self_attn_o_proj_loraA_weight_to_fp16, x = input_787_cast_fp16)[name = tensor("input_789_cast_fp16")]; - tensor var_10584 = const()[name = tensor("op_10584"), val = tensor([1, 1])]; - tensor var_10586 = const()[name = tensor("op_10586"), val = tensor([1, 1])]; - tensor lora_out_1053_pad_type_0 = const()[name = tensor("lora_out_1053_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1053_pad_0 = const()[name = tensor("lora_out_1053_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1055_weight_0_to_fp16 = const()[name = tensor("lora_out_1055_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510505984)))]; - tensor lora_out_1055_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10586, groups = var_10427, pad = lora_out_1053_pad_0, pad_type = lora_out_1053_pad_type_0, strides = var_10584, weight = lora_out_1055_weight_0_to_fp16, x = input_789_cast_fp16)[name = tensor("lora_out_1055_cast_fp16")]; - tensor obj_319_cast_fp16 = add(x = pretrained_out_527_cast_fp16, y = lora_out_1055_cast_fp16)[name = tensor("obj_319_cast_fp16")]; - tensor inputs_159_cast_fp16 = add(x = inputs_157_cast_fp16, y = obj_319_cast_fp16)[name = tensor("inputs_159_cast_fp16")]; - tensor var_10599 = const()[name = tensor("op_10599"), val = tensor([1])]; - tensor channels_mean_159_cast_fp16 = reduce_mean(axes = var_10599, keep_dims = var_10428, 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_10603 = const()[name = tensor("op_10603"), val = tensor([1])]; - tensor var_10604_cast_fp16 = reduce_mean(axes = var_10603, keep_dims = var_10428, x = zero_mean_sq_159_cast_fp16)[name = tensor("op_10604_cast_fp16")]; - tensor var_10605_to_fp16 = const()[name = tensor("op_10605_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_10606_cast_fp16 = add(x = var_10604_cast_fp16, y = var_10605_to_fp16)[name = tensor("op_10606_cast_fp16")]; - tensor denom_159_epsilon_0 = const()[name = tensor("denom_159_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_159_cast_fp16 = rsqrt(epsilon = denom_159_epsilon_0, x = var_10606_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_321_gamma_0_to_fp16 = const()[name = tensor("obj_321_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510547008)))]; - tensor obj_321_beta_0_to_fp16 = const()[name = tensor("obj_321_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510549632)))]; - tensor obj_321_epsilon_0_to_fp16 = const()[name = tensor("obj_321_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_321_cast_fp16 = batch_norm(beta = obj_321_beta_0_to_fp16, epsilon = obj_321_epsilon_0_to_fp16, gamma = obj_321_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_321_cast_fp16")]; - tensor var_10624 = const()[name = tensor("op_10624"), val = tensor([1, 1])]; - tensor var_10626 = const()[name = tensor("op_10626"), val = tensor([1, 1])]; - tensor pretrained_out_529_pad_type_0 = const()[name = tensor("pretrained_out_529_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_529_pad_0 = const()[name = tensor("pretrained_out_529_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510552256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511371520))), name = tensor("layers_26_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_26_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_26_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511371648)))]; - tensor pretrained_out_529_cast_fp16 = conv(bias = layers_26_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_10626, groups = var_10427, pad = pretrained_out_529_pad_0, pad_type = pretrained_out_529_pad_type_0, strides = var_10624, weight = layers_26_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_321_cast_fp16)[name = tensor("pretrained_out_529_cast_fp16")]; - tensor var_10630 = const()[name = tensor("op_10630"), val = tensor([1, 1])]; - tensor var_10632 = const()[name = tensor("op_10632"), val = tensor([1, 1])]; - tensor input_791_pad_type_0 = const()[name = tensor("input_791_pad_type_0"), val = tensor("custom")]; - tensor input_791_pad_0 = const()[name = tensor("input_791_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_26_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511374272)))]; - tensor input_791_cast_fp16 = conv(dilations = var_10632, groups = var_10427, pad = input_791_pad_0, pad_type = input_791_pad_type_0, strides = var_10630, weight = layers_26_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_321_cast_fp16)[name = tensor("input_791_cast_fp16")]; - tensor var_10636 = const()[name = tensor("op_10636"), val = tensor([1, 1])]; - tensor var_10638 = const()[name = tensor("op_10638"), val = tensor([1, 1])]; - tensor lora_out_1057_pad_type_0 = const()[name = tensor("lora_out_1057_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1057_pad_0 = const()[name = tensor("lora_out_1057_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1059_weight_0_to_fp16 = const()[name = tensor("lora_out_1059_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511415296)))]; - tensor lora_out_1059_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10638, groups = var_10427, pad = lora_out_1057_pad_0, pad_type = lora_out_1057_pad_type_0, strides = var_10636, weight = lora_out_1059_weight_0_to_fp16, x = input_791_cast_fp16)[name = tensor("lora_out_1059_cast_fp16")]; - tensor query_107_cast_fp16 = add(x = pretrained_out_529_cast_fp16, y = lora_out_1059_cast_fp16)[name = tensor("query_107_cast_fp16")]; - tensor var_10648 = const()[name = tensor("op_10648"), val = tensor([1, 1])]; - tensor var_10650 = const()[name = tensor("op_10650"), val = tensor([1, 1])]; - tensor pretrained_out_531_pad_type_0 = const()[name = tensor("pretrained_out_531_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_531_pad_0 = const()[name = tensor("pretrained_out_531_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511456320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512275584))), name = tensor("layers_26_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_531_cast_fp16 = conv(dilations = var_10650, groups = var_10427, pad = pretrained_out_531_pad_0, pad_type = pretrained_out_531_pad_type_0, strides = var_10648, weight = layers_26_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_531_cast_fp16")]; - tensor var_10654 = const()[name = tensor("op_10654"), val = tensor([1, 1])]; - tensor var_10656 = const()[name = tensor("op_10656"), val = tensor([1, 1])]; - tensor input_793_pad_type_0 = const()[name = tensor("input_793_pad_type_0"), val = tensor("custom")]; - tensor input_793_pad_0 = const()[name = tensor("input_793_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_26_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512275712)))]; - tensor input_793_cast_fp16 = conv(dilations = var_10656, groups = var_10427, pad = input_793_pad_0, pad_type = input_793_pad_type_0, strides = var_10654, weight = layers_26_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_793_cast_fp16")]; - tensor var_10660 = const()[name = tensor("op_10660"), val = tensor([1, 1])]; - tensor var_10662 = const()[name = tensor("op_10662"), val = tensor([1, 1])]; - tensor lora_out_1061_pad_type_0 = const()[name = tensor("lora_out_1061_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1061_pad_0 = const()[name = tensor("lora_out_1061_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1063_weight_0_to_fp16 = const()[name = tensor("lora_out_1063_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512316736)))]; - tensor lora_out_1063_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10662, groups = var_10427, pad = lora_out_1061_pad_0, pad_type = lora_out_1061_pad_type_0, strides = var_10660, weight = lora_out_1063_weight_0_to_fp16, x = input_793_cast_fp16)[name = tensor("lora_out_1063_cast_fp16")]; - tensor key_107_cast_fp16 = add(x = pretrained_out_531_cast_fp16, y = lora_out_1063_cast_fp16)[name = tensor("key_107_cast_fp16")]; - tensor var_10673 = const()[name = tensor("op_10673"), val = tensor([1, 1])]; - tensor var_10675 = const()[name = tensor("op_10675"), val = tensor([1, 1])]; - tensor pretrained_out_533_pad_type_0 = const()[name = tensor("pretrained_out_533_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_533_pad_0 = const()[name = tensor("pretrained_out_533_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512357760))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513177024))), name = tensor("layers_26_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_26_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_26_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513177152)))]; - tensor pretrained_out_533_cast_fp16 = conv(bias = layers_26_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_10675, groups = var_10427, pad = pretrained_out_533_pad_0, pad_type = pretrained_out_533_pad_type_0, strides = var_10673, weight = layers_26_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_533_cast_fp16")]; - tensor var_10679 = const()[name = tensor("op_10679"), val = tensor([1, 1])]; - tensor var_10681 = const()[name = tensor("op_10681"), val = tensor([1, 1])]; - tensor input_795_pad_type_0 = const()[name = tensor("input_795_pad_type_0"), val = tensor("custom")]; - tensor input_795_pad_0 = const()[name = tensor("input_795_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_26_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513179776)))]; - tensor input_795_cast_fp16 = conv(dilations = var_10681, groups = var_10427, pad = input_795_pad_0, pad_type = input_795_pad_type_0, strides = var_10679, weight = layers_26_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_795_cast_fp16")]; - tensor var_10685 = const()[name = tensor("op_10685"), val = tensor([1, 1])]; - tensor var_10687 = const()[name = tensor("op_10687"), val = tensor([1, 1])]; - tensor lora_out_1065_pad_type_0 = const()[name = tensor("lora_out_1065_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1065_pad_0 = const()[name = tensor("lora_out_1065_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1067_weight_0_to_fp16 = const()[name = tensor("lora_out_1067_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513220800)))]; - tensor lora_out_1067_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10687, groups = var_10427, pad = lora_out_1065_pad_0, pad_type = lora_out_1065_pad_type_0, strides = var_10685, weight = lora_out_1067_weight_0_to_fp16, x = input_795_cast_fp16)[name = tensor("lora_out_1067_cast_fp16")]; - tensor value_107_cast_fp16 = add(x = pretrained_out_533_cast_fp16, y = lora_out_1067_cast_fp16)[name = tensor("value_107_cast_fp16")]; - tensor var_10694 = const()[name = tensor("op_10694"), val = tensor([1, 20, 64, -1])]; - tensor var_10695_cast_fp16 = reshape(shape = var_10694, x = query_107_cast_fp16)[name = tensor("op_10695_cast_fp16")]; - tensor var_10696_to_fp16 = const()[name = tensor("op_10696_to_fp16"), val = tensor(0x1p-3)]; - tensor var_10697_cast_fp16 = mul(x = var_10695_cast_fp16, y = var_10696_to_fp16)[name = tensor("op_10697_cast_fp16")]; - tensor var_10698 = const()[name = tensor("op_10698"), val = tensor([1, 20, 64, -1])]; - tensor var_10699_cast_fp16 = reshape(shape = var_10698, x = key_107_cast_fp16)[name = tensor("op_10699_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_10697_cast_fp16, y = var_10699_cast_fp16)[name = tensor("mh_w_161_cast_fp16")]; - tensor var_10702_cast_fp16 = softmax(axis = var_10420, x = mh_w_161_cast_fp16)[name = tensor("op_10702_cast_fp16")]; - tensor var_10703 = const()[name = tensor("op_10703"), val = tensor([1, 20, 64, -1])]; - tensor var_10704_cast_fp16 = reshape(shape = var_10703, x = value_107_cast_fp16)[name = tensor("op_10704_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_10704_cast_fp16, y = var_10702_cast_fp16)[name = tensor("attn_107_cast_fp16")]; - tensor var_10707 = const()[name = tensor("op_10707"), val = tensor([1, 1280, 1, -1])]; - tensor input_797_cast_fp16 = reshape(shape = var_10707, x = attn_107_cast_fp16)[name = tensor("input_797_cast_fp16")]; - tensor var_10714 = const()[name = tensor("op_10714"), val = tensor([1, 1])]; - tensor var_10716 = const()[name = tensor("op_10716"), val = tensor([1, 1])]; - tensor pretrained_out_535_pad_type_0 = const()[name = tensor("pretrained_out_535_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_535_pad_0 = const()[name = tensor("pretrained_out_535_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513261824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514081088))), name = tensor("layers_26_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_26_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_26_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514081216)))]; - tensor pretrained_out_535_cast_fp16 = conv(bias = layers_26_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_10716, groups = var_10427, pad = pretrained_out_535_pad_0, pad_type = pretrained_out_535_pad_type_0, strides = var_10714, weight = layers_26_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_797_cast_fp16)[name = tensor("pretrained_out_535_cast_fp16")]; - tensor var_10720 = const()[name = tensor("op_10720"), val = tensor([1, 1])]; - tensor var_10722 = const()[name = tensor("op_10722"), val = tensor([1, 1])]; - tensor input_799_pad_type_0 = const()[name = tensor("input_799_pad_type_0"), val = tensor("custom")]; - tensor input_799_pad_0 = const()[name = tensor("input_799_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_26_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514083840)))]; - tensor input_799_cast_fp16 = conv(dilations = var_10722, groups = var_10427, pad = input_799_pad_0, pad_type = input_799_pad_type_0, strides = var_10720, weight = layers_26_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_797_cast_fp16)[name = tensor("input_799_cast_fp16")]; - tensor var_10726 = const()[name = tensor("op_10726"), val = tensor([1, 1])]; - tensor var_10728 = const()[name = tensor("op_10728"), val = tensor([1, 1])]; - tensor lora_out_1069_pad_type_0 = const()[name = tensor("lora_out_1069_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1069_pad_0 = const()[name = tensor("lora_out_1069_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1071_weight_0_to_fp16 = const()[name = tensor("lora_out_1071_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514124864)))]; - tensor lora_out_1071_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10728, groups = var_10427, pad = lora_out_1069_pad_0, pad_type = lora_out_1069_pad_type_0, strides = var_10726, weight = lora_out_1071_weight_0_to_fp16, x = input_799_cast_fp16)[name = tensor("lora_out_1071_cast_fp16")]; - tensor obj_323_cast_fp16 = add(x = pretrained_out_535_cast_fp16, y = lora_out_1071_cast_fp16)[name = tensor("obj_323_cast_fp16")]; - tensor inputs_161_cast_fp16 = add(x = inputs_159_cast_fp16, y = obj_323_cast_fp16)[name = tensor("inputs_161_cast_fp16")]; - tensor var_10737 = const()[name = tensor("op_10737"), val = tensor([1])]; - tensor channels_mean_161_cast_fp16 = reduce_mean(axes = var_10737, keep_dims = var_10428, 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_10741 = const()[name = tensor("op_10741"), val = tensor([1])]; - tensor var_10742_cast_fp16 = reduce_mean(axes = var_10741, keep_dims = var_10428, x = zero_mean_sq_161_cast_fp16)[name = tensor("op_10742_cast_fp16")]; - tensor var_10743_to_fp16 = const()[name = tensor("op_10743_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_10744_cast_fp16 = add(x = var_10742_cast_fp16, y = var_10743_to_fp16)[name = tensor("op_10744_cast_fp16")]; - tensor denom_161_epsilon_0 = const()[name = tensor("denom_161_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_161_cast_fp16 = rsqrt(epsilon = denom_161_epsilon_0, x = var_10744_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_801_gamma_0_to_fp16 = const()[name = tensor("input_801_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514165888)))]; - tensor input_801_beta_0_to_fp16 = const()[name = tensor("input_801_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514168512)))]; - tensor input_801_epsilon_0_to_fp16 = const()[name = tensor("input_801_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_801_cast_fp16 = batch_norm(beta = input_801_beta_0_to_fp16, epsilon = input_801_epsilon_0_to_fp16, gamma = input_801_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_801_cast_fp16")]; - tensor var_10758 = const()[name = tensor("op_10758"), val = tensor([1, 1])]; - tensor var_10760 = const()[name = tensor("op_10760"), val = tensor([1, 1])]; - tensor pretrained_out_537_pad_type_0 = const()[name = tensor("pretrained_out_537_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_537_pad_0 = const()[name = tensor("pretrained_out_537_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514171136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517448000))), name = tensor("layers_26_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_26_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_26_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517448128)))]; - tensor pretrained_out_537_cast_fp16 = conv(bias = layers_26_fc1_pretrained_bias_to_fp16, dilations = var_10760, groups = var_10427, pad = pretrained_out_537_pad_0, pad_type = pretrained_out_537_pad_type_0, strides = var_10758, weight = layers_26_fc1_pretrained_weight_to_fp16_palettized, x = input_801_cast_fp16)[name = tensor("pretrained_out_537_cast_fp16")]; - tensor var_10764 = const()[name = tensor("op_10764"), val = tensor([1, 1])]; - tensor var_10766 = const()[name = tensor("op_10766"), val = tensor([1, 1])]; - tensor input_803_pad_type_0 = const()[name = tensor("input_803_pad_type_0"), val = tensor("custom")]; - tensor input_803_pad_0 = const()[name = tensor("input_803_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_26_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517458432)))]; - tensor input_803_cast_fp16 = conv(dilations = var_10766, groups = var_10427, pad = input_803_pad_0, pad_type = input_803_pad_type_0, strides = var_10764, weight = layers_26_fc1_loraA_weight_to_fp16, x = input_801_cast_fp16)[name = tensor("input_803_cast_fp16")]; - tensor var_10770 = const()[name = tensor("op_10770"), val = tensor([1, 1])]; - tensor var_10772 = const()[name = tensor("op_10772"), val = tensor([1, 1])]; - tensor lora_out_1073_pad_type_0 = const()[name = tensor("lora_out_1073_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1073_pad_0 = const()[name = tensor("lora_out_1073_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1075_weight_0_to_fp16 = const()[name = tensor("lora_out_1075_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517499456)))]; - tensor lora_out_1075_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_10772, groups = var_10427, pad = lora_out_1073_pad_0, pad_type = lora_out_1073_pad_type_0, strides = var_10770, weight = lora_out_1075_weight_0_to_fp16, x = input_803_cast_fp16)[name = tensor("lora_out_1075_cast_fp16")]; - tensor input_805_cast_fp16 = add(x = pretrained_out_537_cast_fp16, y = lora_out_1075_cast_fp16)[name = tensor("input_805_cast_fp16")]; - tensor input_807_mode_0 = const()[name = tensor("input_807_mode_0"), val = tensor("EXACT")]; - tensor input_807_cast_fp16 = gelu(mode = input_807_mode_0, x = input_805_cast_fp16)[name = tensor("input_807_cast_fp16")]; - tensor var_10784 = const()[name = tensor("op_10784"), val = tensor([1, 1])]; - tensor var_10786 = const()[name = tensor("op_10786"), val = tensor([1, 1])]; - tensor pretrained_out_539_pad_type_0 = const()[name = tensor("pretrained_out_539_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_539_pad_0 = const()[name = tensor("pretrained_out_539_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517663360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520940224))), name = tensor("layers_26_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_26_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_26_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520940352)))]; - tensor pretrained_out_539_cast_fp16 = conv(bias = layers_26_fc2_pretrained_bias_to_fp16, dilations = var_10786, groups = var_10427, pad = pretrained_out_539_pad_0, pad_type = pretrained_out_539_pad_type_0, strides = var_10784, weight = layers_26_fc2_pretrained_weight_to_fp16_palettized, x = input_807_cast_fp16)[name = tensor("pretrained_out_539_cast_fp16")]; - tensor var_10790 = const()[name = tensor("op_10790"), val = tensor([1, 1])]; - tensor var_10792 = const()[name = tensor("op_10792"), val = tensor([1, 1])]; - tensor input_809_pad_type_0 = const()[name = tensor("input_809_pad_type_0"), val = tensor("custom")]; - tensor input_809_pad_0 = const()[name = tensor("input_809_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_26_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_26_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520942976)))]; - tensor input_809_cast_fp16 = conv(dilations = var_10792, groups = var_10427, pad = input_809_pad_0, pad_type = input_809_pad_type_0, strides = var_10790, weight = layers_26_fc2_loraA_weight_to_fp16, x = input_807_cast_fp16)[name = tensor("input_809_cast_fp16")]; - tensor var_10796 = const()[name = tensor("op_10796"), val = tensor([1, 1])]; - tensor var_10798 = const()[name = tensor("op_10798"), val = tensor([1, 1])]; - tensor lora_out_1077_pad_type_0 = const()[name = tensor("lora_out_1077_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1077_pad_0 = const()[name = tensor("lora_out_1077_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1079_weight_0_to_fp16 = const()[name = tensor("lora_out_1079_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521106880)))]; - tensor lora_out_1079_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10798, groups = var_10427, pad = lora_out_1077_pad_0, pad_type = lora_out_1077_pad_type_0, strides = var_10796, weight = lora_out_1079_weight_0_to_fp16, x = input_809_cast_fp16)[name = tensor("lora_out_1079_cast_fp16")]; - tensor hidden_states_55_cast_fp16 = add(x = pretrained_out_539_cast_fp16, y = lora_out_1079_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_10814 = const()[name = tensor("op_10814"), val = tensor(3)]; - tensor var_10821 = const()[name = tensor("op_10821"), val = tensor(1)]; - tensor var_10822 = const()[name = tensor("op_10822"), val = tensor(true)]; - tensor var_10834 = const()[name = tensor("op_10834"), val = tensor([1])]; - tensor channels_mean_163_cast_fp16 = reduce_mean(axes = var_10834, keep_dims = var_10822, 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_10838 = const()[name = tensor("op_10838"), val = tensor([1])]; - tensor var_10839_cast_fp16 = reduce_mean(axes = var_10838, keep_dims = var_10822, x = zero_mean_sq_163_cast_fp16)[name = tensor("op_10839_cast_fp16")]; - tensor var_10840_to_fp16 = const()[name = tensor("op_10840_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_10841_cast_fp16 = add(x = var_10839_cast_fp16, y = var_10840_to_fp16)[name = tensor("op_10841_cast_fp16")]; - tensor denom_163_epsilon_0 = const()[name = tensor("denom_163_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_163_cast_fp16 = rsqrt(epsilon = denom_163_epsilon_0, x = var_10841_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_325_gamma_0_to_fp16 = const()[name = tensor("obj_325_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521147904)))]; - tensor obj_325_beta_0_to_fp16 = const()[name = tensor("obj_325_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521150528)))]; - tensor obj_325_epsilon_0_to_fp16 = const()[name = tensor("obj_325_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_325_cast_fp16 = batch_norm(beta = obj_325_beta_0_to_fp16, epsilon = obj_325_epsilon_0_to_fp16, gamma = obj_325_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_325_cast_fp16")]; - tensor var_10859 = const()[name = tensor("op_10859"), val = tensor([1, 1])]; - tensor var_10861 = const()[name = tensor("op_10861"), val = tensor([1, 1])]; - tensor pretrained_out_541_pad_type_0 = const()[name = tensor("pretrained_out_541_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_541_pad_0 = const()[name = tensor("pretrained_out_541_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521153152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521972416))), name = tensor("layers_27_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_27_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521972544)))]; - tensor pretrained_out_541_cast_fp16 = conv(bias = layers_27_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_10861, groups = var_10821, pad = pretrained_out_541_pad_0, pad_type = pretrained_out_541_pad_type_0, strides = var_10859, weight = layers_27_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_325_cast_fp16)[name = tensor("pretrained_out_541_cast_fp16")]; - tensor var_10865 = const()[name = tensor("op_10865"), val = tensor([1, 1])]; - tensor var_10867 = const()[name = tensor("op_10867"), val = tensor([1, 1])]; - tensor input_811_pad_type_0 = const()[name = tensor("input_811_pad_type_0"), val = tensor("custom")]; - tensor input_811_pad_0 = const()[name = tensor("input_811_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521975168)))]; - tensor input_811_cast_fp16 = conv(dilations = var_10867, groups = var_10821, pad = input_811_pad_0, pad_type = input_811_pad_type_0, strides = var_10865, weight = layers_27_self_attn_q_proj_loraA_weight_to_fp16, x = obj_325_cast_fp16)[name = tensor("input_811_cast_fp16")]; - tensor var_10871 = const()[name = tensor("op_10871"), val = tensor([1, 1])]; - tensor var_10873 = const()[name = tensor("op_10873"), val = tensor([1, 1])]; - tensor lora_out_1081_pad_type_0 = const()[name = tensor("lora_out_1081_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1081_pad_0 = const()[name = tensor("lora_out_1081_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1083_weight_0_to_fp16 = const()[name = tensor("lora_out_1083_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522016192)))]; - tensor lora_out_1083_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10873, groups = var_10821, pad = lora_out_1081_pad_0, pad_type = lora_out_1081_pad_type_0, strides = var_10871, weight = lora_out_1083_weight_0_to_fp16, x = input_811_cast_fp16)[name = tensor("lora_out_1083_cast_fp16")]; - tensor query_109_cast_fp16 = add(x = pretrained_out_541_cast_fp16, y = lora_out_1083_cast_fp16)[name = tensor("query_109_cast_fp16")]; - tensor var_10883 = const()[name = tensor("op_10883"), val = tensor([1, 1])]; - tensor var_10885 = const()[name = tensor("op_10885"), val = tensor([1, 1])]; - tensor pretrained_out_543_pad_type_0 = const()[name = tensor("pretrained_out_543_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_543_pad_0 = const()[name = tensor("pretrained_out_543_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522057216))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522876480))), name = tensor("layers_27_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_543_cast_fp16 = conv(dilations = var_10885, groups = var_10821, pad = pretrained_out_543_pad_0, pad_type = pretrained_out_543_pad_type_0, strides = var_10883, weight = layers_27_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_325_cast_fp16)[name = tensor("pretrained_out_543_cast_fp16")]; - tensor var_10889 = const()[name = tensor("op_10889"), val = tensor([1, 1])]; - tensor var_10891 = const()[name = tensor("op_10891"), val = tensor([1, 1])]; - tensor input_813_pad_type_0 = const()[name = tensor("input_813_pad_type_0"), val = tensor("custom")]; - tensor input_813_pad_0 = const()[name = tensor("input_813_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522876608)))]; - tensor input_813_cast_fp16 = conv(dilations = var_10891, groups = var_10821, pad = input_813_pad_0, pad_type = input_813_pad_type_0, strides = var_10889, weight = layers_27_self_attn_k_proj_loraA_weight_to_fp16, x = obj_325_cast_fp16)[name = tensor("input_813_cast_fp16")]; - tensor var_10895 = const()[name = tensor("op_10895"), val = tensor([1, 1])]; - tensor var_10897 = const()[name = tensor("op_10897"), val = tensor([1, 1])]; - tensor lora_out_1085_pad_type_0 = const()[name = tensor("lora_out_1085_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1085_pad_0 = const()[name = tensor("lora_out_1085_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1087_weight_0_to_fp16 = const()[name = tensor("lora_out_1087_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522917632)))]; - tensor lora_out_1087_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10897, groups = var_10821, pad = lora_out_1085_pad_0, pad_type = lora_out_1085_pad_type_0, strides = var_10895, weight = lora_out_1087_weight_0_to_fp16, x = input_813_cast_fp16)[name = tensor("lora_out_1087_cast_fp16")]; - tensor current_key_55_cast_fp16 = add(x = pretrained_out_543_cast_fp16, y = lora_out_1087_cast_fp16)[name = tensor("current_key_55_cast_fp16")]; - tensor var_10908 = const()[name = tensor("op_10908"), val = tensor([1, 1])]; - tensor var_10910 = const()[name = tensor("op_10910"), val = tensor([1, 1])]; - tensor pretrained_out_545_pad_type_0 = const()[name = tensor("pretrained_out_545_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_545_pad_0 = const()[name = tensor("pretrained_out_545_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522958656))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(523777920))), name = tensor("layers_27_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_27_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(523778048)))]; - tensor pretrained_out_545_cast_fp16 = conv(bias = layers_27_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_10910, groups = var_10821, pad = pretrained_out_545_pad_0, pad_type = pretrained_out_545_pad_type_0, strides = var_10908, weight = layers_27_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_325_cast_fp16)[name = tensor("pretrained_out_545_cast_fp16")]; - tensor var_10914 = const()[name = tensor("op_10914"), val = tensor([1, 1])]; - tensor var_10916 = const()[name = tensor("op_10916"), val = tensor([1, 1])]; - tensor input_815_pad_type_0 = const()[name = tensor("input_815_pad_type_0"), val = tensor("custom")]; - tensor input_815_pad_0 = const()[name = tensor("input_815_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(523780672)))]; - tensor input_815_cast_fp16 = conv(dilations = var_10916, groups = var_10821, pad = input_815_pad_0, pad_type = input_815_pad_type_0, strides = var_10914, weight = layers_27_self_attn_v_proj_loraA_weight_to_fp16, x = obj_325_cast_fp16)[name = tensor("input_815_cast_fp16")]; - tensor var_10920 = const()[name = tensor("op_10920"), val = tensor([1, 1])]; - tensor var_10922 = const()[name = tensor("op_10922"), val = tensor([1, 1])]; - tensor lora_out_1089_pad_type_0 = const()[name = tensor("lora_out_1089_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1089_pad_0 = const()[name = tensor("lora_out_1089_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1091_weight_0_to_fp16 = const()[name = tensor("lora_out_1091_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(523821696)))]; - tensor lora_out_1091_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10922, groups = var_10821, pad = lora_out_1089_pad_0, pad_type = lora_out_1089_pad_type_0, strides = var_10920, weight = lora_out_1091_weight_0_to_fp16, x = input_815_cast_fp16)[name = tensor("lora_out_1091_cast_fp16")]; - tensor current_value_55_cast_fp16 = add(x = pretrained_out_545_cast_fp16, y = lora_out_1091_cast_fp16)[name = tensor("current_value_55_cast_fp16")]; - tensor var_10932_cast_fp16 = mul(x = current_key_55_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_10932_cast_fp16")]; - tensor var_10934_cast_fp16 = mul(x = var_103_cast_fp16_27, y = var_295_cast_fp16)[name = tensor("op_10934_cast_fp16")]; - tensor key_109_cast_fp16 = add(x = var_10932_cast_fp16, y = var_10934_cast_fp16)[name = tensor("key_109_cast_fp16")]; - tensor var_10936_cast_fp16 = mul(x = current_value_55_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_10936_cast_fp16")]; - tensor var_10938_cast_fp16 = mul(x = var_138_cast_fp16_27, y = var_295_cast_fp16)[name = tensor("op_10938_cast_fp16")]; - tensor value_109_cast_fp16 = add(x = var_10936_cast_fp16, y = var_10938_cast_fp16)[name = tensor("value_109_cast_fp16")]; - tensor var_10941 = const()[name = tensor("op_10941"), val = tensor([1, 20, 64, -1])]; - tensor var_10942_cast_fp16 = reshape(shape = var_10941, x = query_109_cast_fp16)[name = tensor("op_10942_cast_fp16")]; - tensor var_10943_to_fp16 = const()[name = tensor("op_10943_to_fp16"), val = tensor(0x1p-3)]; - tensor var_10944_cast_fp16 = mul(x = var_10942_cast_fp16, y = var_10943_to_fp16)[name = tensor("op_10944_cast_fp16")]; - tensor var_10945 = const()[name = tensor("op_10945"), val = tensor([1, 20, 64, -1])]; - tensor var_10946_cast_fp16 = reshape(shape = var_10945, x = key_109_cast_fp16)[name = tensor("op_10946_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_10944_cast_fp16, y = var_10946_cast_fp16)[name = tensor("mh_w_163_cast_fp16")]; - tensor mh_w_165_cast_fp16 = add(x = mh_w_163_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_165_cast_fp16")]; - tensor var_10954_cast_fp16 = softmax(axis = var_10814, x = mh_w_165_cast_fp16)[name = tensor("op_10954_cast_fp16")]; - tensor var_10955 = const()[name = tensor("op_10955"), val = tensor([1, 20, 64, -1])]; - tensor var_10956_cast_fp16 = reshape(shape = var_10955, x = value_109_cast_fp16)[name = tensor("op_10956_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_10956_cast_fp16, y = var_10954_cast_fp16)[name = tensor("attn_109_cast_fp16")]; - tensor var_10959 = const()[name = tensor("op_10959"), val = tensor([1, 1280, 1, -1])]; - tensor input_817_cast_fp16 = reshape(shape = var_10959, x = attn_109_cast_fp16)[name = tensor("input_817_cast_fp16")]; - tensor var_10966 = const()[name = tensor("op_10966"), val = tensor([1, 1])]; - tensor var_10968 = const()[name = tensor("op_10968"), val = tensor([1, 1])]; - tensor pretrained_out_547_pad_type_0 = const()[name = tensor("pretrained_out_547_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_547_pad_0 = const()[name = tensor("pretrained_out_547_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(523862720))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524681984))), name = tensor("layers_27_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_27_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524682112)))]; - tensor pretrained_out_547_cast_fp16 = conv(bias = layers_27_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_10968, groups = var_10821, pad = pretrained_out_547_pad_0, pad_type = pretrained_out_547_pad_type_0, strides = var_10966, weight = layers_27_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_817_cast_fp16)[name = tensor("pretrained_out_547_cast_fp16")]; - tensor var_10972 = const()[name = tensor("op_10972"), val = tensor([1, 1])]; - tensor var_10974 = const()[name = tensor("op_10974"), val = tensor([1, 1])]; - tensor input_819_pad_type_0 = const()[name = tensor("input_819_pad_type_0"), val = tensor("custom")]; - tensor input_819_pad_0 = const()[name = tensor("input_819_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524684736)))]; - tensor input_819_cast_fp16 = conv(dilations = var_10974, groups = var_10821, pad = input_819_pad_0, pad_type = input_819_pad_type_0, strides = var_10972, weight = layers_27_self_attn_o_proj_loraA_weight_to_fp16, x = input_817_cast_fp16)[name = tensor("input_819_cast_fp16")]; - tensor var_10978 = const()[name = tensor("op_10978"), val = tensor([1, 1])]; - tensor var_10980 = const()[name = tensor("op_10980"), val = tensor([1, 1])]; - tensor lora_out_1093_pad_type_0 = const()[name = tensor("lora_out_1093_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1093_pad_0 = const()[name = tensor("lora_out_1093_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1095_weight_0_to_fp16 = const()[name = tensor("lora_out_1095_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524725760)))]; - tensor lora_out_1095_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10980, groups = var_10821, pad = lora_out_1093_pad_0, pad_type = lora_out_1093_pad_type_0, strides = var_10978, weight = lora_out_1095_weight_0_to_fp16, x = input_819_cast_fp16)[name = tensor("lora_out_1095_cast_fp16")]; - tensor obj_331_cast_fp16 = add(x = pretrained_out_547_cast_fp16, y = lora_out_1095_cast_fp16)[name = tensor("obj_331_cast_fp16")]; - tensor inputs_165_cast_fp16 = add(x = inputs_163_cast_fp16, y = obj_331_cast_fp16)[name = tensor("inputs_165_cast_fp16")]; - tensor var_10993 = const()[name = tensor("op_10993"), val = tensor([1])]; - tensor channels_mean_165_cast_fp16 = reduce_mean(axes = var_10993, keep_dims = var_10822, 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_10997 = const()[name = tensor("op_10997"), val = tensor([1])]; - tensor var_10998_cast_fp16 = reduce_mean(axes = var_10997, keep_dims = var_10822, x = zero_mean_sq_165_cast_fp16)[name = tensor("op_10998_cast_fp16")]; - tensor var_10999_to_fp16 = const()[name = tensor("op_10999_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_11000_cast_fp16 = add(x = var_10998_cast_fp16, y = var_10999_to_fp16)[name = tensor("op_11000_cast_fp16")]; - tensor denom_165_epsilon_0 = const()[name = tensor("denom_165_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_165_cast_fp16 = rsqrt(epsilon = denom_165_epsilon_0, x = var_11000_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_333_gamma_0_to_fp16 = const()[name = tensor("obj_333_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524766784)))]; - tensor obj_333_beta_0_to_fp16 = const()[name = tensor("obj_333_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524769408)))]; - tensor obj_333_epsilon_0_to_fp16 = const()[name = tensor("obj_333_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_333_cast_fp16 = batch_norm(beta = obj_333_beta_0_to_fp16, epsilon = obj_333_epsilon_0_to_fp16, gamma = obj_333_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_333_cast_fp16")]; - tensor var_11018 = const()[name = tensor("op_11018"), val = tensor([1, 1])]; - tensor var_11020 = const()[name = tensor("op_11020"), val = tensor([1, 1])]; - tensor pretrained_out_549_pad_type_0 = const()[name = tensor("pretrained_out_549_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_549_pad_0 = const()[name = tensor("pretrained_out_549_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524772032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525591296))), name = tensor("layers_27_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_27_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_27_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525591424)))]; - tensor pretrained_out_549_cast_fp16 = conv(bias = layers_27_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_11020, groups = var_10821, pad = pretrained_out_549_pad_0, pad_type = pretrained_out_549_pad_type_0, strides = var_11018, weight = layers_27_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_333_cast_fp16)[name = tensor("pretrained_out_549_cast_fp16")]; - tensor var_11024 = const()[name = tensor("op_11024"), val = tensor([1, 1])]; - tensor var_11026 = const()[name = tensor("op_11026"), val = tensor([1, 1])]; - tensor input_821_pad_type_0 = const()[name = tensor("input_821_pad_type_0"), val = tensor("custom")]; - tensor input_821_pad_0 = const()[name = tensor("input_821_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_27_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525594048)))]; - tensor input_821_cast_fp16 = conv(dilations = var_11026, groups = var_10821, pad = input_821_pad_0, pad_type = input_821_pad_type_0, strides = var_11024, weight = layers_27_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_333_cast_fp16)[name = tensor("input_821_cast_fp16")]; - tensor var_11030 = const()[name = tensor("op_11030"), val = tensor([1, 1])]; - tensor var_11032 = const()[name = tensor("op_11032"), val = tensor([1, 1])]; - tensor lora_out_1097_pad_type_0 = const()[name = tensor("lora_out_1097_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1097_pad_0 = const()[name = tensor("lora_out_1097_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1099_weight_0_to_fp16 = const()[name = tensor("lora_out_1099_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525635072)))]; - tensor lora_out_1099_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11032, groups = var_10821, pad = lora_out_1097_pad_0, pad_type = lora_out_1097_pad_type_0, strides = var_11030, weight = lora_out_1099_weight_0_to_fp16, x = input_821_cast_fp16)[name = tensor("lora_out_1099_cast_fp16")]; - tensor query_111_cast_fp16 = add(x = pretrained_out_549_cast_fp16, y = lora_out_1099_cast_fp16)[name = tensor("query_111_cast_fp16")]; - tensor var_11042 = const()[name = tensor("op_11042"), val = tensor([1, 1])]; - tensor var_11044 = const()[name = tensor("op_11044"), val = tensor([1, 1])]; - tensor pretrained_out_551_pad_type_0 = const()[name = tensor("pretrained_out_551_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_551_pad_0 = const()[name = tensor("pretrained_out_551_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525676096))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526495360))), name = tensor("layers_27_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_551_cast_fp16 = conv(dilations = var_11044, groups = var_10821, pad = pretrained_out_551_pad_0, pad_type = pretrained_out_551_pad_type_0, strides = var_11042, weight = layers_27_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_551_cast_fp16")]; - tensor var_11048 = const()[name = tensor("op_11048"), val = tensor([1, 1])]; - tensor var_11050 = const()[name = tensor("op_11050"), val = tensor([1, 1])]; - tensor input_823_pad_type_0 = const()[name = tensor("input_823_pad_type_0"), val = tensor("custom")]; - tensor input_823_pad_0 = const()[name = tensor("input_823_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_27_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526495488)))]; - tensor input_823_cast_fp16 = conv(dilations = var_11050, groups = var_10821, pad = input_823_pad_0, pad_type = input_823_pad_type_0, strides = var_11048, weight = layers_27_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_823_cast_fp16")]; - tensor var_11054 = const()[name = tensor("op_11054"), val = tensor([1, 1])]; - tensor var_11056 = const()[name = tensor("op_11056"), val = tensor([1, 1])]; - tensor lora_out_1101_pad_type_0 = const()[name = tensor("lora_out_1101_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1101_pad_0 = const()[name = tensor("lora_out_1101_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1103_weight_0_to_fp16 = const()[name = tensor("lora_out_1103_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526536512)))]; - tensor lora_out_1103_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11056, groups = var_10821, pad = lora_out_1101_pad_0, pad_type = lora_out_1101_pad_type_0, strides = var_11054, weight = lora_out_1103_weight_0_to_fp16, x = input_823_cast_fp16)[name = tensor("lora_out_1103_cast_fp16")]; - tensor key_111_cast_fp16 = add(x = pretrained_out_551_cast_fp16, y = lora_out_1103_cast_fp16)[name = tensor("key_111_cast_fp16")]; - tensor var_11067 = const()[name = tensor("op_11067"), val = tensor([1, 1])]; - tensor var_11069 = const()[name = tensor("op_11069"), val = tensor([1, 1])]; - tensor pretrained_out_553_pad_type_0 = const()[name = tensor("pretrained_out_553_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_553_pad_0 = const()[name = tensor("pretrained_out_553_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526577536))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527396800))), name = tensor("layers_27_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_27_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_27_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527396928)))]; - tensor pretrained_out_553_cast_fp16 = conv(bias = layers_27_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_11069, groups = var_10821, pad = pretrained_out_553_pad_0, pad_type = pretrained_out_553_pad_type_0, strides = var_11067, weight = layers_27_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_553_cast_fp16")]; - tensor var_11073 = const()[name = tensor("op_11073"), val = tensor([1, 1])]; - tensor var_11075 = const()[name = tensor("op_11075"), val = tensor([1, 1])]; - tensor input_825_pad_type_0 = const()[name = tensor("input_825_pad_type_0"), val = tensor("custom")]; - tensor input_825_pad_0 = const()[name = tensor("input_825_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_27_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527399552)))]; - tensor input_825_cast_fp16 = conv(dilations = var_11075, groups = var_10821, pad = input_825_pad_0, pad_type = input_825_pad_type_0, strides = var_11073, weight = layers_27_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_825_cast_fp16")]; - tensor var_11079 = const()[name = tensor("op_11079"), val = tensor([1, 1])]; - tensor var_11081 = const()[name = tensor("op_11081"), val = tensor([1, 1])]; - tensor lora_out_1105_pad_type_0 = const()[name = tensor("lora_out_1105_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1105_pad_0 = const()[name = tensor("lora_out_1105_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1107_weight_0_to_fp16 = const()[name = tensor("lora_out_1107_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527440576)))]; - tensor lora_out_1107_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11081, groups = var_10821, pad = lora_out_1105_pad_0, pad_type = lora_out_1105_pad_type_0, strides = var_11079, weight = lora_out_1107_weight_0_to_fp16, x = input_825_cast_fp16)[name = tensor("lora_out_1107_cast_fp16")]; - tensor value_111_cast_fp16 = add(x = pretrained_out_553_cast_fp16, y = lora_out_1107_cast_fp16)[name = tensor("value_111_cast_fp16")]; - tensor var_11088 = const()[name = tensor("op_11088"), val = tensor([1, 20, 64, -1])]; - tensor var_11089_cast_fp16 = reshape(shape = var_11088, x = query_111_cast_fp16)[name = tensor("op_11089_cast_fp16")]; - tensor var_11090_to_fp16 = const()[name = tensor("op_11090_to_fp16"), val = tensor(0x1p-3)]; - tensor var_11091_cast_fp16 = mul(x = var_11089_cast_fp16, y = var_11090_to_fp16)[name = tensor("op_11091_cast_fp16")]; - tensor var_11092 = const()[name = tensor("op_11092"), val = tensor([1, 20, 64, -1])]; - tensor var_11093_cast_fp16 = reshape(shape = var_11092, x = key_111_cast_fp16)[name = tensor("op_11093_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_11091_cast_fp16, y = var_11093_cast_fp16)[name = tensor("mh_w_167_cast_fp16")]; - tensor var_11096_cast_fp16 = softmax(axis = var_10814, x = mh_w_167_cast_fp16)[name = tensor("op_11096_cast_fp16")]; - tensor var_11097 = const()[name = tensor("op_11097"), val = tensor([1, 20, 64, -1])]; - tensor var_11098_cast_fp16 = reshape(shape = var_11097, x = value_111_cast_fp16)[name = tensor("op_11098_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_11098_cast_fp16, y = var_11096_cast_fp16)[name = tensor("attn_111_cast_fp16")]; - tensor var_11101 = const()[name = tensor("op_11101"), val = tensor([1, 1280, 1, -1])]; - tensor input_827_cast_fp16 = reshape(shape = var_11101, x = attn_111_cast_fp16)[name = tensor("input_827_cast_fp16")]; - tensor var_11108 = const()[name = tensor("op_11108"), val = tensor([1, 1])]; - tensor var_11110 = const()[name = tensor("op_11110"), val = tensor([1, 1])]; - tensor pretrained_out_555_pad_type_0 = const()[name = tensor("pretrained_out_555_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_555_pad_0 = const()[name = tensor("pretrained_out_555_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527481600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528300864))), name = tensor("layers_27_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_27_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_27_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528300992)))]; - tensor pretrained_out_555_cast_fp16 = conv(bias = layers_27_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_11110, groups = var_10821, pad = pretrained_out_555_pad_0, pad_type = pretrained_out_555_pad_type_0, strides = var_11108, weight = layers_27_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_827_cast_fp16)[name = tensor("pretrained_out_555_cast_fp16")]; - tensor var_11114 = const()[name = tensor("op_11114"), val = tensor([1, 1])]; - tensor var_11116 = const()[name = tensor("op_11116"), val = tensor([1, 1])]; - tensor input_829_pad_type_0 = const()[name = tensor("input_829_pad_type_0"), val = tensor("custom")]; - tensor input_829_pad_0 = const()[name = tensor("input_829_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_27_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528303616)))]; - tensor input_829_cast_fp16 = conv(dilations = var_11116, groups = var_10821, pad = input_829_pad_0, pad_type = input_829_pad_type_0, strides = var_11114, weight = layers_27_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_827_cast_fp16)[name = tensor("input_829_cast_fp16")]; - tensor var_11120 = const()[name = tensor("op_11120"), val = tensor([1, 1])]; - tensor var_11122 = const()[name = tensor("op_11122"), val = tensor([1, 1])]; - tensor lora_out_1109_pad_type_0 = const()[name = tensor("lora_out_1109_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1109_pad_0 = const()[name = tensor("lora_out_1109_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1111_weight_0_to_fp16 = const()[name = tensor("lora_out_1111_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528344640)))]; - tensor lora_out_1111_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11122, groups = var_10821, pad = lora_out_1109_pad_0, pad_type = lora_out_1109_pad_type_0, strides = var_11120, weight = lora_out_1111_weight_0_to_fp16, x = input_829_cast_fp16)[name = tensor("lora_out_1111_cast_fp16")]; - tensor obj_335_cast_fp16 = add(x = pretrained_out_555_cast_fp16, y = lora_out_1111_cast_fp16)[name = tensor("obj_335_cast_fp16")]; - tensor inputs_167_cast_fp16 = add(x = inputs_165_cast_fp16, y = obj_335_cast_fp16)[name = tensor("inputs_167_cast_fp16")]; - tensor var_11131 = const()[name = tensor("op_11131"), val = tensor([1])]; - tensor channels_mean_167_cast_fp16 = reduce_mean(axes = var_11131, keep_dims = var_10822, 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_11135 = const()[name = tensor("op_11135"), val = tensor([1])]; - tensor var_11136_cast_fp16 = reduce_mean(axes = var_11135, keep_dims = var_10822, x = zero_mean_sq_167_cast_fp16)[name = tensor("op_11136_cast_fp16")]; - tensor var_11137_to_fp16 = const()[name = tensor("op_11137_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_11138_cast_fp16 = add(x = var_11136_cast_fp16, y = var_11137_to_fp16)[name = tensor("op_11138_cast_fp16")]; - tensor denom_167_epsilon_0 = const()[name = tensor("denom_167_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_167_cast_fp16 = rsqrt(epsilon = denom_167_epsilon_0, x = var_11138_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_831_gamma_0_to_fp16 = const()[name = tensor("input_831_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528385664)))]; - tensor input_831_beta_0_to_fp16 = const()[name = tensor("input_831_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528388288)))]; - tensor input_831_epsilon_0_to_fp16 = const()[name = tensor("input_831_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_831_cast_fp16 = batch_norm(beta = input_831_beta_0_to_fp16, epsilon = input_831_epsilon_0_to_fp16, gamma = input_831_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_831_cast_fp16")]; - tensor var_11152 = const()[name = tensor("op_11152"), val = tensor([1, 1])]; - tensor var_11154 = const()[name = tensor("op_11154"), val = tensor([1, 1])]; - tensor pretrained_out_557_pad_type_0 = const()[name = tensor("pretrained_out_557_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_557_pad_0 = const()[name = tensor("pretrained_out_557_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528390912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(531667776))), name = tensor("layers_27_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_27_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_27_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(531667904)))]; - tensor pretrained_out_557_cast_fp16 = conv(bias = layers_27_fc1_pretrained_bias_to_fp16, dilations = var_11154, groups = var_10821, pad = pretrained_out_557_pad_0, pad_type = pretrained_out_557_pad_type_0, strides = var_11152, weight = layers_27_fc1_pretrained_weight_to_fp16_palettized, x = input_831_cast_fp16)[name = tensor("pretrained_out_557_cast_fp16")]; - tensor var_11158 = const()[name = tensor("op_11158"), val = tensor([1, 1])]; - tensor var_11160 = const()[name = tensor("op_11160"), val = tensor([1, 1])]; - tensor input_833_pad_type_0 = const()[name = tensor("input_833_pad_type_0"), val = tensor("custom")]; - tensor input_833_pad_0 = const()[name = tensor("input_833_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_27_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(531678208)))]; - tensor input_833_cast_fp16 = conv(dilations = var_11160, groups = var_10821, pad = input_833_pad_0, pad_type = input_833_pad_type_0, strides = var_11158, weight = layers_27_fc1_loraA_weight_to_fp16, x = input_831_cast_fp16)[name = tensor("input_833_cast_fp16")]; - tensor var_11164 = const()[name = tensor("op_11164"), val = tensor([1, 1])]; - tensor var_11166 = const()[name = tensor("op_11166"), val = tensor([1, 1])]; - tensor lora_out_1113_pad_type_0 = const()[name = tensor("lora_out_1113_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1113_pad_0 = const()[name = tensor("lora_out_1113_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1115_weight_0_to_fp16 = const()[name = tensor("lora_out_1115_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(531719232)))]; - tensor lora_out_1115_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_11166, groups = var_10821, pad = lora_out_1113_pad_0, pad_type = lora_out_1113_pad_type_0, strides = var_11164, weight = lora_out_1115_weight_0_to_fp16, x = input_833_cast_fp16)[name = tensor("lora_out_1115_cast_fp16")]; - tensor input_835_cast_fp16 = add(x = pretrained_out_557_cast_fp16, y = lora_out_1115_cast_fp16)[name = tensor("input_835_cast_fp16")]; - tensor input_837_mode_0 = const()[name = tensor("input_837_mode_0"), val = tensor("EXACT")]; - tensor input_837_cast_fp16 = gelu(mode = input_837_mode_0, x = input_835_cast_fp16)[name = tensor("input_837_cast_fp16")]; - tensor var_11178 = const()[name = tensor("op_11178"), val = tensor([1, 1])]; - tensor var_11180 = const()[name = tensor("op_11180"), val = tensor([1, 1])]; - tensor pretrained_out_559_pad_type_0 = const()[name = tensor("pretrained_out_559_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_559_pad_0 = const()[name = tensor("pretrained_out_559_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(531883136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535160000))), name = tensor("layers_27_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_27_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_27_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535160128)))]; - tensor pretrained_out_559_cast_fp16 = conv(bias = layers_27_fc2_pretrained_bias_to_fp16, dilations = var_11180, groups = var_10821, pad = pretrained_out_559_pad_0, pad_type = pretrained_out_559_pad_type_0, strides = var_11178, weight = layers_27_fc2_pretrained_weight_to_fp16_palettized, x = input_837_cast_fp16)[name = tensor("pretrained_out_559_cast_fp16")]; - tensor var_11184 = const()[name = tensor("op_11184"), val = tensor([1, 1])]; - tensor var_11186 = const()[name = tensor("op_11186"), val = tensor([1, 1])]; - tensor input_839_pad_type_0 = const()[name = tensor("input_839_pad_type_0"), val = tensor("custom")]; - tensor input_839_pad_0 = const()[name = tensor("input_839_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_27_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_27_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535162752)))]; - tensor input_839_cast_fp16 = conv(dilations = var_11186, groups = var_10821, pad = input_839_pad_0, pad_type = input_839_pad_type_0, strides = var_11184, weight = layers_27_fc2_loraA_weight_to_fp16, x = input_837_cast_fp16)[name = tensor("input_839_cast_fp16")]; - tensor var_11190 = const()[name = tensor("op_11190"), val = tensor([1, 1])]; - tensor var_11192 = const()[name = tensor("op_11192"), val = tensor([1, 1])]; - tensor lora_out_1117_pad_type_0 = const()[name = tensor("lora_out_1117_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1117_pad_0 = const()[name = tensor("lora_out_1117_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1119_weight_0_to_fp16 = const()[name = tensor("lora_out_1119_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535326656)))]; - tensor lora_out_1119_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11192, groups = var_10821, pad = lora_out_1117_pad_0, pad_type = lora_out_1117_pad_type_0, strides = var_11190, weight = lora_out_1119_weight_0_to_fp16, x = input_839_cast_fp16)[name = tensor("lora_out_1119_cast_fp16")]; - tensor hidden_states_57_cast_fp16 = add(x = pretrained_out_559_cast_fp16, y = lora_out_1119_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_11208 = const()[name = tensor("op_11208"), val = tensor(3)]; - tensor var_11215 = const()[name = tensor("op_11215"), val = tensor(1)]; - tensor var_11216 = const()[name = tensor("op_11216"), val = tensor(true)]; - tensor var_11228 = const()[name = tensor("op_11228"), val = tensor([1])]; - tensor channels_mean_169_cast_fp16 = reduce_mean(axes = var_11228, keep_dims = var_11216, 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_11232 = const()[name = tensor("op_11232"), val = tensor([1])]; - tensor var_11233_cast_fp16 = reduce_mean(axes = var_11232, keep_dims = var_11216, x = zero_mean_sq_169_cast_fp16)[name = tensor("op_11233_cast_fp16")]; - tensor var_11234_to_fp16 = const()[name = tensor("op_11234_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_11235_cast_fp16 = add(x = var_11233_cast_fp16, y = var_11234_to_fp16)[name = tensor("op_11235_cast_fp16")]; - tensor denom_169_epsilon_0 = const()[name = tensor("denom_169_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_169_cast_fp16 = rsqrt(epsilon = denom_169_epsilon_0, x = var_11235_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_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(535367680)))]; - 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(535370304)))]; - 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_169_cast_fp16)[name = tensor("obj_337_cast_fp16")]; - tensor var_11253 = const()[name = tensor("op_11253"), val = tensor([1, 1])]; - tensor var_11255 = const()[name = tensor("op_11255"), val = tensor([1, 1])]; - tensor pretrained_out_561_pad_type_0 = const()[name = tensor("pretrained_out_561_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_561_pad_0 = const()[name = tensor("pretrained_out_561_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535372928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536192192))), name = tensor("layers_28_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_28_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536192320)))]; - tensor pretrained_out_561_cast_fp16 = conv(bias = layers_28_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_11255, groups = var_11215, pad = pretrained_out_561_pad_0, pad_type = pretrained_out_561_pad_type_0, strides = var_11253, weight = layers_28_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_337_cast_fp16)[name = tensor("pretrained_out_561_cast_fp16")]; - tensor var_11259 = const()[name = tensor("op_11259"), val = tensor([1, 1])]; - tensor var_11261 = const()[name = tensor("op_11261"), val = tensor([1, 1])]; - tensor input_841_pad_type_0 = const()[name = tensor("input_841_pad_type_0"), val = tensor("custom")]; - tensor input_841_pad_0 = const()[name = tensor("input_841_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536194944)))]; - tensor input_841_cast_fp16 = conv(dilations = var_11261, groups = var_11215, pad = input_841_pad_0, pad_type = input_841_pad_type_0, strides = var_11259, weight = layers_28_self_attn_q_proj_loraA_weight_to_fp16, x = obj_337_cast_fp16)[name = tensor("input_841_cast_fp16")]; - tensor var_11265 = const()[name = tensor("op_11265"), val = tensor([1, 1])]; - tensor var_11267 = const()[name = tensor("op_11267"), val = tensor([1, 1])]; - tensor lora_out_1121_pad_type_0 = const()[name = tensor("lora_out_1121_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1121_pad_0 = const()[name = tensor("lora_out_1121_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1123_weight_0_to_fp16 = const()[name = tensor("lora_out_1123_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536235968)))]; - tensor lora_out_1123_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11267, groups = var_11215, pad = lora_out_1121_pad_0, pad_type = lora_out_1121_pad_type_0, strides = var_11265, weight = lora_out_1123_weight_0_to_fp16, x = input_841_cast_fp16)[name = tensor("lora_out_1123_cast_fp16")]; - tensor query_113_cast_fp16 = add(x = pretrained_out_561_cast_fp16, y = lora_out_1123_cast_fp16)[name = tensor("query_113_cast_fp16")]; - tensor var_11277 = const()[name = tensor("op_11277"), val = tensor([1, 1])]; - tensor var_11279 = const()[name = tensor("op_11279"), val = tensor([1, 1])]; - tensor pretrained_out_563_pad_type_0 = const()[name = tensor("pretrained_out_563_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_563_pad_0 = const()[name = tensor("pretrained_out_563_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536276992))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537096256))), name = tensor("layers_28_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_563_cast_fp16 = conv(dilations = var_11279, groups = var_11215, pad = pretrained_out_563_pad_0, pad_type = pretrained_out_563_pad_type_0, strides = var_11277, weight = layers_28_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_337_cast_fp16)[name = tensor("pretrained_out_563_cast_fp16")]; - tensor var_11283 = const()[name = tensor("op_11283"), val = tensor([1, 1])]; - tensor var_11285 = const()[name = tensor("op_11285"), val = tensor([1, 1])]; - tensor input_843_pad_type_0 = const()[name = tensor("input_843_pad_type_0"), val = tensor("custom")]; - tensor input_843_pad_0 = const()[name = tensor("input_843_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537096384)))]; - tensor input_843_cast_fp16 = conv(dilations = var_11285, groups = var_11215, pad = input_843_pad_0, pad_type = input_843_pad_type_0, strides = var_11283, weight = layers_28_self_attn_k_proj_loraA_weight_to_fp16, x = obj_337_cast_fp16)[name = tensor("input_843_cast_fp16")]; - tensor var_11289 = const()[name = tensor("op_11289"), val = tensor([1, 1])]; - tensor var_11291 = const()[name = tensor("op_11291"), val = tensor([1, 1])]; - tensor lora_out_1125_pad_type_0 = const()[name = tensor("lora_out_1125_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1125_pad_0 = const()[name = tensor("lora_out_1125_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1127_weight_0_to_fp16 = const()[name = tensor("lora_out_1127_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537137408)))]; - tensor lora_out_1127_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11291, groups = var_11215, pad = lora_out_1125_pad_0, pad_type = lora_out_1125_pad_type_0, strides = var_11289, weight = lora_out_1127_weight_0_to_fp16, x = input_843_cast_fp16)[name = tensor("lora_out_1127_cast_fp16")]; - tensor current_key_57_cast_fp16 = add(x = pretrained_out_563_cast_fp16, y = lora_out_1127_cast_fp16)[name = tensor("current_key_57_cast_fp16")]; - tensor var_11302 = const()[name = tensor("op_11302"), val = tensor([1, 1])]; - tensor var_11304 = const()[name = tensor("op_11304"), val = tensor([1, 1])]; - tensor pretrained_out_565_pad_type_0 = const()[name = tensor("pretrained_out_565_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_565_pad_0 = const()[name = tensor("pretrained_out_565_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537178432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537997696))), name = tensor("layers_28_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_28_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537997824)))]; - tensor pretrained_out_565_cast_fp16 = conv(bias = layers_28_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_11304, groups = var_11215, pad = pretrained_out_565_pad_0, pad_type = pretrained_out_565_pad_type_0, strides = var_11302, weight = layers_28_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_337_cast_fp16)[name = tensor("pretrained_out_565_cast_fp16")]; - tensor var_11308 = const()[name = tensor("op_11308"), val = tensor([1, 1])]; - tensor var_11310 = const()[name = tensor("op_11310"), val = tensor([1, 1])]; - tensor input_845_pad_type_0 = const()[name = tensor("input_845_pad_type_0"), val = tensor("custom")]; - tensor input_845_pad_0 = const()[name = tensor("input_845_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538000448)))]; - tensor input_845_cast_fp16 = conv(dilations = var_11310, groups = var_11215, pad = input_845_pad_0, pad_type = input_845_pad_type_0, strides = var_11308, weight = layers_28_self_attn_v_proj_loraA_weight_to_fp16, x = obj_337_cast_fp16)[name = tensor("input_845_cast_fp16")]; - tensor var_11314 = const()[name = tensor("op_11314"), val = tensor([1, 1])]; - tensor var_11316 = const()[name = tensor("op_11316"), val = tensor([1, 1])]; - tensor lora_out_1129_pad_type_0 = const()[name = tensor("lora_out_1129_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1129_pad_0 = const()[name = tensor("lora_out_1129_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1131_weight_0_to_fp16 = const()[name = tensor("lora_out_1131_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538041472)))]; - tensor lora_out_1131_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11316, groups = var_11215, pad = lora_out_1129_pad_0, pad_type = lora_out_1129_pad_type_0, strides = var_11314, weight = lora_out_1131_weight_0_to_fp16, x = input_845_cast_fp16)[name = tensor("lora_out_1131_cast_fp16")]; - tensor current_value_57_cast_fp16 = add(x = pretrained_out_565_cast_fp16, y = lora_out_1131_cast_fp16)[name = tensor("current_value_57_cast_fp16")]; - tensor var_11326_cast_fp16 = mul(x = current_key_57_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_11326_cast_fp16")]; - tensor var_11328_cast_fp16 = mul(x = var_103_cast_fp16_28, y = var_295_cast_fp16)[name = tensor("op_11328_cast_fp16")]; - tensor key_113_cast_fp16 = add(x = var_11326_cast_fp16, y = var_11328_cast_fp16)[name = tensor("key_113_cast_fp16")]; - tensor var_11330_cast_fp16 = mul(x = current_value_57_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_11330_cast_fp16")]; - tensor var_11332_cast_fp16 = mul(x = var_138_cast_fp16_28, y = var_295_cast_fp16)[name = tensor("op_11332_cast_fp16")]; - tensor value_113_cast_fp16 = add(x = var_11330_cast_fp16, y = var_11332_cast_fp16)[name = tensor("value_113_cast_fp16")]; - tensor var_11335 = const()[name = tensor("op_11335"), val = tensor([1, 20, 64, -1])]; - tensor var_11336_cast_fp16 = reshape(shape = var_11335, x = query_113_cast_fp16)[name = tensor("op_11336_cast_fp16")]; - tensor var_11337_to_fp16 = const()[name = tensor("op_11337_to_fp16"), val = tensor(0x1p-3)]; - tensor var_11338_cast_fp16 = mul(x = var_11336_cast_fp16, y = var_11337_to_fp16)[name = tensor("op_11338_cast_fp16")]; - tensor var_11339 = const()[name = tensor("op_11339"), val = tensor([1, 20, 64, -1])]; - tensor var_11340_cast_fp16 = reshape(shape = var_11339, x = key_113_cast_fp16)[name = tensor("op_11340_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_11338_cast_fp16, y = var_11340_cast_fp16)[name = tensor("mh_w_169_cast_fp16")]; - tensor mh_w_171_cast_fp16 = add(x = mh_w_169_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_171_cast_fp16")]; - tensor var_11348_cast_fp16 = softmax(axis = var_11208, x = mh_w_171_cast_fp16)[name = tensor("op_11348_cast_fp16")]; - tensor var_11349 = const()[name = tensor("op_11349"), val = tensor([1, 20, 64, -1])]; - tensor var_11350_cast_fp16 = reshape(shape = var_11349, x = value_113_cast_fp16)[name = tensor("op_11350_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_11350_cast_fp16, y = var_11348_cast_fp16)[name = tensor("attn_113_cast_fp16")]; - tensor var_11353 = const()[name = tensor("op_11353"), val = tensor([1, 1280, 1, -1])]; - tensor input_847_cast_fp16 = reshape(shape = var_11353, x = attn_113_cast_fp16)[name = tensor("input_847_cast_fp16")]; - tensor var_11360 = const()[name = tensor("op_11360"), val = tensor([1, 1])]; - tensor var_11362 = const()[name = tensor("op_11362"), val = tensor([1, 1])]; - tensor pretrained_out_567_pad_type_0 = const()[name = tensor("pretrained_out_567_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_567_pad_0 = const()[name = tensor("pretrained_out_567_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538082496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538901760))), name = tensor("layers_28_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_28_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538901888)))]; - tensor pretrained_out_567_cast_fp16 = conv(bias = layers_28_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_11362, groups = var_11215, pad = pretrained_out_567_pad_0, pad_type = pretrained_out_567_pad_type_0, strides = var_11360, weight = layers_28_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_847_cast_fp16)[name = tensor("pretrained_out_567_cast_fp16")]; - tensor var_11366 = const()[name = tensor("op_11366"), val = tensor([1, 1])]; - tensor var_11368 = const()[name = tensor("op_11368"), val = tensor([1, 1])]; - tensor input_849_pad_type_0 = const()[name = tensor("input_849_pad_type_0"), val = tensor("custom")]; - tensor input_849_pad_0 = const()[name = tensor("input_849_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538904512)))]; - tensor input_849_cast_fp16 = conv(dilations = var_11368, groups = var_11215, pad = input_849_pad_0, pad_type = input_849_pad_type_0, strides = var_11366, weight = layers_28_self_attn_o_proj_loraA_weight_to_fp16, x = input_847_cast_fp16)[name = tensor("input_849_cast_fp16")]; - tensor var_11372 = const()[name = tensor("op_11372"), val = tensor([1, 1])]; - tensor var_11374 = const()[name = tensor("op_11374"), val = tensor([1, 1])]; - tensor lora_out_1133_pad_type_0 = const()[name = tensor("lora_out_1133_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1133_pad_0 = const()[name = tensor("lora_out_1133_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1135_weight_0_to_fp16 = const()[name = tensor("lora_out_1135_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538945536)))]; - tensor lora_out_1135_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11374, groups = var_11215, pad = lora_out_1133_pad_0, pad_type = lora_out_1133_pad_type_0, strides = var_11372, weight = lora_out_1135_weight_0_to_fp16, x = input_849_cast_fp16)[name = tensor("lora_out_1135_cast_fp16")]; - tensor obj_343_cast_fp16 = add(x = pretrained_out_567_cast_fp16, y = lora_out_1135_cast_fp16)[name = tensor("obj_343_cast_fp16")]; - tensor inputs_171_cast_fp16 = add(x = inputs_169_cast_fp16, y = obj_343_cast_fp16)[name = tensor("inputs_171_cast_fp16")]; - tensor var_11387 = const()[name = tensor("op_11387"), val = tensor([1])]; - tensor channels_mean_171_cast_fp16 = reduce_mean(axes = var_11387, keep_dims = var_11216, 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_11391 = const()[name = tensor("op_11391"), val = tensor([1])]; - tensor var_11392_cast_fp16 = reduce_mean(axes = var_11391, keep_dims = var_11216, x = zero_mean_sq_171_cast_fp16)[name = tensor("op_11392_cast_fp16")]; - tensor var_11393_to_fp16 = const()[name = tensor("op_11393_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_11394_cast_fp16 = add(x = var_11392_cast_fp16, y = var_11393_to_fp16)[name = tensor("op_11394_cast_fp16")]; - tensor denom_171_epsilon_0 = const()[name = tensor("denom_171_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_171_cast_fp16 = rsqrt(epsilon = denom_171_epsilon_0, x = var_11394_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_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(538986560)))]; - 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(538989184)))]; - 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_171_cast_fp16)[name = tensor("obj_345_cast_fp16")]; - tensor var_11412 = const()[name = tensor("op_11412"), val = tensor([1, 1])]; - tensor var_11414 = const()[name = tensor("op_11414"), val = tensor([1, 1])]; - tensor pretrained_out_569_pad_type_0 = const()[name = tensor("pretrained_out_569_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_569_pad_0 = const()[name = tensor("pretrained_out_569_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538991808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539811072))), name = tensor("layers_28_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_28_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_28_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539811200)))]; - tensor pretrained_out_569_cast_fp16 = conv(bias = layers_28_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_11414, groups = var_11215, pad = pretrained_out_569_pad_0, pad_type = pretrained_out_569_pad_type_0, strides = var_11412, weight = layers_28_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_345_cast_fp16)[name = tensor("pretrained_out_569_cast_fp16")]; - tensor var_11418 = const()[name = tensor("op_11418"), val = tensor([1, 1])]; - tensor var_11420 = const()[name = tensor("op_11420"), val = tensor([1, 1])]; - tensor input_851_pad_type_0 = const()[name = tensor("input_851_pad_type_0"), val = tensor("custom")]; - tensor input_851_pad_0 = const()[name = tensor("input_851_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_28_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539813824)))]; - tensor input_851_cast_fp16 = conv(dilations = var_11420, groups = var_11215, pad = input_851_pad_0, pad_type = input_851_pad_type_0, strides = var_11418, weight = layers_28_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_345_cast_fp16)[name = tensor("input_851_cast_fp16")]; - tensor var_11424 = const()[name = tensor("op_11424"), val = tensor([1, 1])]; - tensor var_11426 = const()[name = tensor("op_11426"), val = tensor([1, 1])]; - tensor lora_out_1137_pad_type_0 = const()[name = tensor("lora_out_1137_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1137_pad_0 = const()[name = tensor("lora_out_1137_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1139_weight_0_to_fp16 = const()[name = tensor("lora_out_1139_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539854848)))]; - tensor