diff --git "a/LocalDoc_azerbaijani-whisper-small/TextDecoder.mlmodelc/model.mil" "b/LocalDoc_azerbaijani-whisper-small/TextDecoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/LocalDoc_azerbaijani-whisper-small/TextDecoder.mlmodelc/model.mil" @@ -0,0 +1,1904 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + 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_40_axis_0 = const()[name = tensor("op_40_axis_0"), val = tensor(0)]; + tensor var_40_batch_dims_0 = const()[name = tensor("op_40_batch_dims_0"), val = tensor(0)]; + tensor embed_tokens_weight_to_fp16 = const()[name = tensor("embed_tokens_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor var_40_cast_fp16 = gather(axis = var_40_axis_0, batch_dims = var_40_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor("op_40_cast_fp16")]; + tensor var_44_axis_0 = const()[name = tensor("op_44_axis_0"), val = tensor(0)]; + tensor var_44_batch_dims_0 = const()[name = tensor("op_44_batch_dims_0"), val = tensor(0)]; + tensor embed_positions_weight_to_fp16 = const()[name = tensor("embed_positions_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79664768)))]; + tensor var_44_cast_fp16 = gather(axis = var_44_axis_0, batch_dims = var_44_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor("op_44_cast_fp16")]; + tensor hidden_states_1_cast_fp16 = add(x = var_40_cast_fp16, y = var_44_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; + tensor var_58_axes_0 = const()[name = tensor("op_58_axes_0"), val = tensor([2])]; + tensor var_58_cast_fp16 = expand_dims(axes = var_58_axes_0, x = hidden_states_1_cast_fp16)[name = tensor("op_58_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_58_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([768, 768, 768, 768, 768, 768, 768, 768, 768, 768, 768, 768])]; + tensor var_63_axis_0 = const()[name = tensor("op_63_axis_0"), val = tensor(1)]; + tensor var_63_cast_fp16_0, tensor var_63_cast_fp16_1, tensor var_63_cast_fp16_2, tensor var_63_cast_fp16_3, tensor var_63_cast_fp16_4, tensor var_63_cast_fp16_5, tensor var_63_cast_fp16_6, tensor var_63_cast_fp16_7, tensor var_63_cast_fp16_8, tensor var_63_cast_fp16_9, tensor var_63_cast_fp16_10, tensor var_63_cast_fp16_11 = split(axis = var_63_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor("op_63_cast_fp16")]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([768, 768, 768, 768, 768, 768, 768, 768, 768, 768, 768, 768])]; + tensor var_78_axis_0 = const()[name = tensor("op_78_axis_0"), val = tensor(1)]; + tensor var_78_cast_fp16_0, tensor var_78_cast_fp16_1, tensor var_78_cast_fp16_2, tensor var_78_cast_fp16_3, tensor var_78_cast_fp16_4, tensor var_78_cast_fp16_5, tensor var_78_cast_fp16_6, tensor var_78_cast_fp16_7, tensor var_78_cast_fp16_8, tensor var_78_cast_fp16_9, tensor var_78_cast_fp16_10, tensor var_78_cast_fp16_11 = split(axis = var_78_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor("op_78_cast_fp16")]; + tensor var_96 = const()[name = tensor("op_96"), val = tensor(3)]; + tensor out_1_axes_0 = const()[name = tensor("out_1_axes_0"), val = tensor([1])]; + tensor var_121_to_fp16 = const()[name = tensor("op_121_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_121_to_fp16, x = inputs_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(80352960)))]; + 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(80354560)))]; + 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(80356160)))]; + 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(80357760)))]; + 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 query_1_pad_type_0 = const()[name = tensor("query_1_pad_type_0"), val = tensor("valid")]; + tensor query_1_strides_0 = const()[name = tensor("query_1_strides_0"), val = tensor([1, 1])]; + tensor query_1_pad_0 = const()[name = tensor("query_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_1_dilations_0 = const()[name = tensor("query_1_dilations_0"), val = tensor([1, 1])]; + tensor query_1_groups_0 = const()[name = tensor("query_1_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80359360)))]; + tensor layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81539072)))]; + tensor query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = query_1_dilations_0, groups = query_1_groups_0, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = query_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("query_1_cast_fp16")]; + tensor current_key_1_pad_type_0 = const()[name = tensor("current_key_1_pad_type_0"), val = tensor("valid")]; + tensor current_key_1_strides_0 = const()[name = tensor("current_key_1_strides_0"), val = tensor([1, 1])]; + tensor current_key_1_pad_0 = const()[name = tensor("current_key_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_1_dilations_0 = const()[name = tensor("current_key_1_dilations_0"), val = tensor([1, 1])]; + tensor current_key_1_groups_0 = const()[name = tensor("current_key_1_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81540672)))]; + tensor current_key_1_cast_fp16 = conv(dilations = current_key_1_dilations_0, groups = current_key_1_groups_0, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = current_key_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("current_key_1_cast_fp16")]; + tensor current_value_1_pad_type_0 = const()[name = tensor("current_value_1_pad_type_0"), val = tensor("valid")]; + tensor current_value_1_strides_0 = const()[name = tensor("current_value_1_strides_0"), val = tensor([1, 1])]; + tensor current_value_1_pad_0 = const()[name = tensor("current_value_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_1_dilations_0 = const()[name = tensor("current_value_1_dilations_0"), val = tensor([1, 1])]; + tensor current_value_1_groups_0 = const()[name = tensor("current_value_1_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82720384)))]; + tensor layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83900096)))]; + tensor current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = current_value_1_dilations_0, groups = current_value_1_groups_0, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = current_value_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("current_value_1_cast_fp16")]; + tensor var_156_axes_0 = const()[name = tensor("op_156_axes_0"), val = tensor([1])]; + tensor var_156_cast_fp16 = expand_dims(axes = var_156_axes_0, x = kv_cache_update_mask)[name = tensor("op_156_cast_fp16")]; + tensor var_157_axes_0 = const()[name = tensor("op_157_axes_0"), val = tensor([2])]; + tensor var_157_cast_fp16 = expand_dims(axes = var_157_axes_0, x = var_156_cast_fp16)[name = tensor("op_157_cast_fp16")]; + tensor var_97_to_fp16 = const()[name = tensor("op_97_to_fp16"), val = tensor(0x1p+0)]; + tensor var_159_cast_fp16 = sub(x = var_97_to_fp16, y = var_157_cast_fp16)[name = tensor("op_159_cast_fp16")]; + tensor var_160_cast_fp16 = mul(x = var_63_cast_fp16_0, y = var_159_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor var_161_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_161_cast_fp16")]; + tensor key_1_cast_fp16 = add(x = var_160_cast_fp16, y = var_161_cast_fp16)[name = tensor("key_1_cast_fp16")]; + tensor var_164_cast_fp16 = mul(x = var_78_cast_fp16_0, y = var_159_cast_fp16)[name = tensor("op_164_cast_fp16")]; + tensor var_165_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_165_cast_fp16")]; + tensor value_1_cast_fp16 = add(x = var_164_cast_fp16, y = var_165_cast_fp16)[name = tensor("value_1_cast_fp16")]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_1_cast_fp16 = reshape(shape = var_169, x = query_1_cast_fp16)[name = tensor("mh_q_1_cast_fp16")]; + tensor var_171_to_fp16 = const()[name = tensor("op_171_to_fp16"), val = tensor(0x1p-3)]; + tensor var_172_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_171_to_fp16)[name = tensor("op_172_cast_fp16")]; + tensor var_175 = const()[name = tensor("op_175"), val = tensor([1, 12, 64, 448])]; + tensor var_176_cast_fp16 = reshape(shape = var_175, x = key_1_cast_fp16)[name = tensor("op_176_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_172_cast_fp16, y = var_176_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; + tensor var_180_axes_0 = const()[name = tensor("op_180_axes_0"), val = tensor([1])]; + tensor var_180_cast_fp16 = expand_dims(axes = var_180_axes_0, x = decoder_key_padding_mask)[name = tensor("op_180_cast_fp16")]; + tensor var_181_axes_0 = const()[name = tensor("op_181_axes_0"), val = tensor([2])]; + tensor var_181_cast_fp16 = expand_dims(axes = var_181_axes_0, x = var_180_cast_fp16)[name = tensor("op_181_cast_fp16")]; + tensor mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_181_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; + tensor var_184_cast_fp16 = softmax(axis = var_96, x = mh_w_3_cast_fp16)[name = tensor("op_184_cast_fp16")]; + tensor var_185 = const()[name = tensor("op_185"), val = tensor([1, 12, 64, 448])]; + tensor var_186_cast_fp16 = reshape(shape = var_185, x = value_1_cast_fp16)[name = tensor("op_186_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_186_cast_fp16, y = var_184_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_189 = const()[name = tensor("op_189"), val = tensor([1, 768, 1, 1])]; + tensor input_1_cast_fp16 = reshape(shape = var_189, x = attn_1_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor obj_7_pad_type_0 = const()[name = tensor("obj_7_pad_type_0"), val = tensor("valid")]; + tensor obj_7_strides_0 = const()[name = tensor("obj_7_strides_0"), val = tensor([1, 1])]; + tensor obj_7_pad_0 = const()[name = tensor("obj_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_7_dilations_0 = const()[name = tensor("obj_7_dilations_0"), val = tensor([1, 1])]; + tensor obj_7_groups_0 = const()[name = tensor("obj_7_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83901696)))]; + tensor layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85081408)))]; + tensor obj_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = obj_7_dilations_0, groups = obj_7_groups_0, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = obj_7_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("obj_7_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([1])]; + tensor var_211_to_fp16 = const()[name = tensor("op_211_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_211_to_fp16, x = inputs_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(85083008)))]; + 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(85084608)))]; + 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 query_3_pad_type_0 = const()[name = tensor("query_3_pad_type_0"), val = tensor("valid")]; + tensor query_3_strides_0 = const()[name = tensor("query_3_strides_0"), val = tensor([1, 1])]; + tensor query_3_pad_0 = const()[name = tensor("query_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_3_dilations_0 = const()[name = tensor("query_3_dilations_0"), val = tensor([1, 1])]; + tensor query_3_groups_0 = const()[name = tensor("query_3_groups_0"), val = tensor(1)]; + tensor layers_0_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85086208)))]; + tensor layers_0_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86265920)))]; + tensor query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = query_3_dilations_0, groups = query_3_groups_0, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = query_3_strides_0, weight = layers_0_encoder_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("query_3_cast_fp16")]; + tensor key_3_pad_type_0 = const()[name = tensor("key_3_pad_type_0"), val = tensor("valid")]; + tensor key_3_strides_0 = const()[name = tensor("key_3_strides_0"), val = tensor([1, 1])]; + tensor key_3_pad_0 = const()[name = tensor("key_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_3_dilations_0 = const()[name = tensor("key_3_dilations_0"), val = tensor([1, 1])]; + tensor key_3_groups_0 = const()[name = tensor("key_3_groups_0"), val = tensor(1)]; + tensor layers_0_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86267520)))]; + tensor key_3_cast_fp16 = conv(dilations = key_3_dilations_0, groups = key_3_groups_0, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = key_3_strides_0, weight = layers_0_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_3_cast_fp16")]; + tensor value_3_pad_type_0 = const()[name = tensor("value_3_pad_type_0"), val = tensor("valid")]; + tensor value_3_strides_0 = const()[name = tensor("value_3_strides_0"), val = tensor([1, 1])]; + tensor value_3_pad_0 = const()[name = tensor("value_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_3_dilations_0 = const()[name = tensor("value_3_dilations_0"), val = tensor([1, 1])]; + tensor value_3_groups_0 = const()[name = tensor("value_3_groups_0"), val = tensor(1)]; + tensor layers_0_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87447232)))]; + tensor layers_0_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88626944)))]; + tensor value_3_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_bias_to_fp16, dilations = value_3_dilations_0, groups = value_3_groups_0, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = value_3_strides_0, weight = layers_0_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_3_cast_fp16")]; + tensor var_247 = const()[name = tensor("op_247"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_3_cast_fp16 = reshape(shape = var_247, x = query_3_cast_fp16)[name = tensor("mh_q_3_cast_fp16")]; + tensor var_249_to_fp16 = const()[name = tensor("op_249_to_fp16"), val = tensor(0x1p-3)]; + tensor var_250_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_249_to_fp16)[name = tensor("op_250_cast_fp16")]; + tensor var_253 = const()[name = tensor("op_253"), val = tensor([1, 12, 64, 1500])]; + tensor var_254_cast_fp16 = reshape(shape = var_253, x = key_3_cast_fp16)[name = tensor("op_254_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_250_cast_fp16, y = var_254_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; + tensor obj_13_cast_fp16 = softmax(axis = var_96, x = mh_w_5_cast_fp16)[name = tensor("obj_13_cast_fp16")]; + tensor var_258 = const()[name = tensor("op_258"), val = tensor([1, 12, 64, 1500])]; + tensor var_259_cast_fp16 = reshape(shape = var_258, x = value_3_cast_fp16)[name = tensor("op_259_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_259_cast_fp16, y = obj_13_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_262 = const()[name = tensor("op_262"), val = tensor([1, 768, 1, 1])]; + tensor input_3_cast_fp16 = reshape(shape = var_262, x = attn_3_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor obj_11_pad_type_0 = const()[name = tensor("obj_11_pad_type_0"), val = tensor("valid")]; + tensor obj_11_strides_0 = const()[name = tensor("obj_11_strides_0"), val = tensor([1, 1])]; + tensor obj_11_pad_0 = const()[name = tensor("obj_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_11_dilations_0 = const()[name = tensor("obj_11_dilations_0"), val = tensor([1, 1])]; + tensor obj_11_groups_0 = const()[name = tensor("obj_11_groups_0"), val = tensor(1)]; + tensor layers_0_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88628544)))]; + tensor layers_0_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89808256)))]; + tensor obj_11_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = obj_11_dilations_0, groups = obj_11_groups_0, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = obj_11_strides_0, weight = layers_0_encoder_attn_o_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("obj_11_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor out_5_axes_0 = const()[name = tensor("out_5_axes_0"), val = tensor([1])]; + tensor var_280_to_fp16 = const()[name = tensor("op_280_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_280_to_fp16, x = inputs_5_cast_fp16)[name = tensor("out_5_cast_fp16")]; + tensor input_5_gamma_0_to_fp16 = const()[name = tensor("input_5_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89809856)))]; + tensor input_5_beta_0_to_fp16 = const()[name = tensor("input_5_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89811456)))]; + tensor input_5_epsilon_0_to_fp16 = const()[name = tensor("input_5_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = batch_norm(beta = input_5_beta_0_to_fp16, epsilon = input_5_epsilon_0_to_fp16, gamma = input_5_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor input_7_pad_type_0 = const()[name = tensor("input_7_pad_type_0"), val = tensor("valid")]; + tensor input_7_strides_0 = const()[name = tensor("input_7_strides_0"), val = tensor([1, 1])]; + tensor input_7_pad_0 = const()[name = tensor("input_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_7_dilations_0 = const()[name = tensor("input_7_dilations_0"), val = tensor([1, 1])]; + tensor input_7_groups_0 = const()[name = tensor("input_7_groups_0"), val = tensor(1)]; + tensor layers_0_fc1_weight_to_fp16 = const()[name = tensor("layers_0_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89813056)))]; + tensor layers_0_fc1_bias_to_fp16 = const()[name = tensor("layers_0_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94531712)))]; + tensor input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_0_fc1_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_mode_0 = const()[name = tensor("input_9_mode_0"), val = tensor("EXACT")]; + tensor input_9_cast_fp16 = gelu(mode = input_9_mode_0, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor hidden_states_3_pad_type_0 = const()[name = tensor("hidden_states_3_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_3_strides_0 = const()[name = tensor("hidden_states_3_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_3_pad_0 = const()[name = tensor("hidden_states_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_3_dilations_0 = const()[name = tensor("hidden_states_3_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_3_groups_0 = const()[name = tensor("hidden_states_3_groups_0"), val = tensor(1)]; + tensor layers_0_fc2_weight_to_fp16 = const()[name = tensor("layers_0_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94537920)))]; + tensor layers_0_fc2_bias_to_fp16 = const()[name = tensor("layers_0_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99256576)))]; + tensor hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_fc2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("hidden_states_3_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor var_315 = const()[name = tensor("op_315"), val = tensor(3)]; + tensor out_7_axes_0 = const()[name = tensor("out_7_axes_0"), val = tensor([1])]; + tensor var_340_to_fp16 = const()[name = tensor("op_340_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_340_to_fp16, x = inputs_7_cast_fp16)[name = tensor("out_7_cast_fp16")]; + tensor obj_15_gamma_0_to_fp16 = const()[name = tensor("obj_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99258176)))]; + tensor obj_15_beta_0_to_fp16 = const()[name = tensor("obj_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99259776)))]; + tensor obj_15_epsilon_0_to_fp16 = const()[name = tensor("obj_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor("obj_15_cast_fp16")]; + tensor query_5_pad_type_0 = const()[name = tensor("query_5_pad_type_0"), val = tensor("valid")]; + tensor query_5_strides_0 = const()[name = tensor("query_5_strides_0"), val = tensor([1, 1])]; + tensor query_5_pad_0 = const()[name = tensor("query_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_5_dilations_0 = const()[name = tensor("query_5_dilations_0"), val = tensor([1, 1])]; + tensor query_5_groups_0 = const()[name = tensor("query_5_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99261376)))]; + tensor layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100441088)))]; + tensor query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = query_5_dilations_0, groups = query_5_groups_0, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = query_5_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("query_5_cast_fp16")]; + tensor current_key_3_pad_type_0 = const()[name = tensor("current_key_3_pad_type_0"), val = tensor("valid")]; + tensor current_key_3_strides_0 = const()[name = tensor("current_key_3_strides_0"), val = tensor([1, 1])]; + tensor current_key_3_pad_0 = const()[name = tensor("current_key_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_3_dilations_0 = const()[name = tensor("current_key_3_dilations_0"), val = tensor([1, 1])]; + tensor current_key_3_groups_0 = const()[name = tensor("current_key_3_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100442688)))]; + tensor current_key_3_cast_fp16 = conv(dilations = current_key_3_dilations_0, groups = current_key_3_groups_0, pad = current_key_3_pad_0, pad_type = current_key_3_pad_type_0, strides = current_key_3_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("current_key_3_cast_fp16")]; + tensor current_value_3_pad_type_0 = const()[name = tensor("current_value_3_pad_type_0"), val = tensor("valid")]; + tensor current_value_3_strides_0 = const()[name = tensor("current_value_3_strides_0"), val = tensor([1, 1])]; + tensor current_value_3_pad_0 = const()[name = tensor("current_value_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_3_dilations_0 = const()[name = tensor("current_value_3_dilations_0"), val = tensor([1, 1])]; + tensor current_value_3_groups_0 = const()[name = tensor("current_value_3_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101622400)))]; + tensor layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102802112)))]; + tensor current_value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = current_value_3_dilations_0, groups = current_value_3_groups_0, pad = current_value_3_pad_0, pad_type = current_value_3_pad_type_0, strides = current_value_3_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("current_value_3_cast_fp16")]; + tensor var_379_cast_fp16 = mul(x = var_63_cast_fp16_1, y = var_159_cast_fp16)[name = tensor("op_379_cast_fp16")]; + tensor var_380_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_380_cast_fp16")]; + tensor key_5_cast_fp16 = add(x = var_379_cast_fp16, y = var_380_cast_fp16)[name = tensor("key_5_cast_fp16")]; + tensor var_383_cast_fp16 = mul(x = var_78_cast_fp16_1, y = var_159_cast_fp16)[name = tensor("op_383_cast_fp16")]; + tensor var_384_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor value_5_cast_fp16 = add(x = var_383_cast_fp16, y = var_384_cast_fp16)[name = tensor("value_5_cast_fp16")]; + tensor var_388 = const()[name = tensor("op_388"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_5_cast_fp16 = reshape(shape = var_388, x = query_5_cast_fp16)[name = tensor("mh_q_5_cast_fp16")]; + tensor var_390_to_fp16 = const()[name = tensor("op_390_to_fp16"), val = tensor(0x1p-3)]; + tensor var_391_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_390_to_fp16)[name = tensor("op_391_cast_fp16")]; + tensor var_394 = const()[name = tensor("op_394"), val = tensor([1, 12, 64, 448])]; + tensor var_395_cast_fp16 = reshape(shape = var_394, x = key_5_cast_fp16)[name = tensor("op_395_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_391_cast_fp16, y = var_395_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; + tensor mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_181_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; + tensor var_403_cast_fp16 = softmax(axis = var_315, x = mh_w_9_cast_fp16)[name = tensor("op_403_cast_fp16")]; + tensor var_404 = const()[name = tensor("op_404"), val = tensor([1, 12, 64, 448])]; + tensor var_405_cast_fp16 = reshape(shape = var_404, x = value_5_cast_fp16)[name = tensor("op_405_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_405_cast_fp16, y = var_403_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_408 = const()[name = tensor("op_408"), val = tensor([1, 768, 1, 1])]; + tensor input_11_cast_fp16 = reshape(shape = var_408, x = attn_5_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor obj_21_pad_type_0 = const()[name = tensor("obj_21_pad_type_0"), val = tensor("valid")]; + tensor obj_21_strides_0 = const()[name = tensor("obj_21_strides_0"), val = tensor([1, 1])]; + tensor obj_21_pad_0 = const()[name = tensor("obj_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_21_dilations_0 = const()[name = tensor("obj_21_dilations_0"), val = tensor([1, 1])]; + tensor obj_21_groups_0 = const()[name = tensor("obj_21_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102803712)))]; + tensor layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103983424)))]; + tensor obj_21_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = obj_21_dilations_0, groups = obj_21_groups_0, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = obj_21_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("obj_21_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_21_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([1])]; + tensor var_430_to_fp16 = const()[name = tensor("op_430_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_430_to_fp16, x = inputs_9_cast_fp16)[name = tensor("out_9_cast_fp16")]; + tensor obj_23_gamma_0_to_fp16 = const()[name = tensor("obj_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103985024)))]; + tensor obj_23_beta_0_to_fp16 = const()[name = tensor("obj_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103986624)))]; + tensor obj_23_epsilon_0_to_fp16 = const()[name = tensor("obj_23_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor("obj_23_cast_fp16")]; + tensor query_7_pad_type_0 = const()[name = tensor("query_7_pad_type_0"), val = tensor("valid")]; + tensor query_7_strides_0 = const()[name = tensor("query_7_strides_0"), val = tensor([1, 1])]; + tensor query_7_pad_0 = const()[name = tensor("query_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_7_dilations_0 = const()[name = tensor("query_7_dilations_0"), val = tensor([1, 1])]; + tensor query_7_groups_0 = const()[name = tensor("query_7_groups_0"), val = tensor(1)]; + tensor layers_1_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103988224)))]; + tensor layers_1_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105167936)))]; + tensor query_7_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = query_7_dilations_0, groups = query_7_groups_0, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = query_7_strides_0, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor("query_7_cast_fp16")]; + tensor key_7_pad_type_0 = const()[name = tensor("key_7_pad_type_0"), val = tensor("valid")]; + tensor key_7_strides_0 = const()[name = tensor("key_7_strides_0"), val = tensor([1, 1])]; + tensor key_7_pad_0 = const()[name = tensor("key_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_7_dilations_0 = const()[name = tensor("key_7_dilations_0"), val = tensor([1, 1])]; + tensor key_7_groups_0 = const()[name = tensor("key_7_groups_0"), val = tensor(1)]; + tensor layers_1_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105169536)))]; + tensor key_7_cast_fp16 = conv(dilations = key_7_dilations_0, groups = key_7_groups_0, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = key_7_strides_0, weight = layers_1_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_7_cast_fp16")]; + tensor value_7_pad_type_0 = const()[name = tensor("value_7_pad_type_0"), val = tensor("valid")]; + tensor value_7_strides_0 = const()[name = tensor("value_7_strides_0"), val = tensor([1, 1])]; + tensor value_7_pad_0 = const()[name = tensor("value_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_7_dilations_0 = const()[name = tensor("value_7_dilations_0"), val = tensor([1, 1])]; + tensor value_7_groups_0 = const()[name = tensor("value_7_groups_0"), val = tensor(1)]; + tensor layers_1_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106349248)))]; + tensor layers_1_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107528960)))]; + tensor value_7_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_bias_to_fp16, dilations = value_7_dilations_0, groups = value_7_groups_0, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = value_7_strides_0, weight = layers_1_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_7_cast_fp16")]; + tensor var_466 = const()[name = tensor("op_466"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_7_cast_fp16 = reshape(shape = var_466, x = query_7_cast_fp16)[name = tensor("mh_q_7_cast_fp16")]; + tensor var_468_to_fp16 = const()[name = tensor("op_468_to_fp16"), val = tensor(0x1p-3)]; + tensor var_469_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_468_to_fp16)[name = tensor("op_469_cast_fp16")]; + tensor var_472 = const()[name = tensor("op_472"), val = tensor([1, 12, 64, 1500])]; + tensor var_473_cast_fp16 = reshape(shape = var_472, x = key_7_cast_fp16)[name = tensor("op_473_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_469_cast_fp16, y = var_473_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; + tensor obj_27_cast_fp16 = softmax(axis = var_315, x = mh_w_11_cast_fp16)[name = tensor("obj_27_cast_fp16")]; + tensor var_477 = const()[name = tensor("op_477"), val = tensor([1, 12, 64, 1500])]; + tensor var_478_cast_fp16 = reshape(shape = var_477, x = value_7_cast_fp16)[name = tensor("op_478_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_478_cast_fp16, y = obj_27_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_481 = const()[name = tensor("op_481"), val = tensor([1, 768, 1, 1])]; + tensor input_13_cast_fp16 = reshape(shape = var_481, x = attn_7_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor obj_25_pad_type_0 = const()[name = tensor("obj_25_pad_type_0"), val = tensor("valid")]; + tensor obj_25_strides_0 = const()[name = tensor("obj_25_strides_0"), val = tensor([1, 1])]; + tensor obj_25_pad_0 = const()[name = tensor("obj_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_25_dilations_0 = const()[name = tensor("obj_25_dilations_0"), val = tensor([1, 1])]; + tensor obj_25_groups_0 = const()[name = tensor("obj_25_groups_0"), val = tensor(1)]; + tensor layers_1_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107530560)))]; + tensor layers_1_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108710272)))]; + tensor obj_25_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = obj_25_dilations_0, groups = obj_25_groups_0, pad = obj_25_pad_0, pad_type = obj_25_pad_type_0, strides = obj_25_strides_0, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("obj_25_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_25_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor out_11_axes_0 = const()[name = tensor("out_11_axes_0"), val = tensor([1])]; + tensor var_499_to_fp16 = const()[name = tensor("op_499_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_499_to_fp16, x = inputs_11_cast_fp16)[name = tensor("out_11_cast_fp16")]; + tensor input_15_gamma_0_to_fp16 = const()[name = tensor("input_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108711872)))]; + tensor input_15_beta_0_to_fp16 = const()[name = tensor("input_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108713472)))]; + tensor input_15_epsilon_0_to_fp16 = const()[name = tensor("input_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = batch_norm(beta = input_15_beta_0_to_fp16, epsilon = input_15_epsilon_0_to_fp16, gamma = input_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_pad_type_0 = const()[name = tensor("input_17_pad_type_0"), val = tensor("valid")]; + tensor input_17_strides_0 = const()[name = tensor("input_17_strides_0"), val = tensor([1, 1])]; + tensor input_17_pad_0 = const()[name = tensor("input_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_17_dilations_0 = const()[name = tensor("input_17_dilations_0"), val = tensor([1, 1])]; + tensor input_17_groups_0 = const()[name = tensor("input_17_groups_0"), val = tensor(1)]; + tensor layers_1_fc1_weight_to_fp16 = const()[name = tensor("layers_1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108715072)))]; + tensor layers_1_fc1_bias_to_fp16 = const()[name = tensor("layers_1_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113433728)))]; + tensor input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_mode_0 = const()[name = tensor("input_19_mode_0"), val = tensor("EXACT")]; + tensor input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor hidden_states_5_pad_type_0 = const()[name = tensor("hidden_states_5_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_5_strides_0 = const()[name = tensor("hidden_states_5_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_5_pad_0 = const()[name = tensor("hidden_states_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_5_dilations_0 = const()[name = tensor("hidden_states_5_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_5_groups_0 = const()[name = tensor("hidden_states_5_groups_0"), val = tensor(1)]; + tensor layers_1_fc2_weight_to_fp16 = const()[name = tensor("layers_1_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113439936)))]; + tensor layers_1_fc2_bias_to_fp16 = const()[name = tensor("layers_1_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118158592)))]; + tensor hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = hidden_states_5_dilations_0, groups = hidden_states_5_groups_0, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = hidden_states_5_strides_0, weight = layers_1_fc2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("hidden_states_5_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_534 = const()[name = tensor("op_534"), val = tensor(3)]; + tensor out_13_axes_0 = const()[name = tensor("out_13_axes_0"), val = tensor([1])]; + tensor var_559_to_fp16 = const()[name = tensor("op_559_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_559_to_fp16, x = inputs_13_cast_fp16)[name = tensor("out_13_cast_fp16")]; + tensor obj_29_gamma_0_to_fp16 = const()[name = tensor("obj_29_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118160192)))]; + tensor obj_29_beta_0_to_fp16 = const()[name = tensor("obj_29_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118161792)))]; + tensor obj_29_epsilon_0_to_fp16 = const()[name = tensor("obj_29_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor("obj_29_cast_fp16")]; + tensor query_9_pad_type_0 = const()[name = tensor("query_9_pad_type_0"), val = tensor("valid")]; + tensor query_9_strides_0 = const()[name = tensor("query_9_strides_0"), val = tensor([1, 1])]; + tensor query_9_pad_0 = const()[name = tensor("query_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_9_dilations_0 = const()[name = tensor("query_9_dilations_0"), val = tensor([1, 1])]; + tensor query_9_groups_0 = const()[name = tensor("query_9_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118163392)))]; + tensor layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119343104)))]; + tensor query_9_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = query_9_dilations_0, groups = query_9_groups_0, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = query_9_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("query_9_cast_fp16")]; + tensor current_key_5_pad_type_0 = const()[name = tensor("current_key_5_pad_type_0"), val = tensor("valid")]; + tensor current_key_5_strides_0 = const()[name = tensor("current_key_5_strides_0"), val = tensor([1, 1])]; + tensor current_key_5_pad_0 = const()[name = tensor("current_key_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_5_dilations_0 = const()[name = tensor("current_key_5_dilations_0"), val = tensor([1, 1])]; + tensor current_key_5_groups_0 = const()[name = tensor("current_key_5_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119344704)))]; + tensor current_key_5_cast_fp16 = conv(dilations = current_key_5_dilations_0, groups = current_key_5_groups_0, pad = current_key_5_pad_0, pad_type = current_key_5_pad_type_0, strides = current_key_5_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("current_key_5_cast_fp16")]; + tensor current_value_5_pad_type_0 = const()[name = tensor("current_value_5_pad_type_0"), val = tensor("valid")]; + tensor current_value_5_strides_0 = const()[name = tensor("current_value_5_strides_0"), val = tensor([1, 1])]; + tensor current_value_5_pad_0 = const()[name = tensor("current_value_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_5_dilations_0 = const()[name = tensor("current_value_5_dilations_0"), val = tensor([1, 1])]; + tensor current_value_5_groups_0 = const()[name = tensor("current_value_5_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120524416)))]; + tensor layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121704128)))]; + tensor current_value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = current_value_5_dilations_0, groups = current_value_5_groups_0, pad = current_value_5_pad_0, pad_type = current_value_5_pad_type_0, strides = current_value_5_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("current_value_5_cast_fp16")]; + tensor var_598_cast_fp16 = mul(x = var_63_cast_fp16_2, y = var_159_cast_fp16)[name = tensor("op_598_cast_fp16")]; + tensor var_599_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_599_cast_fp16")]; + tensor key_9_cast_fp16 = add(x = var_598_cast_fp16, y = var_599_cast_fp16)[name = tensor("key_9_cast_fp16")]; + tensor var_602_cast_fp16 = mul(x = var_78_cast_fp16_2, y = var_159_cast_fp16)[name = tensor("op_602_cast_fp16")]; + tensor var_603_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor value_9_cast_fp16 = add(x = var_602_cast_fp16, y = var_603_cast_fp16)[name = tensor("value_9_cast_fp16")]; + tensor var_607 = const()[name = tensor("op_607"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_9_cast_fp16 = reshape(shape = var_607, x = query_9_cast_fp16)[name = tensor("mh_q_9_cast_fp16")]; + tensor var_609_to_fp16 = const()[name = tensor("op_609_to_fp16"), val = tensor(0x1p-3)]; + tensor var_610_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_609_to_fp16)[name = tensor("op_610_cast_fp16")]; + tensor var_613 = const()[name = tensor("op_613"), val = tensor([1, 12, 64, 448])]; + tensor var_614_cast_fp16 = reshape(shape = var_613, x = key_9_cast_fp16)[name = tensor("op_614_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_610_cast_fp16, y = var_614_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; + tensor mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_181_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; + tensor var_622_cast_fp16 = softmax(axis = var_534, x = mh_w_15_cast_fp16)[name = tensor("op_622_cast_fp16")]; + tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 12, 64, 448])]; + tensor var_624_cast_fp16 = reshape(shape = var_623, x = value_9_cast_fp16)[name = tensor("op_624_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_624_cast_fp16, y = var_622_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor([1, 768, 1, 1])]; + tensor input_21_cast_fp16 = reshape(shape = var_627, x = attn_9_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor obj_35_pad_type_0 = const()[name = tensor("obj_35_pad_type_0"), val = tensor("valid")]; + tensor obj_35_strides_0 = const()[name = tensor("obj_35_strides_0"), val = tensor([1, 1])]; + tensor obj_35_pad_0 = const()[name = tensor("obj_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_35_dilations_0 = const()[name = tensor("obj_35_dilations_0"), val = tensor([1, 1])]; + tensor obj_35_groups_0 = const()[name = tensor("obj_35_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121705728)))]; + tensor layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122885440)))]; + tensor obj_35_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = obj_35_dilations_0, groups = obj_35_groups_0, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = obj_35_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("obj_35_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_35_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([1])]; + tensor var_649_to_fp16 = const()[name = tensor("op_649_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_649_to_fp16, x = inputs_15_cast_fp16)[name = tensor("out_15_cast_fp16")]; + tensor obj_37_gamma_0_to_fp16 = const()[name = tensor("obj_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122887040)))]; + 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(122888640)))]; + tensor obj_37_epsilon_0_to_fp16 = const()[name = tensor("obj_37_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor("obj_37_cast_fp16")]; + tensor query_11_pad_type_0 = const()[name = tensor("query_11_pad_type_0"), val = tensor("valid")]; + tensor query_11_strides_0 = const()[name = tensor("query_11_strides_0"), val = tensor([1, 1])]; + tensor query_11_pad_0 = const()[name = tensor("query_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_11_dilations_0 = const()[name = tensor("query_11_dilations_0"), val = tensor([1, 1])]; + tensor query_11_groups_0 = const()[name = tensor("query_11_groups_0"), val = tensor(1)]; + tensor layers_2_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122890240)))]; + tensor layers_2_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124069952)))]; + tensor query_11_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_bias_to_fp16, dilations = query_11_dilations_0, groups = query_11_groups_0, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = query_11_strides_0, weight = layers_2_encoder_attn_q_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("query_11_cast_fp16")]; + tensor key_11_pad_type_0 = const()[name = tensor("key_11_pad_type_0"), val = tensor("valid")]; + tensor key_11_strides_0 = const()[name = tensor("key_11_strides_0"), val = tensor([1, 1])]; + tensor key_11_pad_0 = const()[name = tensor("key_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_11_dilations_0 = const()[name = tensor("key_11_dilations_0"), val = tensor([1, 1])]; + tensor key_11_groups_0 = const()[name = tensor("key_11_groups_0"), val = tensor(1)]; + tensor layers_2_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124071552)))]; + tensor key_11_cast_fp16 = conv(dilations = key_11_dilations_0, groups = key_11_groups_0, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = key_11_strides_0, weight = layers_2_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_11_cast_fp16")]; + tensor value_11_pad_type_0 = const()[name = tensor("value_11_pad_type_0"), val = tensor("valid")]; + tensor value_11_strides_0 = const()[name = tensor("value_11_strides_0"), val = tensor([1, 1])]; + tensor value_11_pad_0 = const()[name = tensor("value_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_11_dilations_0 = const()[name = tensor("value_11_dilations_0"), val = tensor([1, 1])]; + tensor value_11_groups_0 = const()[name = tensor("value_11_groups_0"), val = tensor(1)]; + tensor layers_2_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125251264)))]; + tensor layers_2_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126430976)))]; + tensor value_11_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_bias_to_fp16, dilations = value_11_dilations_0, groups = value_11_groups_0, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = value_11_strides_0, weight = layers_2_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_11_cast_fp16")]; + tensor var_685 = const()[name = tensor("op_685"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_11_cast_fp16 = reshape(shape = var_685, x = query_11_cast_fp16)[name = tensor("mh_q_11_cast_fp16")]; + tensor var_687_to_fp16 = const()[name = tensor("op_687_to_fp16"), val = tensor(0x1p-3)]; + tensor var_688_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_687_to_fp16)[name = tensor("op_688_cast_fp16")]; + tensor var_691 = const()[name = tensor("op_691"), val = tensor([1, 12, 64, 1500])]; + tensor var_692_cast_fp16 = reshape(shape = var_691, x = key_11_cast_fp16)[name = tensor("op_692_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_688_cast_fp16, y = var_692_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; + tensor obj_41_cast_fp16 = softmax(axis = var_534, x = mh_w_17_cast_fp16)[name = tensor("obj_41_cast_fp16")]; + tensor var_696 = const()[name = tensor("op_696"), val = tensor([1, 12, 64, 1500])]; + tensor var_697_cast_fp16 = reshape(shape = var_696, x = value_11_cast_fp16)[name = tensor("op_697_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_697_cast_fp16, y = obj_41_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_700 = const()[name = tensor("op_700"), val = tensor([1, 768, 1, 1])]; + tensor input_23_cast_fp16 = reshape(shape = var_700, x = attn_11_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor obj_39_pad_type_0 = const()[name = tensor("obj_39_pad_type_0"), val = tensor("valid")]; + tensor obj_39_strides_0 = const()[name = tensor("obj_39_strides_0"), val = tensor([1, 1])]; + tensor obj_39_pad_0 = const()[name = tensor("obj_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_39_dilations_0 = const()[name = tensor("obj_39_dilations_0"), val = tensor([1, 1])]; + tensor obj_39_groups_0 = const()[name = tensor("obj_39_groups_0"), val = tensor(1)]; + tensor layers_2_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126432576)))]; + tensor layers_2_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127612288)))]; + tensor obj_39_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = obj_39_dilations_0, groups = obj_39_groups_0, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = obj_39_strides_0, weight = layers_2_encoder_attn_o_proj_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("obj_39_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_39_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor out_17_axes_0 = const()[name = tensor("out_17_axes_0"), val = tensor([1])]; + tensor var_718_to_fp16 = const()[name = tensor("op_718_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_718_to_fp16, x = inputs_17_cast_fp16)[name = tensor("out_17_cast_fp16")]; + tensor input_25_gamma_0_to_fp16 = const()[name = tensor("input_25_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127613888)))]; + tensor input_25_beta_0_to_fp16 = const()[name = tensor("input_25_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127615488)))]; + tensor input_25_epsilon_0_to_fp16 = const()[name = tensor("input_25_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_25_cast_fp16 = batch_norm(beta = input_25_beta_0_to_fp16, epsilon = input_25_epsilon_0_to_fp16, gamma = input_25_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_pad_type_0 = const()[name = tensor("input_27_pad_type_0"), val = tensor("valid")]; + tensor input_27_strides_0 = const()[name = tensor("input_27_strides_0"), val = tensor([1, 1])]; + tensor input_27_pad_0 = const()[name = tensor("input_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_27_dilations_0 = const()[name = tensor("input_27_dilations_0"), val = tensor([1, 1])]; + tensor input_27_groups_0 = const()[name = tensor("input_27_groups_0"), val = tensor(1)]; + tensor layers_2_fc1_weight_to_fp16 = const()[name = tensor("layers_2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127617088)))]; + tensor layers_2_fc1_bias_to_fp16 = const()[name = tensor("layers_2_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132335744)))]; + tensor input_27_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = input_27_dilations_0, groups = input_27_groups_0, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = input_27_strides_0, weight = layers_2_fc1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_mode_0 = const()[name = tensor("input_29_mode_0"), val = tensor("EXACT")]; + tensor input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor hidden_states_7_pad_type_0 = const()[name = tensor("hidden_states_7_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_7_strides_0 = const()[name = tensor("hidden_states_7_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_7_pad_0 = const()[name = tensor("hidden_states_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_7_dilations_0 = const()[name = tensor("hidden_states_7_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_7_groups_0 = const()[name = tensor("hidden_states_7_groups_0"), val = tensor(1)]; + tensor layers_2_fc2_weight_to_fp16 = const()[name = tensor("layers_2_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132341952)))]; + tensor layers_2_fc2_bias_to_fp16 = const()[name = tensor("layers_2_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137060608)))]; + tensor hidden_states_7_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_2_fc2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("hidden_states_7_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor var_753 = const()[name = tensor("op_753"), val = tensor(3)]; + tensor out_19_axes_0 = const()[name = tensor("out_19_axes_0"), val = tensor([1])]; + tensor var_778_to_fp16 = const()[name = tensor("op_778_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_778_to_fp16, x = inputs_19_cast_fp16)[name = tensor("out_19_cast_fp16")]; + tensor obj_43_gamma_0_to_fp16 = const()[name = tensor("obj_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137062208)))]; + tensor obj_43_beta_0_to_fp16 = const()[name = tensor("obj_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137063808)))]; + tensor obj_43_epsilon_0_to_fp16 = const()[name = tensor("obj_43_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor("obj_43_cast_fp16")]; + tensor query_13_pad_type_0 = const()[name = tensor("query_13_pad_type_0"), val = tensor("valid")]; + tensor query_13_strides_0 = const()[name = tensor("query_13_strides_0"), val = tensor([1, 1])]; + tensor query_13_pad_0 = const()[name = tensor("query_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_13_dilations_0 = const()[name = tensor("query_13_dilations_0"), val = tensor([1, 1])]; + tensor query_13_groups_0 = const()[name = tensor("query_13_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137065408)))]; + tensor layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138245120)))]; + tensor query_13_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = query_13_dilations_0, groups = query_13_groups_0, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = query_13_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("query_13_cast_fp16")]; + tensor current_key_7_pad_type_0 = const()[name = tensor("current_key_7_pad_type_0"), val = tensor("valid")]; + tensor current_key_7_strides_0 = const()[name = tensor("current_key_7_strides_0"), val = tensor([1, 1])]; + tensor current_key_7_pad_0 = const()[name = tensor("current_key_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_7_dilations_0 = const()[name = tensor("current_key_7_dilations_0"), val = tensor([1, 1])]; + tensor current_key_7_groups_0 = const()[name = tensor("current_key_7_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138246720)))]; + tensor current_key_7_cast_fp16 = conv(dilations = current_key_7_dilations_0, groups = current_key_7_groups_0, pad = current_key_7_pad_0, pad_type = current_key_7_pad_type_0, strides = current_key_7_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("current_key_7_cast_fp16")]; + tensor current_value_7_pad_type_0 = const()[name = tensor("current_value_7_pad_type_0"), val = tensor("valid")]; + tensor current_value_7_strides_0 = const()[name = tensor("current_value_7_strides_0"), val = tensor([1, 1])]; + tensor current_value_7_pad_0 = const()[name = tensor("current_value_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_7_dilations_0 = const()[name = tensor("current_value_7_dilations_0"), val = tensor([1, 1])]; + tensor current_value_7_groups_0 = const()[name = tensor("current_value_7_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139426432)))]; + tensor layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140606144)))]; + tensor current_value_7_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = current_value_7_dilations_0, groups = current_value_7_groups_0, pad = current_value_7_pad_0, pad_type = current_value_7_pad_type_0, strides = current_value_7_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("current_value_7_cast_fp16")]; + tensor var_817_cast_fp16 = mul(x = var_63_cast_fp16_3, y = var_159_cast_fp16)[name = tensor("op_817_cast_fp16")]; + tensor var_818_cast_fp16 = mul(x = current_key_7_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_818_cast_fp16")]; + tensor key_13_cast_fp16 = add(x = var_817_cast_fp16, y = var_818_cast_fp16)[name = tensor("key_13_cast_fp16")]; + tensor var_821_cast_fp16 = mul(x = var_78_cast_fp16_3, y = var_159_cast_fp16)[name = tensor("op_821_cast_fp16")]; + tensor var_822_cast_fp16 = mul(x = current_value_7_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_822_cast_fp16")]; + tensor value_13_cast_fp16 = add(x = var_821_cast_fp16, y = var_822_cast_fp16)[name = tensor("value_13_cast_fp16")]; + tensor var_826 = const()[name = tensor("op_826"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_13_cast_fp16 = reshape(shape = var_826, x = query_13_cast_fp16)[name = tensor("mh_q_13_cast_fp16")]; + tensor var_828_to_fp16 = const()[name = tensor("op_828_to_fp16"), val = tensor(0x1p-3)]; + tensor var_829_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_828_to_fp16)[name = tensor("op_829_cast_fp16")]; + tensor var_832 = const()[name = tensor("op_832"), val = tensor([1, 12, 64, 448])]; + tensor var_833_cast_fp16 = reshape(shape = var_832, x = key_13_cast_fp16)[name = tensor("op_833_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_829_cast_fp16, y = var_833_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; + tensor mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_181_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; + tensor var_841_cast_fp16 = softmax(axis = var_753, x = mh_w_21_cast_fp16)[name = tensor("op_841_cast_fp16")]; + tensor var_842 = const()[name = tensor("op_842"), val = tensor([1, 12, 64, 448])]; + tensor var_843_cast_fp16 = reshape(shape = var_842, x = value_13_cast_fp16)[name = tensor("op_843_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_843_cast_fp16, y = var_841_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_846 = const()[name = tensor("op_846"), val = tensor([1, 768, 1, 1])]; + tensor input_31_cast_fp16 = reshape(shape = var_846, x = attn_13_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor obj_49_pad_type_0 = const()[name = tensor("obj_49_pad_type_0"), val = tensor("valid")]; + tensor obj_49_strides_0 = const()[name = tensor("obj_49_strides_0"), val = tensor([1, 1])]; + tensor obj_49_pad_0 = const()[name = tensor("obj_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_49_dilations_0 = const()[name = tensor("obj_49_dilations_0"), val = tensor([1, 1])]; + tensor obj_49_groups_0 = const()[name = tensor("obj_49_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140607744)))]; + tensor layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141787456)))]; + tensor obj_49_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = obj_49_dilations_0, groups = obj_49_groups_0, pad = obj_49_pad_0, pad_type = obj_49_pad_type_0, strides = obj_49_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("obj_49_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_49_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([1])]; + tensor var_868_to_fp16 = const()[name = tensor("op_868_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_868_to_fp16, x = inputs_21_cast_fp16)[name = tensor("out_21_cast_fp16")]; + tensor obj_51_gamma_0_to_fp16 = const()[name = tensor("obj_51_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141789056)))]; + tensor obj_51_beta_0_to_fp16 = const()[name = tensor("obj_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141790656)))]; + tensor obj_51_epsilon_0_to_fp16 = const()[name = tensor("obj_51_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor("obj_51_cast_fp16")]; + tensor query_15_pad_type_0 = const()[name = tensor("query_15_pad_type_0"), val = tensor("valid")]; + tensor query_15_strides_0 = const()[name = tensor("query_15_strides_0"), val = tensor([1, 1])]; + tensor query_15_pad_0 = const()[name = tensor("query_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_15_dilations_0 = const()[name = tensor("query_15_dilations_0"), val = tensor([1, 1])]; + tensor query_15_groups_0 = const()[name = tensor("query_15_groups_0"), val = tensor(1)]; + tensor layers_3_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141792256)))]; + tensor layers_3_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142971968)))]; + tensor query_15_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_bias_to_fp16, dilations = query_15_dilations_0, groups = query_15_groups_0, pad = query_15_pad_0, pad_type = query_15_pad_type_0, strides = query_15_strides_0, weight = layers_3_encoder_attn_q_proj_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor("query_15_cast_fp16")]; + tensor key_15_pad_type_0 = const()[name = tensor("key_15_pad_type_0"), val = tensor("valid")]; + tensor key_15_strides_0 = const()[name = tensor("key_15_strides_0"), val = tensor([1, 1])]; + tensor key_15_pad_0 = const()[name = tensor("key_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_15_dilations_0 = const()[name = tensor("key_15_dilations_0"), val = tensor([1, 1])]; + tensor key_15_groups_0 = const()[name = tensor("key_15_groups_0"), val = tensor(1)]; + tensor layers_3_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142973568)))]; + tensor key_15_cast_fp16 = conv(dilations = key_15_dilations_0, groups = key_15_groups_0, pad = key_15_pad_0, pad_type = key_15_pad_type_0, strides = key_15_strides_0, weight = layers_3_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_15_cast_fp16")]; + tensor value_15_pad_type_0 = const()[name = tensor("value_15_pad_type_0"), val = tensor("valid")]; + tensor value_15_strides_0 = const()[name = tensor("value_15_strides_0"), val = tensor([1, 1])]; + tensor value_15_pad_0 = const()[name = tensor("value_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_15_dilations_0 = const()[name = tensor("value_15_dilations_0"), val = tensor([1, 1])]; + tensor value_15_groups_0 = const()[name = tensor("value_15_groups_0"), val = tensor(1)]; + tensor layers_3_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144153280)))]; + tensor layers_3_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145332992)))]; + tensor value_15_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_bias_to_fp16, dilations = value_15_dilations_0, groups = value_15_groups_0, pad = value_15_pad_0, pad_type = value_15_pad_type_0, strides = value_15_strides_0, weight = layers_3_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_15_cast_fp16")]; + tensor var_904 = const()[name = tensor("op_904"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_15_cast_fp16 = reshape(shape = var_904, x = query_15_cast_fp16)[name = tensor("mh_q_15_cast_fp16")]; + tensor var_906_to_fp16 = const()[name = tensor("op_906_to_fp16"), val = tensor(0x1p-3)]; + tensor var_907_cast_fp16 = mul(x = mh_q_15_cast_fp16, y = var_906_to_fp16)[name = tensor("op_907_cast_fp16")]; + tensor var_910 = const()[name = tensor("op_910"), val = tensor([1, 12, 64, 1500])]; + tensor var_911_cast_fp16 = reshape(shape = var_910, x = key_15_cast_fp16)[name = tensor("op_911_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_907_cast_fp16, y = var_911_cast_fp16)[name = tensor("mh_w_23_cast_fp16")]; + tensor obj_55_cast_fp16 = softmax(axis = var_753, x = mh_w_23_cast_fp16)[name = tensor("obj_55_cast_fp16")]; + tensor var_915 = const()[name = tensor("op_915"), val = tensor([1, 12, 64, 1500])]; + tensor var_916_cast_fp16 = reshape(shape = var_915, x = value_15_cast_fp16)[name = tensor("op_916_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_916_cast_fp16, y = obj_55_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_919 = const()[name = tensor("op_919"), val = tensor([1, 768, 1, 1])]; + tensor input_33_cast_fp16 = reshape(shape = var_919, x = attn_15_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor obj_53_pad_type_0 = const()[name = tensor("obj_53_pad_type_0"), val = tensor("valid")]; + tensor obj_53_strides_0 = const()[name = tensor("obj_53_strides_0"), val = tensor([1, 1])]; + tensor obj_53_pad_0 = const()[name = tensor("obj_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_53_dilations_0 = const()[name = tensor("obj_53_dilations_0"), val = tensor([1, 1])]; + tensor obj_53_groups_0 = const()[name = tensor("obj_53_groups_0"), val = tensor(1)]; + tensor layers_3_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145334592)))]; + tensor layers_3_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146514304)))]; + tensor obj_53_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = obj_53_dilations_0, groups = obj_53_groups_0, pad = obj_53_pad_0, pad_type = obj_53_pad_type_0, strides = obj_53_strides_0, weight = layers_3_encoder_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("obj_53_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_53_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor out_23_axes_0 = const()[name = tensor("out_23_axes_0"), val = tensor([1])]; + tensor var_937_to_fp16 = const()[name = tensor("op_937_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_937_to_fp16, x = inputs_23_cast_fp16)[name = tensor("out_23_cast_fp16")]; + tensor input_35_gamma_0_to_fp16 = const()[name = tensor("input_35_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146515904)))]; + tensor input_35_beta_0_to_fp16 = const()[name = tensor("input_35_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146517504)))]; + tensor input_35_epsilon_0_to_fp16 = const()[name = tensor("input_35_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = batch_norm(beta = input_35_beta_0_to_fp16, epsilon = input_35_epsilon_0_to_fp16, gamma = input_35_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor input_37_pad_type_0 = const()[name = tensor("input_37_pad_type_0"), val = tensor("valid")]; + tensor input_37_strides_0 = const()[name = tensor("input_37_strides_0"), val = tensor([1, 1])]; + tensor input_37_pad_0 = const()[name = tensor("input_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_37_dilations_0 = const()[name = tensor("input_37_dilations_0"), val = tensor([1, 1])]; + tensor input_37_groups_0 = const()[name = tensor("input_37_groups_0"), val = tensor(1)]; + tensor layers_3_fc1_weight_to_fp16 = const()[name = tensor("layers_3_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146519104)))]; + tensor layers_3_fc1_bias_to_fp16 = const()[name = tensor("layers_3_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151237760)))]; + tensor input_37_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = input_37_dilations_0, groups = input_37_groups_0, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = input_37_strides_0, weight = layers_3_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_mode_0 = const()[name = tensor("input_39_mode_0"), val = tensor("EXACT")]; + tensor input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor hidden_states_9_pad_type_0 = const()[name = tensor("hidden_states_9_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_9_strides_0 = const()[name = tensor("hidden_states_9_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_9_pad_0 = const()[name = tensor("hidden_states_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_9_dilations_0 = const()[name = tensor("hidden_states_9_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_9_groups_0 = const()[name = tensor("hidden_states_9_groups_0"), val = tensor(1)]; + tensor layers_3_fc2_weight_to_fp16 = const()[name = tensor("layers_3_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151243968)))]; + tensor layers_3_fc2_bias_to_fp16 = const()[name = tensor("layers_3_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155962624)))]; + tensor hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = hidden_states_9_dilations_0, groups = hidden_states_9_groups_0, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = hidden_states_9_strides_0, weight = layers_3_fc2_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("hidden_states_9_cast_fp16")]; + tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor("inputs_25_cast_fp16")]; + tensor var_972 = const()[name = tensor("op_972"), val = tensor(3)]; + tensor out_25_axes_0 = const()[name = tensor("out_25_axes_0"), val = tensor([1])]; + tensor var_997_to_fp16 = const()[name = tensor("op_997_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_997_to_fp16, x = inputs_25_cast_fp16)[name = tensor("out_25_cast_fp16")]; + tensor obj_57_gamma_0_to_fp16 = const()[name = tensor("obj_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155964224)))]; + 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(155965824)))]; + tensor obj_57_epsilon_0_to_fp16 = const()[name = tensor("obj_57_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_57_cast_fp16 = batch_norm(beta = obj_57_beta_0_to_fp16, epsilon = obj_57_epsilon_0_to_fp16, gamma = obj_57_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor("obj_57_cast_fp16")]; + tensor query_17_pad_type_0 = const()[name = tensor("query_17_pad_type_0"), val = tensor("valid")]; + tensor query_17_strides_0 = const()[name = tensor("query_17_strides_0"), val = tensor([1, 1])]; + tensor query_17_pad_0 = const()[name = tensor("query_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_17_dilations_0 = const()[name = tensor("query_17_dilations_0"), val = tensor([1, 1])]; + tensor query_17_groups_0 = const()[name = tensor("query_17_groups_0"), val = tensor(1)]; + tensor layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155967424)))]; + tensor layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157147136)))]; + tensor query_17_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_bias_to_fp16, dilations = query_17_dilations_0, groups = query_17_groups_0, pad = query_17_pad_0, pad_type = query_17_pad_type_0, strides = query_17_strides_0, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("query_17_cast_fp16")]; + tensor current_key_9_pad_type_0 = const()[name = tensor("current_key_9_pad_type_0"), val = tensor("valid")]; + tensor current_key_9_strides_0 = const()[name = tensor("current_key_9_strides_0"), val = tensor([1, 1])]; + tensor current_key_9_pad_0 = const()[name = tensor("current_key_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_9_dilations_0 = const()[name = tensor("current_key_9_dilations_0"), val = tensor([1, 1])]; + tensor current_key_9_groups_0 = const()[name = tensor("current_key_9_groups_0"), val = tensor(1)]; + tensor layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157148736)))]; + tensor current_key_9_cast_fp16 = conv(dilations = current_key_9_dilations_0, groups = current_key_9_groups_0, pad = current_key_9_pad_0, pad_type = current_key_9_pad_type_0, strides = current_key_9_strides_0, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("current_key_9_cast_fp16")]; + tensor current_value_9_pad_type_0 = const()[name = tensor("current_value_9_pad_type_0"), val = tensor("valid")]; + tensor current_value_9_strides_0 = const()[name = tensor("current_value_9_strides_0"), val = tensor([1, 1])]; + tensor current_value_9_pad_0 = const()[name = tensor("current_value_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_9_dilations_0 = const()[name = tensor("current_value_9_dilations_0"), val = tensor([1, 1])]; + tensor current_value_9_groups_0 = const()[name = tensor("current_value_9_groups_0"), val = tensor(1)]; + tensor layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158328448)))]; + tensor layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159508160)))]; + tensor current_value_9_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_bias_to_fp16, dilations = current_value_9_dilations_0, groups = current_value_9_groups_0, pad = current_value_9_pad_0, pad_type = current_value_9_pad_type_0, strides = current_value_9_strides_0, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("current_value_9_cast_fp16")]; + tensor var_1036_cast_fp16 = mul(x = var_63_cast_fp16_4, y = var_159_cast_fp16)[name = tensor("op_1036_cast_fp16")]; + tensor var_1037_cast_fp16 = mul(x = current_key_9_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_1037_cast_fp16")]; + tensor key_17_cast_fp16 = add(x = var_1036_cast_fp16, y = var_1037_cast_fp16)[name = tensor("key_17_cast_fp16")]; + tensor var_1040_cast_fp16 = mul(x = var_78_cast_fp16_4, y = var_159_cast_fp16)[name = tensor("op_1040_cast_fp16")]; + tensor var_1041_cast_fp16 = mul(x = current_value_9_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_1041_cast_fp16")]; + tensor value_17_cast_fp16 = add(x = var_1040_cast_fp16, y = var_1041_cast_fp16)[name = tensor("value_17_cast_fp16")]; + tensor var_1045 = const()[name = tensor("op_1045"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_17_cast_fp16 = reshape(shape = var_1045, x = query_17_cast_fp16)[name = tensor("mh_q_17_cast_fp16")]; + tensor var_1047_to_fp16 = const()[name = tensor("op_1047_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1048_cast_fp16 = mul(x = mh_q_17_cast_fp16, y = var_1047_to_fp16)[name = tensor("op_1048_cast_fp16")]; + tensor var_1051 = const()[name = tensor("op_1051"), val = tensor([1, 12, 64, 448])]; + tensor var_1052_cast_fp16 = reshape(shape = var_1051, x = key_17_cast_fp16)[name = tensor("op_1052_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_1048_cast_fp16, y = var_1052_cast_fp16)[name = tensor("mh_w_25_cast_fp16")]; + tensor mh_w_27_cast_fp16 = add(x = mh_w_25_cast_fp16, y = var_181_cast_fp16)[name = tensor("mh_w_27_cast_fp16")]; + tensor var_1060_cast_fp16 = softmax(axis = var_972, x = mh_w_27_cast_fp16)[name = tensor("op_1060_cast_fp16")]; + tensor var_1061 = const()[name = tensor("op_1061"), val = tensor([1, 12, 64, 448])]; + tensor var_1062_cast_fp16 = reshape(shape = var_1061, x = value_17_cast_fp16)[name = tensor("op_1062_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_1062_cast_fp16, y = var_1060_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_1065 = const()[name = tensor("op_1065"), val = tensor([1, 768, 1, 1])]; + tensor input_41_cast_fp16 = reshape(shape = var_1065, x = attn_17_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor obj_63_pad_type_0 = const()[name = tensor("obj_63_pad_type_0"), val = tensor("valid")]; + tensor obj_63_strides_0 = const()[name = tensor("obj_63_strides_0"), val = tensor([1, 1])]; + tensor obj_63_pad_0 = const()[name = tensor("obj_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_63_dilations_0 = const()[name = tensor("obj_63_dilations_0"), val = tensor([1, 1])]; + tensor obj_63_groups_0 = const()[name = tensor("obj_63_groups_0"), val = tensor(1)]; + tensor layers_4_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159509760)))]; + tensor layers_4_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160689472)))]; + tensor obj_63_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_bias_to_fp16, dilations = obj_63_dilations_0, groups = obj_63_groups_0, pad = obj_63_pad_0, pad_type = obj_63_pad_type_0, strides = obj_63_strides_0, weight = layers_4_self_attn_o_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("obj_63_cast_fp16")]; + tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_63_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([1])]; + tensor var_1087_to_fp16 = const()[name = tensor("op_1087_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_1087_to_fp16, x = inputs_27_cast_fp16)[name = tensor("out_27_cast_fp16")]; + tensor obj_65_gamma_0_to_fp16 = const()[name = tensor("obj_65_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160691072)))]; + tensor obj_65_beta_0_to_fp16 = const()[name = tensor("obj_65_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160692672)))]; + tensor obj_65_epsilon_0_to_fp16 = const()[name = tensor("obj_65_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_65_cast_fp16 = batch_norm(beta = obj_65_beta_0_to_fp16, epsilon = obj_65_epsilon_0_to_fp16, gamma = obj_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor("obj_65_cast_fp16")]; + tensor query_19_pad_type_0 = const()[name = tensor("query_19_pad_type_0"), val = tensor("valid")]; + tensor query_19_strides_0 = const()[name = tensor("query_19_strides_0"), val = tensor([1, 1])]; + tensor query_19_pad_0 = const()[name = tensor("query_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_19_dilations_0 = const()[name = tensor("query_19_dilations_0"), val = tensor([1, 1])]; + tensor query_19_groups_0 = const()[name = tensor("query_19_groups_0"), val = tensor(1)]; + tensor layers_4_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160694272)))]; + tensor layers_4_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161873984)))]; + tensor query_19_cast_fp16 = conv(bias = layers_4_encoder_attn_q_proj_bias_to_fp16, dilations = query_19_dilations_0, groups = query_19_groups_0, pad = query_19_pad_0, pad_type = query_19_pad_type_0, strides = query_19_strides_0, weight = layers_4_encoder_attn_q_proj_weight_to_fp16, x = obj_65_cast_fp16)[name = tensor("query_19_cast_fp16")]; + tensor key_19_pad_type_0 = const()[name = tensor("key_19_pad_type_0"), val = tensor("valid")]; + tensor key_19_strides_0 = const()[name = tensor("key_19_strides_0"), val = tensor([1, 1])]; + tensor key_19_pad_0 = const()[name = tensor("key_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_19_dilations_0 = const()[name = tensor("key_19_dilations_0"), val = tensor([1, 1])]; + tensor key_19_groups_0 = const()[name = tensor("key_19_groups_0"), val = tensor(1)]; + tensor layers_4_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161875584)))]; + tensor key_19_cast_fp16 = conv(dilations = key_19_dilations_0, groups = key_19_groups_0, pad = key_19_pad_0, pad_type = key_19_pad_type_0, strides = key_19_strides_0, weight = layers_4_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_19_cast_fp16")]; + tensor value_19_pad_type_0 = const()[name = tensor("value_19_pad_type_0"), val = tensor("valid")]; + tensor value_19_strides_0 = const()[name = tensor("value_19_strides_0"), val = tensor([1, 1])]; + tensor value_19_pad_0 = const()[name = tensor("value_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_19_dilations_0 = const()[name = tensor("value_19_dilations_0"), val = tensor([1, 1])]; + tensor value_19_groups_0 = const()[name = tensor("value_19_groups_0"), val = tensor(1)]; + tensor layers_4_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163055296)))]; + tensor layers_4_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164235008)))]; + tensor value_19_cast_fp16 = conv(bias = layers_4_encoder_attn_v_proj_bias_to_fp16, dilations = value_19_dilations_0, groups = value_19_groups_0, pad = value_19_pad_0, pad_type = value_19_pad_type_0, strides = value_19_strides_0, weight = layers_4_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_19_cast_fp16")]; + tensor var_1123 = const()[name = tensor("op_1123"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_19_cast_fp16 = reshape(shape = var_1123, x = query_19_cast_fp16)[name = tensor("mh_q_19_cast_fp16")]; + tensor var_1125_to_fp16 = const()[name = tensor("op_1125_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1126_cast_fp16 = mul(x = mh_q_19_cast_fp16, y = var_1125_to_fp16)[name = tensor("op_1126_cast_fp16")]; + tensor var_1129 = const()[name = tensor("op_1129"), val = tensor([1, 12, 64, 1500])]; + tensor var_1130_cast_fp16 = reshape(shape = var_1129, x = key_19_cast_fp16)[name = tensor("op_1130_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_1126_cast_fp16, y = var_1130_cast_fp16)[name = tensor("mh_w_29_cast_fp16")]; + tensor obj_69_cast_fp16 = softmax(axis = var_972, x = mh_w_29_cast_fp16)[name = tensor("obj_69_cast_fp16")]; + tensor var_1134 = const()[name = tensor("op_1134"), val = tensor([1, 12, 64, 1500])]; + tensor var_1135_cast_fp16 = reshape(shape = var_1134, x = value_19_cast_fp16)[name = tensor("op_1135_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_1135_cast_fp16, y = obj_69_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_1138 = const()[name = tensor("op_1138"), val = tensor([1, 768, 1, 1])]; + tensor input_43_cast_fp16 = reshape(shape = var_1138, x = attn_19_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor obj_67_pad_type_0 = const()[name = tensor("obj_67_pad_type_0"), val = tensor("valid")]; + tensor obj_67_strides_0 = const()[name = tensor("obj_67_strides_0"), val = tensor([1, 1])]; + tensor obj_67_pad_0 = const()[name = tensor("obj_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_67_dilations_0 = const()[name = tensor("obj_67_dilations_0"), val = tensor([1, 1])]; + tensor obj_67_groups_0 = const()[name = tensor("obj_67_groups_0"), val = tensor(1)]; + tensor layers_4_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164236608)))]; + tensor layers_4_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165416320)))]; + tensor obj_67_cast_fp16 = conv(bias = layers_4_encoder_attn_o_proj_bias_to_fp16, dilations = obj_67_dilations_0, groups = obj_67_groups_0, pad = obj_67_pad_0, pad_type = obj_67_pad_type_0, strides = obj_67_strides_0, weight = layers_4_encoder_attn_o_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("obj_67_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = obj_67_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; + tensor out_29_axes_0 = const()[name = tensor("out_29_axes_0"), val = tensor([1])]; + tensor var_1156_to_fp16 = const()[name = tensor("op_1156_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_1156_to_fp16, x = inputs_29_cast_fp16)[name = tensor("out_29_cast_fp16")]; + tensor input_45_gamma_0_to_fp16 = const()[name = tensor("input_45_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165417920)))]; + tensor input_45_beta_0_to_fp16 = const()[name = tensor("input_45_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165419520)))]; + tensor input_45_epsilon_0_to_fp16 = const()[name = tensor("input_45_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = batch_norm(beta = input_45_beta_0_to_fp16, epsilon = input_45_epsilon_0_to_fp16, gamma = input_45_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_pad_type_0 = const()[name = tensor("input_47_pad_type_0"), val = tensor("valid")]; + tensor input_47_strides_0 = const()[name = tensor("input_47_strides_0"), val = tensor([1, 1])]; + tensor input_47_pad_0 = const()[name = tensor("input_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_47_dilations_0 = const()[name = tensor("input_47_dilations_0"), val = tensor([1, 1])]; + tensor input_47_groups_0 = const()[name = tensor("input_47_groups_0"), val = tensor(1)]; + tensor layers_4_fc1_weight_to_fp16 = const()[name = tensor("layers_4_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165421120)))]; + tensor layers_4_fc1_bias_to_fp16 = const()[name = tensor("layers_4_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170139776)))]; + tensor input_47_cast_fp16 = conv(bias = layers_4_fc1_bias_to_fp16, dilations = input_47_dilations_0, groups = input_47_groups_0, pad = input_47_pad_0, pad_type = input_47_pad_type_0, strides = input_47_strides_0, weight = layers_4_fc1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor hidden_states_11_pad_type_0 = const()[name = tensor("hidden_states_11_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_11_strides_0 = const()[name = tensor("hidden_states_11_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_11_pad_0 = const()[name = tensor("hidden_states_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_11_dilations_0 = const()[name = tensor("hidden_states_11_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_11_groups_0 = const()[name = tensor("hidden_states_11_groups_0"), val = tensor(1)]; + tensor layers_4_fc2_weight_to_fp16 = const()[name = tensor("layers_4_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170145984)))]; + tensor layers_4_fc2_bias_to_fp16 = const()[name = tensor("layers_4_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174864640)))]; + tensor hidden_states_11_cast_fp16 = conv(bias = layers_4_fc2_bias_to_fp16, dilations = hidden_states_11_dilations_0, groups = hidden_states_11_groups_0, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = hidden_states_11_strides_0, weight = layers_4_fc2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("hidden_states_11_cast_fp16")]; + tensor inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = hidden_states_11_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; + tensor var_1191 = const()[name = tensor("op_1191"), val = tensor(3)]; + tensor out_31_axes_0 = const()[name = tensor("out_31_axes_0"), val = tensor([1])]; + tensor var_1216_to_fp16 = const()[name = tensor("op_1216_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1216_to_fp16, x = inputs_31_cast_fp16)[name = tensor("out_31_cast_fp16")]; + tensor obj_71_gamma_0_to_fp16 = const()[name = tensor("obj_71_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174866240)))]; + tensor obj_71_beta_0_to_fp16 = const()[name = tensor("obj_71_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174867840)))]; + tensor obj_71_epsilon_0_to_fp16 = const()[name = tensor("obj_71_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_71_cast_fp16 = batch_norm(beta = obj_71_beta_0_to_fp16, epsilon = obj_71_epsilon_0_to_fp16, gamma = obj_71_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor("obj_71_cast_fp16")]; + tensor query_21_pad_type_0 = const()[name = tensor("query_21_pad_type_0"), val = tensor("valid")]; + tensor query_21_strides_0 = const()[name = tensor("query_21_strides_0"), val = tensor([1, 1])]; + tensor query_21_pad_0 = const()[name = tensor("query_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_21_dilations_0 = const()[name = tensor("query_21_dilations_0"), val = tensor([1, 1])]; + tensor query_21_groups_0 = const()[name = tensor("query_21_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174869440)))]; + tensor layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176049152)))]; + tensor query_21_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_bias_to_fp16, dilations = query_21_dilations_0, groups = query_21_groups_0, pad = query_21_pad_0, pad_type = query_21_pad_type_0, strides = query_21_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("query_21_cast_fp16")]; + tensor current_key_11_pad_type_0 = const()[name = tensor("current_key_11_pad_type_0"), val = tensor("valid")]; + tensor current_key_11_strides_0 = const()[name = tensor("current_key_11_strides_0"), val = tensor([1, 1])]; + tensor current_key_11_pad_0 = const()[name = tensor("current_key_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_11_dilations_0 = const()[name = tensor("current_key_11_dilations_0"), val = tensor([1, 1])]; + tensor current_key_11_groups_0 = const()[name = tensor("current_key_11_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176050752)))]; + tensor current_key_11_cast_fp16 = conv(dilations = current_key_11_dilations_0, groups = current_key_11_groups_0, pad = current_key_11_pad_0, pad_type = current_key_11_pad_type_0, strides = current_key_11_strides_0, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("current_key_11_cast_fp16")]; + tensor current_value_11_pad_type_0 = const()[name = tensor("current_value_11_pad_type_0"), val = tensor("valid")]; + tensor current_value_11_strides_0 = const()[name = tensor("current_value_11_strides_0"), val = tensor([1, 1])]; + tensor current_value_11_pad_0 = const()[name = tensor("current_value_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_11_dilations_0 = const()[name = tensor("current_value_11_dilations_0"), val = tensor([1, 1])]; + tensor current_value_11_groups_0 = const()[name = tensor("current_value_11_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177230464)))]; + tensor layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178410176)))]; + tensor current_value_11_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_bias_to_fp16, dilations = current_value_11_dilations_0, groups = current_value_11_groups_0, pad = current_value_11_pad_0, pad_type = current_value_11_pad_type_0, strides = current_value_11_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("current_value_11_cast_fp16")]; + tensor var_1255_cast_fp16 = mul(x = var_63_cast_fp16_5, y = var_159_cast_fp16)[name = tensor("op_1255_cast_fp16")]; + tensor var_1256_cast_fp16 = mul(x = current_key_11_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_1256_cast_fp16")]; + tensor key_21_cast_fp16 = add(x = var_1255_cast_fp16, y = var_1256_cast_fp16)[name = tensor("key_21_cast_fp16")]; + tensor var_1259_cast_fp16 = mul(x = var_78_cast_fp16_5, y = var_159_cast_fp16)[name = tensor("op_1259_cast_fp16")]; + tensor var_1260_cast_fp16 = mul(x = current_value_11_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_1260_cast_fp16")]; + tensor value_21_cast_fp16 = add(x = var_1259_cast_fp16, y = var_1260_cast_fp16)[name = tensor("value_21_cast_fp16")]; + tensor var_1264 = const()[name = tensor("op_1264"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_21_cast_fp16 = reshape(shape = var_1264, x = query_21_cast_fp16)[name = tensor("mh_q_21_cast_fp16")]; + tensor var_1266_to_fp16 = const()[name = tensor("op_1266_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1267_cast_fp16 = mul(x = mh_q_21_cast_fp16, y = var_1266_to_fp16)[name = tensor("op_1267_cast_fp16")]; + tensor var_1270 = const()[name = tensor("op_1270"), val = tensor([1, 12, 64, 448])]; + tensor var_1271_cast_fp16 = reshape(shape = var_1270, x = key_21_cast_fp16)[name = tensor("op_1271_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_1267_cast_fp16, y = var_1271_cast_fp16)[name = tensor("mh_w_31_cast_fp16")]; + tensor mh_w_33_cast_fp16 = add(x = mh_w_31_cast_fp16, y = var_181_cast_fp16)[name = tensor("mh_w_33_cast_fp16")]; + tensor var_1279_cast_fp16 = softmax(axis = var_1191, x = mh_w_33_cast_fp16)[name = tensor("op_1279_cast_fp16")]; + tensor var_1280 = const()[name = tensor("op_1280"), val = tensor([1, 12, 64, 448])]; + tensor var_1281_cast_fp16 = reshape(shape = var_1280, x = value_21_cast_fp16)[name = tensor("op_1281_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_1281_cast_fp16, y = var_1279_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_1284 = const()[name = tensor("op_1284"), val = tensor([1, 768, 1, 1])]; + tensor input_51_cast_fp16 = reshape(shape = var_1284, x = attn_21_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor obj_77_pad_type_0 = const()[name = tensor("obj_77_pad_type_0"), val = tensor("valid")]; + tensor obj_77_strides_0 = const()[name = tensor("obj_77_strides_0"), val = tensor([1, 1])]; + tensor obj_77_pad_0 = const()[name = tensor("obj_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_77_dilations_0 = const()[name = tensor("obj_77_dilations_0"), val = tensor([1, 1])]; + tensor obj_77_groups_0 = const()[name = tensor("obj_77_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178411776)))]; + tensor layers_5_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179591488)))]; + tensor obj_77_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_bias_to_fp16, dilations = obj_77_dilations_0, groups = obj_77_groups_0, pad = obj_77_pad_0, pad_type = obj_77_pad_type_0, strides = obj_77_strides_0, weight = layers_5_self_attn_o_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("obj_77_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = obj_77_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([1])]; + tensor var_1306_to_fp16 = const()[name = tensor("op_1306_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1306_to_fp16, x = inputs_33_cast_fp16)[name = tensor("out_33_cast_fp16")]; + tensor obj_79_gamma_0_to_fp16 = const()[name = tensor("obj_79_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179593088)))]; + tensor obj_79_beta_0_to_fp16 = const()[name = tensor("obj_79_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179594688)))]; + tensor obj_79_epsilon_0_to_fp16 = const()[name = tensor("obj_79_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_79_cast_fp16 = batch_norm(beta = obj_79_beta_0_to_fp16, epsilon = obj_79_epsilon_0_to_fp16, gamma = obj_79_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor("obj_79_cast_fp16")]; + tensor query_23_pad_type_0 = const()[name = tensor("query_23_pad_type_0"), val = tensor("valid")]; + tensor query_23_strides_0 = const()[name = tensor("query_23_strides_0"), val = tensor([1, 1])]; + tensor query_23_pad_0 = const()[name = tensor("query_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_23_dilations_0 = const()[name = tensor("query_23_dilations_0"), val = tensor([1, 1])]; + tensor query_23_groups_0 = const()[name = tensor("query_23_groups_0"), val = tensor(1)]; + tensor layers_5_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179596288)))]; + tensor layers_5_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180776000)))]; + tensor query_23_cast_fp16 = conv(bias = layers_5_encoder_attn_q_proj_bias_to_fp16, dilations = query_23_dilations_0, groups = query_23_groups_0, pad = query_23_pad_0, pad_type = query_23_pad_type_0, strides = query_23_strides_0, weight = layers_5_encoder_attn_q_proj_weight_to_fp16, x = obj_79_cast_fp16)[name = tensor("query_23_cast_fp16")]; + tensor key_23_pad_type_0 = const()[name = tensor("key_23_pad_type_0"), val = tensor("valid")]; + tensor key_23_strides_0 = const()[name = tensor("key_23_strides_0"), val = tensor([1, 1])]; + tensor key_23_pad_0 = const()[name = tensor("key_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_23_dilations_0 = const()[name = tensor("key_23_dilations_0"), val = tensor([1, 1])]; + tensor key_23_groups_0 = const()[name = tensor("key_23_groups_0"), val = tensor(1)]; + tensor layers_5_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180777600)))]; + tensor key_23_cast_fp16 = conv(dilations = key_23_dilations_0, groups = key_23_groups_0, pad = key_23_pad_0, pad_type = key_23_pad_type_0, strides = key_23_strides_0, weight = layers_5_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_23_cast_fp16")]; + tensor value_23_pad_type_0 = const()[name = tensor("value_23_pad_type_0"), val = tensor("valid")]; + tensor value_23_strides_0 = const()[name = tensor("value_23_strides_0"), val = tensor([1, 1])]; + tensor value_23_pad_0 = const()[name = tensor("value_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_23_dilations_0 = const()[name = tensor("value_23_dilations_0"), val = tensor([1, 1])]; + tensor value_23_groups_0 = const()[name = tensor("value_23_groups_0"), val = tensor(1)]; + tensor layers_5_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181957312)))]; + tensor layers_5_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183137024)))]; + tensor