lora_out_1139_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11426, groups = var_11215, pad = lora_out_1137_pad_0, pad_type = lora_out_1137_pad_type_0, strides = var_11424, weight = lora_out_1139_weight_0_to_fp16, x = input_851_cast_fp16)[name = tensor("lora_out_1139_cast_fp16")]; - tensor query_115_cast_fp16 = add(x = pretrained_out_569_cast_fp16, y = lora_out_1139_cast_fp16)[name = tensor("query_115_cast_fp16")]; - tensor var_11436 = const()[name = tensor("op_11436"), val = tensor([1, 1])]; - tensor var_11438 = const()[name = tensor("op_11438"), val = tensor([1, 1])]; - tensor pretrained_out_571_pad_type_0 = const()[name = tensor("pretrained_out_571_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_571_pad_0 = const()[name = tensor("pretrained_out_571_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539895872))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540715136))), name = tensor("layers_28_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_571_cast_fp16 = conv(dilations = var_11438, groups = var_11215, pad = pretrained_out_571_pad_0, pad_type = pretrained_out_571_pad_type_0, strides = var_11436, weight = layers_28_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_571_cast_fp16")]; - tensor var_11442 = const()[name = tensor("op_11442"), val = tensor([1, 1])]; - tensor var_11444 = const()[name = tensor("op_11444"), val = tensor([1, 1])]; - tensor input_853_pad_type_0 = const()[name = tensor("input_853_pad_type_0"), val = tensor("custom")]; - tensor input_853_pad_0 = const()[name = tensor("input_853_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_28_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540715264)))]; - tensor input_853_cast_fp16 = conv(dilations = var_11444, groups = var_11215, pad = input_853_pad_0, pad_type = input_853_pad_type_0, strides = var_11442, weight = layers_28_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_853_cast_fp16")]; - tensor var_11448 = const()[name = tensor("op_11448"), val = tensor([1, 1])]; - tensor var_11450 = const()[name = tensor("op_11450"), val = tensor([1, 1])]; - tensor lora_out_1141_pad_type_0 = const()[name = tensor("lora_out_1141_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1141_pad_0 = const()[name = tensor("lora_out_1141_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1143_weight_0_to_fp16 = const()[name = tensor("lora_out_1143_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540756288)))]; - tensor lora_out_1143_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11450, groups = var_11215, pad = lora_out_1141_pad_0, pad_type = lora_out_1141_pad_type_0, strides = var_11448, weight = lora_out_1143_weight_0_to_fp16, x = input_853_cast_fp16)[name = tensor("lora_out_1143_cast_fp16")]; - tensor key_115_cast_fp16 = add(x = pretrained_out_571_cast_fp16, y = lora_out_1143_cast_fp16)[name = tensor("key_115_cast_fp16")]; - tensor var_11461 = const()[name = tensor("op_11461"), val = tensor([1, 1])]; - tensor var_11463 = const()[name = tensor("op_11463"), val = tensor([1, 1])]; - tensor pretrained_out_573_pad_type_0 = const()[name = tensor("pretrained_out_573_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_573_pad_0 = const()[name = tensor("pretrained_out_573_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540797312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(541616576))), name = tensor("layers_28_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_28_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_28_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(541616704)))]; - tensor pretrained_out_573_cast_fp16 = conv(bias = layers_28_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_11463, groups = var_11215, pad = pretrained_out_573_pad_0, pad_type = pretrained_out_573_pad_type_0, strides = var_11461, weight = layers_28_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_573_cast_fp16")]; - tensor var_11467 = const()[name = tensor("op_11467"), val = tensor([1, 1])]; - tensor var_11469 = const()[name = tensor("op_11469"), val = tensor([1, 1])]; - tensor input_855_pad_type_0 = const()[name = tensor("input_855_pad_type_0"), val = tensor("custom")]; - tensor input_855_pad_0 = const()[name = tensor("input_855_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_28_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(541619328)))]; - tensor input_855_cast_fp16 = conv(dilations = var_11469, groups = var_11215, pad = input_855_pad_0, pad_type = input_855_pad_type_0, strides = var_11467, weight = layers_28_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_855_cast_fp16")]; - tensor var_11473 = const()[name = tensor("op_11473"), val = tensor([1, 1])]; - tensor var_11475 = const()[name = tensor("op_11475"), val = tensor([1, 1])]; - tensor lora_out_1145_pad_type_0 = const()[name = tensor("lora_out_1145_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1145_pad_0 = const()[name = tensor("lora_out_1145_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1147_weight_0_to_fp16 = const()[name = tensor("lora_out_1147_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(541660352)))]; - tensor lora_out_1147_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11475, groups = var_11215, pad = lora_out_1145_pad_0, pad_type = lora_out_1145_pad_type_0, strides = var_11473, weight = lora_out_1147_weight_0_to_fp16, x = input_855_cast_fp16)[name = tensor("lora_out_1147_cast_fp16")]; - tensor value_115_cast_fp16 = add(x = pretrained_out_573_cast_fp16, y = lora_out_1147_cast_fp16)[name = tensor("value_115_cast_fp16")]; - tensor var_11482 = const()[name = tensor("op_11482"), val = tensor([1, 20, 64, -1])]; - tensor var_11483_cast_fp16 = reshape(shape = var_11482, x = query_115_cast_fp16)[name = tensor("op_11483_cast_fp16")]; - tensor var_11484_to_fp16 = const()[name = tensor("op_11484_to_fp16"), val = tensor(0x1p-3)]; - tensor var_11485_cast_fp16 = mul(x = var_11483_cast_fp16, y = var_11484_to_fp16)[name = tensor("op_11485_cast_fp16")]; - tensor var_11486 = const()[name = tensor("op_11486"), val = tensor([1, 20, 64, -1])]; - tensor var_11487_cast_fp16 = reshape(shape = var_11486, x = key_115_cast_fp16)[name = tensor("op_11487_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_11485_cast_fp16, y = var_11487_cast_fp16)[name = tensor("mh_w_173_cast_fp16")]; - tensor var_11490_cast_fp16 = softmax(axis = var_11208, x = mh_w_173_cast_fp16)[name = tensor("op_11490_cast_fp16")]; - tensor var_11491 = const()[name = tensor("op_11491"), val = tensor([1, 20, 64, -1])]; - tensor var_11492_cast_fp16 = reshape(shape = var_11491, x = value_115_cast_fp16)[name = tensor("op_11492_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_11492_cast_fp16, y = var_11490_cast_fp16)[name = tensor("attn_115_cast_fp16")]; - tensor var_11495 = const()[name = tensor("op_11495"), val = tensor([1, 1280, 1, -1])]; - tensor input_857_cast_fp16 = reshape(shape = var_11495, x = attn_115_cast_fp16)[name = tensor("input_857_cast_fp16")]; - tensor var_11502 = const()[name = tensor("op_11502"), val = tensor([1, 1])]; - tensor var_11504 = const()[name = tensor("op_11504"), val = tensor([1, 1])]; - tensor pretrained_out_575_pad_type_0 = const()[name = tensor("pretrained_out_575_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_575_pad_0 = const()[name = tensor("pretrained_out_575_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(541701376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(542520640))), name = tensor("layers_28_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_28_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_28_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(542520768)))]; - tensor pretrained_out_575_cast_fp16 = conv(bias = layers_28_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_11504, groups = var_11215, pad = pretrained_out_575_pad_0, pad_type = pretrained_out_575_pad_type_0, strides = var_11502, weight = layers_28_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = tensor("pretrained_out_575_cast_fp16")]; - tensor var_11508 = const()[name = tensor("op_11508"), val = tensor([1, 1])]; - tensor var_11510 = const()[name = tensor("op_11510"), val = tensor([1, 1])]; - tensor input_859_pad_type_0 = const()[name = tensor("input_859_pad_type_0"), val = tensor("custom")]; - tensor input_859_pad_0 = const()[name = tensor("input_859_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_28_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(542523392)))]; - tensor input_859_cast_fp16 = conv(dilations = var_11510, groups = var_11215, pad = input_859_pad_0, pad_type = input_859_pad_type_0, strides = var_11508, weight = layers_28_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_857_cast_fp16)[name = tensor("input_859_cast_fp16")]; - tensor var_11514 = const()[name = tensor("op_11514"), val = tensor([1, 1])]; - tensor var_11516 = const()[name = tensor("op_11516"), val = tensor([1, 1])]; - tensor lora_out_1149_pad_type_0 = const()[name = tensor("lora_out_1149_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1149_pad_0 = const()[name = tensor("lora_out_1149_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1151_weight_0_to_fp16 = const()[name = tensor("lora_out_1151_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(542564416)))]; - tensor lora_out_1151_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11516, groups = var_11215, pad = lora_out_1149_pad_0, pad_type = lora_out_1149_pad_type_0, strides = var_11514, weight = lora_out_1151_weight_0_to_fp16, x = input_859_cast_fp16)[name = tensor("lora_out_1151_cast_fp16")]; - tensor obj_347_cast_fp16 = add(x = pretrained_out_575_cast_fp16, y = lora_out_1151_cast_fp16)[name = tensor("obj_347_cast_fp16")]; - tensor inputs_173_cast_fp16 = add(x = inputs_171_cast_fp16, y = obj_347_cast_fp16)[name = tensor("inputs_173_cast_fp16")]; - tensor var_11525 = const()[name = tensor("op_11525"), val = tensor([1])]; - tensor channels_mean_173_cast_fp16 = reduce_mean(axes = var_11525, keep_dims = var_11216, 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_11529 = const()[name = tensor("op_11529"), val = tensor([1])]; - tensor var_11530_cast_fp16 = reduce_mean(axes = var_11529, keep_dims = var_11216, x = zero_mean_sq_173_cast_fp16)[name = tensor("op_11530_cast_fp16")]; - tensor var_11531_to_fp16 = const()[name = tensor("op_11531_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_11532_cast_fp16 = add(x = var_11530_cast_fp16, y = var_11531_to_fp16)[name = tensor("op_11532_cast_fp16")]; - tensor denom_173_epsilon_0 = const()[name = tensor("denom_173_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_173_cast_fp16 = rsqrt(epsilon = denom_173_epsilon_0, x = var_11532_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_861_gamma_0_to_fp16 = const()[name = tensor("input_861_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(542605440)))]; - tensor input_861_beta_0_to_fp16 = const()[name = tensor("input_861_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(542608064)))]; - tensor input_861_epsilon_0_to_fp16 = const()[name = tensor("input_861_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_861_cast_fp16 = batch_norm(beta = input_861_beta_0_to_fp16, epsilon = input_861_epsilon_0_to_fp16, gamma = input_861_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_861_cast_fp16")]; - tensor var_11546 = const()[name = tensor("op_11546"), val = tensor([1, 1])]; - tensor var_11548 = const()[name = tensor("op_11548"), val = tensor([1, 1])]; - tensor pretrained_out_577_pad_type_0 = const()[name = tensor("pretrained_out_577_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_577_pad_0 = const()[name = tensor("pretrained_out_577_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(542610688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545887552))), name = tensor("layers_28_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_28_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_28_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545887680)))]; - tensor pretrained_out_577_cast_fp16 = conv(bias = layers_28_fc1_pretrained_bias_to_fp16, dilations = var_11548, groups = var_11215, pad = pretrained_out_577_pad_0, pad_type = pretrained_out_577_pad_type_0, strides = var_11546, weight = layers_28_fc1_pretrained_weight_to_fp16_palettized, x = input_861_cast_fp16)[name = tensor("pretrained_out_577_cast_fp16")]; - tensor var_11552 = const()[name = tensor("op_11552"), val = tensor([1, 1])]; - tensor var_11554 = const()[name = tensor("op_11554"), val = tensor([1, 1])]; - tensor input_863_pad_type_0 = const()[name = tensor("input_863_pad_type_0"), val = tensor("custom")]; - tensor input_863_pad_0 = const()[name = tensor("input_863_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_28_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545897984)))]; - tensor input_863_cast_fp16 = conv(dilations = var_11554, groups = var_11215, pad = input_863_pad_0, pad_type = input_863_pad_type_0, strides = var_11552, weight = layers_28_fc1_loraA_weight_to_fp16, x = input_861_cast_fp16)[name = tensor("input_863_cast_fp16")]; - tensor var_11558 = const()[name = tensor("op_11558"), val = tensor([1, 1])]; - tensor var_11560 = const()[name = tensor("op_11560"), val = tensor([1, 1])]; - tensor lora_out_1153_pad_type_0 = const()[name = tensor("lora_out_1153_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1153_pad_0 = const()[name = tensor("lora_out_1153_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1155_weight_0_to_fp16 = const()[name = tensor("lora_out_1155_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545939008)))]; - tensor lora_out_1155_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_11560, groups = var_11215, pad = lora_out_1153_pad_0, pad_type = lora_out_1153_pad_type_0, strides = var_11558, weight = lora_out_1155_weight_0_to_fp16, x = input_863_cast_fp16)[name = tensor("lora_out_1155_cast_fp16")]; - tensor input_865_cast_fp16 = add(x = pretrained_out_577_cast_fp16, y = lora_out_1155_cast_fp16)[name = tensor("input_865_cast_fp16")]; - tensor input_867_mode_0 = const()[name = tensor("input_867_mode_0"), val = tensor("EXACT")]; - tensor input_867_cast_fp16 = gelu(mode = input_867_mode_0, x = input_865_cast_fp16)[name = tensor("input_867_cast_fp16")]; - tensor var_11572 = const()[name = tensor("op_11572"), val = tensor([1, 1])]; - tensor var_11574 = const()[name = tensor("op_11574"), val = tensor([1, 1])]; - tensor pretrained_out_579_pad_type_0 = const()[name = tensor("pretrained_out_579_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_579_pad_0 = const()[name = tensor("pretrained_out_579_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546102912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549379776))), name = tensor("layers_28_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_28_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_28_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549379904)))]; - tensor pretrained_out_579_cast_fp16 = conv(bias = layers_28_fc2_pretrained_bias_to_fp16, dilations = var_11574, groups = var_11215, pad = pretrained_out_579_pad_0, pad_type = pretrained_out_579_pad_type_0, strides = var_11572, weight = layers_28_fc2_pretrained_weight_to_fp16_palettized, x = input_867_cast_fp16)[name = tensor("pretrained_out_579_cast_fp16")]; - tensor var_11578 = const()[name = tensor("op_11578"), val = tensor([1, 1])]; - tensor var_11580 = const()[name = tensor("op_11580"), val = tensor([1, 1])]; - tensor input_869_pad_type_0 = const()[name = tensor("input_869_pad_type_0"), val = tensor("custom")]; - tensor input_869_pad_0 = const()[name = tensor("input_869_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_28_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_28_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549382528)))]; - tensor input_869_cast_fp16 = conv(dilations = var_11580, groups = var_11215, pad = input_869_pad_0, pad_type = input_869_pad_type_0, strides = var_11578, weight = layers_28_fc2_loraA_weight_to_fp16, x = input_867_cast_fp16)[name = tensor("input_869_cast_fp16")]; - tensor var_11584 = const()[name = tensor("op_11584"), val = tensor([1, 1])]; - tensor var_11586 = const()[name = tensor("op_11586"), val = tensor([1, 1])]; - tensor lora_out_1157_pad_type_0 = const()[name = tensor("lora_out_1157_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1157_pad_0 = const()[name = tensor("lora_out_1157_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1159_weight_0_to_fp16 = const()[name = tensor("lora_out_1159_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549546432)))]; - tensor lora_out_1159_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11586, groups = var_11215, pad = lora_out_1157_pad_0, pad_type = lora_out_1157_pad_type_0, strides = var_11584, weight = lora_out_1159_weight_0_to_fp16, x = input_869_cast_fp16)[name = tensor("lora_out_1159_cast_fp16")]; - tensor hidden_states_59_cast_fp16 = add(x = pretrained_out_579_cast_fp16, y = lora_out_1159_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_11602 = const()[name = tensor("op_11602"), val = tensor(3)]; - tensor var_11609 = const()[name = tensor("op_11609"), val = tensor(1)]; - tensor var_11610 = const()[name = tensor("op_11610"), val = tensor(true)]; - tensor var_11622 = const()[name = tensor("op_11622"), val = tensor([1])]; - tensor channels_mean_175_cast_fp16 = reduce_mean(axes = var_11622, keep_dims = var_11610, 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_11626 = const()[name = tensor("op_11626"), val = tensor([1])]; - tensor var_11627_cast_fp16 = reduce_mean(axes = var_11626, keep_dims = var_11610, x = zero_mean_sq_175_cast_fp16)[name = tensor("op_11627_cast_fp16")]; - tensor var_11628_to_fp16 = const()[name = tensor("op_11628_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_11629_cast_fp16 = add(x = var_11627_cast_fp16, y = var_11628_to_fp16)[name = tensor("op_11629_cast_fp16")]; - tensor denom_175_epsilon_0 = const()[name = tensor("denom_175_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_175_cast_fp16 = rsqrt(epsilon = denom_175_epsilon_0, x = var_11629_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_349_gamma_0_to_fp16 = const()[name = tensor("obj_349_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549587456)))]; - tensor obj_349_beta_0_to_fp16 = const()[name = tensor("obj_349_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549590080)))]; - tensor obj_349_epsilon_0_to_fp16 = const()[name = tensor("obj_349_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_349_cast_fp16 = batch_norm(beta = obj_349_beta_0_to_fp16, epsilon = obj_349_epsilon_0_to_fp16, gamma = obj_349_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_349_cast_fp16")]; - tensor var_11647 = const()[name = tensor("op_11647"), val = tensor([1, 1])]; - tensor var_11649 = const()[name = tensor("op_11649"), val = tensor([1, 1])]; - tensor pretrained_out_581_pad_type_0 = const()[name = tensor("pretrained_out_581_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_581_pad_0 = const()[name = tensor("pretrained_out_581_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549592704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550411968))), name = tensor("layers_29_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_29_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550412096)))]; - tensor pretrained_out_581_cast_fp16 = conv(bias = layers_29_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_11649, groups = var_11609, pad = pretrained_out_581_pad_0, pad_type = pretrained_out_581_pad_type_0, strides = var_11647, weight = layers_29_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_349_cast_fp16)[name = tensor("pretrained_out_581_cast_fp16")]; - tensor var_11653 = const()[name = tensor("op_11653"), val = tensor([1, 1])]; - tensor var_11655 = const()[name = tensor("op_11655"), val = tensor([1, 1])]; - tensor input_871_pad_type_0 = const()[name = tensor("input_871_pad_type_0"), val = tensor("custom")]; - tensor input_871_pad_0 = const()[name = tensor("input_871_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550414720)))]; - tensor input_871_cast_fp16 = conv(dilations = var_11655, groups = var_11609, pad = input_871_pad_0, pad_type = input_871_pad_type_0, strides = var_11653, weight = layers_29_self_attn_q_proj_loraA_weight_to_fp16, x = obj_349_cast_fp16)[name = tensor("input_871_cast_fp16")]; - tensor var_11659 = const()[name = tensor("op_11659"), val = tensor([1, 1])]; - tensor var_11661 = const()[name = tensor("op_11661"), val = tensor([1, 1])]; - tensor lora_out_1161_pad_type_0 = const()[name = tensor("lora_out_1161_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1161_pad_0 = const()[name = tensor("lora_out_1161_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1163_weight_0_to_fp16 = const()[name = tensor("lora_out_1163_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550455744)))]; - tensor lora_out_1163_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11661, groups = var_11609, pad = lora_out_1161_pad_0, pad_type = lora_out_1161_pad_type_0, strides = var_11659, weight = lora_out_1163_weight_0_to_fp16, x = input_871_cast_fp16)[name = tensor("lora_out_1163_cast_fp16")]; - tensor query_117_cast_fp16 = add(x = pretrained_out_581_cast_fp16, y = lora_out_1163_cast_fp16)[name = tensor("query_117_cast_fp16")]; - tensor var_11671 = const()[name = tensor("op_11671"), val = tensor([1, 1])]; - tensor var_11673 = const()[name = tensor("op_11673"), val = tensor([1, 1])]; - tensor pretrained_out_583_pad_type_0 = const()[name = tensor("pretrained_out_583_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_583_pad_0 = const()[name = tensor("pretrained_out_583_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550496768))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(551316032))), name = tensor("layers_29_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_583_cast_fp16 = conv(dilations = var_11673, groups = var_11609, pad = pretrained_out_583_pad_0, pad_type = pretrained_out_583_pad_type_0, strides = var_11671, weight = layers_29_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_349_cast_fp16)[name = tensor("pretrained_out_583_cast_fp16")]; - tensor var_11677 = const()[name = tensor("op_11677"), val = tensor([1, 1])]; - tensor var_11679 = const()[name = tensor("op_11679"), val = tensor([1, 1])]; - tensor input_873_pad_type_0 = const()[name = tensor("input_873_pad_type_0"), val = tensor("custom")]; - tensor input_873_pad_0 = const()[name = tensor("input_873_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(551316160)))]; - tensor input_873_cast_fp16 = conv(dilations = var_11679, groups = var_11609, pad = input_873_pad_0, pad_type = input_873_pad_type_0, strides = var_11677, weight = layers_29_self_attn_k_proj_loraA_weight_to_fp16, x = obj_349_cast_fp16)[name = tensor("input_873_cast_fp16")]; - tensor var_11683 = const()[name = tensor("op_11683"), val = tensor([1, 1])]; - tensor var_11685 = const()[name = tensor("op_11685"), val = tensor([1, 1])]; - tensor lora_out_1165_pad_type_0 = const()[name = tensor("lora_out_1165_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1165_pad_0 = const()[name = tensor("lora_out_1165_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1167_weight_0_to_fp16 = const()[name = tensor("lora_out_1167_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(551357184)))]; - tensor lora_out_1167_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11685, groups = var_11609, pad = lora_out_1165_pad_0, pad_type = lora_out_1165_pad_type_0, strides = var_11683, weight = lora_out_1167_weight_0_to_fp16, x = input_873_cast_fp16)[name = tensor("lora_out_1167_cast_fp16")]; - tensor current_key_59_cast_fp16 = add(x = pretrained_out_583_cast_fp16, y = lora_out_1167_cast_fp16)[name = tensor("current_key_59_cast_fp16")]; - tensor var_11696 = const()[name = tensor("op_11696"), val = tensor([1, 1])]; - tensor var_11698 = const()[name = tensor("op_11698"), val = tensor([1, 1])]; - tensor pretrained_out_585_pad_type_0 = const()[name = tensor("pretrained_out_585_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_585_pad_0 = const()[name = tensor("pretrained_out_585_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(551398208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552217472))), name = tensor("layers_29_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_29_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552217600)))]; - tensor pretrained_out_585_cast_fp16 = conv(bias = layers_29_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_11698, groups = var_11609, pad = pretrained_out_585_pad_0, pad_type = pretrained_out_585_pad_type_0, strides = var_11696, weight = layers_29_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_349_cast_fp16)[name = tensor("pretrained_out_585_cast_fp16")]; - tensor var_11702 = const()[name = tensor("op_11702"), val = tensor([1, 1])]; - tensor var_11704 = const()[name = tensor("op_11704"), val = tensor([1, 1])]; - tensor input_875_pad_type_0 = const()[name = tensor("input_875_pad_type_0"), val = tensor("custom")]; - tensor input_875_pad_0 = const()[name = tensor("input_875_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552220224)))]; - tensor input_875_cast_fp16 = conv(dilations = var_11704, groups = var_11609, pad = input_875_pad_0, pad_type = input_875_pad_type_0, strides = var_11702, weight = layers_29_self_attn_v_proj_loraA_weight_to_fp16, x = obj_349_cast_fp16)[name = tensor("input_875_cast_fp16")]; - tensor var_11708 = const()[name = tensor("op_11708"), val = tensor([1, 1])]; - tensor var_11710 = const()[name = tensor("op_11710"), val = tensor([1, 1])]; - tensor lora_out_1169_pad_type_0 = const()[name = tensor("lora_out_1169_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1169_pad_0 = const()[name = tensor("lora_out_1169_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1171_weight_0_to_fp16 = const()[name = tensor("lora_out_1171_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552261248)))]; - tensor lora_out_1171_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11710, groups = var_11609, pad = lora_out_1169_pad_0, pad_type = lora_out_1169_pad_type_0, strides = var_11708, weight = lora_out_1171_weight_0_to_fp16, x = input_875_cast_fp16)[name = tensor("lora_out_1171_cast_fp16")]; - tensor current_value_59_cast_fp16 = add(x = pretrained_out_585_cast_fp16, y = lora_out_1171_cast_fp16)[name = tensor("current_value_59_cast_fp16")]; - tensor var_11720_cast_fp16 = mul(x = current_key_59_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_11720_cast_fp16")]; - tensor var_11722_cast_fp16 = mul(x = var_103_cast_fp16_29, y = var_295_cast_fp16)[name = tensor("op_11722_cast_fp16")]; - tensor key_117_cast_fp16 = add(x = var_11720_cast_fp16, y = var_11722_cast_fp16)[name = tensor("key_117_cast_fp16")]; - tensor var_11724_cast_fp16 = mul(x = current_value_59_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_11724_cast_fp16")]; - tensor var_11726_cast_fp16 = mul(x = var_138_cast_fp16_29, y = var_295_cast_fp16)[name = tensor("op_11726_cast_fp16")]; - tensor value_117_cast_fp16 = add(x = var_11724_cast_fp16, y = var_11726_cast_fp16)[name = tensor("value_117_cast_fp16")]; - tensor var_11729 = const()[name = tensor("op_11729"), val = tensor([1, 20, 64, -1])]; - tensor var_11730_cast_fp16 = reshape(shape = var_11729, x = query_117_cast_fp16)[name = tensor("op_11730_cast_fp16")]; - tensor var_11731_to_fp16 = const()[name = tensor("op_11731_to_fp16"), val = tensor(0x1p-3)]; - tensor var_11732_cast_fp16 = mul(x = var_11730_cast_fp16, y = var_11731_to_fp16)[name = tensor("op_11732_cast_fp16")]; - tensor var_11733 = const()[name = tensor("op_11733"), val = tensor([1, 20, 64, -1])]; - tensor var_11734_cast_fp16 = reshape(shape = var_11733, x = key_117_cast_fp16)[name = tensor("op_11734_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_11732_cast_fp16, y = var_11734_cast_fp16)[name = tensor("mh_w_175_cast_fp16")]; - tensor mh_w_177_cast_fp16 = add(x = mh_w_175_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_177_cast_fp16")]; - tensor var_11742_cast_fp16 = softmax(axis = var_11602, x = mh_w_177_cast_fp16)[name = tensor("op_11742_cast_fp16")]; - tensor var_11743 = const()[name = tensor("op_11743"), val = tensor([1, 20, 64, -1])]; - tensor var_11744_cast_fp16 = reshape(shape = var_11743, x = value_117_cast_fp16)[name = tensor("op_11744_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_11744_cast_fp16, y = var_11742_cast_fp16)[name = tensor("attn_117_cast_fp16")]; - tensor var_11747 = const()[name = tensor("op_11747"), val = tensor([1, 1280, 1, -1])]; - tensor input_877_cast_fp16 = reshape(shape = var_11747, x = attn_117_cast_fp16)[name = tensor("input_877_cast_fp16")]; - tensor var_11754 = const()[name = tensor("op_11754"), val = tensor([1, 1])]; - tensor var_11756 = const()[name = tensor("op_11756"), val = tensor([1, 1])]; - tensor pretrained_out_587_pad_type_0 = const()[name = tensor("pretrained_out_587_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_587_pad_0 = const()[name = tensor("pretrained_out_587_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552302272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553121536))), name = tensor("layers_29_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_29_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553121664)))]; - tensor pretrained_out_587_cast_fp16 = conv(bias = layers_29_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_11756, groups = var_11609, pad = pretrained_out_587_pad_0, pad_type = pretrained_out_587_pad_type_0, strides = var_11754, weight = layers_29_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_877_cast_fp16)[name = tensor("pretrained_out_587_cast_fp16")]; - tensor var_11760 = const()[name = tensor("op_11760"), val = tensor([1, 1])]; - tensor var_11762 = const()[name = tensor("op_11762"), val = tensor([1, 1])]; - tensor input_879_pad_type_0 = const()[name = tensor("input_879_pad_type_0"), val = tensor("custom")]; - tensor input_879_pad_0 = const()[name = tensor("input_879_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553124288)))]; - tensor input_879_cast_fp16 = conv(dilations = var_11762, groups = var_11609, pad = input_879_pad_0, pad_type = input_879_pad_type_0, strides = var_11760, weight = layers_29_self_attn_o_proj_loraA_weight_to_fp16, x = input_877_cast_fp16)[name = tensor("input_879_cast_fp16")]; - tensor var_11766 = const()[name = tensor("op_11766"), val = tensor([1, 1])]; - tensor var_11768 = const()[name = tensor("op_11768"), val = tensor([1, 1])]; - tensor lora_out_1173_pad_type_0 = const()[name = tensor("lora_out_1173_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1173_pad_0 = const()[name = tensor("lora_out_1173_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1175_weight_0_to_fp16 = const()[name = tensor("lora_out_1175_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553165312)))]; - tensor lora_out_1175_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11768, groups = var_11609, pad = lora_out_1173_pad_0, pad_type = lora_out_1173_pad_type_0, strides = var_11766, weight = lora_out_1175_weight_0_to_fp16, x = input_879_cast_fp16)[name = tensor("lora_out_1175_cast_fp16")]; - tensor obj_355_cast_fp16 = add(x = pretrained_out_587_cast_fp16, y = lora_out_1175_cast_fp16)[name = tensor("obj_355_cast_fp16")]; - tensor inputs_177_cast_fp16 = add(x = inputs_175_cast_fp16, y = obj_355_cast_fp16)[name = tensor("inputs_177_cast_fp16")]; - tensor var_11781 = const()[name = tensor("op_11781"), val = tensor([1])]; - tensor channels_mean_177_cast_fp16 = reduce_mean(axes = var_11781, keep_dims = var_11610, 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_11785 = const()[name = tensor("op_11785"), val = tensor([1])]; - tensor var_11786_cast_fp16 = reduce_mean(axes = var_11785, keep_dims = var_11610, x = zero_mean_sq_177_cast_fp16)[name = tensor("op_11786_cast_fp16")]; - tensor var_11787_to_fp16 = const()[name = tensor("op_11787_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_11788_cast_fp16 = add(x = var_11786_cast_fp16, y = var_11787_to_fp16)[name = tensor("op_11788_cast_fp16")]; - tensor denom_177_epsilon_0 = const()[name = tensor("denom_177_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_177_cast_fp16 = rsqrt(epsilon = denom_177_epsilon_0, x = var_11788_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_357_gamma_0_to_fp16 = const()[name = tensor("obj_357_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553206336)))]; - tensor obj_357_beta_0_to_fp16 = const()[name = tensor("obj_357_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553208960)))]; - tensor obj_357_epsilon_0_to_fp16 = const()[name = tensor("obj_357_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_357_cast_fp16 = batch_norm(beta = obj_357_beta_0_to_fp16, epsilon = obj_357_epsilon_0_to_fp16, gamma = obj_357_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_357_cast_fp16")]; - tensor var_11806 = const()[name = tensor("op_11806"), val = tensor([1, 1])]; - tensor var_11808 = const()[name = tensor("op_11808"), val = tensor([1, 1])]; - tensor pretrained_out_589_pad_type_0 = const()[name = tensor("pretrained_out_589_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_589_pad_0 = const()[name = tensor("pretrained_out_589_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553211584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554030848))), name = tensor("layers_29_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_29_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_29_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554030976)))]; - tensor pretrained_out_589_cast_fp16 = conv(bias = layers_29_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_11808, groups = var_11609, pad = pretrained_out_589_pad_0, pad_type = pretrained_out_589_pad_type_0, strides = var_11806, weight = layers_29_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_357_cast_fp16)[name = tensor("pretrained_out_589_cast_fp16")]; - tensor var_11812 = const()[name = tensor("op_11812"), val = tensor([1, 1])]; - tensor var_11814 = const()[name = tensor("op_11814"), val = tensor([1, 1])]; - tensor input_881_pad_type_0 = const()[name = tensor("input_881_pad_type_0"), val = tensor("custom")]; - tensor input_881_pad_0 = const()[name = tensor("input_881_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_29_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554033600)))]; - tensor input_881_cast_fp16 = conv(dilations = var_11814, groups = var_11609, pad = input_881_pad_0, pad_type = input_881_pad_type_0, strides = var_11812, weight = layers_29_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_357_cast_fp16)[name = tensor("input_881_cast_fp16")]; - tensor var_11818 = const()[name = tensor("op_11818"), val = tensor([1, 1])]; - tensor var_11820 = const()[name = tensor("op_11820"), val = tensor([1, 1])]; - tensor lora_out_1177_pad_type_0 = const()[name = tensor("lora_out_1177_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1177_pad_0 = const()[name = tensor("lora_out_1177_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1179_weight_0_to_fp16 = const()[name = tensor("lora_out_1179_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554074624)))]; - tensor lora_out_1179_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11820, groups = var_11609, pad = lora_out_1177_pad_0, pad_type = lora_out_1177_pad_type_0, strides = var_11818, weight = lora_out_1179_weight_0_to_fp16, x = input_881_cast_fp16)[name = tensor("lora_out_1179_cast_fp16")]; - tensor query_119_cast_fp16 = add(x = pretrained_out_589_cast_fp16, y = lora_out_1179_cast_fp16)[name = tensor("query_119_cast_fp16")]; - tensor var_11830 = const()[name = tensor("op_11830"), val = tensor([1, 1])]; - tensor var_11832 = const()[name = tensor("op_11832"), val = tensor([1, 1])]; - tensor pretrained_out_591_pad_type_0 = const()[name = tensor("pretrained_out_591_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_591_pad_0 = const()[name = tensor("pretrained_out_591_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554115648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554934912))), name = tensor("layers_29_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_591_cast_fp16 = conv(dilations = var_11832, groups = var_11609, pad = pretrained_out_591_pad_0, pad_type = pretrained_out_591_pad_type_0, strides = var_11830, weight = layers_29_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_591_cast_fp16")]; - tensor var_11836 = const()[name = tensor("op_11836"), val = tensor([1, 1])]; - tensor var_11838 = const()[name = tensor("op_11838"), val = tensor([1, 1])]; - tensor input_883_pad_type_0 = const()[name = tensor("input_883_pad_type_0"), val = tensor("custom")]; - tensor input_883_pad_0 = const()[name = tensor("input_883_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_29_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554935040)))]; - tensor input_883_cast_fp16 = conv(dilations = var_11838, groups = var_11609, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = var_11836, weight = layers_29_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_883_cast_fp16")]; - tensor var_11842 = const()[name = tensor("op_11842"), val = tensor([1, 1])]; - tensor var_11844 = const()[name = tensor("op_11844"), val = tensor([1, 1])]; - tensor lora_out_1181_pad_type_0 = const()[name = tensor("lora_out_1181_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1181_pad_0 = const()[name = tensor("lora_out_1181_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1183_weight_0_to_fp16 = const()[name = tensor("lora_out_1183_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554976064)))]; - tensor lora_out_1183_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11844, groups = var_11609, pad = lora_out_1181_pad_0, pad_type = lora_out_1181_pad_type_0, strides = var_11842, weight = lora_out_1183_weight_0_to_fp16, x = input_883_cast_fp16)[name = tensor("lora_out_1183_cast_fp16")]; - tensor key_119_cast_fp16 = add(x = pretrained_out_591_cast_fp16, y = lora_out_1183_cast_fp16)[name = tensor("key_119_cast_fp16")]; - tensor var_11855 = const()[name = tensor("op_11855"), val = tensor([1, 1])]; - tensor var_11857 = const()[name = tensor("op_11857"), val = tensor([1, 1])]; - tensor pretrained_out_593_pad_type_0 = const()[name = tensor("pretrained_out_593_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_593_pad_0 = const()[name = tensor("pretrained_out_593_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555017088))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555836352))), name = tensor("layers_29_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_29_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_29_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555836480)))]; - tensor pretrained_out_593_cast_fp16 = conv(bias = layers_29_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_11857, groups = var_11609, pad = pretrained_out_593_pad_0, pad_type = pretrained_out_593_pad_type_0, strides = var_11855, weight = layers_29_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_593_cast_fp16")]; - tensor var_11861 = const()[name = tensor("op_11861"), val = tensor([1, 1])]; - tensor var_11863 = const()[name = tensor("op_11863"), val = tensor([1, 1])]; - tensor input_885_pad_type_0 = const()[name = tensor("input_885_pad_type_0"), val = tensor("custom")]; - tensor input_885_pad_0 = const()[name = tensor("input_885_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_29_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555839104)))]; - tensor input_885_cast_fp16 = conv(dilations = var_11863, groups = var_11609, pad = input_885_pad_0, pad_type = input_885_pad_type_0, strides = var_11861, weight = layers_29_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_885_cast_fp16")]; - tensor var_11867 = const()[name = tensor("op_11867"), val = tensor([1, 1])]; - tensor var_11869 = const()[name = tensor("op_11869"), val = tensor([1, 1])]; - tensor lora_out_1185_pad_type_0 = const()[name = tensor("lora_out_1185_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1185_pad_0 = const()[name = tensor("lora_out_1185_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1187_weight_0_to_fp16 = const()[name = tensor("lora_out_1187_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555880128)))]; - tensor lora_out_1187_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11869, groups = var_11609, pad = lora_out_1185_pad_0, pad_type = lora_out_1185_pad_type_0, strides = var_11867, weight = lora_out_1187_weight_0_to_fp16, x = input_885_cast_fp16)[name = tensor("lora_out_1187_cast_fp16")]; - tensor value_119_cast_fp16 = add(x = pretrained_out_593_cast_fp16, y = lora_out_1187_cast_fp16)[name = tensor("value_119_cast_fp16")]; - tensor var_11876 = const()[name = tensor("op_11876"), val = tensor([1, 20, 64, -1])]; - tensor var_11877_cast_fp16 = reshape(shape = var_11876, x = query_119_cast_fp16)[name = tensor("op_11877_cast_fp16")]; - tensor var_11878_to_fp16 = const()[name = tensor("op_11878_to_fp16"), val = tensor(0x1p-3)]; - tensor var_11879_cast_fp16 = mul(x = var_11877_cast_fp16, y = var_11878_to_fp16)[name = tensor("op_11879_cast_fp16")]; - tensor var_11880 = const()[name = tensor("op_11880"), val = tensor([1, 20, 64, -1])]; - tensor var_11881_cast_fp16 = reshape(shape = var_11880, x = key_119_cast_fp16)[name = tensor("op_11881_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_11879_cast_fp16, y = var_11881_cast_fp16)[name = tensor("mh_w_179_cast_fp16")]; - tensor var_11884_cast_fp16 = softmax(axis = var_11602, x = mh_w_179_cast_fp16)[name = tensor("op_11884_cast_fp16")]; - tensor var_11885 = const()[name = tensor("op_11885"), val = tensor([1, 20, 64, -1])]; - tensor var_11886_cast_fp16 = reshape(shape = var_11885, x = value_119_cast_fp16)[name = tensor("op_11886_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_11886_cast_fp16, y = var_11884_cast_fp16)[name = tensor("attn_119_cast_fp16")]; - tensor var_11889 = const()[name = tensor("op_11889"), val = tensor([1, 1280, 1, -1])]; - tensor input_887_cast_fp16 = reshape(shape = var_11889, x = attn_119_cast_fp16)[name = tensor("input_887_cast_fp16")]; - tensor var_11896 = const()[name = tensor("op_11896"), val = tensor([1, 1])]; - tensor var_11898 = const()[name = tensor("op_11898"), val = tensor([1, 1])]; - tensor pretrained_out_595_pad_type_0 = const()[name = tensor("pretrained_out_595_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_595_pad_0 = const()[name = tensor("pretrained_out_595_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555921152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556740416))), name = tensor("layers_29_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_29_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_29_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556740544)))]; - tensor pretrained_out_595_cast_fp16 = conv(bias = layers_29_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_11898, groups = var_11609, pad = pretrained_out_595_pad_0, pad_type = pretrained_out_595_pad_type_0, strides = var_11896, weight = layers_29_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_887_cast_fp16)[name = tensor("pretrained_out_595_cast_fp16")]; - tensor var_11902 = const()[name = tensor("op_11902"), val = tensor([1, 1])]; - tensor var_11904 = const()[name = tensor("op_11904"), val = tensor([1, 1])]; - tensor input_889_pad_type_0 = const()[name = tensor("input_889_pad_type_0"), val = tensor("custom")]; - tensor input_889_pad_0 = const()[name = tensor("input_889_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_29_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556743168)))]; - tensor input_889_cast_fp16 = conv(dilations = var_11904, groups = var_11609, pad = input_889_pad_0, pad_type = input_889_pad_type_0, strides = var_11902, weight = layers_29_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_887_cast_fp16)[name = tensor("input_889_cast_fp16")]; - tensor var_11908 = const()[name = tensor("op_11908"), val = tensor([1, 1])]; - tensor var_11910 = const()[name = tensor("op_11910"), val = tensor([1, 1])]; - tensor lora_out_1189_pad_type_0 = const()[name = tensor("lora_out_1189_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1189_pad_0 = const()[name = tensor("lora_out_1189_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1191_weight_0_to_fp16 = const()[name = tensor("lora_out_1191_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556784192)))]; - tensor lora_out_1191_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11910, groups = var_11609, pad = lora_out_1189_pad_0, pad_type = lora_out_1189_pad_type_0, strides = var_11908, weight = lora_out_1191_weight_0_to_fp16, x = input_889_cast_fp16)[name = tensor("lora_out_1191_cast_fp16")]; - tensor obj_359_cast_fp16 = add(x = pretrained_out_595_cast_fp16, y = lora_out_1191_cast_fp16)[name = tensor("obj_359_cast_fp16")]; - tensor inputs_179_cast_fp16 = add(x = inputs_177_cast_fp16, y = obj_359_cast_fp16)[name = tensor("inputs_179_cast_fp16")]; - tensor var_11919 = const()[name = tensor("op_11919"), val = tensor([1])]; - tensor channels_mean_179_cast_fp16 = reduce_mean(axes = var_11919, keep_dims = var_11610, 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_11923 = const()[name = tensor("op_11923"), val = tensor([1])]; - tensor var_11924_cast_fp16 = reduce_mean(axes = var_11923, keep_dims = var_11610, x = zero_mean_sq_179_cast_fp16)[name = tensor("op_11924_cast_fp16")]; - tensor var_11925_to_fp16 = const()[name = tensor("op_11925_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_11926_cast_fp16 = add(x = var_11924_cast_fp16, y = var_11925_to_fp16)[name = tensor("op_11926_cast_fp16")]; - tensor denom_179_epsilon_0 = const()[name = tensor("denom_179_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_179_cast_fp16 = rsqrt(epsilon = denom_179_epsilon_0, x = var_11926_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_891_gamma_0_to_fp16 = const()[name = tensor("input_891_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556825216)))]; - tensor input_891_beta_0_to_fp16 = const()[name = tensor("input_891_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556827840)))]; - tensor input_891_epsilon_0_to_fp16 = const()[name = tensor("input_891_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_891_cast_fp16 = batch_norm(beta = input_891_beta_0_to_fp16, epsilon = input_891_epsilon_0_to_fp16, gamma = input_891_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_891_cast_fp16")]; - tensor var_11940 = const()[name = tensor("op_11940"), val = tensor([1, 1])]; - tensor var_11942 = const()[name = tensor("op_11942"), val = tensor([1, 1])]; - tensor pretrained_out_597_pad_type_0 = const()[name = tensor("pretrained_out_597_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_597_pad_0 = const()[name = tensor("pretrained_out_597_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556830464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560107328))), name = tensor("layers_29_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_29_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_29_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560107456)))]; - tensor pretrained_out_597_cast_fp16 = conv(bias = layers_29_fc1_pretrained_bias_to_fp16, dilations = var_11942, groups = var_11609, pad = pretrained_out_597_pad_0, pad_type = pretrained_out_597_pad_type_0, strides = var_11940, weight = layers_29_fc1_pretrained_weight_to_fp16_palettized, x = input_891_cast_fp16)[name = tensor("pretrained_out_597_cast_fp16")]; - tensor var_11946 = const()[name = tensor("op_11946"), val = tensor([1, 1])]; - tensor var_11948 = const()[name = tensor("op_11948"), val = tensor([1, 1])]; - tensor input_893_pad_type_0 = const()[name = tensor("input_893_pad_type_0"), val = tensor("custom")]; - tensor input_893_pad_0 = const()[name = tensor("input_893_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_29_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560117760)))]; - tensor input_893_cast_fp16 = conv(dilations = var_11948, groups = var_11609, pad = input_893_pad_0, pad_type = input_893_pad_type_0, strides = var_11946, weight = layers_29_fc1_loraA_weight_to_fp16, x = input_891_cast_fp16)[name = tensor("input_893_cast_fp16")]; - tensor var_11952 = const()[name = tensor("op_11952"), val = tensor([1, 1])]; - tensor var_11954 = const()[name = tensor("op_11954"), val = tensor([1, 1])]; - tensor lora_out_1193_pad_type_0 = const()[name = tensor("lora_out_1193_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1193_pad_0 = const()[name = tensor("lora_out_1193_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1195_weight_0_to_fp16 = const()[name = tensor("lora_out_1195_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560158784)))]; - tensor lora_out_1195_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_11954, groups = var_11609, pad = lora_out_1193_pad_0, pad_type = lora_out_1193_pad_type_0, strides = var_11952, weight = lora_out_1195_weight_0_to_fp16, x = input_893_cast_fp16)[name = tensor("lora_out_1195_cast_fp16")]; - tensor input_895_cast_fp16 = add(x = pretrained_out_597_cast_fp16, y = lora_out_1195_cast_fp16)[name = tensor("input_895_cast_fp16")]; - tensor input_897_mode_0 = const()[name = tensor("input_897_mode_0"), val = tensor("EXACT")]; - tensor input_897_cast_fp16 = gelu(mode = input_897_mode_0, x = input_895_cast_fp16)[name = tensor("input_897_cast_fp16")]; - tensor var_11966 = const()[name = tensor("op_11966"), val = tensor([1, 1])]; - tensor var_11968 = const()[name = tensor("op_11968"), val = tensor([1, 1])]; - tensor pretrained_out_599_pad_type_0 = const()[name = tensor("pretrained_out_599_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_599_pad_0 = const()[name = tensor("pretrained_out_599_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560322688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563599552))), name = tensor("layers_29_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_29_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_29_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563599680)))]; - tensor pretrained_out_599_cast_fp16 = conv(bias = layers_29_fc2_pretrained_bias_to_fp16, dilations = var_11968, groups = var_11609, pad = pretrained_out_599_pad_0, pad_type = pretrained_out_599_pad_type_0, strides = var_11966, weight = layers_29_fc2_pretrained_weight_to_fp16_palettized, x = input_897_cast_fp16)[name = tensor("pretrained_out_599_cast_fp16")]; - tensor var_11972 = const()[name = tensor("op_11972"), val = tensor([1, 1])]; - tensor var_11974 = const()[name = tensor("op_11974"), val = tensor([1, 1])]; - tensor input_899_pad_type_0 = const()[name = tensor("input_899_pad_type_0"), val = tensor("custom")]; - tensor input_899_pad_0 = const()[name = tensor("input_899_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_29_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_29_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563602304)))]; - tensor input_899_cast_fp16 = conv(dilations = var_11974, groups = var_11609, pad = input_899_pad_0, pad_type = input_899_pad_type_0, strides = var_11972, weight = layers_29_fc2_loraA_weight_to_fp16, x = input_897_cast_fp16)[name = tensor("input_899_cast_fp16")]; - tensor var_11978 = const()[name = tensor("op_11978"), val = tensor([1, 1])]; - tensor var_11980 = const()[name = tensor("op_11980"), val = tensor([1, 1])]; - tensor lora_out_1197_pad_type_0 = const()[name = tensor("lora_out_1197_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1197_pad_0 = const()[name = tensor("lora_out_1197_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1199_weight_0_to_fp16 = const()[name = tensor("lora_out_1199_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563766208)))]; - tensor lora_out_1199_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11980, groups = var_11609, pad = lora_out_1197_pad_0, pad_type = lora_out_1197_pad_type_0, strides = var_11978, weight = lora_out_1199_weight_0_to_fp16, x = input_899_cast_fp16)[name = tensor("lora_out_1199_cast_fp16")]; - tensor hidden_states_61_cast_fp16 = add(x = pretrained_out_599_cast_fp16, y = lora_out_1199_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_11996 = const()[name = tensor("op_11996"), val = tensor(3)]; - tensor var_12003 = const()[name = tensor("op_12003"), val = tensor(1)]; - tensor var_12004 = const()[name = tensor("op_12004"), val = tensor(true)]; - tensor var_12016 = const()[name = tensor("op_12016"), val = tensor([1])]; - tensor channels_mean_181_cast_fp16 = reduce_mean(axes = var_12016, keep_dims = var_12004, 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_12020 = const()[name = tensor("op_12020"), val = tensor([1])]; - tensor var_12021_cast_fp16 = reduce_mean(axes = var_12020, keep_dims = var_12004, x = zero_mean_sq_181_cast_fp16)[name = tensor("op_12021_cast_fp16")]; - tensor var_12022_to_fp16 = const()[name = tensor("op_12022_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_12023_cast_fp16 = add(x = var_12021_cast_fp16, y = var_12022_to_fp16)[name = tensor("op_12023_cast_fp16")]; - tensor denom_181_epsilon_0 = const()[name = tensor("denom_181_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_181_cast_fp16 = rsqrt(epsilon = denom_181_epsilon_0, x = var_12023_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_361_gamma_0_to_fp16 = const()[name = tensor("obj_361_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563807232)))]; - tensor obj_361_beta_0_to_fp16 = const()[name = tensor("obj_361_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563809856)))]; - tensor obj_361_epsilon_0_to_fp16 = const()[name = tensor("obj_361_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_361_cast_fp16 = batch_norm(beta = obj_361_beta_0_to_fp16, epsilon = obj_361_epsilon_0_to_fp16, gamma = obj_361_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_361_cast_fp16")]; - tensor var_12041 = const()[name = tensor("op_12041"), val = tensor([1, 1])]; - tensor var_12043 = const()[name = tensor("op_12043"), val = tensor([1, 1])]; - tensor pretrained_out_601_pad_type_0 = const()[name = tensor("pretrained_out_601_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_601_pad_0 = const()[name = tensor("pretrained_out_601_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563812480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564631744))), name = tensor("layers_30_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_30_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564631872)))]; - tensor pretrained_out_601_cast_fp16 = conv(bias = layers_30_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_12043, groups = var_12003, pad = pretrained_out_601_pad_0, pad_type = pretrained_out_601_pad_type_0, strides = var_12041, weight = layers_30_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_361_cast_fp16)[name = tensor("pretrained_out_601_cast_fp16")]; - tensor var_12047 = const()[name = tensor("op_12047"), val = tensor([1, 1])]; - tensor var_12049 = const()[name = tensor("op_12049"), val = tensor([1, 1])]; - tensor input_901_pad_type_0 = const()[name = tensor("input_901_pad_type_0"), val = tensor("custom")]; - tensor input_901_pad_0 = const()[name = tensor("input_901_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564634496)))]; - tensor input_901_cast_fp16 = conv(dilations = var_12049, groups = var_12003, pad = input_901_pad_0, pad_type = input_901_pad_type_0, strides = var_12047, weight = layers_30_self_attn_q_proj_loraA_weight_to_fp16, x = obj_361_cast_fp16)[name = tensor("input_901_cast_fp16")]; - tensor var_12053 = const()[name = tensor("op_12053"), val = tensor([1, 1])]; - tensor var_12055 = const()[name = tensor("op_12055"), val = tensor([1, 1])]; - tensor lora_out_1201_pad_type_0 = const()[name = tensor("lora_out_1201_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1201_pad_0 = const()[name = tensor("lora_out_1201_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1203_weight_0_to_fp16 = const()[name = tensor("lora_out_1203_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564675520)))]; - tensor lora_out_1203_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12055, groups = var_12003, pad = lora_out_1201_pad_0, pad_type = lora_out_1201_pad_type_0, strides = var_12053, weight = lora_out_1203_weight_0_to_fp16, x = input_901_cast_fp16)[name = tensor("lora_out_1203_cast_fp16")]; - tensor query_121_cast_fp16 = add(x = pretrained_out_601_cast_fp16, y = lora_out_1203_cast_fp16)[name = tensor("query_121_cast_fp16")]; - tensor var_12065 = const()[name = tensor("op_12065"), val = tensor([1, 1])]; - tensor var_12067 = const()[name = tensor("op_12067"), val = tensor([1, 1])]; - tensor pretrained_out_603_pad_type_0 = const()[name = tensor("pretrained_out_603_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_603_pad_0 = const()[name = tensor("pretrained_out_603_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564716544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565535808))), name = tensor("layers_30_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_603_cast_fp16 = conv(dilations = var_12067, groups = var_12003, pad = pretrained_out_603_pad_0, pad_type = pretrained_out_603_pad_type_0, strides = var_12065, weight = layers_30_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_361_cast_fp16)[name = tensor("pretrained_out_603_cast_fp16")]; - tensor var_12071 = const()[name = tensor("op_12071"), val = tensor([1, 1])]; - tensor var_12073 = const()[name = tensor("op_12073"), val = tensor([1, 1])]; - tensor input_903_pad_type_0 = const()[name = tensor("input_903_pad_type_0"), val = tensor("custom")]; - tensor input_903_pad_0 = const()[name = tensor("input_903_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565535936)))]; - tensor input_903_cast_fp16 = conv(dilations = var_12073, groups = var_12003, pad = input_903_pad_0, pad_type = input_903_pad_type_0, strides = var_12071, weight = layers_30_self_attn_k_proj_loraA_weight_to_fp16, x = obj_361_cast_fp16)[name = tensor("input_903_cast_fp16")]; - tensor var_12077 = const()[name = tensor("op_12077"), val = tensor([1, 1])]; - tensor var_12079 = const()[name = tensor("op_12079"), val = tensor([1, 1])]; - tensor lora_out_1205_pad_type_0 = const()[name = tensor("lora_out_1205_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1205_pad_0 = const()[name = tensor("lora_out_1205_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1207_weight_0_to_fp16 = const()[name = tensor("lora_out_1207_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565576960)))]; - tensor lora_out_1207_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12079, groups = var_12003, pad = lora_out_1205_pad_0, pad_type = lora_out_1205_pad_type_0, strides = var_12077, weight = lora_out_1207_weight_0_to_fp16, x = input_903_cast_fp16)[name = tensor("lora_out_1207_cast_fp16")]; - tensor current_key_61_cast_fp16 = add(x = pretrained_out_603_cast_fp16, y = lora_out_1207_cast_fp16)[name = tensor("current_key_61_cast_fp16")]; - tensor var_12090 = const()[name = tensor("op_12090"), val = tensor([1, 1])]; - tensor var_12092 = const()[name = tensor("op_12092"), val = tensor([1, 1])]; - tensor pretrained_out_605_pad_type_0 = const()[name = tensor("pretrained_out_605_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_605_pad_0 = const()[name = tensor("pretrained_out_605_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565617984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566437248))), name = tensor("layers_30_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_30_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566437376)))]; - tensor pretrained_out_605_cast_fp16 = conv(bias = layers_30_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_12092, groups = var_12003, pad = pretrained_out_605_pad_0, pad_type = pretrained_out_605_pad_type_0, strides = var_12090, weight = layers_30_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_361_cast_fp16)[name = tensor("pretrained_out_605_cast_fp16")]; - tensor var_12096 = const()[name = tensor("op_12096"), val = tensor([1, 1])]; - tensor var_12098 = const()[name = tensor("op_12098"), val = tensor([1, 1])]; - tensor input_905_pad_type_0 = const()[name = tensor("input_905_pad_type_0"), val = tensor("custom")]; - tensor input_905_pad_0 = const()[name = tensor("input_905_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566440000)))]; - tensor input_905_cast_fp16 = conv(dilations = var_12098, groups = var_12003, pad = input_905_pad_0, pad_type = input_905_pad_type_0, strides = var_12096, weight = layers_30_self_attn_v_proj_loraA_weight_to_fp16, x = obj_361_cast_fp16)[name = tensor("input_905_cast_fp16")]; - tensor var_12102 = const()[name = tensor("op_12102"), val = tensor([1, 1])]; - tensor var_12104 = const()[name = tensor("op_12104"), val = tensor([1, 1])]; - tensor lora_out_1209_pad_type_0 = const()[name = tensor("lora_out_1209_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1209_pad_0 = const()[name = tensor("lora_out_1209_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1211_weight_0_to_fp16 = const()[name = tensor("lora_out_1211_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566481024)))]; - tensor lora_out_1211_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12104, groups = var_12003, pad = lora_out_1209_pad_0, pad_type = lora_out_1209_pad_type_0, strides = var_12102, weight = lora_out_1211_weight_0_to_fp16, x = input_905_cast_fp16)[name = tensor("lora_out_1211_cast_fp16")]; - tensor current_value_61_cast_fp16 = add(x = pretrained_out_605_cast_fp16, y = lora_out_1211_cast_fp16)[name = tensor("current_value_61_cast_fp16")]; - tensor var_12114_cast_fp16 = mul(x = current_key_61_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_12114_cast_fp16")]; - tensor var_12116_cast_fp16 = mul(x = var_103_cast_fp16_30, y = var_295_cast_fp16)[name = tensor("op_12116_cast_fp16")]; - tensor key_121_cast_fp16 = add(x = var_12114_cast_fp16, y = var_12116_cast_fp16)[name = tensor("key_121_cast_fp16")]; - tensor var_12118_cast_fp16 = mul(x = current_value_61_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_12118_cast_fp16")]; - tensor var_12120_cast_fp16 = mul(x = var_138_cast_fp16_30, y = var_295_cast_fp16)[name = tensor("op_12120_cast_fp16")]; - tensor value_121_cast_fp16 = add(x = var_12118_cast_fp16, y = var_12120_cast_fp16)[name = tensor("value_121_cast_fp16")]; - tensor var_12123 = const()[name = tensor("op_12123"), val = tensor([1, 20, 64, -1])]; - tensor var_12124_cast_fp16 = reshape(shape = var_12123, x = query_121_cast_fp16)[name = tensor("op_12124_cast_fp16")]; - tensor var_12125_to_fp16 = const()[name = tensor("op_12125_to_fp16"), val = tensor(0x1p-3)]; - tensor var_12126_cast_fp16 = mul(x = var_12124_cast_fp16, y = var_12125_to_fp16)[name = tensor("op_12126_cast_fp16")]; - tensor var_12127 = const()[name = tensor("op_12127"), val = tensor([1, 20, 64, -1])]; - tensor var_12128_cast_fp16 = reshape(shape = var_12127, x = key_121_cast_fp16)[name = tensor("op_12128_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_12126_cast_fp16, y = var_12128_cast_fp16)[name = tensor("mh_w_181_cast_fp16")]; - tensor mh_w_183_cast_fp16 = add(x = mh_w_181_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_183_cast_fp16")]; - tensor var_12136_cast_fp16 = softmax(axis = var_11996, x = mh_w_183_cast_fp16)[name = tensor("op_12136_cast_fp16")]; - tensor var_12137 = const()[name = tensor("op_12137"), val = tensor([1, 20, 64, -1])]; - tensor var_12138_cast_fp16 = reshape(shape = var_12137, x = value_121_cast_fp16)[name = tensor("op_12138_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_12138_cast_fp16, y = var_12136_cast_fp16)[name = tensor("attn_121_cast_fp16")]; - tensor var_12141 = const()[name = tensor("op_12141"), val = tensor([1, 1280, 1, -1])]; - tensor input_907_cast_fp16 = reshape(shape = var_12141, x = attn_121_cast_fp16)[name = tensor("input_907_cast_fp16")]; - tensor var_12148 = const()[name = tensor("op_12148"), val = tensor([1, 1])]; - tensor var_12150 = const()[name = tensor("op_12150"), val = tensor([1, 1])]; - tensor pretrained_out_607_pad_type_0 = const()[name = tensor("pretrained_out_607_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_607_pad_0 = const()[name = tensor("pretrained_out_607_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566522048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(567341312))), name = tensor("layers_30_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_30_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(567341440)))]; - tensor pretrained_out_607_cast_fp16 = conv(bias = layers_30_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_12150, groups = var_12003, pad = pretrained_out_607_pad_0, pad_type = pretrained_out_607_pad_type_0, strides = var_12148, weight = layers_30_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_907_cast_fp16)[name = tensor("pretrained_out_607_cast_fp16")]; - tensor var_12154 = const()[name = tensor("op_12154"), val = tensor([1, 1])]; - tensor var_12156 = const()[name = tensor("op_12156"), val = tensor([1, 1])]; - tensor input_909_pad_type_0 = const()[name = tensor("input_909_pad_type_0"), val = tensor("custom")]; - tensor input_909_pad_0 = const()[name = tensor("input_909_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(567344064)))]; - tensor input_909_cast_fp16 = conv(dilations = var_12156, groups = var_12003, pad = input_909_pad_0, pad_type = input_909_pad_type_0, strides = var_12154, weight = layers_30_self_attn_o_proj_loraA_weight_to_fp16, x = input_907_cast_fp16)[name = tensor("input_909_cast_fp16")]; - tensor var_12160 = const()[name = tensor("op_12160"), val = tensor([1, 1])]; - tensor var_12162 = const()[name = tensor("op_12162"), val = tensor([1, 1])]; - tensor lora_out_1213_pad_type_0 = const()[name = tensor("lora_out_1213_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1213_pad_0 = const()[name = tensor("lora_out_1213_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1215_weight_0_to_fp16 = const()[name = tensor("lora_out_1215_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(567385088)))]; - tensor lora_out_1215_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12162, groups = var_12003, pad = lora_out_1213_pad_0, pad_type = lora_out_1213_pad_type_0, strides = var_12160, weight = lora_out_1215_weight_0_to_fp16, x = input_909_cast_fp16)[name = tensor("lora_out_1215_cast_fp16")]; - tensor obj_367_cast_fp16 = add(x = pretrained_out_607_cast_fp16, y = lora_out_1215_cast_fp16)[name = tensor("obj_367_cast_fp16")]; - tensor inputs_183_cast_fp16 = add(x = inputs_181_cast_fp16, y = obj_367_cast_fp16)[name = tensor("inputs_183_cast_fp16")]; - tensor var_12175 = const()[name = tensor("op_12175"), val = tensor([1])]; - tensor channels_mean_183_cast_fp16 = reduce_mean(axes = var_12175, keep_dims = var_12004, 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_12179 = const()[name = tensor("op_12179"), val = tensor([1])]; - tensor var_12180_cast_fp16 = reduce_mean(axes = var_12179, keep_dims = var_12004, x = zero_mean_sq_183_cast_fp16)[name = tensor("op_12180_cast_fp16")]; - tensor var_12181_to_fp16 = const()[name = tensor("op_12181_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_12182_cast_fp16 = add(x = var_12180_cast_fp16, y = var_12181_to_fp16)[name = tensor("op_12182_cast_fp16")]; - tensor denom_183_epsilon_0 = const()[name = tensor("denom_183_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_183_cast_fp16 = rsqrt(epsilon = denom_183_epsilon_0, x = var_12182_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_369_gamma_0_to_fp16 = const()[name = tensor("obj_369_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(567426112)))]; - tensor obj_369_beta_0_to_fp16 = const()[name = tensor("obj_369_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(567428736)))]; - tensor obj_369_epsilon_0_to_fp16 = const()[name = tensor("obj_369_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_369_cast_fp16 = batch_norm(beta = obj_369_beta_0_to_fp16, epsilon = obj_369_epsilon_0_to_fp16, gamma = obj_369_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_369_cast_fp16")]; - tensor var_12200 = const()[name = tensor("op_12200"), val = tensor([1, 1])]; - tensor var_12202 = const()[name = tensor("op_12202"), val = tensor([1, 1])]; - tensor pretrained_out_609_pad_type_0 = const()[name = tensor("pretrained_out_609_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_609_pad_0 = const()[name = tensor("pretrained_out_609_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(567431360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568250624))), name = tensor("layers_30_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_30_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_30_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568250752)))]; - tensor pretrained_out_609_cast_fp16 = conv(bias = layers_30_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_12202, groups = var_12003, pad = pretrained_out_609_pad_0, pad_type = pretrained_out_609_pad_type_0, strides = var_12200, weight = layers_30_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_369_cast_fp16)[name = tensor("pretrained_out_609_cast_fp16")]; - tensor var_12206 = const()[name = tensor("op_12206"), val = tensor([1, 1])]; - tensor var_12208 = const()[name = tensor("op_12208"), val = tensor([1, 1])]; - tensor input_911_pad_type_0 = const()[name = tensor("input_911_pad_type_0"), val = tensor("custom")]; - tensor input_911_pad_0 = const()[name = tensor("input_911_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_30_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568253376)))]; - tensor input_911_cast_fp16 = conv(dilations = var_12208, groups = var_12003, pad = input_911_pad_0, pad_type = input_911_pad_type_0, strides = var_12206, weight = layers_30_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_369_cast_fp16)[name = tensor("input_911_cast_fp16")]; - tensor var_12212 = const()[name = tensor("op_12212"), val = tensor([1, 1])]; - tensor var_12214 = const()[name = tensor("op_12214"), val = tensor([1, 1])]; - tensor lora_out_1217_pad_type_0 = const()[name = tensor("lora_out_1217_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1217_pad_0 = const()[name = tensor("lora_out_1217_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1219_weight_0_to_fp16 = const()[name = tensor("lora_out_1219_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568294400)))]; - tensor lora_out_1219_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12214, groups = var_12003, pad = lora_out_1217_pad_0, pad_type = lora_out_1217_pad_type_0, strides = var_12212, weight = lora_out_1219_weight_0_to_fp16, x = input_911_cast_fp16)[name = tensor("lora_out_1219_cast_fp16")]; - tensor query_123_cast_fp16 = add(x = pretrained_out_609_cast_fp16, y = lora_out_1219_cast_fp16)[name = tensor("query_123_cast_fp16")]; - tensor var_12224 = const()[name = tensor("op_12224"), val = tensor([1, 1])]; - tensor var_12226 = const()[name = tensor("op_12226"), val = tensor([1, 1])]; - tensor pretrained_out_611_pad_type_0 = const()[name = tensor("pretrained_out_611_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_611_pad_0 = const()[name = tensor("pretrained_out_611_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568335424))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(569154688))), name = tensor("layers_30_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_611_cast_fp16 = conv(dilations = var_12226, groups = var_12003, pad = pretrained_out_611_pad_0, pad_type = pretrained_out_611_pad_type_0, strides = var_12224, weight = layers_30_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_611_cast_fp16")]; - tensor var_12230 = const()[name = tensor("op_12230"), val = tensor([1, 1])]; - tensor var_12232 = const()[name = tensor("op_12232"), val = tensor([1, 1])]; - tensor input_913_pad_type_0 = const()[name = tensor("input_913_pad_type_0"), val = tensor("custom")]; - tensor input_913_pad_0 = const()[name = tensor("input_913_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_30_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(569154816)))]; - tensor input_913_cast_fp16 = conv(dilations = var_12232, groups = var_12003, pad = input_913_pad_0, pad_type = input_913_pad_type_0, strides = var_12230, weight = layers_30_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_913_cast_fp16")]; - tensor var_12236 = const()[name = tensor("op_12236"), val = tensor([1, 1])]; - tensor var_12238 = const()[name = tensor("op_12238"), val = tensor([1, 1])]; - tensor lora_out_1221_pad_type_0 = const()[name = tensor("lora_out_1221_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1221_pad_0 = const()[name = tensor("lora_out_1221_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1223_weight_0_to_fp16 = const()[name = tensor("lora_out_1223_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(569195840)))]; - tensor lora_out_1223_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12238, groups = var_12003, pad = lora_out_1221_pad_0, pad_type = lora_out_1221_pad_type_0, strides = var_12236, weight = lora_out_1223_weight_0_to_fp16, x = input_913_cast_fp16)[name = tensor("lora_out_1223_cast_fp16")]; - tensor key_123_cast_fp16 = add(x = pretrained_out_611_cast_fp16, y = lora_out_1223_cast_fp16)[name = tensor("key_123_cast_fp16")]; - tensor var_12249 = const()[name = tensor("op_12249"), val = tensor([1, 1])]; - tensor var_12251 = const()[name = tensor("op_12251"), val = tensor([1, 1])]; - tensor pretrained_out_613_pad_type_0 = const()[name = tensor("pretrained_out_613_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_613_pad_0 = const()[name = tensor("pretrained_out_613_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(569236864))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570056128))), name = tensor("layers_30_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_30_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_30_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570056256)))]; - tensor pretrained_out_613_cast_fp16 = conv(bias = layers_30_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_12251, groups = var_12003, pad = pretrained_out_613_pad_0, pad_type = pretrained_out_613_pad_type_0, strides = var_12249, weight = layers_30_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_613_cast_fp16")]; - tensor var_12255 = const()[name = tensor("op_12255"), val = tensor([1, 1])]; - tensor var_12257 = const()[name = tensor("op_12257"), val = tensor([1, 1])]; - tensor input_915_pad_type_0 = const()[name = tensor("input_915_pad_type_0"), val = tensor("custom")]; - tensor input_915_pad_0 = const()[name = tensor("input_915_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_30_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570058880)))]; - tensor