value_23_cast_fp16 = conv(bias = layers_5_encoder_attn_v_proj_bias_to_fp16, dilations = value_23_dilations_0, groups = value_23_groups_0, pad = value_23_pad_0, pad_type = value_23_pad_type_0, strides = value_23_strides_0, weight = layers_5_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_23_cast_fp16")]; + tensor var_1342 = const()[name = tensor("op_1342"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_23_cast_fp16 = reshape(shape = var_1342, x = query_23_cast_fp16)[name = tensor("mh_q_23_cast_fp16")]; + tensor var_1344_to_fp16 = const()[name = tensor("op_1344_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1345_cast_fp16 = mul(x = mh_q_23_cast_fp16, y = var_1344_to_fp16)[name = tensor("op_1345_cast_fp16")]; + tensor var_1348 = const()[name = tensor("op_1348"), val = tensor([1, 12, 64, 1500])]; + tensor var_1349_cast_fp16 = reshape(shape = var_1348, x = key_23_cast_fp16)[name = tensor("op_1349_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_1345_cast_fp16, y = var_1349_cast_fp16)[name = tensor("mh_w_35_cast_fp16")]; + tensor obj_83_cast_fp16 = softmax(axis = var_1191, x = mh_w_35_cast_fp16)[name = tensor("obj_83_cast_fp16")]; + tensor var_1353 = const()[name = tensor("op_1353"), val = tensor([1, 12, 64, 1500])]; + tensor var_1354_cast_fp16 = reshape(shape = var_1353, x = value_23_cast_fp16)[name = tensor("op_1354_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_1354_cast_fp16, y = obj_83_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_1357 = const()[name = tensor("op_1357"), val = tensor([1, 768, 1, 1])]; + tensor input_53_cast_fp16 = reshape(shape = var_1357, x = attn_23_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor obj_81_pad_type_0 = const()[name = tensor("obj_81_pad_type_0"), val = tensor("valid")]; + tensor obj_81_strides_0 = const()[name = tensor("obj_81_strides_0"), val = tensor([1, 1])]; + tensor obj_81_pad_0 = const()[name = tensor("obj_81_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_81_dilations_0 = const()[name = tensor("obj_81_dilations_0"), val = tensor([1, 1])]; + tensor obj_81_groups_0 = const()[name = tensor("obj_81_groups_0"), val = tensor(1)]; + tensor layers_5_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183138624)))]; + tensor layers_5_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184318336)))]; + tensor obj_81_cast_fp16 = conv(bias = layers_5_encoder_attn_o_proj_bias_to_fp16, dilations = obj_81_dilations_0, groups = obj_81_groups_0, pad = obj_81_pad_0, pad_type = obj_81_pad_type_0, strides = obj_81_strides_0, weight = layers_5_encoder_attn_o_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("obj_81_cast_fp16")]; + tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_81_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; + tensor out_35_axes_0 = const()[name = tensor("out_35_axes_0"), val = tensor([1])]; + tensor var_1378_to_fp16 = const()[name = tensor("op_1378_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1378_to_fp16, x = inputs_35_cast_fp16)[name = tensor("out_35_cast_fp16")]; + tensor input_55_gamma_0_to_fp16 = const()[name = tensor("input_55_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184319936)))]; + tensor input_55_beta_0_to_fp16 = const()[name = tensor("input_55_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184321536)))]; + tensor input_55_epsilon_0_to_fp16 = const()[name = tensor("input_55_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_55_cast_fp16 = batch_norm(beta = input_55_beta_0_to_fp16, epsilon = input_55_epsilon_0_to_fp16, gamma = input_55_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_pad_type_0 = const()[name = tensor("input_57_pad_type_0"), val = tensor("valid")]; + tensor input_57_strides_0 = const()[name = tensor("input_57_strides_0"), val = tensor([1, 1])]; + tensor input_57_pad_0 = const()[name = tensor("input_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_57_dilations_0 = const()[name = tensor("input_57_dilations_0"), val = tensor([1, 1])]; + tensor input_57_groups_0 = const()[name = tensor("input_57_groups_0"), val = tensor(1)]; + tensor layers_5_fc1_weight_to_fp16 = const()[name = tensor("layers_5_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184323136)))]; + tensor layers_5_fc1_bias_to_fp16 = const()[name = tensor("layers_5_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189041792)))]; + tensor input_57_cast_fp16 = conv(bias = layers_5_fc1_bias_to_fp16, dilations = input_57_dilations_0, groups = input_57_groups_0, pad = input_57_pad_0, pad_type = input_57_pad_type_0, strides = input_57_strides_0, weight = layers_5_fc1_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor hidden_states_13_pad_type_0 = const()[name = tensor("hidden_states_13_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_13_strides_0 = const()[name = tensor("hidden_states_13_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_13_pad_0 = const()[name = tensor("hidden_states_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_13_dilations_0 = const()[name = tensor("hidden_states_13_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_13_groups_0 = const()[name = tensor("hidden_states_13_groups_0"), val = tensor(1)]; + tensor layers_5_fc2_weight_to_fp16 = const()[name = tensor("layers_5_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189048000)))]; + tensor layers_5_fc2_bias_to_fp16 = const()[name = tensor("layers_5_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193766656)))]; + tensor hidden_states_13_cast_fp16 = conv(bias = layers_5_fc2_bias_to_fp16, dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_5_fc2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("hidden_states_13_cast_fp16")]; + tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor("inputs_37_cast_fp16")]; + tensor var_1414 = const()[name = tensor("op_1414"), val = tensor(3)]; + tensor out_37_axes_0 = const()[name = tensor("out_37_axes_0"), val = tensor([1])]; + tensor var_1439_to_fp16 = const()[name = tensor("op_1439_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_1439_to_fp16, x = inputs_37_cast_fp16)[name = tensor("out_37_cast_fp16")]; + tensor obj_85_gamma_0_to_fp16 = const()[name = tensor("obj_85_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193768256)))]; + 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(193769856)))]; + tensor obj_85_epsilon_0_to_fp16 = const()[name = tensor("obj_85_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_85_cast_fp16 = batch_norm(beta = obj_85_beta_0_to_fp16, epsilon = obj_85_epsilon_0_to_fp16, gamma = obj_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_37_cast_fp16)[name = tensor("obj_85_cast_fp16")]; + tensor query_25_pad_type_0 = const()[name = tensor("query_25_pad_type_0"), val = tensor("valid")]; + tensor query_25_strides_0 = const()[name = tensor("query_25_strides_0"), val = tensor([1, 1])]; + tensor query_25_pad_0 = const()[name = tensor("query_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_25_dilations_0 = const()[name = tensor("query_25_dilations_0"), val = tensor([1, 1])]; + tensor query_25_groups_0 = const()[name = tensor("query_25_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193771456)))]; + tensor layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194951168)))]; + tensor query_25_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_bias_to_fp16, dilations = query_25_dilations_0, groups = query_25_groups_0, pad = query_25_pad_0, pad_type = query_25_pad_type_0, strides = query_25_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("query_25_cast_fp16")]; + tensor current_key_13_pad_type_0 = const()[name = tensor("current_key_13_pad_type_0"), val = tensor("valid")]; + tensor current_key_13_strides_0 = const()[name = tensor("current_key_13_strides_0"), val = tensor([1, 1])]; + tensor current_key_13_pad_0 = const()[name = tensor("current_key_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_13_dilations_0 = const()[name = tensor("current_key_13_dilations_0"), val = tensor([1, 1])]; + tensor current_key_13_groups_0 = const()[name = tensor("current_key_13_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194952768)))]; + tensor current_key_13_cast_fp16 = conv(dilations = current_key_13_dilations_0, groups = current_key_13_groups_0, pad = current_key_13_pad_0, pad_type = current_key_13_pad_type_0, strides = current_key_13_strides_0, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("current_key_13_cast_fp16")]; + tensor current_value_13_pad_type_0 = const()[name = tensor("current_value_13_pad_type_0"), val = tensor("valid")]; + tensor current_value_13_strides_0 = const()[name = tensor("current_value_13_strides_0"), val = tensor([1, 1])]; + tensor current_value_13_pad_0 = const()[name = tensor("current_value_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_13_dilations_0 = const()[name = tensor("current_value_13_dilations_0"), val = tensor([1, 1])]; + tensor current_value_13_groups_0 = const()[name = tensor("current_value_13_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196132480)))]; + tensor layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197312192)))]; + tensor current_value_13_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_bias_to_fp16, dilations = current_value_13_dilations_0, groups = current_value_13_groups_0, pad = current_value_13_pad_0, pad_type = current_value_13_pad_type_0, strides = current_value_13_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("current_value_13_cast_fp16")]; + tensor var_1478_cast_fp16 = mul(x = var_63_cast_fp16_6, y = var_159_cast_fp16)[name = tensor("op_1478_cast_fp16")]; + tensor var_1479_cast_fp16 = mul(x = current_key_13_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_1479_cast_fp16")]; + tensor key_25_cast_fp16 = add(x = var_1478_cast_fp16, y = var_1479_cast_fp16)[name = tensor("key_25_cast_fp16")]; + tensor var_1482_cast_fp16 = mul(x = var_78_cast_fp16_6, y = var_159_cast_fp16)[name = tensor("op_1482_cast_fp16")]; + tensor var_1483_cast_fp16 = mul(x = current_value_13_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_1483_cast_fp16")]; + tensor value_25_cast_fp16 = add(x = var_1482_cast_fp16, y = var_1483_cast_fp16)[name = tensor("value_25_cast_fp16")]; + tensor var_1487 = const()[name = tensor("op_1487"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_25_cast_fp16 = reshape(shape = var_1487, x = query_25_cast_fp16)[name = tensor("mh_q_25_cast_fp16")]; + tensor var_1489_to_fp16 = const()[name = tensor("op_1489_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1490_cast_fp16 = mul(x = mh_q_25_cast_fp16, y = var_1489_to_fp16)[name = tensor("op_1490_cast_fp16")]; + tensor var_1493 = const()[name = tensor("op_1493"), val = tensor([1, 12, 64, 448])]; + tensor var_1494_cast_fp16 = reshape(shape = var_1493, x = key_25_cast_fp16)[name = tensor("op_1494_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_1490_cast_fp16, y = var_1494_cast_fp16)[name = tensor("mh_w_37_cast_fp16")]; + tensor mh_w_39_cast_fp16 = add(x = mh_w_37_cast_fp16, y = var_181_cast_fp16)[name = tensor("mh_w_39_cast_fp16")]; + tensor var_1502_cast_fp16 = softmax(axis = var_1414, x = mh_w_39_cast_fp16)[name = tensor("op_1502_cast_fp16")]; + tensor var_1503 = const()[name = tensor("op_1503"), val = tensor([1, 12, 64, 448])]; + tensor var_1504_cast_fp16 = reshape(shape = var_1503, x = value_25_cast_fp16)[name = tensor("op_1504_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_1504_cast_fp16, y = var_1502_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_1507 = const()[name = tensor("op_1507"), val = tensor([1, 768, 1, 1])]; + tensor input_61_cast_fp16 = reshape(shape = var_1507, x = attn_25_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor obj_91_pad_type_0 = const()[name = tensor("obj_91_pad_type_0"), val = tensor("valid")]; + tensor obj_91_strides_0 = const()[name = tensor("obj_91_strides_0"), val = tensor([1, 1])]; + tensor obj_91_pad_0 = const()[name = tensor("obj_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_91_dilations_0 = const()[name = tensor("obj_91_dilations_0"), val = tensor([1, 1])]; + tensor obj_91_groups_0 = const()[name = tensor("obj_91_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197313792)))]; + tensor layers_6_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198493504)))]; + tensor obj_91_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_bias_to_fp16, dilations = obj_91_dilations_0, groups = obj_91_groups_0, pad = obj_91_pad_0, pad_type = obj_91_pad_type_0, strides = obj_91_strides_0, weight = layers_6_self_attn_o_proj_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("obj_91_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_91_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; + tensor out_39_axes_0 = const()[name = tensor("out_39_axes_0"), val = tensor([1])]; + tensor var_1529_to_fp16 = const()[name = tensor("op_1529_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_1529_to_fp16, x = inputs_39_cast_fp16)[name = tensor("out_39_cast_fp16")]; + tensor obj_93_gamma_0_to_fp16 = const()[name = tensor("obj_93_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198495104)))]; + 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(198496704)))]; + tensor obj_93_epsilon_0_to_fp16 = const()[name = tensor("obj_93_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_93_cast_fp16 = batch_norm(beta = obj_93_beta_0_to_fp16, epsilon = obj_93_epsilon_0_to_fp16, gamma = obj_93_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_39_cast_fp16)[name = tensor("obj_93_cast_fp16")]; + tensor query_27_pad_type_0 = const()[name = tensor("query_27_pad_type_0"), val = tensor("valid")]; + tensor query_27_strides_0 = const()[name = tensor("query_27_strides_0"), val = tensor([1, 1])]; + tensor query_27_pad_0 = const()[name = tensor("query_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_27_dilations_0 = const()[name = tensor("query_27_dilations_0"), val = tensor([1, 1])]; + tensor query_27_groups_0 = const()[name = tensor("query_27_groups_0"), val = tensor(1)]; + tensor layers_6_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198498304)))]; + tensor layers_6_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199678016)))]; + tensor query_27_cast_fp16 = conv(bias = layers_6_encoder_attn_q_proj_bias_to_fp16, dilations = query_27_dilations_0, groups = query_27_groups_0, pad = query_27_pad_0, pad_type = query_27_pad_type_0, strides = query_27_strides_0, weight = layers_6_encoder_attn_q_proj_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor("query_27_cast_fp16")]; + tensor key_27_pad_type_0 = const()[name = tensor("key_27_pad_type_0"), val = tensor("valid")]; + tensor key_27_strides_0 = const()[name = tensor("key_27_strides_0"), val = tensor([1, 1])]; + tensor key_27_pad_0 = const()[name = tensor("key_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_27_dilations_0 = const()[name = tensor("key_27_dilations_0"), val = tensor([1, 1])]; + tensor key_27_groups_0 = const()[name = tensor("key_27_groups_0"), val = tensor(1)]; + tensor layers_6_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199679616)))]; + tensor key_27_cast_fp16 = conv(dilations = key_27_dilations_0, groups = key_27_groups_0, pad = key_27_pad_0, pad_type = key_27_pad_type_0, strides = key_27_strides_0, weight = layers_6_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_27_cast_fp16")]; + tensor value_27_pad_type_0 = const()[name = tensor("value_27_pad_type_0"), val = tensor("valid")]; + tensor value_27_strides_0 = const()[name = tensor("value_27_strides_0"), val = tensor([1, 1])]; + tensor value_27_pad_0 = const()[name = tensor("value_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_27_dilations_0 = const()[name = tensor("value_27_dilations_0"), val = tensor([1, 1])]; + tensor value_27_groups_0 = const()[name = tensor("value_27_groups_0"), val = tensor(1)]; + tensor layers_6_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200859328)))]; + tensor layers_6_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202039040)))]; + tensor value_27_cast_fp16 = conv(bias = layers_6_encoder_attn_v_proj_bias_to_fp16, dilations = value_27_dilations_0, groups = value_27_groups_0, pad = value_27_pad_0, pad_type = value_27_pad_type_0, strides = value_27_strides_0, weight = layers_6_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_27_cast_fp16")]; + tensor var_1565 = const()[name = tensor("op_1565"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_27_cast_fp16 = reshape(shape = var_1565, x = query_27_cast_fp16)[name = tensor("mh_q_27_cast_fp16")]; + tensor var_1567_to_fp16 = const()[name = tensor("op_1567_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1568_cast_fp16 = mul(x = mh_q_27_cast_fp16, y = var_1567_to_fp16)[name = tensor("op_1568_cast_fp16")]; + tensor var_1571 = const()[name = tensor("op_1571"), val = tensor([1, 12, 64, 1500])]; + tensor var_1572_cast_fp16 = reshape(shape = var_1571, x = key_27_cast_fp16)[name = tensor("op_1572_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_1568_cast_fp16, y = var_1572_cast_fp16)[name = tensor("mh_w_41_cast_fp16")]; + tensor obj_97_cast_fp16 = softmax(axis = var_1414, x = mh_w_41_cast_fp16)[name = tensor("obj_97_cast_fp16")]; + tensor var_1576 = const()[name = tensor("op_1576"), val = tensor([1, 12, 64, 1500])]; + tensor var_1577_cast_fp16 = reshape(shape = var_1576, x = value_27_cast_fp16)[name = tensor("op_1577_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_1577_cast_fp16, y = obj_97_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_1580 = const()[name = tensor("op_1580"), val = tensor([1, 768, 1, 1])]; + tensor input_63_cast_fp16 = reshape(shape = var_1580, x = attn_27_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor obj_95_pad_type_0 = const()[name = tensor("obj_95_pad_type_0"), val = tensor("valid")]; + tensor obj_95_strides_0 = const()[name = tensor("obj_95_strides_0"), val = tensor([1, 1])]; + tensor obj_95_pad_0 = const()[name = tensor("obj_95_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_95_dilations_0 = const()[name = tensor("obj_95_dilations_0"), val = tensor([1, 1])]; + tensor obj_95_groups_0 = const()[name = tensor("obj_95_groups_0"), val = tensor(1)]; + tensor layers_6_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202040640)))]; + tensor layers_6_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203220352)))]; + tensor obj_95_cast_fp16 = conv(bias = layers_6_encoder_attn_o_proj_bias_to_fp16, dilations = obj_95_dilations_0, groups = obj_95_groups_0, pad = obj_95_pad_0, pad_type = obj_95_pad_type_0, strides = obj_95_strides_0, weight = layers_6_encoder_attn_o_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("obj_95_cast_fp16")]; + tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = obj_95_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; + tensor out_41_axes_0 = const()[name = tensor("out_41_axes_0"), val = tensor([1])]; + tensor var_1598_to_fp16 = const()[name = tensor("op_1598_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_1598_to_fp16, x = inputs_41_cast_fp16)[name = tensor("out_41_cast_fp16")]; + tensor input_65_gamma_0_to_fp16 = const()[name = tensor("input_65_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203221952)))]; + tensor input_65_beta_0_to_fp16 = const()[name = tensor("input_65_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203223552)))]; + tensor input_65_epsilon_0_to_fp16 = const()[name = tensor("input_65_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = batch_norm(beta = input_65_beta_0_to_fp16, epsilon = input_65_epsilon_0_to_fp16, gamma = input_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_41_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor input_67_pad_type_0 = const()[name = tensor("input_67_pad_type_0"), val = tensor("valid")]; + tensor input_67_strides_0 = const()[name = tensor("input_67_strides_0"), val = tensor([1, 1])]; + tensor input_67_pad_0 = const()[name = tensor("input_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_67_dilations_0 = const()[name = tensor("input_67_dilations_0"), val = tensor([1, 1])]; + tensor input_67_groups_0 = const()[name = tensor("input_67_groups_0"), val = tensor(1)]; + tensor layers_6_fc1_weight_to_fp16 = const()[name = tensor("layers_6_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203225152)))]; + tensor layers_6_fc1_bias_to_fp16 = const()[name = tensor("layers_6_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207943808)))]; + tensor input_67_cast_fp16 = conv(bias = layers_6_fc1_bias_to_fp16, dilations = input_67_dilations_0, groups = input_67_groups_0, pad = input_67_pad_0, pad_type = input_67_pad_type_0, strides = input_67_strides_0, weight = layers_6_fc1_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_mode_0 = const()[name = tensor("input_69_mode_0"), val = tensor("EXACT")]; + tensor input_69_cast_fp16 = gelu(mode = input_69_mode_0, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor hidden_states_15_pad_type_0 = const()[name = tensor("hidden_states_15_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_15_strides_0 = const()[name = tensor("hidden_states_15_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_15_pad_0 = const()[name = tensor("hidden_states_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_15_dilations_0 = const()[name = tensor("hidden_states_15_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_15_groups_0 = const()[name = tensor("hidden_states_15_groups_0"), val = tensor(1)]; + tensor layers_6_fc2_weight_to_fp16 = const()[name = tensor("layers_6_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207950016)))]; + tensor layers_6_fc2_bias_to_fp16 = const()[name = tensor("layers_6_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212668672)))]; + tensor hidden_states_15_cast_fp16 = conv(bias = layers_6_fc2_bias_to_fp16, dilations = hidden_states_15_dilations_0, groups = hidden_states_15_groups_0, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = hidden_states_15_strides_0, weight = layers_6_fc2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("hidden_states_15_cast_fp16")]; + tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = hidden_states_15_cast_fp16)[name = tensor("inputs_43_cast_fp16")]; + tensor var_1633 = const()[name = tensor("op_1633"), val = tensor(3)]; + tensor out_43_axes_0 = const()[name = tensor("out_43_axes_0"), val = tensor([1])]; + tensor var_1658_to_fp16 = const()[name = tensor("op_1658_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_1658_to_fp16, x = inputs_43_cast_fp16)[name = tensor("out_43_cast_fp16")]; + tensor obj_99_gamma_0_to_fp16 = const()[name = tensor("obj_99_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212670272)))]; + tensor obj_99_beta_0_to_fp16 = const()[name = tensor("obj_99_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212671872)))]; + tensor obj_99_epsilon_0_to_fp16 = const()[name = tensor("obj_99_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_99_cast_fp16 = batch_norm(beta = obj_99_beta_0_to_fp16, epsilon = obj_99_epsilon_0_to_fp16, gamma = obj_99_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_43_cast_fp16)[name = tensor("obj_99_cast_fp16")]; + tensor query_29_pad_type_0 = const()[name = tensor("query_29_pad_type_0"), val = tensor("valid")]; + tensor query_29_strides_0 = const()[name = tensor("query_29_strides_0"), val = tensor([1, 1])]; + tensor query_29_pad_0 = const()[name = tensor("query_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_29_dilations_0 = const()[name = tensor("query_29_dilations_0"), val = tensor([1, 1])]; + tensor query_29_groups_0 = const()[name = tensor("query_29_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212673472)))]; + tensor layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213853184)))]; + tensor query_29_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_bias_to_fp16, dilations = query_29_dilations_0, groups = query_29_groups_0, pad = query_29_pad_0, pad_type = query_29_pad_type_0, strides = query_29_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("query_29_cast_fp16")]; + tensor current_key_15_pad_type_0 = const()[name = tensor("current_key_15_pad_type_0"), val = tensor("valid")]; + tensor current_key_15_strides_0 = const()[name = tensor("current_key_15_strides_0"), val = tensor([1, 1])]; + tensor current_key_15_pad_0 = const()[name = tensor("current_key_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_15_dilations_0 = const()[name = tensor("current_key_15_dilations_0"), val = tensor([1, 1])]; + tensor current_key_15_groups_0 = const()[name = tensor("current_key_15_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213854784)))]; + tensor current_key_15_cast_fp16 = conv(dilations = current_key_15_dilations_0, groups = current_key_15_groups_0, pad = current_key_15_pad_0, pad_type = current_key_15_pad_type_0, strides = current_key_15_strides_0, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("current_key_15_cast_fp16")]; + tensor current_value_15_pad_type_0 = const()[name = tensor("current_value_15_pad_type_0"), val = tensor("valid")]; + tensor current_value_15_strides_0 = const()[name = tensor("current_value_15_strides_0"), val = tensor([1, 1])]; + tensor current_value_15_pad_0 = const()[name = tensor("current_value_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_15_dilations_0 = const()[name = tensor("current_value_15_dilations_0"), val = tensor([1, 1])]; + tensor current_value_15_groups_0 = const()[name = tensor("current_value_15_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215034496)))]; + tensor layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216214208)))]; + tensor current_value_15_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_bias_to_fp16, dilations = current_value_15_dilations_0, groups = current_value_15_groups_0, pad = current_value_15_pad_0, pad_type = current_value_15_pad_type_0, strides = current_value_15_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("current_value_15_cast_fp16")]; + tensor var_1697_cast_fp16 = mul(x = var_63_cast_fp16_7, y = var_159_cast_fp16)[name = tensor("op_1697_cast_fp16")]; + tensor var_1698_cast_fp16 = mul(x = current_key_15_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_1698_cast_fp16")]; + tensor key_29_cast_fp16 = add(x = var_1697_cast_fp16, y = var_1698_cast_fp16)[name = tensor("key_29_cast_fp16")]; + tensor var_1701_cast_fp16 = mul(x = var_78_cast_fp16_7, y = var_159_cast_fp16)[name = tensor("op_1701_cast_fp16")]; + tensor var_1702_cast_fp16 = mul(x = current_value_15_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_1702_cast_fp16")]; + tensor value_29_cast_fp16 = add(x = var_1701_cast_fp16, y = var_1702_cast_fp16)[name = tensor("value_29_cast_fp16")]; + tensor var_1706 = const()[name = tensor("op_1706"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_29_cast_fp16 = reshape(shape = var_1706, x = query_29_cast_fp16)[name = tensor("mh_q_29_cast_fp16")]; + tensor var_1708_to_fp16 = const()[name = tensor("op_1708_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1709_cast_fp16 = mul(x = mh_q_29_cast_fp16, y = var_1708_to_fp16)[name = tensor("op_1709_cast_fp16")]; + tensor var_1712 = const()[name = tensor("op_1712"), val = tensor([1, 12, 64, 448])]; + tensor var_1713_cast_fp16 = reshape(shape = var_1712, x = key_29_cast_fp16)[name = tensor("op_1713_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_1709_cast_fp16, y = var_1713_cast_fp16)[name = tensor("mh_w_43_cast_fp16")]; + tensor mh_w_45_cast_fp16 = add(x = mh_w_43_cast_fp16, y = var_181_cast_fp16)[name = tensor("mh_w_45_cast_fp16")]; + tensor var_1721_cast_fp16 = softmax(axis = var_1633, x = mh_w_45_cast_fp16)[name = tensor("op_1721_cast_fp16")]; + tensor var_1722 = const()[name = tensor("op_1722"), val = tensor([1, 12, 64, 448])]; + tensor var_1723_cast_fp16 = reshape(shape = var_1722, x = value_29_cast_fp16)[name = tensor("op_1723_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_1723_cast_fp16, y = var_1721_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_1726 = const()[name = tensor("op_1726"), val = tensor([1, 768, 1, 1])]; + tensor input_71_cast_fp16 = reshape(shape = var_1726, x = attn_29_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor obj_105_pad_type_0 = const()[name = tensor("obj_105_pad_type_0"), val = tensor("valid")]; + tensor obj_105_strides_0 = const()[name = tensor("obj_105_strides_0"), val = tensor([1, 1])]; + tensor obj_105_pad_0 = const()[name = tensor("obj_105_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_105_dilations_0 = const()[name = tensor("obj_105_dilations_0"), val = tensor([1, 1])]; + tensor obj_105_groups_0 = const()[name = tensor("obj_105_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216215808)))]; + tensor layers_7_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217395520)))]; + tensor obj_105_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_bias_to_fp16, dilations = obj_105_dilations_0, groups = obj_105_groups_0, pad = obj_105_pad_0, pad_type = obj_105_pad_type_0, strides = obj_105_strides_0, weight = layers_7_self_attn_o_proj_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("obj_105_cast_fp16")]; + tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = obj_105_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; + tensor out_45_axes_0 = const()[name = tensor("out_45_axes_0"), val = tensor([1])]; + tensor var_1748_to_fp16 = const()[name = tensor("op_1748_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_1748_to_fp16, x = inputs_45_cast_fp16)[name = tensor("out_45_cast_fp16")]; + tensor obj_107_gamma_0_to_fp16 = const()[name = tensor("obj_107_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217397120)))]; + tensor obj_107_beta_0_to_fp16 = const()[name = tensor("obj_107_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217398720)))]; + tensor obj_107_epsilon_0_to_fp16 = const()[name = tensor("obj_107_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_107_cast_fp16 = batch_norm(beta = obj_107_beta_0_to_fp16, epsilon = obj_107_epsilon_0_to_fp16, gamma = obj_107_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor("obj_107_cast_fp16")]; + tensor query_31_pad_type_0 = const()[name = tensor("query_31_pad_type_0"), val = tensor("valid")]; + tensor query_31_strides_0 = const()[name = tensor("query_31_strides_0"), val = tensor([1, 1])]; + tensor query_31_pad_0 = const()[name = tensor("query_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_31_dilations_0 = const()[name = tensor("query_31_dilations_0"), val = tensor([1, 1])]; + tensor query_31_groups_0 = const()[name = tensor("query_31_groups_0"), val = tensor(1)]; + tensor layers_7_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217400320)))]; + tensor layers_7_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218580032)))]; + tensor query_31_cast_fp16 = conv(bias = layers_7_encoder_attn_q_proj_bias_to_fp16, dilations = query_31_dilations_0, groups = query_31_groups_0, pad = query_31_pad_0, pad_type = query_31_pad_type_0, strides = query_31_strides_0, weight = layers_7_encoder_attn_q_proj_weight_to_fp16, x = obj_107_cast_fp16)[name = tensor("query_31_cast_fp16")]; + tensor key_31_pad_type_0 = const()[name = tensor("key_31_pad_type_0"), val = tensor("valid")]; + tensor key_31_strides_0 = const()[name = tensor("key_31_strides_0"), val = tensor([1, 1])]; + tensor key_31_pad_0 = const()[name = tensor("key_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_31_dilations_0 = const()[name = tensor("key_31_dilations_0"), val = tensor([1, 1])]; + tensor key_31_groups_0 = const()[name = tensor("key_31_groups_0"), val = tensor(1)]; + tensor layers_7_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218581632)))]; + tensor key_31_cast_fp16 = conv(dilations = key_31_dilations_0, groups = key_31_groups_0, pad = key_31_pad_0, pad_type = key_31_pad_type_0, strides = key_31_strides_0, weight = layers_7_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_31_cast_fp16")]; + tensor value_31_pad_type_0 = const()[name = tensor("value_31_pad_type_0"), val = tensor("valid")]; + tensor value_31_strides_0 = const()[name = tensor("value_31_strides_0"), val = tensor([1, 1])]; + tensor value_31_pad_0 = const()[name = tensor("value_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_31_dilations_0 = const()[name = tensor("value_31_dilations_0"), val = tensor([1, 1])]; + tensor value_31_groups_0 = const()[name = tensor("value_31_groups_0"), val = tensor(1)]; + tensor layers_7_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219761344)))]; + tensor layers_7_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220941056)))]; + tensor value_31_cast_fp16 = conv(bias = layers_7_encoder_attn_v_proj_bias_to_fp16, dilations = value_31_dilations_0, groups = value_31_groups_0, pad = value_31_pad_0, pad_type = value_31_pad_type_0, strides = value_31_strides_0, weight = layers_7_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_31_cast_fp16")]; + tensor var_1784 = const()[name = tensor("op_1784"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_31_cast_fp16 = reshape(shape = var_1784, x = query_31_cast_fp16)[name = tensor("mh_q_31_cast_fp16")]; + tensor var_1786_to_fp16 = const()[name = tensor("op_1786_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1787_cast_fp16 = mul(x = mh_q_31_cast_fp16, y = var_1786_to_fp16)[name = tensor("op_1787_cast_fp16")]; + tensor var_1790 = const()[name = tensor("op_1790"), val = tensor([1, 12, 64, 1500])]; + tensor var_1791_cast_fp16 = reshape(shape = var_1790, x = key_31_cast_fp16)[name = tensor("op_1791_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_1787_cast_fp16, y = var_1791_cast_fp16)[name = tensor("mh_w_47_cast_fp16")]; + tensor obj_111_cast_fp16 = softmax(axis = var_1633, x = mh_w_47_cast_fp16)[name = tensor("obj_111_cast_fp16")]; + tensor var_1795 = const()[name = tensor("op_1795"), val = tensor([1, 12, 64, 1500])]; + tensor var_1796_cast_fp16 = reshape(shape = var_1795, x = value_31_cast_fp16)[name = tensor("op_1796_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_1796_cast_fp16, y = obj_111_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_1799 = const()[name = tensor("op_1799"), val = tensor([1, 768, 1, 1])]; + tensor input_73_cast_fp16 = reshape(shape = var_1799, x = attn_31_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor obj_109_pad_type_0 = const()[name = tensor("obj_109_pad_type_0"), val = tensor("valid")]; + tensor obj_109_strides_0 = const()[name = tensor("obj_109_strides_0"), val = tensor([1, 1])]; + tensor obj_109_pad_0 = const()[name = tensor("obj_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_109_dilations_0 = const()[name = tensor("obj_109_dilations_0"), val = tensor([1, 1])]; + tensor obj_109_groups_0 = const()[name = tensor("obj_109_groups_0"), val = tensor(1)]; + tensor layers_7_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220942656)))]; + tensor layers_7_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222122368)))]; + tensor obj_109_cast_fp16 = conv(bias = layers_7_encoder_attn_o_proj_bias_to_fp16, dilations = obj_109_dilations_0, groups = obj_109_groups_0, pad = obj_109_pad_0, pad_type = obj_109_pad_type_0, strides = obj_109_strides_0, weight = layers_7_encoder_attn_o_proj_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("obj_109_cast_fp16")]; + tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_109_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; + tensor out_47_axes_0 = const()[name = tensor("out_47_axes_0"), val = tensor([1])]; + tensor var_1817_to_fp16 = const()[name = tensor("op_1817_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_1817_to_fp16, x = inputs_47_cast_fp16)[name = tensor("out_47_cast_fp16")]; + tensor input_75_gamma_0_to_fp16 = const()[name = tensor("input_75_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222123968)))]; + tensor input_75_beta_0_to_fp16 = const()[name = tensor("input_75_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222125568)))]; + tensor input_75_epsilon_0_to_fp16 = const()[name = tensor("input_75_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_75_cast_fp16 = batch_norm(beta = input_75_beta_0_to_fp16, epsilon = input_75_epsilon_0_to_fp16, gamma = input_75_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_47_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_77_pad_type_0 = const()[name = tensor("input_77_pad_type_0"), val = tensor("valid")]; + tensor input_77_strides_0 = const()[name = tensor("input_77_strides_0"), val = tensor([1, 1])]; + tensor input_77_pad_0 = const()[name = tensor("input_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_77_dilations_0 = const()[name = tensor("input_77_dilations_0"), val = tensor([1, 1])]; + tensor input_77_groups_0 = const()[name = tensor("input_77_groups_0"), val = tensor(1)]; + tensor layers_7_fc1_weight_to_fp16 = const()[name = tensor("layers_7_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222127168)))]; + tensor layers_7_fc1_bias_to_fp16 = const()[name = tensor("layers_7_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226845824)))]; + tensor input_77_cast_fp16 = conv(bias = layers_7_fc1_bias_to_fp16, dilations = input_77_dilations_0, groups = input_77_groups_0, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = input_77_strides_0, weight = layers_7_fc1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_mode_0 = const()[name = tensor("input_79_mode_0"), val = tensor("EXACT")]; + tensor input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor hidden_states_17_pad_type_0 = const()[name = tensor("hidden_states_17_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_17_strides_0 = const()[name = tensor("hidden_states_17_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_17_pad_0 = const()[name = tensor("hidden_states_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_17_dilations_0 = const()[name = tensor("hidden_states_17_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_17_groups_0 = const()[name = tensor("hidden_states_17_groups_0"), val = tensor(1)]; + tensor layers_7_fc2_weight_to_fp16 = const()[name = tensor("layers_7_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226852032)))]; + tensor layers_7_fc2_bias_to_fp16 = const()[name = tensor("layers_7_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231570688)))]; + tensor hidden_states_17_cast_fp16 = conv(bias = layers_7_fc2_bias_to_fp16, dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_7_fc2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("hidden_states_17_cast_fp16")]; + tensor inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = hidden_states_17_cast_fp16)[name = tensor("inputs_49_cast_fp16")]; + tensor var_1852 = const()[name = tensor("op_1852"), val = tensor(3)]; + tensor out_49_axes_0 = const()[name = tensor("out_49_axes_0"), val = tensor([1])]; + tensor var_1877_to_fp16 = const()[name = tensor("op_1877_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_1877_to_fp16, x = inputs_49_cast_fp16)[name = tensor("out_49_cast_fp16")]; + tensor obj_113_gamma_0_to_fp16 = const()[name = tensor("obj_113_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231572288)))]; + tensor obj_113_beta_0_to_fp16 = const()[name = tensor("obj_113_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231573888)))]; + tensor obj_113_epsilon_0_to_fp16 = const()[name = tensor("obj_113_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_113_cast_fp16 = batch_norm(beta = obj_113_beta_0_to_fp16, epsilon = obj_113_epsilon_0_to_fp16, gamma = obj_113_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_49_cast_fp16)[name = tensor("obj_113_cast_fp16")]; + tensor query_33_pad_type_0 = const()[name = tensor("query_33_pad_type_0"), val = tensor("valid")]; + tensor query_33_strides_0 = const()[name = tensor("query_33_strides_0"), val = tensor([1, 1])]; + tensor query_33_pad_0 = const()[name = tensor("query_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_33_dilations_0 = const()[name = tensor("query_33_dilations_0"), val = tensor([1, 1])]; + tensor query_33_groups_0 = const()[name = tensor("query_33_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231575488)))]; + tensor layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232755200)))]; + tensor query_33_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_bias_to_fp16, dilations = query_33_dilations_0, groups = query_33_groups_0, pad = query_33_pad_0, pad_type = query_33_pad_type_0, strides = query_33_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("query_33_cast_fp16")]; + tensor current_key_17_pad_type_0 = const()[name = tensor("current_key_17_pad_type_0"), val = tensor("valid")]; + tensor current_key_17_strides_0 = const()[name = tensor("current_key_17_strides_0"), val = tensor([1, 1])]; + tensor current_key_17_pad_0 = const()[name = tensor("current_key_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_17_dilations_0 = const()[name = tensor("current_key_17_dilations_0"), val = tensor([1, 1])]; + tensor current_key_17_groups_0 = const()[name = tensor("current_key_17_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232756800)))]; + tensor current_key_17_cast_fp16 = conv(dilations = current_key_17_dilations_0, groups = current_key_17_groups_0, pad = current_key_17_pad_0, pad_type = current_key_17_pad_type_0, strides = current_key_17_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("current_key_17_cast_fp16")]; + tensor current_value_17_pad_type_0 = const()[name = tensor("current_value_17_pad_type_0"), val = tensor("valid")]; + tensor current_value_17_strides_0 = const()[name = tensor("current_value_17_strides_0"), val = tensor([1, 1])]; + tensor current_value_17_pad_0 = const()[name = tensor("current_value_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_17_dilations_0 = const()[name = tensor("current_value_17_dilations_0"), val = tensor([1, 1])]; + tensor current_value_17_groups_0 = const()[name = tensor("current_value_17_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233936512)))]; + tensor layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235116224)))]; + tensor current_value_17_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_bias_to_fp16, dilations = current_value_17_dilations_0, groups = current_value_17_groups_0, pad = current_value_17_pad_0, pad_type = current_value_17_pad_type_0, strides = current_value_17_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("current_value_17_cast_fp16")]; + tensor var_1916_cast_fp16 = mul(x = var_63_cast_fp16_8, y = var_159_cast_fp16)[name = tensor("op_1916_cast_fp16")]; + tensor var_1917_cast_fp16 = mul(x = current_key_17_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_1917_cast_fp16")]; + tensor key_33_cast_fp16 = add(x = var_1916_cast_fp16, y = var_1917_cast_fp16)[name = tensor("key_33_cast_fp16")]; + tensor var_1920_cast_fp16 = mul(x = var_78_cast_fp16_8, y = var_159_cast_fp16)[name = tensor("op_1920_cast_fp16")]; + tensor var_1921_cast_fp16 = mul(x = current_value_17_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_1921_cast_fp16")]; + tensor value_33_cast_fp16 = add(x = var_1920_cast_fp16, y = var_1921_cast_fp16)[name = tensor("value_33_cast_fp16")]; + tensor var_1925 = const()[name = tensor("op_1925"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_33_cast_fp16 = reshape(shape = var_1925, x = query_33_cast_fp16)[name = tensor("mh_q_33_cast_fp16")]; + tensor var_1927_to_fp16 = const()[name = tensor("op_1927_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1928_cast_fp16 = mul(x = mh_q_33_cast_fp16, y = var_1927_to_fp16)[name = tensor("op_1928_cast_fp16")]; + tensor var_1931 = const()[name = tensor("op_1931"), val = tensor([1, 12, 64, 448])]; + tensor var_1932_cast_fp16 = reshape(shape = var_1931, x = key_33_cast_fp16)[name = tensor("op_1932_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_1928_cast_fp16, y = var_1932_cast_fp16)[name = tensor("mh_w_49_cast_fp16")]; + tensor mh_w_51_cast_fp16 = add(x = mh_w_49_cast_fp16, y = var_181_cast_fp16)[name = tensor("mh_w_51_cast_fp16")]; + tensor var_1940_cast_fp16 = softmax(axis = var_1852, x = mh_w_51_cast_fp16)[name = tensor("op_1940_cast_fp16")]; + tensor var_1941 = const()[name = tensor("op_1941"), val = tensor([1, 12, 64, 448])]; + tensor var_1942_cast_fp16 = reshape(shape = var_1941, x = value_33_cast_fp16)[name = tensor("op_1942_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_1942_cast_fp16, y = var_1940_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_1945 = const()[name = tensor("op_1945"), val = tensor([1, 768, 1, 1])]; + tensor input_81_cast_fp16 = reshape(shape = var_1945, x = attn_33_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor obj_119_pad_type_0 = const()[name = tensor("obj_119_pad_type_0"), val = tensor("valid")]; + tensor obj_119_strides_0 = const()[name = tensor("obj_119_strides_0"), val = tensor([1, 1])]; + tensor obj_119_pad_0 = const()[name = tensor("obj_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_119_dilations_0 = const()[name = tensor("obj_119_dilations_0"), val = tensor([1, 1])]; + tensor obj_119_groups_0 = const()[name = tensor("obj_119_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235117824)))]; + tensor layers_8_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236297536)))]; + tensor