input_915_cast_fp16 = conv(dilations = var_12257, groups = var_12003, pad = input_915_pad_0, pad_type = input_915_pad_type_0, strides = var_12255, weight = layers_30_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_915_cast_fp16")]; - tensor var_12261 = const()[name = tensor("op_12261"), val = tensor([1, 1])]; - tensor var_12263 = const()[name = tensor("op_12263"), val = tensor([1, 1])]; - tensor lora_out_1225_pad_type_0 = const()[name = tensor("lora_out_1225_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1225_pad_0 = const()[name = tensor("lora_out_1225_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1227_weight_0_to_fp16 = const()[name = tensor("lora_out_1227_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570099904)))]; - tensor lora_out_1227_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12263, groups = var_12003, pad = lora_out_1225_pad_0, pad_type = lora_out_1225_pad_type_0, strides = var_12261, weight = lora_out_1227_weight_0_to_fp16, x = input_915_cast_fp16)[name = tensor("lora_out_1227_cast_fp16")]; - tensor value_123_cast_fp16 = add(x = pretrained_out_613_cast_fp16, y = lora_out_1227_cast_fp16)[name = tensor("value_123_cast_fp16")]; - tensor var_12270 = const()[name = tensor("op_12270"), val = tensor([1, 20, 64, -1])]; - tensor var_12271_cast_fp16 = reshape(shape = var_12270, x = query_123_cast_fp16)[name = tensor("op_12271_cast_fp16")]; - tensor var_12272_to_fp16 = const()[name = tensor("op_12272_to_fp16"), val = tensor(0x1p-3)]; - tensor var_12273_cast_fp16 = mul(x = var_12271_cast_fp16, y = var_12272_to_fp16)[name = tensor("op_12273_cast_fp16")]; - tensor var_12274 = const()[name = tensor("op_12274"), val = tensor([1, 20, 64, -1])]; - tensor var_12275_cast_fp16 = reshape(shape = var_12274, x = key_123_cast_fp16)[name = tensor("op_12275_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_12273_cast_fp16, y = var_12275_cast_fp16)[name = tensor("mh_w_185_cast_fp16")]; - tensor var_12278_cast_fp16 = softmax(axis = var_11996, x = mh_w_185_cast_fp16)[name = tensor("op_12278_cast_fp16")]; - tensor var_12279 = const()[name = tensor("op_12279"), val = tensor([1, 20, 64, -1])]; - tensor var_12280_cast_fp16 = reshape(shape = var_12279, x = value_123_cast_fp16)[name = tensor("op_12280_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_12280_cast_fp16, y = var_12278_cast_fp16)[name = tensor("attn_123_cast_fp16")]; - tensor var_12283 = const()[name = tensor("op_12283"), val = tensor([1, 1280, 1, -1])]; - tensor input_917_cast_fp16 = reshape(shape = var_12283, x = attn_123_cast_fp16)[name = tensor("input_917_cast_fp16")]; - tensor var_12290 = const()[name = tensor("op_12290"), val = tensor([1, 1])]; - tensor var_12292 = const()[name = tensor("op_12292"), val = tensor([1, 1])]; - tensor pretrained_out_615_pad_type_0 = const()[name = tensor("pretrained_out_615_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_615_pad_0 = const()[name = tensor("pretrained_out_615_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570140928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570960192))), name = tensor("layers_30_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_30_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_30_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570960320)))]; - tensor pretrained_out_615_cast_fp16 = conv(bias = layers_30_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_12292, groups = var_12003, pad = pretrained_out_615_pad_0, pad_type = pretrained_out_615_pad_type_0, strides = var_12290, weight = layers_30_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_917_cast_fp16)[name = tensor("pretrained_out_615_cast_fp16")]; - tensor var_12296 = const()[name = tensor("op_12296"), val = tensor([1, 1])]; - tensor var_12298 = const()[name = tensor("op_12298"), val = tensor([1, 1])]; - tensor input_919_pad_type_0 = const()[name = tensor("input_919_pad_type_0"), val = tensor("custom")]; - tensor input_919_pad_0 = const()[name = tensor("input_919_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_30_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570962944)))]; - tensor input_919_cast_fp16 = conv(dilations = var_12298, groups = var_12003, pad = input_919_pad_0, pad_type = input_919_pad_type_0, strides = var_12296, weight = layers_30_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_917_cast_fp16)[name = tensor("input_919_cast_fp16")]; - tensor var_12302 = const()[name = tensor("op_12302"), val = tensor([1, 1])]; - tensor var_12304 = const()[name = tensor("op_12304"), val = tensor([1, 1])]; - tensor lora_out_1229_pad_type_0 = const()[name = tensor("lora_out_1229_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1229_pad_0 = const()[name = tensor("lora_out_1229_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1231_weight_0_to_fp16 = const()[name = tensor("lora_out_1231_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571003968)))]; - tensor lora_out_1231_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12304, groups = var_12003, pad = lora_out_1229_pad_0, pad_type = lora_out_1229_pad_type_0, strides = var_12302, weight = lora_out_1231_weight_0_to_fp16, x = input_919_cast_fp16)[name = tensor("lora_out_1231_cast_fp16")]; - tensor obj_371_cast_fp16 = add(x = pretrained_out_615_cast_fp16, y = lora_out_1231_cast_fp16)[name = tensor("obj_371_cast_fp16")]; - tensor inputs_185_cast_fp16 = add(x = inputs_183_cast_fp16, y = obj_371_cast_fp16)[name = tensor("inputs_185_cast_fp16")]; - tensor var_12313 = const()[name = tensor("op_12313"), val = tensor([1])]; - tensor channels_mean_185_cast_fp16 = reduce_mean(axes = var_12313, keep_dims = var_12004, 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_12317 = const()[name = tensor("op_12317"), val = tensor([1])]; - tensor var_12318_cast_fp16 = reduce_mean(axes = var_12317, keep_dims = var_12004, x = zero_mean_sq_185_cast_fp16)[name = tensor("op_12318_cast_fp16")]; - tensor var_12319_to_fp16 = const()[name = tensor("op_12319_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_12320_cast_fp16 = add(x = var_12318_cast_fp16, y = var_12319_to_fp16)[name = tensor("op_12320_cast_fp16")]; - tensor denom_185_epsilon_0 = const()[name = tensor("denom_185_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_185_cast_fp16 = rsqrt(epsilon = denom_185_epsilon_0, x = var_12320_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_921_gamma_0_to_fp16 = const()[name = tensor("input_921_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571044992)))]; - tensor input_921_beta_0_to_fp16 = const()[name = tensor("input_921_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571047616)))]; - tensor input_921_epsilon_0_to_fp16 = const()[name = tensor("input_921_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_921_cast_fp16 = batch_norm(beta = input_921_beta_0_to_fp16, epsilon = input_921_epsilon_0_to_fp16, gamma = input_921_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_921_cast_fp16")]; - tensor var_12334 = const()[name = tensor("op_12334"), val = tensor([1, 1])]; - tensor var_12336 = const()[name = tensor("op_12336"), val = tensor([1, 1])]; - tensor pretrained_out_617_pad_type_0 = const()[name = tensor("pretrained_out_617_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_617_pad_0 = const()[name = tensor("pretrained_out_617_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571050240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(574327104))), name = tensor("layers_30_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_30_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_30_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(574327232)))]; - tensor pretrained_out_617_cast_fp16 = conv(bias = layers_30_fc1_pretrained_bias_to_fp16, dilations = var_12336, groups = var_12003, pad = pretrained_out_617_pad_0, pad_type = pretrained_out_617_pad_type_0, strides = var_12334, weight = layers_30_fc1_pretrained_weight_to_fp16_palettized, x = input_921_cast_fp16)[name = tensor("pretrained_out_617_cast_fp16")]; - tensor var_12340 = const()[name = tensor("op_12340"), val = tensor([1, 1])]; - tensor var_12342 = const()[name = tensor("op_12342"), val = tensor([1, 1])]; - tensor input_923_pad_type_0 = const()[name = tensor("input_923_pad_type_0"), val = tensor("custom")]; - tensor input_923_pad_0 = const()[name = tensor("input_923_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_30_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(574337536)))]; - tensor input_923_cast_fp16 = conv(dilations = var_12342, groups = var_12003, pad = input_923_pad_0, pad_type = input_923_pad_type_0, strides = var_12340, weight = layers_30_fc1_loraA_weight_to_fp16, x = input_921_cast_fp16)[name = tensor("input_923_cast_fp16")]; - tensor var_12346 = const()[name = tensor("op_12346"), val = tensor([1, 1])]; - tensor var_12348 = const()[name = tensor("op_12348"), val = tensor([1, 1])]; - tensor lora_out_1233_pad_type_0 = const()[name = tensor("lora_out_1233_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1233_pad_0 = const()[name = tensor("lora_out_1233_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1235_weight_0_to_fp16 = const()[name = tensor("lora_out_1235_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(574378560)))]; - tensor lora_out_1235_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_12348, groups = var_12003, pad = lora_out_1233_pad_0, pad_type = lora_out_1233_pad_type_0, strides = var_12346, weight = lora_out_1235_weight_0_to_fp16, x = input_923_cast_fp16)[name = tensor("lora_out_1235_cast_fp16")]; - tensor input_925_cast_fp16 = add(x = pretrained_out_617_cast_fp16, y = lora_out_1235_cast_fp16)[name = tensor("input_925_cast_fp16")]; - tensor input_927_mode_0 = const()[name = tensor("input_927_mode_0"), val = tensor("EXACT")]; - tensor input_927_cast_fp16 = gelu(mode = input_927_mode_0, x = input_925_cast_fp16)[name = tensor("input_927_cast_fp16")]; - tensor var_12360 = const()[name = tensor("op_12360"), val = tensor([1, 1])]; - tensor var_12362 = const()[name = tensor("op_12362"), val = tensor([1, 1])]; - tensor pretrained_out_619_pad_type_0 = const()[name = tensor("pretrained_out_619_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_619_pad_0 = const()[name = tensor("pretrained_out_619_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(574542464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(577819328))), name = tensor("layers_30_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_30_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_30_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(577819456)))]; - tensor pretrained_out_619_cast_fp16 = conv(bias = layers_30_fc2_pretrained_bias_to_fp16, dilations = var_12362, groups = var_12003, pad = pretrained_out_619_pad_0, pad_type = pretrained_out_619_pad_type_0, strides = var_12360, weight = layers_30_fc2_pretrained_weight_to_fp16_palettized, x = input_927_cast_fp16)[name = tensor("pretrained_out_619_cast_fp16")]; - tensor var_12366 = const()[name = tensor("op_12366"), val = tensor([1, 1])]; - tensor var_12368 = const()[name = tensor("op_12368"), val = tensor([1, 1])]; - tensor input_929_pad_type_0 = const()[name = tensor("input_929_pad_type_0"), val = tensor("custom")]; - tensor input_929_pad_0 = const()[name = tensor("input_929_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_30_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_30_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(577822080)))]; - tensor input_929_cast_fp16 = conv(dilations = var_12368, groups = var_12003, pad = input_929_pad_0, pad_type = input_929_pad_type_0, strides = var_12366, weight = layers_30_fc2_loraA_weight_to_fp16, x = input_927_cast_fp16)[name = tensor("input_929_cast_fp16")]; - tensor var_12372 = const()[name = tensor("op_12372"), val = tensor([1, 1])]; - tensor var_12374 = const()[name = tensor("op_12374"), val = tensor([1, 1])]; - tensor lora_out_1237_pad_type_0 = const()[name = tensor("lora_out_1237_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1237_pad_0 = const()[name = tensor("lora_out_1237_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1239_weight_0_to_fp16 = const()[name = tensor("lora_out_1239_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(577985984)))]; - tensor lora_out_1239_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12374, groups = var_12003, pad = lora_out_1237_pad_0, pad_type = lora_out_1237_pad_type_0, strides = var_12372, weight = lora_out_1239_weight_0_to_fp16, x = input_929_cast_fp16)[name = tensor("lora_out_1239_cast_fp16")]; - tensor hidden_states_63_cast_fp16 = add(x = pretrained_out_619_cast_fp16, y = lora_out_1239_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_12390 = const()[name = tensor("op_12390"), val = tensor(3)]; - tensor var_12397 = const()[name = tensor("op_12397"), val = tensor(1)]; - tensor var_12398 = const()[name = tensor("op_12398"), val = tensor(true)]; - tensor var_12410 = const()[name = tensor("op_12410"), val = tensor([1])]; - tensor channels_mean_187_cast_fp16 = reduce_mean(axes = var_12410, keep_dims = var_12398, 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_12414 = const()[name = tensor("op_12414"), val = tensor([1])]; - tensor var_12415_cast_fp16 = reduce_mean(axes = var_12414, keep_dims = var_12398, x = zero_mean_sq_187_cast_fp16)[name = tensor("op_12415_cast_fp16")]; - tensor var_12416_to_fp16 = const()[name = tensor("op_12416_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_12417_cast_fp16 = add(x = var_12415_cast_fp16, y = var_12416_to_fp16)[name = tensor("op_12417_cast_fp16")]; - tensor denom_187_epsilon_0 = const()[name = tensor("denom_187_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_187_cast_fp16 = rsqrt(epsilon = denom_187_epsilon_0, x = var_12417_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_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(578027008)))]; - 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(578029632)))]; - 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_187_cast_fp16)[name = tensor("obj_373_cast_fp16")]; - tensor var_12435 = const()[name = tensor("op_12435"), val = tensor([1, 1])]; - tensor var_12437 = const()[name = tensor("op_12437"), val = tensor([1, 1])]; - tensor pretrained_out_621_pad_type_0 = const()[name = tensor("pretrained_out_621_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_621_pad_0 = const()[name = tensor("pretrained_out_621_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578032256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578851520))), name = tensor("layers_31_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_31_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578851648)))]; - tensor pretrained_out_621_cast_fp16 = conv(bias = layers_31_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_12437, groups = var_12397, pad = pretrained_out_621_pad_0, pad_type = pretrained_out_621_pad_type_0, strides = var_12435, weight = layers_31_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_373_cast_fp16)[name = tensor("pretrained_out_621_cast_fp16")]; - tensor var_12441 = const()[name = tensor("op_12441"), val = tensor([1, 1])]; - tensor var_12443 = const()[name = tensor("op_12443"), val = tensor([1, 1])]; - tensor input_931_pad_type_0 = const()[name = tensor("input_931_pad_type_0"), val = tensor("custom")]; - tensor input_931_pad_0 = const()[name = tensor("input_931_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578854272)))]; - tensor input_931_cast_fp16 = conv(dilations = var_12443, groups = var_12397, pad = input_931_pad_0, pad_type = input_931_pad_type_0, strides = var_12441, weight = layers_31_self_attn_q_proj_loraA_weight_to_fp16, x = obj_373_cast_fp16)[name = tensor("input_931_cast_fp16")]; - tensor var_12447 = const()[name = tensor("op_12447"), val = tensor([1, 1])]; - tensor var_12449 = const()[name = tensor("op_12449"), val = tensor([1, 1])]; - tensor lora_out_1241_pad_type_0 = const()[name = tensor("lora_out_1241_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1241_pad_0 = const()[name = tensor("lora_out_1241_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1243_weight_0_to_fp16 = const()[name = tensor("lora_out_1243_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578895296)))]; - tensor lora_out_1243_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12449, groups = var_12397, pad = lora_out_1241_pad_0, pad_type = lora_out_1241_pad_type_0, strides = var_12447, weight = lora_out_1243_weight_0_to_fp16, x = input_931_cast_fp16)[name = tensor("lora_out_1243_cast_fp16")]; - tensor query_125_cast_fp16 = add(x = pretrained_out_621_cast_fp16, y = lora_out_1243_cast_fp16)[name = tensor("query_125_cast_fp16")]; - tensor var_12459 = const()[name = tensor("op_12459"), val = tensor([1, 1])]; - tensor var_12461 = const()[name = tensor("op_12461"), val = tensor([1, 1])]; - tensor pretrained_out_623_pad_type_0 = const()[name = tensor("pretrained_out_623_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_623_pad_0 = const()[name = tensor("pretrained_out_623_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578936320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579755584))), name = tensor("layers_31_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_623_cast_fp16 = conv(dilations = var_12461, groups = var_12397, pad = pretrained_out_623_pad_0, pad_type = pretrained_out_623_pad_type_0, strides = var_12459, weight = layers_31_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_373_cast_fp16)[name = tensor("pretrained_out_623_cast_fp16")]; - tensor var_12465 = const()[name = tensor("op_12465"), val = tensor([1, 1])]; - tensor var_12467 = const()[name = tensor("op_12467"), val = tensor([1, 1])]; - tensor input_933_pad_type_0 = const()[name = tensor("input_933_pad_type_0"), val = tensor("custom")]; - tensor input_933_pad_0 = const()[name = tensor("input_933_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579755712)))]; - tensor input_933_cast_fp16 = conv(dilations = var_12467, groups = var_12397, pad = input_933_pad_0, pad_type = input_933_pad_type_0, strides = var_12465, weight = layers_31_self_attn_k_proj_loraA_weight_to_fp16, x = obj_373_cast_fp16)[name = tensor("input_933_cast_fp16")]; - tensor var_12471 = const()[name = tensor("op_12471"), val = tensor([1, 1])]; - tensor var_12473 = const()[name = tensor("op_12473"), val = tensor([1, 1])]; - tensor lora_out_1245_pad_type_0 = const()[name = tensor("lora_out_1245_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1245_pad_0 = const()[name = tensor("lora_out_1245_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1247_weight_0_to_fp16 = const()[name = tensor("lora_out_1247_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579796736)))]; - tensor lora_out_1247_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12473, groups = var_12397, pad = lora_out_1245_pad_0, pad_type = lora_out_1245_pad_type_0, strides = var_12471, weight = lora_out_1247_weight_0_to_fp16, x = input_933_cast_fp16)[name = tensor("lora_out_1247_cast_fp16")]; - tensor current_key_cast_fp16 = add(x = pretrained_out_623_cast_fp16, y = lora_out_1247_cast_fp16)[name = tensor("current_key_cast_fp16")]; - tensor var_12484 = const()[name = tensor("op_12484"), val = tensor([1, 1])]; - tensor var_12486 = const()[name = tensor("op_12486"), val = tensor([1, 1])]; - tensor pretrained_out_625_pad_type_0 = const()[name = tensor("pretrained_out_625_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_625_pad_0 = const()[name = tensor("pretrained_out_625_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579837760))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580657024))), name = tensor("layers_31_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_31_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580657152)))]; - tensor pretrained_out_625_cast_fp16 = conv(bias = layers_31_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_12486, groups = var_12397, pad = pretrained_out_625_pad_0, pad_type = pretrained_out_625_pad_type_0, strides = var_12484, weight = layers_31_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_373_cast_fp16)[name = tensor("pretrained_out_625_cast_fp16")]; - tensor var_12490 = const()[name = tensor("op_12490"), val = tensor([1, 1])]; - tensor var_12492 = const()[name = tensor("op_12492"), val = tensor([1, 1])]; - tensor input_935_pad_type_0 = const()[name = tensor("input_935_pad_type_0"), val = tensor("custom")]; - tensor input_935_pad_0 = const()[name = tensor("input_935_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580659776)))]; - tensor input_935_cast_fp16 = conv(dilations = var_12492, groups = var_12397, pad = input_935_pad_0, pad_type = input_935_pad_type_0, strides = var_12490, weight = layers_31_self_attn_v_proj_loraA_weight_to_fp16, x = obj_373_cast_fp16)[name = tensor("input_935_cast_fp16")]; - tensor var_12496 = const()[name = tensor("op_12496"), val = tensor([1, 1])]; - tensor var_12498 = const()[name = tensor("op_12498"), val = tensor([1, 1])]; - tensor lora_out_1249_pad_type_0 = const()[name = tensor("lora_out_1249_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1249_pad_0 = const()[name = tensor("lora_out_1249_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1251_weight_0_to_fp16 = const()[name = tensor("lora_out_1251_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580700800)))]; - tensor lora_out_1251_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12498, groups = var_12397, pad = lora_out_1249_pad_0, pad_type = lora_out_1249_pad_type_0, strides = var_12496, weight = lora_out_1251_weight_0_to_fp16, x = input_935_cast_fp16)[name = tensor("lora_out_1251_cast_fp16")]; - tensor current_value_cast_fp16 = add(x = pretrained_out_625_cast_fp16, y = lora_out_1251_cast_fp16)[name = tensor("current_value_cast_fp16")]; - tensor var_12508_cast_fp16 = mul(x = current_key_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_12508_cast_fp16")]; - tensor var_12510_cast_fp16 = mul(x = var_103_cast_fp16_31, y = var_295_cast_fp16)[name = tensor("op_12510_cast_fp16")]; - tensor key_125_cast_fp16 = add(x = var_12508_cast_fp16, y = var_12510_cast_fp16)[name = tensor("key_125_cast_fp16")]; - tensor var_12512_cast_fp16 = mul(x = current_value_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_12512_cast_fp16")]; - tensor var_12514_cast_fp16 = mul(x = var_138_cast_fp16_31, y = var_295_cast_fp16)[name = tensor("op_12514_cast_fp16")]; - tensor value_125_cast_fp16 = add(x = var_12512_cast_fp16, y = var_12514_cast_fp16)[name = tensor("value_125_cast_fp16")]; - tensor var_12517 = const()[name = tensor("op_12517"), val = tensor([1, 20, 64, -1])]; - tensor var_12518_cast_fp16 = reshape(shape = var_12517, x = query_125_cast_fp16)[name = tensor("op_12518_cast_fp16")]; - tensor var_12519_to_fp16 = const()[name = tensor("op_12519_to_fp16"), val = tensor(0x1p-3)]; - tensor var_12520_cast_fp16 = mul(x = var_12518_cast_fp16, y = var_12519_to_fp16)[name = tensor("op_12520_cast_fp16")]; - tensor var_12521 = const()[name = tensor("op_12521"), val = tensor([1, 20, 64, -1])]; - tensor var_12522_cast_fp16 = reshape(shape = var_12521, x = key_125_cast_fp16)[name = tensor("op_12522_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_12520_cast_fp16, y = var_12522_cast_fp16)[name = tensor("mh_w_187_cast_fp16")]; - tensor mh_w_189_cast_fp16 = add(x = mh_w_187_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_189_cast_fp16")]; - tensor var_12530_cast_fp16 = softmax(axis = var_12390, x = mh_w_189_cast_fp16)[name = tensor("op_12530_cast_fp16")]; - tensor var_12531 = const()[name = tensor("op_12531"), val = tensor([1, 20, 64, -1])]; - tensor var_12532_cast_fp16 = reshape(shape = var_12531, x = value_125_cast_fp16)[name = tensor("op_12532_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_12532_cast_fp16, y = var_12530_cast_fp16)[name = tensor("attn_125_cast_fp16")]; - tensor var_12535 = const()[name = tensor("op_12535"), val = tensor([1, 1280, 1, -1])]; - tensor input_937_cast_fp16 = reshape(shape = var_12535, x = attn_125_cast_fp16)[name = tensor("input_937_cast_fp16")]; - tensor var_12542 = const()[name = tensor("op_12542"), val = tensor([1, 1])]; - tensor var_12544 = const()[name = tensor("op_12544"), val = tensor([1, 1])]; - tensor pretrained_out_627_pad_type_0 = const()[name = tensor("pretrained_out_627_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_627_pad_0 = const()[name = tensor("pretrained_out_627_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580741824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(581561088))), name = tensor("layers_31_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_31_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(581561216)))]; - tensor pretrained_out_627_cast_fp16 = conv(bias = layers_31_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_12544, groups = var_12397, pad = pretrained_out_627_pad_0, pad_type = pretrained_out_627_pad_type_0, strides = var_12542, weight = layers_31_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_937_cast_fp16)[name = tensor("pretrained_out_627_cast_fp16")]; - tensor var_12548 = const()[name = tensor("op_12548"), val = tensor([1, 1])]; - tensor var_12550 = const()[name = tensor("op_12550"), val = tensor([1, 1])]; - tensor input_939_pad_type_0 = const()[name = tensor("input_939_pad_type_0"), val = tensor("custom")]; - tensor input_939_pad_0 = const()[name = tensor("input_939_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(581563840)))]; - tensor input_939_cast_fp16 = conv(dilations = var_12550, groups = var_12397, pad = input_939_pad_0, pad_type = input_939_pad_type_0, strides = var_12548, weight = layers_31_self_attn_o_proj_loraA_weight_to_fp16, x = input_937_cast_fp16)[name = tensor("input_939_cast_fp16")]; - tensor var_12554 = const()[name = tensor("op_12554"), val = tensor([1, 1])]; - tensor var_12556 = const()[name = tensor("op_12556"), val = tensor([1, 1])]; - tensor lora_out_1253_pad_type_0 = const()[name = tensor("lora_out_1253_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1253_pad_0 = const()[name = tensor("lora_out_1253_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1255_weight_0_to_fp16 = const()[name = tensor("lora_out_1255_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(581604864)))]; - tensor lora_out_1255_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12556, groups = var_12397, pad = lora_out_1253_pad_0, pad_type = lora_out_1253_pad_type_0, strides = var_12554, weight = lora_out_1255_weight_0_to_fp16, x = input_939_cast_fp16)[name = tensor("lora_out_1255_cast_fp16")]; - tensor obj_379_cast_fp16 = add(x = pretrained_out_627_cast_fp16, y = lora_out_1255_cast_fp16)[name = tensor("obj_379_cast_fp16")]; - tensor inputs_189_cast_fp16 = add(x = inputs_187_cast_fp16, y = obj_379_cast_fp16)[name = tensor("inputs_189_cast_fp16")]; - tensor var_12569 = const()[name = tensor("op_12569"), val = tensor([1])]; - tensor channels_mean_189_cast_fp16 = reduce_mean(axes = var_12569, keep_dims = var_12398, 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_12573 = const()[name = tensor("op_12573"), val = tensor([1])]; - tensor var_12574_cast_fp16 = reduce_mean(axes = var_12573, keep_dims = var_12398, x = zero_mean_sq_189_cast_fp16)[name = tensor("op_12574_cast_fp16")]; - tensor var_12575_to_fp16 = const()[name = tensor("op_12575_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_12576_cast_fp16 = add(x = var_12574_cast_fp16, y = var_12575_to_fp16)[name = tensor("op_12576_cast_fp16")]; - tensor denom_189_epsilon_0 = const()[name = tensor("denom_189_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_189_cast_fp16 = rsqrt(epsilon = denom_189_epsilon_0, x = var_12576_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_381_gamma_0_to_fp16 = const()[name = tensor("obj_381_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(581645888)))]; - tensor obj_381_beta_0_to_fp16 = const()[name = tensor("obj_381_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(581648512)))]; - tensor obj_381_epsilon_0_to_fp16 = const()[name = tensor("obj_381_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor obj_381_cast_fp16 = batch_norm(beta = obj_381_beta_0_to_fp16, epsilon = obj_381_epsilon_0_to_fp16, gamma = obj_381_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_381_cast_fp16")]; - tensor var_12594 = const()[name = tensor("op_12594"), val = tensor([1, 1])]; - tensor var_12596 = const()[name = tensor("op_12596"), val = tensor([1, 1])]; - tensor pretrained_out_629_pad_type_0 = const()[name = tensor("pretrained_out_629_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_629_pad_0 = const()[name = tensor("pretrained_out_629_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(581651136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(582470400))), name = tensor("layers_31_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_31_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_31_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(582470528)))]; - tensor pretrained_out_629_cast_fp16 = conv(bias = layers_31_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_12596, groups = var_12397, pad = pretrained_out_629_pad_0, pad_type = pretrained_out_629_pad_type_0, strides = var_12594, weight = layers_31_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_381_cast_fp16)[name = tensor("pretrained_out_629_cast_fp16")]; - tensor var_12600 = const()[name = tensor("op_12600"), val = tensor([1, 1])]; - tensor var_12602 = const()[name = tensor("op_12602"), val = tensor([1, 1])]; - tensor input_941_pad_type_0 = const()[name = tensor("input_941_pad_type_0"), val = tensor("custom")]; - tensor input_941_pad_0 = const()[name = tensor("input_941_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_31_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(582473152)))]; - tensor input_941_cast_fp16 = conv(dilations = var_12602, groups = var_12397, pad = input_941_pad_0, pad_type = input_941_pad_type_0, strides = var_12600, weight = layers_31_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_381_cast_fp16)[name = tensor("input_941_cast_fp16")]; - tensor var_12606 = const()[name = tensor("op_12606"), val = tensor([1, 1])]; - tensor var_12608 = const()[name = tensor("op_12608"), val = tensor([1, 1])]; - tensor lora_out_1257_pad_type_0 = const()[name = tensor("lora_out_1257_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1257_pad_0 = const()[name = tensor("lora_out_1257_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1259_weight_0_to_fp16 = const()[name = tensor("lora_out_1259_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(582514176)))]; - tensor lora_out_1259_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12608, groups = var_12397, pad = lora_out_1257_pad_0, pad_type = lora_out_1257_pad_type_0, strides = var_12606, weight = lora_out_1259_weight_0_to_fp16, x = input_941_cast_fp16)[name = tensor("lora_out_1259_cast_fp16")]; - tensor query_cast_fp16 = add(x = pretrained_out_629_cast_fp16, y = lora_out_1259_cast_fp16)[name = tensor("query_cast_fp16")]; - tensor var_12618 = const()[name = tensor("op_12618"), val = tensor([1, 1])]; - tensor var_12620 = const()[name = tensor("op_12620"), val = tensor([1, 1])]; - tensor pretrained_out_631_pad_type_0 = const()[name = tensor("pretrained_out_631_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_631_pad_0 = const()[name = tensor("pretrained_out_631_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(582555200))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(583374464))), name = tensor("layers_31_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor pretrained_out_631_cast_fp16 = conv(dilations = var_12620, groups = var_12397, pad = pretrained_out_631_pad_0, pad_type = pretrained_out_631_pad_type_0, strides = var_12618, weight = layers_31_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_631_cast_fp16")]; - tensor var_12624 = const()[name = tensor("op_12624"), val = tensor([1, 1])]; - tensor var_12626 = const()[name = tensor("op_12626"), val = tensor([1, 1])]; - tensor input_943_pad_type_0 = const()[name = tensor("input_943_pad_type_0"), val = tensor("custom")]; - tensor input_943_pad_0 = const()[name = tensor("input_943_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_31_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(583374592)))]; - tensor input_943_cast_fp16 = conv(dilations = var_12626, groups = var_12397, pad = input_943_pad_0, pad_type = input_943_pad_type_0, strides = var_12624, weight = layers_31_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_943_cast_fp16")]; - tensor var_12630 = const()[name = tensor("op_12630"), val = tensor([1, 1])]; - tensor var_12632 = const()[name = tensor("op_12632"), val = tensor([1, 1])]; - tensor lora_out_1261_pad_type_0 = const()[name = tensor("lora_out_1261_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1261_pad_0 = const()[name = tensor("lora_out_1261_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1263_weight_0_to_fp16 = const()[name = tensor("lora_out_1263_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(583415616)))]; - tensor lora_out_1263_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12632, groups = var_12397, pad = lora_out_1261_pad_0, pad_type = lora_out_1261_pad_type_0, strides = var_12630, weight = lora_out_1263_weight_0_to_fp16, x = input_943_cast_fp16)[name = tensor("lora_out_1263_cast_fp16")]; - tensor key_cast_fp16 = add(x = pretrained_out_631_cast_fp16, y = lora_out_1263_cast_fp16)[name = tensor("key_cast_fp16")]; - tensor var_12643 = const()[name = tensor("op_12643"), val = tensor([1, 1])]; - tensor var_12645 = const()[name = tensor("op_12645"), val = tensor([1, 1])]; - tensor pretrained_out_633_pad_type_0 = const()[name = tensor("pretrained_out_633_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_633_pad_0 = const()[name = tensor("pretrained_out_633_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(583456640))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(584275904))), name = tensor("layers_31_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_31_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_31_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(584276032)))]; - tensor pretrained_out_633_cast_fp16 = conv(bias = layers_31_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_12645, groups = var_12397, pad = pretrained_out_633_pad_0, pad_type = pretrained_out_633_pad_type_0, strides = var_12643, weight = layers_31_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_633_cast_fp16")]; - tensor var_12649 = const()[name = tensor("op_12649"), val = tensor([1, 1])]; - tensor var_12651 = const()[name = tensor("op_12651"), val = tensor([1, 1])]; - tensor input_945_pad_type_0 = const()[name = tensor("input_945_pad_type_0"), val = tensor("custom")]; - tensor input_945_pad_0 = const()[name = tensor("input_945_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_31_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(584278656)))]; - tensor input_945_cast_fp16 = conv(dilations = var_12651, groups = var_12397, pad = input_945_pad_0, pad_type = input_945_pad_type_0, strides = var_12649, weight = layers_31_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_945_cast_fp16")]; - tensor var_12655 = const()[name = tensor("op_12655"), val = tensor([1, 1])]; - tensor var_12657 = const()[name = tensor("op_12657"), val = tensor([1, 1])]; - tensor lora_out_1265_pad_type_0 = const()[name = tensor("lora_out_1265_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1265_pad_0 = const()[name = tensor("lora_out_1265_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1267_weight_0_to_fp16 = const()[name = tensor("lora_out_1267_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(584319680)))]; - tensor lora_out_1267_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12657, groups = var_12397, pad = lora_out_1265_pad_0, pad_type = lora_out_1265_pad_type_0, strides = var_12655, weight = lora_out_1267_weight_0_to_fp16, x = input_945_cast_fp16)[name = tensor("lora_out_1267_cast_fp16")]; - tensor value_cast_fp16 = add(x = pretrained_out_633_cast_fp16, y = lora_out_1267_cast_fp16)[name = tensor("value_cast_fp16")]; - tensor var_12664 = const()[name = tensor("op_12664"), val = tensor([1, 20, 64, -1])]; - tensor var_12665_cast_fp16 = reshape(shape = var_12664, x = query_cast_fp16)[name = tensor("op_12665_cast_fp16")]; - tensor var_12666_to_fp16 = const()[name = tensor("op_12666_to_fp16"), val = tensor(0x1p-3)]; - tensor var_12667_cast_fp16 = mul(x = var_12665_cast_fp16, y = var_12666_to_fp16)[name = tensor("op_12667_cast_fp16")]; - tensor var_12668 = const()[name = tensor("op_12668"), val = tensor([1, 20, 64, -1])]; - tensor var_12669_cast_fp16 = reshape(shape = var_12668, x = key_cast_fp16)[name = tensor("op_12669_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_12667_cast_fp16, y = var_12669_cast_fp16)[name = tensor("mh_w_cast_fp16")]; - tensor var_12672_cast_fp16 = softmax(axis = var_12390, x = mh_w_cast_fp16)[name = tensor("op_12672_cast_fp16")]; - tensor var_12673 = const()[name = tensor("op_12673"), val = tensor([1, 20, 64, -1])]; - tensor var_12674_cast_fp16 = reshape(shape = var_12673, x = value_cast_fp16)[name = tensor("op_12674_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_12674_cast_fp16, y = var_12672_cast_fp16)[name = tensor("attn_cast_fp16")]; - tensor var_12677 = const()[name = tensor("op_12677"), val = tensor([1, 1280, 1, -1])]; - tensor input_947_cast_fp16 = reshape(shape = var_12677, x = attn_cast_fp16)[name = tensor("input_947_cast_fp16")]; - tensor var_12684 = const()[name = tensor("op_12684"), val = tensor([1, 1])]; - tensor var_12686 = const()[name = tensor("op_12686"), val = tensor([1, 1])]; - tensor pretrained_out_635_pad_type_0 = const()[name = tensor("pretrained_out_635_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_635_pad_0 = const()[name = tensor("pretrained_out_635_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(584360704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585179968))), name = tensor("layers_31_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; - tensor layers_31_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_31_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585180096)))]; - tensor pretrained_out_635_cast_fp16 = conv(bias = layers_31_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_12686, groups = var_12397, pad = pretrained_out_635_pad_0, pad_type = pretrained_out_635_pad_type_0, strides = var_12684, weight = layers_31_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_947_cast_fp16)[name = tensor("pretrained_out_635_cast_fp16")]; - tensor var_12690 = const()[name = tensor("op_12690"), val = tensor([1, 1])]; - tensor var_12692 = const()[name = tensor("op_12692"), val = tensor([1, 1])]; - tensor input_949_pad_type_0 = const()[name = tensor("input_949_pad_type_0"), val = tensor("custom")]; - tensor input_949_pad_0 = const()[name = tensor("input_949_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_31_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585182720)))]; - tensor input_949_cast_fp16 = conv(dilations = var_12692, groups = var_12397, pad = input_949_pad_0, pad_type = input_949_pad_type_0, strides = var_12690, weight = layers_31_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_947_cast_fp16)[name = tensor("input_949_cast_fp16")]; - tensor var_12696 = const()[name = tensor("op_12696"), val = tensor([1, 1])]; - tensor var_12698 = const()[name = tensor("op_12698"), val = tensor([1, 1])]; - tensor lora_out_1269_pad_type_0 = const()[name = tensor("lora_out_1269_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1269_pad_0 = const()[name = tensor("lora_out_1269_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1271_weight_0_to_fp16 = const()[name = tensor("lora_out_1271_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585223744)))]; - tensor lora_out_1271_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12698, groups = var_12397, pad = lora_out_1269_pad_0, pad_type = lora_out_1269_pad_type_0, strides = var_12696, weight = lora_out_1271_weight_0_to_fp16, x = input_949_cast_fp16)[name = tensor("lora_out_1271_cast_fp16")]; - tensor obj_383_cast_fp16 = add(x = pretrained_out_635_cast_fp16, y = lora_out_1271_cast_fp16)[name = tensor("obj_383_cast_fp16")]; - tensor inputs_191_cast_fp16 = add(x = inputs_189_cast_fp16, y = obj_383_cast_fp16)[name = tensor("inputs_191_cast_fp16")]; - tensor var_12707 = const()[name = tensor("op_12707"), val = tensor([1])]; - tensor channels_mean_191_cast_fp16 = reduce_mean(axes = var_12707, keep_dims = var_12398, 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_12711 = const()[name = tensor("op_12711"), val = tensor([1])]; - tensor var_12712_cast_fp16 = reduce_mean(axes = var_12711, keep_dims = var_12398, x = zero_mean_sq_191_cast_fp16)[name = tensor("op_12712_cast_fp16")]; - tensor var_12713_to_fp16 = const()[name = tensor("op_12713_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_12714_cast_fp16 = add(x = var_12712_cast_fp16, y = var_12713_to_fp16)[name = tensor("op_12714_cast_fp16")]; - tensor denom_191_epsilon_0 = const()[name = tensor("denom_191_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_191_cast_fp16 = rsqrt(epsilon = denom_191_epsilon_0, x = var_12714_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_951_gamma_0_to_fp16 = const()[name = tensor("input_951_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585264768)))]; - tensor input_951_beta_0_to_fp16 = const()[name = tensor("input_951_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585267392)))]; - tensor input_951_epsilon_0_to_fp16 = const()[name = tensor("input_951_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; - tensor input_951_cast_fp16 = batch_norm(beta = input_951_beta_0_to_fp16, epsilon = input_951_epsilon_0_to_fp16, gamma = input_951_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_951_cast_fp16")]; - tensor var_12728 = const()[name = tensor("op_12728"), val = tensor([1, 1])]; - tensor var_12730 = const()[name = tensor("op_12730"), val = tensor([1, 1])]; - tensor pretrained_out_637_pad_type_0 = const()[name = tensor("pretrained_out_637_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_637_pad_0 = const()[name = tensor("pretrained_out_637_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585270016))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588546880))), name = tensor("layers_31_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; - tensor layers_31_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_31_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588547008)))]; - tensor pretrained_out_637_cast_fp16 = conv(bias = layers_31_fc1_pretrained_bias_to_fp16, dilations = var_12730, groups = var_12397, pad = pretrained_out_637_pad_0, pad_type = pretrained_out_637_pad_type_0, strides = var_12728, weight = layers_31_fc1_pretrained_weight_to_fp16_palettized, x = input_951_cast_fp16)[name = tensor("pretrained_out_637_cast_fp16")]; - tensor var_12734 = const()[name = tensor("op_12734"), val = tensor([1, 1])]; - tensor var_12736 = const()[name = tensor("op_12736"), val = tensor([1, 1])]; - tensor input_953_pad_type_0 = const()[name = tensor("input_953_pad_type_0"), val = tensor("custom")]; - tensor input_953_pad_0 = const()[name = tensor("input_953_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_31_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588557312)))]; - tensor input_953_cast_fp16 = conv(dilations = var_12736, groups = var_12397, pad = input_953_pad_0, pad_type = input_953_pad_type_0, strides = var_12734, weight = layers_31_fc1_loraA_weight_to_fp16, x = input_951_cast_fp16)[name = tensor("input_953_cast_fp16")]; - tensor var_12740 = const()[name = tensor("op_12740"), val = tensor([1, 1])]; - tensor var_12742 = const()[name = tensor("op_12742"), val = tensor([1, 1])]; - tensor lora_out_1273_pad_type_0 = const()[name = tensor("lora_out_1273_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1273_pad_0 = const()[name = tensor("lora_out_1273_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_1275_weight_0_to_fp16 = const()[name = tensor("lora_out_1275_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588598336)))]; - tensor lora_out_1275_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_12742, groups = var_12397, pad = lora_out_1273_pad_0, pad_type = lora_out_1273_pad_type_0, strides = var_12740, weight = lora_out_1275_weight_0_to_fp16, x = input_953_cast_fp16)[name = tensor("lora_out_1275_cast_fp16")]; - tensor input_955_cast_fp16 = add(x = pretrained_out_637_cast_fp16, y = lora_out_1275_cast_fp16)[name = tensor("input_955_cast_fp16")]; - tensor input_957_mode_0 = const()[name = tensor("input_957_mode_0"), val = tensor("EXACT")]; - tensor input_957_cast_fp16 = gelu(mode = input_957_mode_0, x = input_955_cast_fp16)[name = tensor("input_957_cast_fp16")]; - tensor var_12754 = const()[name = tensor("op_12754"), val = tensor([1, 1])]; - tensor var_12756 = const()[name = tensor("op_12756"), val = tensor([1, 1])]; - tensor pretrained_out_pad_type_0 = const()[name = tensor("pretrained_out_pad_type_0"), val = tensor("custom")]; - tensor pretrained_out_pad_0 = const()[name = tensor("pretrained_out_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588762240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(592039104))), name = tensor("layers_31_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; - tensor layers_31_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_31_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(592039232)))]; - tensor pretrained_out_cast_fp16 = conv(bias = layers_31_fc2_pretrained_bias_to_fp16, dilations = var_12756, groups = var_12397, pad = pretrained_out_pad_0, pad_type = pretrained_out_pad_type_0, strides = var_12754, weight = layers_31_fc2_pretrained_weight_to_fp16_palettized, x = input_957_cast_fp16)[name = tensor("pretrained_out_cast_fp16")]; - tensor var_12760 = const()[name = tensor("op_12760"), val = tensor([1, 1])]; - tensor var_12762 = const()[name = tensor("op_12762"), val = tensor([1, 1])]; - tensor input_pad_type_0 = const()[name = tensor("input_pad_type_0"), val = tensor("custom")]; - tensor input_pad_0 = const()[name = tensor("input_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor layers_31_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_31_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(592041856)))]; - tensor input_cast_fp16 = conv(dilations = var_12762, groups = var_12397, pad = input_pad_0, pad_type = input_pad_type_0, strides = var_12760, weight = layers_31_fc2_loraA_weight_to_fp16, x = input_957_cast_fp16)[name = tensor("input_cast_fp16")]; - tensor var_12766 = const()[name = tensor("op_12766"), val = tensor([1, 1])]; - tensor var_12768 = const()[name = tensor("op_12768"), val = tensor([1, 1])]; - tensor lora_out_1277_pad_type_0 = const()[name = tensor("lora_out_1277_pad_type_0"), val = tensor("custom")]; - tensor lora_out_1277_pad_0 = const()[name = tensor("lora_out_1277_pad_0"), val = tensor([0, 0, 0, 0])]; - tensor lora_out_weight_0_to_fp16 = const()[name = tensor("lora_out_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(592205760)))]; - tensor lora_out_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12768, groups = var_12397, pad = lora_out_1277_pad_0, pad_type = lora_out_1277_pad_type_0, strides = var_12766, weight = lora_out_weight_0_to_fp16, x = input_cast_fp16)[name = tensor("lora_out_cast_fp16")]; - tensor hidden_states_65_cast_fp16 = add(x = pretrained_out_cast_fp16, y = lora_out_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_12781 = const()[name = tensor("op_12781"), val = tensor(true)]; - tensor var_12785 = const()[name = tensor("op_12785"), val = tensor([1])]; - tensor channels_mean_cast_fp16 = reduce_mean(axes = var_12785, keep_dims = var_12781, 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_12789 = const()[name = tensor("op_12789"), val = tensor([1])]; - tensor var_12790_cast_fp16 = reduce_mean(axes = var_12789, keep_dims = var_12781, x = zero_mean_sq_cast_fp16)[name = tensor("op_12790_cast_fp16")]; - tensor var_12791_to_fp16 = const()[name = tensor("op_12791_to_fp16"), val = tensor(0x1.5p-17)]; - tensor var_12792_cast_fp16 = add(x = var_12790_cast_fp16, y = var_12791_to_fp16)[name = tensor("op_12792_cast_fp16")]; - tensor denom_epsilon_0 = const()[name = tensor("denom_epsilon_0"), val = tensor(0x1.197998p-40)]; - tensor denom_cast_fp16 = rsqrt(epsilon = denom_epsilon_0, x = var_12792_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(592246784)))]; - 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(592249408)))]; - 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_12802_axes_0 = const()[name = tensor("op_12802_axes_0"), val = tensor([2])]; - tensor var_12802_cast_fp16 = squeeze(axes = var_12802_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_12802_cast_fp16")]; - tensor var_12805_perm_0 = const()[name = tensor("op_12805_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(592252032)))]; - tensor transpose_0 = transpose(perm = var_12805_perm_0, x = var_12802_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_12809 = const()[name = tensor("op_12809"), val = tensor(1)]; - tensor obj_387_interleave_0 = const()[name = tensor("obj_387_interleave_0"), val = tensor(false)]; - tensor key_cache_updates = concat(axis = var_12809, interleave = obj_387_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_387_cast_fp16")]; - tensor var_12812 = const()[name = tensor("op_12812"), val = tensor(1)]; - tensor obj_interleave_0 = const()[name = tensor("obj_interleave_0"), val = tensor(false)]; - tensor value_cache_updates = concat(axis = var_12812, interleave = obj_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_cast_fp16")]; - } -> (logits, key_cache_updates, value_cache_updates); -} \ No newline at end of file