obj_119_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_bias_to_fp16, dilations = obj_119_dilations_0, groups = obj_119_groups_0, pad = obj_119_pad_0, pad_type = obj_119_pad_type_0, strides = obj_119_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("obj_119_cast_fp16")]; + tensor inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_119_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; + tensor out_51_axes_0 = const()[name = tensor("out_51_axes_0"), val = tensor([1])]; + tensor var_1967_to_fp16 = const()[name = tensor("op_1967_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_1967_to_fp16, x = inputs_51_cast_fp16)[name = tensor("out_51_cast_fp16")]; + tensor obj_121_gamma_0_to_fp16 = const()[name = tensor("obj_121_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236299136)))]; + 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(236300736)))]; + tensor obj_121_epsilon_0_to_fp16 = const()[name = tensor("obj_121_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_121_cast_fp16 = batch_norm(beta = obj_121_beta_0_to_fp16, epsilon = obj_121_epsilon_0_to_fp16, gamma = obj_121_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_51_cast_fp16)[name = tensor("obj_121_cast_fp16")]; + tensor query_35_pad_type_0 = const()[name = tensor("query_35_pad_type_0"), val = tensor("valid")]; + tensor query_35_strides_0 = const()[name = tensor("query_35_strides_0"), val = tensor([1, 1])]; + tensor query_35_pad_0 = const()[name = tensor("query_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_35_dilations_0 = const()[name = tensor("query_35_dilations_0"), val = tensor([1, 1])]; + tensor query_35_groups_0 = const()[name = tensor("query_35_groups_0"), val = tensor(1)]; + tensor layers_8_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236302336)))]; + tensor layers_8_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237482048)))]; + tensor query_35_cast_fp16 = conv(bias = layers_8_encoder_attn_q_proj_bias_to_fp16, dilations = query_35_dilations_0, groups = query_35_groups_0, pad = query_35_pad_0, pad_type = query_35_pad_type_0, strides = query_35_strides_0, weight = layers_8_encoder_attn_q_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("query_35_cast_fp16")]; + tensor key_35_pad_type_0 = const()[name = tensor("key_35_pad_type_0"), val = tensor("valid")]; + tensor key_35_strides_0 = const()[name = tensor("key_35_strides_0"), val = tensor([1, 1])]; + tensor key_35_pad_0 = const()[name = tensor("key_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_35_dilations_0 = const()[name = tensor("key_35_dilations_0"), val = tensor([1, 1])]; + tensor key_35_groups_0 = const()[name = tensor("key_35_groups_0"), val = tensor(1)]; + tensor layers_8_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237483648)))]; + tensor key_35_cast_fp16 = conv(dilations = key_35_dilations_0, groups = key_35_groups_0, pad = key_35_pad_0, pad_type = key_35_pad_type_0, strides = key_35_strides_0, weight = layers_8_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_35_cast_fp16")]; + tensor value_35_pad_type_0 = const()[name = tensor("value_35_pad_type_0"), val = tensor("valid")]; + tensor value_35_strides_0 = const()[name = tensor("value_35_strides_0"), val = tensor([1, 1])]; + tensor value_35_pad_0 = const()[name = tensor("value_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_35_dilations_0 = const()[name = tensor("value_35_dilations_0"), val = tensor([1, 1])]; + tensor value_35_groups_0 = const()[name = tensor("value_35_groups_0"), val = tensor(1)]; + tensor layers_8_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238663360)))]; + tensor layers_8_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239843072)))]; + tensor value_35_cast_fp16 = conv(bias = layers_8_encoder_attn_v_proj_bias_to_fp16, dilations = value_35_dilations_0, groups = value_35_groups_0, pad = value_35_pad_0, pad_type = value_35_pad_type_0, strides = value_35_strides_0, weight = layers_8_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_35_cast_fp16")]; + tensor var_2003 = const()[name = tensor("op_2003"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_35_cast_fp16 = reshape(shape = var_2003, x = query_35_cast_fp16)[name = tensor("mh_q_35_cast_fp16")]; + tensor var_2005_to_fp16 = const()[name = tensor("op_2005_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2006_cast_fp16 = mul(x = mh_q_35_cast_fp16, y = var_2005_to_fp16)[name = tensor("op_2006_cast_fp16")]; + tensor var_2009 = const()[name = tensor("op_2009"), val = tensor([1, 12, 64, 1500])]; + tensor var_2010_cast_fp16 = reshape(shape = var_2009, x = key_35_cast_fp16)[name = tensor("op_2010_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_2006_cast_fp16, y = var_2010_cast_fp16)[name = tensor("mh_w_53_cast_fp16")]; + tensor obj_125_cast_fp16 = softmax(axis = var_1852, x = mh_w_53_cast_fp16)[name = tensor("obj_125_cast_fp16")]; + tensor var_2014 = const()[name = tensor("op_2014"), val = tensor([1, 12, 64, 1500])]; + tensor var_2015_cast_fp16 = reshape(shape = var_2014, x = value_35_cast_fp16)[name = tensor("op_2015_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_2015_cast_fp16, y = obj_125_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_2018 = const()[name = tensor("op_2018"), val = tensor([1, 768, 1, 1])]; + tensor input_83_cast_fp16 = reshape(shape = var_2018, x = attn_35_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor obj_123_pad_type_0 = const()[name = tensor("obj_123_pad_type_0"), val = tensor("valid")]; + tensor obj_123_strides_0 = const()[name = tensor("obj_123_strides_0"), val = tensor([1, 1])]; + tensor obj_123_pad_0 = const()[name = tensor("obj_123_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_123_dilations_0 = const()[name = tensor("obj_123_dilations_0"), val = tensor([1, 1])]; + tensor obj_123_groups_0 = const()[name = tensor("obj_123_groups_0"), val = tensor(1)]; + tensor layers_8_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239844672)))]; + tensor layers_8_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(241024384)))]; + tensor obj_123_cast_fp16 = conv(bias = layers_8_encoder_attn_o_proj_bias_to_fp16, dilations = obj_123_dilations_0, groups = obj_123_groups_0, pad = obj_123_pad_0, pad_type = obj_123_pad_type_0, strides = obj_123_strides_0, weight = layers_8_encoder_attn_o_proj_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("obj_123_cast_fp16")]; + tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = obj_123_cast_fp16)[name = tensor("inputs_53_cast_fp16")]; + tensor out_53_axes_0 = const()[name = tensor("out_53_axes_0"), val = tensor([1])]; + tensor var_2039_to_fp16 = const()[name = tensor("op_2039_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_2039_to_fp16, x = inputs_53_cast_fp16)[name = tensor("out_53_cast_fp16")]; + tensor input_85_gamma_0_to_fp16 = const()[name = tensor("input_85_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(241025984)))]; + tensor input_85_beta_0_to_fp16 = const()[name = tensor("input_85_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(241027584)))]; + tensor input_85_epsilon_0_to_fp16 = const()[name = tensor("input_85_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_85_cast_fp16 = batch_norm(beta = input_85_beta_0_to_fp16, epsilon = input_85_epsilon_0_to_fp16, gamma = input_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_53_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor input_87_pad_type_0 = const()[name = tensor("input_87_pad_type_0"), val = tensor("valid")]; + tensor input_87_strides_0 = const()[name = tensor("input_87_strides_0"), val = tensor([1, 1])]; + tensor input_87_pad_0 = const()[name = tensor("input_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_87_dilations_0 = const()[name = tensor("input_87_dilations_0"), val = tensor([1, 1])]; + tensor input_87_groups_0 = const()[name = tensor("input_87_groups_0"), val = tensor(1)]; + tensor layers_8_fc1_weight_to_fp16 = const()[name = tensor("layers_8_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(241029184)))]; + tensor layers_8_fc1_bias_to_fp16 = const()[name = tensor("layers_8_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245747840)))]; + tensor input_87_cast_fp16 = conv(bias = layers_8_fc1_bias_to_fp16, dilations = input_87_dilations_0, groups = input_87_groups_0, pad = input_87_pad_0, pad_type = input_87_pad_type_0, strides = input_87_strides_0, weight = layers_8_fc1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_mode_0 = const()[name = tensor("input_89_mode_0"), val = tensor("EXACT")]; + tensor input_89_cast_fp16 = gelu(mode = input_89_mode_0, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor hidden_states_19_pad_type_0 = const()[name = tensor("hidden_states_19_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_19_strides_0 = const()[name = tensor("hidden_states_19_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_19_pad_0 = const()[name = tensor("hidden_states_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_19_dilations_0 = const()[name = tensor("hidden_states_19_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_19_groups_0 = const()[name = tensor("hidden_states_19_groups_0"), val = tensor(1)]; + tensor layers_8_fc2_weight_to_fp16 = const()[name = tensor("layers_8_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245754048)))]; + tensor layers_8_fc2_bias_to_fp16 = const()[name = tensor("layers_8_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250472704)))]; + tensor hidden_states_19_cast_fp16 = conv(bias = layers_8_fc2_bias_to_fp16, dilations = hidden_states_19_dilations_0, groups = hidden_states_19_groups_0, pad = hidden_states_19_pad_0, pad_type = hidden_states_19_pad_type_0, strides = hidden_states_19_strides_0, weight = layers_8_fc2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("hidden_states_19_cast_fp16")]; + tensor inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = hidden_states_19_cast_fp16)[name = tensor("inputs_55_cast_fp16")]; + tensor var_2075 = const()[name = tensor("op_2075"), val = tensor(3)]; + tensor out_55_axes_0 = const()[name = tensor("out_55_axes_0"), val = tensor([1])]; + tensor var_2100_to_fp16 = const()[name = tensor("op_2100_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_2100_to_fp16, x = inputs_55_cast_fp16)[name = tensor("out_55_cast_fp16")]; + tensor obj_127_gamma_0_to_fp16 = const()[name = tensor("obj_127_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250474304)))]; + tensor obj_127_beta_0_to_fp16 = const()[name = tensor("obj_127_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250475904)))]; + tensor obj_127_epsilon_0_to_fp16 = const()[name = tensor("obj_127_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_127_cast_fp16 = batch_norm(beta = obj_127_beta_0_to_fp16, epsilon = obj_127_epsilon_0_to_fp16, gamma = obj_127_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_55_cast_fp16)[name = tensor("obj_127_cast_fp16")]; + tensor query_37_pad_type_0 = const()[name = tensor("query_37_pad_type_0"), val = tensor("valid")]; + tensor query_37_strides_0 = const()[name = tensor("query_37_strides_0"), val = tensor([1, 1])]; + tensor query_37_pad_0 = const()[name = tensor("query_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_37_dilations_0 = const()[name = tensor("query_37_dilations_0"), val = tensor([1, 1])]; + tensor query_37_groups_0 = const()[name = tensor("query_37_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250477504)))]; + tensor layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251657216)))]; + tensor query_37_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_bias_to_fp16, dilations = query_37_dilations_0, groups = query_37_groups_0, pad = query_37_pad_0, pad_type = query_37_pad_type_0, strides = query_37_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("query_37_cast_fp16")]; + tensor current_key_19_pad_type_0 = const()[name = tensor("current_key_19_pad_type_0"), val = tensor("valid")]; + tensor current_key_19_strides_0 = const()[name = tensor("current_key_19_strides_0"), val = tensor([1, 1])]; + tensor current_key_19_pad_0 = const()[name = tensor("current_key_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_19_dilations_0 = const()[name = tensor("current_key_19_dilations_0"), val = tensor([1, 1])]; + tensor current_key_19_groups_0 = const()[name = tensor("current_key_19_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251658816)))]; + tensor current_key_19_cast_fp16 = conv(dilations = current_key_19_dilations_0, groups = current_key_19_groups_0, pad = current_key_19_pad_0, pad_type = current_key_19_pad_type_0, strides = current_key_19_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("current_key_19_cast_fp16")]; + tensor current_value_19_pad_type_0 = const()[name = tensor("current_value_19_pad_type_0"), val = tensor("valid")]; + tensor current_value_19_strides_0 = const()[name = tensor("current_value_19_strides_0"), val = tensor([1, 1])]; + tensor current_value_19_pad_0 = const()[name = tensor("current_value_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_19_dilations_0 = const()[name = tensor("current_value_19_dilations_0"), val = tensor([1, 1])]; + tensor current_value_19_groups_0 = const()[name = tensor("current_value_19_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252838528)))]; + tensor layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254018240)))]; + tensor current_value_19_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_bias_to_fp16, dilations = current_value_19_dilations_0, groups = current_value_19_groups_0, pad = current_value_19_pad_0, pad_type = current_value_19_pad_type_0, strides = current_value_19_strides_0, weight = layers_9_self_attn_v_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("current_value_19_cast_fp16")]; + tensor var_2139_cast_fp16 = mul(x = var_63_cast_fp16_9, y = var_159_cast_fp16)[name = tensor("op_2139_cast_fp16")]; + tensor var_2140_cast_fp16 = mul(x = current_key_19_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_2140_cast_fp16")]; + tensor key_37_cast_fp16 = add(x = var_2139_cast_fp16, y = var_2140_cast_fp16)[name = tensor("key_37_cast_fp16")]; + tensor var_2143_cast_fp16 = mul(x = var_78_cast_fp16_9, y = var_159_cast_fp16)[name = tensor("op_2143_cast_fp16")]; + tensor var_2144_cast_fp16 = mul(x = current_value_19_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_2144_cast_fp16")]; + tensor value_37_cast_fp16 = add(x = var_2143_cast_fp16, y = var_2144_cast_fp16)[name = tensor("value_37_cast_fp16")]; + tensor var_2148 = const()[name = tensor("op_2148"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_37_cast_fp16 = reshape(shape = var_2148, x = query_37_cast_fp16)[name = tensor("mh_q_37_cast_fp16")]; + tensor var_2150_to_fp16 = const()[name = tensor("op_2150_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2151_cast_fp16 = mul(x = mh_q_37_cast_fp16, y = var_2150_to_fp16)[name = tensor("op_2151_cast_fp16")]; + tensor var_2154 = const()[name = tensor("op_2154"), val = tensor([1, 12, 64, 448])]; + tensor var_2155_cast_fp16 = reshape(shape = var_2154, x = key_37_cast_fp16)[name = tensor("op_2155_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_2151_cast_fp16, y = var_2155_cast_fp16)[name = tensor("mh_w_55_cast_fp16")]; + tensor mh_w_57_cast_fp16 = add(x = mh_w_55_cast_fp16, y = var_181_cast_fp16)[name = tensor("mh_w_57_cast_fp16")]; + tensor var_2163_cast_fp16 = softmax(axis = var_2075, x = mh_w_57_cast_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor var_2164 = const()[name = tensor("op_2164"), val = tensor([1, 12, 64, 448])]; + tensor var_2165_cast_fp16 = reshape(shape = var_2164, x = value_37_cast_fp16)[name = tensor("op_2165_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_2165_cast_fp16, y = var_2163_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_2168 = const()[name = tensor("op_2168"), val = tensor([1, 768, 1, 1])]; + tensor input_91_cast_fp16 = reshape(shape = var_2168, x = attn_37_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor obj_133_pad_type_0 = const()[name = tensor("obj_133_pad_type_0"), val = tensor("valid")]; + tensor obj_133_strides_0 = const()[name = tensor("obj_133_strides_0"), val = tensor([1, 1])]; + tensor obj_133_pad_0 = const()[name = tensor("obj_133_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_133_dilations_0 = const()[name = tensor("obj_133_dilations_0"), val = tensor([1, 1])]; + tensor obj_133_groups_0 = const()[name = tensor("obj_133_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254019840)))]; + tensor layers_9_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255199552)))]; + tensor obj_133_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_bias_to_fp16, dilations = obj_133_dilations_0, groups = obj_133_groups_0, pad = obj_133_pad_0, pad_type = obj_133_pad_type_0, strides = obj_133_strides_0, weight = layers_9_self_attn_o_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("obj_133_cast_fp16")]; + tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = obj_133_cast_fp16)[name = tensor("inputs_57_cast_fp16")]; + tensor out_57_axes_0 = const()[name = tensor("out_57_axes_0"), val = tensor([1])]; + tensor var_2190_to_fp16 = const()[name = tensor("op_2190_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_57_cast_fp16 = layer_norm(axes = out_57_axes_0, epsilon = var_2190_to_fp16, x = inputs_57_cast_fp16)[name = tensor("out_57_cast_fp16")]; + tensor obj_135_gamma_0_to_fp16 = const()[name = tensor("obj_135_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255201152)))]; + tensor obj_135_beta_0_to_fp16 = const()[name = tensor("obj_135_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255202752)))]; + tensor obj_135_epsilon_0_to_fp16 = const()[name = tensor("obj_135_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_135_cast_fp16 = batch_norm(beta = obj_135_beta_0_to_fp16, epsilon = obj_135_epsilon_0_to_fp16, gamma = obj_135_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_57_cast_fp16)[name = tensor("obj_135_cast_fp16")]; + tensor query_39_pad_type_0 = const()[name = tensor("query_39_pad_type_0"), val = tensor("valid")]; + tensor query_39_strides_0 = const()[name = tensor("query_39_strides_0"), val = tensor([1, 1])]; + tensor query_39_pad_0 = const()[name = tensor("query_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_39_dilations_0 = const()[name = tensor("query_39_dilations_0"), val = tensor([1, 1])]; + tensor query_39_groups_0 = const()[name = tensor("query_39_groups_0"), val = tensor(1)]; + tensor layers_9_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255204352)))]; + tensor layers_9_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256384064)))]; + tensor query_39_cast_fp16 = conv(bias = layers_9_encoder_attn_q_proj_bias_to_fp16, dilations = query_39_dilations_0, groups = query_39_groups_0, pad = query_39_pad_0, pad_type = query_39_pad_type_0, strides = query_39_strides_0, weight = layers_9_encoder_attn_q_proj_weight_to_fp16, x = obj_135_cast_fp16)[name = tensor("query_39_cast_fp16")]; + tensor key_39_pad_type_0 = const()[name = tensor("key_39_pad_type_0"), val = tensor("valid")]; + tensor key_39_strides_0 = const()[name = tensor("key_39_strides_0"), val = tensor([1, 1])]; + tensor key_39_pad_0 = const()[name = tensor("key_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_39_dilations_0 = const()[name = tensor("key_39_dilations_0"), val = tensor([1, 1])]; + tensor key_39_groups_0 = const()[name = tensor("key_39_groups_0"), val = tensor(1)]; + tensor layers_9_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256385664)))]; + tensor key_39_cast_fp16 = conv(dilations = key_39_dilations_0, groups = key_39_groups_0, pad = key_39_pad_0, pad_type = key_39_pad_type_0, strides = key_39_strides_0, weight = layers_9_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_39_cast_fp16")]; + tensor value_39_pad_type_0 = const()[name = tensor("value_39_pad_type_0"), val = tensor("valid")]; + tensor value_39_strides_0 = const()[name = tensor("value_39_strides_0"), val = tensor([1, 1])]; + tensor value_39_pad_0 = const()[name = tensor("value_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_39_dilations_0 = const()[name = tensor("value_39_dilations_0"), val = tensor([1, 1])]; + tensor value_39_groups_0 = const()[name = tensor("value_39_groups_0"), val = tensor(1)]; + tensor layers_9_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257565376)))]; + tensor layers_9_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258745088)))]; + tensor value_39_cast_fp16 = conv(bias = layers_9_encoder_attn_v_proj_bias_to_fp16, dilations = value_39_dilations_0, groups = value_39_groups_0, pad = value_39_pad_0, pad_type = value_39_pad_type_0, strides = value_39_strides_0, weight = layers_9_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_39_cast_fp16")]; + tensor var_2226 = const()[name = tensor("op_2226"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_39_cast_fp16 = reshape(shape = var_2226, x = query_39_cast_fp16)[name = tensor("mh_q_39_cast_fp16")]; + tensor var_2228_to_fp16 = const()[name = tensor("op_2228_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2229_cast_fp16 = mul(x = mh_q_39_cast_fp16, y = var_2228_to_fp16)[name = tensor("op_2229_cast_fp16")]; + tensor var_2232 = const()[name = tensor("op_2232"), val = tensor([1, 12, 64, 1500])]; + tensor var_2233_cast_fp16 = reshape(shape = var_2232, x = key_39_cast_fp16)[name = tensor("op_2233_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_2229_cast_fp16, y = var_2233_cast_fp16)[name = tensor("mh_w_59_cast_fp16")]; + tensor obj_139_cast_fp16 = softmax(axis = var_2075, x = mh_w_59_cast_fp16)[name = tensor("obj_139_cast_fp16")]; + tensor var_2237 = const()[name = tensor("op_2237"), val = tensor([1, 12, 64, 1500])]; + tensor var_2238_cast_fp16 = reshape(shape = var_2237, x = value_39_cast_fp16)[name = tensor("op_2238_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_2238_cast_fp16, y = obj_139_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_2241 = const()[name = tensor("op_2241"), val = tensor([1, 768, 1, 1])]; + tensor input_93_cast_fp16 = reshape(shape = var_2241, x = attn_39_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor obj_137_pad_type_0 = const()[name = tensor("obj_137_pad_type_0"), val = tensor("valid")]; + tensor obj_137_strides_0 = const()[name = tensor("obj_137_strides_0"), val = tensor([1, 1])]; + tensor obj_137_pad_0 = const()[name = tensor("obj_137_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_137_dilations_0 = const()[name = tensor("obj_137_dilations_0"), val = tensor([1, 1])]; + tensor obj_137_groups_0 = const()[name = tensor("obj_137_groups_0"), val = tensor(1)]; + tensor layers_9_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258746688)))]; + tensor layers_9_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259926400)))]; + tensor obj_137_cast_fp16 = conv(bias = layers_9_encoder_attn_o_proj_bias_to_fp16, dilations = obj_137_dilations_0, groups = obj_137_groups_0, pad = obj_137_pad_0, pad_type = obj_137_pad_type_0, strides = obj_137_strides_0, weight = layers_9_encoder_attn_o_proj_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("obj_137_cast_fp16")]; + tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_137_cast_fp16)[name = tensor("inputs_59_cast_fp16")]; + tensor out_59_axes_0 = const()[name = tensor("out_59_axes_0"), val = tensor([1])]; + tensor var_2262_to_fp16 = const()[name = tensor("op_2262_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_59_cast_fp16 = layer_norm(axes = out_59_axes_0, epsilon = var_2262_to_fp16, x = inputs_59_cast_fp16)[name = tensor("out_59_cast_fp16")]; + tensor input_95_gamma_0_to_fp16 = const()[name = tensor("input_95_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259928000)))]; + tensor input_95_beta_0_to_fp16 = const()[name = tensor("input_95_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259929600)))]; + tensor input_95_epsilon_0_to_fp16 = const()[name = tensor("input_95_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_95_cast_fp16 = batch_norm(beta = input_95_beta_0_to_fp16, epsilon = input_95_epsilon_0_to_fp16, gamma = input_95_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_59_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor input_97_pad_type_0 = const()[name = tensor("input_97_pad_type_0"), val = tensor("valid")]; + tensor input_97_strides_0 = const()[name = tensor("input_97_strides_0"), val = tensor([1, 1])]; + tensor input_97_pad_0 = const()[name = tensor("input_97_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_97_dilations_0 = const()[name = tensor("input_97_dilations_0"), val = tensor([1, 1])]; + tensor input_97_groups_0 = const()[name = tensor("input_97_groups_0"), val = tensor(1)]; + tensor layers_9_fc1_weight_to_fp16 = const()[name = tensor("layers_9_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259931200)))]; + tensor layers_9_fc1_bias_to_fp16 = const()[name = tensor("layers_9_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264649856)))]; + tensor input_97_cast_fp16 = conv(bias = layers_9_fc1_bias_to_fp16, dilations = input_97_dilations_0, groups = input_97_groups_0, pad = input_97_pad_0, pad_type = input_97_pad_type_0, strides = input_97_strides_0, weight = layers_9_fc1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_mode_0 = const()[name = tensor("input_99_mode_0"), val = tensor("EXACT")]; + tensor input_99_cast_fp16 = gelu(mode = input_99_mode_0, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor hidden_states_21_pad_type_0 = const()[name = tensor("hidden_states_21_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_21_strides_0 = const()[name = tensor("hidden_states_21_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_21_pad_0 = const()[name = tensor("hidden_states_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_21_dilations_0 = const()[name = tensor("hidden_states_21_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_21_groups_0 = const()[name = tensor("hidden_states_21_groups_0"), val = tensor(1)]; + tensor layers_9_fc2_weight_to_fp16 = const()[name = tensor("layers_9_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264656064)))]; + tensor layers_9_fc2_bias_to_fp16 = const()[name = tensor("layers_9_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269374720)))]; + tensor hidden_states_21_cast_fp16 = conv(bias = layers_9_fc2_bias_to_fp16, dilations = hidden_states_21_dilations_0, groups = hidden_states_21_groups_0, pad = hidden_states_21_pad_0, pad_type = hidden_states_21_pad_type_0, strides = hidden_states_21_strides_0, weight = layers_9_fc2_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("hidden_states_21_cast_fp16")]; + tensor inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = hidden_states_21_cast_fp16)[name = tensor("inputs_61_cast_fp16")]; + tensor var_2298 = const()[name = tensor("op_2298"), val = tensor(3)]; + tensor out_61_axes_0 = const()[name = tensor("out_61_axes_0"), val = tensor([1])]; + tensor var_2323_to_fp16 = const()[name = tensor("op_2323_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_61_cast_fp16 = layer_norm(axes = out_61_axes_0, epsilon = var_2323_to_fp16, x = inputs_61_cast_fp16)[name = tensor("out_61_cast_fp16")]; + tensor obj_141_gamma_0_to_fp16 = const()[name = tensor("obj_141_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269376320)))]; + 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(269377920)))]; + tensor obj_141_epsilon_0_to_fp16 = const()[name = tensor("obj_141_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_141_cast_fp16 = batch_norm(beta = obj_141_beta_0_to_fp16, epsilon = obj_141_epsilon_0_to_fp16, gamma = obj_141_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_61_cast_fp16)[name = tensor("obj_141_cast_fp16")]; + tensor query_41_pad_type_0 = const()[name = tensor("query_41_pad_type_0"), val = tensor("valid")]; + tensor query_41_strides_0 = const()[name = tensor("query_41_strides_0"), val = tensor([1, 1])]; + tensor query_41_pad_0 = const()[name = tensor("query_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_41_dilations_0 = const()[name = tensor("query_41_dilations_0"), val = tensor([1, 1])]; + tensor query_41_groups_0 = const()[name = tensor("query_41_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269379520)))]; + tensor layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270559232)))]; + tensor query_41_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_bias_to_fp16, dilations = query_41_dilations_0, groups = query_41_groups_0, pad = query_41_pad_0, pad_type = query_41_pad_type_0, strides = query_41_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("query_41_cast_fp16")]; + tensor current_key_21_pad_type_0 = const()[name = tensor("current_key_21_pad_type_0"), val = tensor("valid")]; + tensor current_key_21_strides_0 = const()[name = tensor("current_key_21_strides_0"), val = tensor([1, 1])]; + tensor current_key_21_pad_0 = const()[name = tensor("current_key_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_21_dilations_0 = const()[name = tensor("current_key_21_dilations_0"), val = tensor([1, 1])]; + tensor current_key_21_groups_0 = const()[name = tensor("current_key_21_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270560832)))]; + tensor current_key_21_cast_fp16 = conv(dilations = current_key_21_dilations_0, groups = current_key_21_groups_0, pad = current_key_21_pad_0, pad_type = current_key_21_pad_type_0, strides = current_key_21_strides_0, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("current_key_21_cast_fp16")]; + tensor current_value_21_pad_type_0 = const()[name = tensor("current_value_21_pad_type_0"), val = tensor("valid")]; + tensor current_value_21_strides_0 = const()[name = tensor("current_value_21_strides_0"), val = tensor([1, 1])]; + tensor current_value_21_pad_0 = const()[name = tensor("current_value_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_21_dilations_0 = const()[name = tensor("current_value_21_dilations_0"), val = tensor([1, 1])]; + tensor current_value_21_groups_0 = const()[name = tensor("current_value_21_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(271740544)))]; + tensor layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272920256)))]; + tensor current_value_21_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_bias_to_fp16, dilations = current_value_21_dilations_0, groups = current_value_21_groups_0, pad = current_value_21_pad_0, pad_type = current_value_21_pad_type_0, strides = current_value_21_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("current_value_21_cast_fp16")]; + tensor var_2362_cast_fp16 = mul(x = var_63_cast_fp16_10, y = var_159_cast_fp16)[name = tensor("op_2362_cast_fp16")]; + tensor var_2363_cast_fp16 = mul(x = current_key_21_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_2363_cast_fp16")]; + tensor key_41_cast_fp16 = add(x = var_2362_cast_fp16, y = var_2363_cast_fp16)[name = tensor("key_41_cast_fp16")]; + tensor var_2366_cast_fp16 = mul(x = var_78_cast_fp16_10, y = var_159_cast_fp16)[name = tensor("op_2366_cast_fp16")]; + tensor var_2367_cast_fp16 = mul(x = current_value_21_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_2367_cast_fp16")]; + tensor value_41_cast_fp16 = add(x = var_2366_cast_fp16, y = var_2367_cast_fp16)[name = tensor("value_41_cast_fp16")]; + tensor var_2371 = const()[name = tensor("op_2371"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_41_cast_fp16 = reshape(shape = var_2371, x = query_41_cast_fp16)[name = tensor("mh_q_41_cast_fp16")]; + tensor var_2373_to_fp16 = const()[name = tensor("op_2373_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2374_cast_fp16 = mul(x = mh_q_41_cast_fp16, y = var_2373_to_fp16)[name = tensor("op_2374_cast_fp16")]; + tensor var_2377 = const()[name = tensor("op_2377"), val = tensor([1, 12, 64, 448])]; + tensor var_2378_cast_fp16 = reshape(shape = var_2377, x = key_41_cast_fp16)[name = tensor("op_2378_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_2374_cast_fp16, y = var_2378_cast_fp16)[name = tensor("mh_w_61_cast_fp16")]; + tensor mh_w_63_cast_fp16 = add(x = mh_w_61_cast_fp16, y = var_181_cast_fp16)[name = tensor("mh_w_63_cast_fp16")]; + tensor var_2386_cast_fp16 = softmax(axis = var_2298, x = mh_w_63_cast_fp16)[name = tensor("op_2386_cast_fp16")]; + tensor var_2387 = const()[name = tensor("op_2387"), val = tensor([1, 12, 64, 448])]; + tensor var_2388_cast_fp16 = reshape(shape = var_2387, x = value_41_cast_fp16)[name = tensor("op_2388_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_2388_cast_fp16, y = var_2386_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_2391 = const()[name = tensor("op_2391"), val = tensor([1, 768, 1, 1])]; + tensor input_101_cast_fp16 = reshape(shape = var_2391, x = attn_41_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor obj_147_pad_type_0 = const()[name = tensor("obj_147_pad_type_0"), val = tensor("valid")]; + tensor obj_147_strides_0 = const()[name = tensor("obj_147_strides_0"), val = tensor([1, 1])]; + tensor obj_147_pad_0 = const()[name = tensor("obj_147_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_147_dilations_0 = const()[name = tensor("obj_147_dilations_0"), val = tensor([1, 1])]; + tensor obj_147_groups_0 = const()[name = tensor("obj_147_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272921856)))]; + tensor layers_10_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(274101568)))]; + tensor obj_147_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_bias_to_fp16, dilations = obj_147_dilations_0, groups = obj_147_groups_0, pad = obj_147_pad_0, pad_type = obj_147_pad_type_0, strides = obj_147_strides_0, weight = layers_10_self_attn_o_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("obj_147_cast_fp16")]; + tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_147_cast_fp16)[name = tensor("inputs_63_cast_fp16")]; + tensor out_63_axes_0 = const()[name = tensor("out_63_axes_0"), val = tensor([1])]; + tensor var_2413_to_fp16 = const()[name = tensor("op_2413_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_63_cast_fp16 = layer_norm(axes = out_63_axes_0, epsilon = var_2413_to_fp16, x = inputs_63_cast_fp16)[name = tensor("out_63_cast_fp16")]; + tensor obj_149_gamma_0_to_fp16 = const()[name = tensor("obj_149_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(274103168)))]; + tensor obj_149_beta_0_to_fp16 = const()[name = tensor("obj_149_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(274104768)))]; + tensor obj_149_epsilon_0_to_fp16 = const()[name = tensor("obj_149_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_149_cast_fp16 = batch_norm(beta = obj_149_beta_0_to_fp16, epsilon = obj_149_epsilon_0_to_fp16, gamma = obj_149_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_63_cast_fp16)[name = tensor("obj_149_cast_fp16")]; + tensor query_43_pad_type_0 = const()[name = tensor("query_43_pad_type_0"), val = tensor("valid")]; + tensor query_43_strides_0 = const()[name = tensor("query_43_strides_0"), val = tensor([1, 1])]; + tensor query_43_pad_0 = const()[name = tensor("query_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_43_dilations_0 = const()[name = tensor("query_43_dilations_0"), val = tensor([1, 1])]; + tensor query_43_groups_0 = const()[name = tensor("query_43_groups_0"), val = tensor(1)]; + tensor layers_10_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(274106368)))]; + tensor layers_10_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(275286080)))]; + tensor query_43_cast_fp16 = conv(bias = layers_10_encoder_attn_q_proj_bias_to_fp16, dilations = query_43_dilations_0, groups = query_43_groups_0, pad = query_43_pad_0, pad_type = query_43_pad_type_0, strides = query_43_strides_0, weight = layers_10_encoder_attn_q_proj_weight_to_fp16, x = obj_149_cast_fp16)[name = tensor("query_43_cast_fp16")]; + tensor key_43_pad_type_0 = const()[name = tensor("key_43_pad_type_0"), val = tensor("valid")]; + tensor key_43_strides_0 = const()[name = tensor("key_43_strides_0"), val = tensor([1, 1])]; + tensor key_43_pad_0 = const()[name = tensor("key_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_43_dilations_0 = const()[name = tensor("key_43_dilations_0"), val = tensor([1, 1])]; + tensor key_43_groups_0 = const()[name = tensor("key_43_groups_0"), val = tensor(1)]; + tensor layers_10_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(275287680)))]; + tensor key_43_cast_fp16 = conv(dilations = key_43_dilations_0, groups = key_43_groups_0, pad = key_43_pad_0, pad_type = key_43_pad_type_0, strides = key_43_strides_0, weight = layers_10_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_43_cast_fp16")]; + tensor value_43_pad_type_0 = const()[name = tensor("value_43_pad_type_0"), val = tensor("valid")]; + tensor value_43_strides_0 = const()[name = tensor("value_43_strides_0"), val = tensor([1, 1])]; + tensor value_43_pad_0 = const()[name = tensor("value_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_43_dilations_0 = const()[name = tensor("value_43_dilations_0"), val = tensor([1, 1])]; + tensor value_43_groups_0 = const()[name = tensor("value_43_groups_0"), val = tensor(1)]; + tensor layers_10_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276467392)))]; + tensor layers_10_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277647104)))]; + tensor value_43_cast_fp16 = conv(bias = layers_10_encoder_attn_v_proj_bias_to_fp16, dilations = value_43_dilations_0, groups = value_43_groups_0, pad = value_43_pad_0, pad_type = value_43_pad_type_0, strides = value_43_strides_0, weight = layers_10_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_43_cast_fp16")]; + tensor var_2449 = const()[name = tensor("op_2449"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_43_cast_fp16 = reshape(shape = var_2449, x = query_43_cast_fp16)[name = tensor("mh_q_43_cast_fp16")]; + tensor var_2451_to_fp16 = const()[name = tensor("op_2451_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2452_cast_fp16 = mul(x = mh_q_43_cast_fp16, y = var_2451_to_fp16)[name = tensor("op_2452_cast_fp16")]; + tensor var_2455 = const()[name = tensor("op_2455"), val = tensor([1, 12, 64, 1500])]; + tensor var_2456_cast_fp16 = reshape(shape = var_2455, x = key_43_cast_fp16)[name = tensor("op_2456_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_2452_cast_fp16, y = var_2456_cast_fp16)[name = tensor("mh_w_65_cast_fp16")]; + tensor obj_153_cast_fp16 = softmax(axis = var_2298, x = mh_w_65_cast_fp16)[name = tensor("obj_153_cast_fp16")]; + tensor var_2460 = const()[name = tensor("op_2460"), val = tensor([1, 12, 64, 1500])]; + tensor var_2461_cast_fp16 = reshape(shape = var_2460, x = value_43_cast_fp16)[name = tensor("op_2461_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_2461_cast_fp16, y = obj_153_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_2464 = const()[name = tensor("op_2464"), val = tensor([1, 768, 1, 1])]; + tensor input_103_cast_fp16 = reshape(shape = var_2464, x = attn_43_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor obj_151_pad_type_0 = const()[name = tensor("obj_151_pad_type_0"), val = tensor("valid")]; + tensor obj_151_strides_0 = const()[name = tensor("obj_151_strides_0"), val = tensor([1, 1])]; + tensor obj_151_pad_0 = const()[name = tensor("obj_151_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_151_dilations_0 = const()[name = tensor("obj_151_dilations_0"), val = tensor([1, 1])]; + tensor obj_151_groups_0 = const()[name = tensor("obj_151_groups_0"), val = tensor(1)]; + tensor layers_10_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277648704)))]; + tensor layers_10_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278828416)))]; + tensor obj_151_cast_fp16 = conv(bias = layers_10_encoder_attn_o_proj_bias_to_fp16, dilations = obj_151_dilations_0, groups = obj_151_groups_0, pad = obj_151_pad_0, pad_type = obj_151_pad_type_0, strides = obj_151_strides_0, weight = layers_10_encoder_attn_o_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("obj_151_cast_fp16")]; + tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = obj_151_cast_fp16)[name = tensor("inputs_65_cast_fp16")]; + tensor out_65_axes_0 = const()[name = tensor("out_65_axes_0"), val = tensor([1])]; + tensor var_2485_to_fp16 = const()[name = tensor("op_2485_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_65_cast_fp16 = layer_norm(axes = out_65_axes_0, epsilon = var_2485_to_fp16, x = inputs_65_cast_fp16)[name = tensor("out_65_cast_fp16")]; + tensor input_105_gamma_0_to_fp16 = const()[name = tensor("input_105_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278830016)))]; + tensor input_105_beta_0_to_fp16 = const()[name = tensor("input_105_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278831616)))]; + tensor input_105_epsilon_0_to_fp16 = const()[name = tensor("input_105_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_105_cast_fp16 = batch_norm(beta = input_105_beta_0_to_fp16, epsilon = input_105_epsilon_0_to_fp16, gamma = input_105_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_65_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor input_107_pad_type_0 = const()[name = tensor("input_107_pad_type_0"), val = tensor("valid")]; + tensor input_107_strides_0 = const()[name = tensor("input_107_strides_0"), val = tensor([1, 1])]; + tensor input_107_pad_0 = const()[name = tensor("input_107_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_107_dilations_0 = const()[name = tensor("input_107_dilations_0"), val = tensor([1, 1])]; + tensor input_107_groups_0 = const()[name = tensor("input_107_groups_0"), val = tensor(1)]; + tensor layers_10_fc1_weight_to_fp16 = const()[name = tensor("layers_10_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278833216)))]; + tensor layers_10_fc1_bias_to_fp16 = const()[name = tensor("layers_10_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283551872)))]; + tensor input_107_cast_fp16 = conv(bias = layers_10_fc1_bias_to_fp16, dilations = input_107_dilations_0, groups = input_107_groups_0, pad = input_107_pad_0, pad_type = input_107_pad_type_0, strides = input_107_strides_0, weight = layers_10_fc1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_mode_0 = const()[name = tensor("input_109_mode_0"), val = tensor("EXACT")]; + tensor input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor hidden_states_23_pad_type_0 = const()[name = tensor("hidden_states_23_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_23_strides_0 = const()[name = tensor("hidden_states_23_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_23_pad_0 = const()[name = tensor("hidden_states_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_23_dilations_0 = const()[name = tensor("hidden_states_23_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_23_groups_0 = const()[name = tensor("hidden_states_23_groups_0"), val = tensor(1)]; + tensor layers_10_fc2_weight_to_fp16 = const()[name = tensor("layers_10_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283558080)))]; + tensor layers_10_fc2_bias_to_fp16 = const()[name = tensor("layers_10_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288276736)))]; + tensor hidden_states_23_cast_fp16 = conv(bias = layers_10_fc2_bias_to_fp16, dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_10_fc2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("hidden_states_23_cast_fp16")]; + tensor inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = hidden_states_23_cast_fp16)[name = tensor("inputs_67_cast_fp16")]; + tensor var_2521 = const()[name = tensor("op_2521"), val = tensor(3)]; + tensor out_67_axes_0 = const()[name = tensor("out_67_axes_0"), val = tensor([1])]; + tensor var_2546_to_fp16 = const()[name = tensor("op_2546_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_67_cast_fp16 = layer_norm(axes = out_67_axes_0, epsilon = var_2546_to_fp16, x = inputs_67_cast_fp16)[name = tensor("out_67_cast_fp16")]; + tensor obj_155_gamma_0_to_fp16 = const()[name = tensor("obj_155_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288278336)))]; + tensor obj_155_beta_0_to_fp16 = const()[name = tensor("obj_155_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288279936)))]; + tensor obj_155_epsilon_0_to_fp16 = const()[name = tensor("obj_155_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_155_cast_fp16 = batch_norm(beta = obj_155_beta_0_to_fp16, epsilon = obj_155_epsilon_0_to_fp16, gamma = obj_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_67_cast_fp16)[name = tensor("obj_155_cast_fp16")]; + tensor query_45_pad_type_0 = const()[name = tensor("query_45_pad_type_0"), val = tensor("valid")]; + tensor query_45_strides_0 = const()[name = tensor("query_45_strides_0"), val = tensor([1, 1])]; + tensor query_45_pad_0 = const()[name = tensor("query_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_45_dilations_0 = const()[name = tensor("query_45_dilations_0"), val = tensor([1, 1])]; + tensor query_45_groups_0 = const()[name = tensor("query_45_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288281536)))]; + tensor layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289461248)))]; + tensor query_45_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_bias_to_fp16, dilations = query_45_dilations_0, groups = query_45_groups_0, pad = query_45_pad_0, pad_type = query_45_pad_type_0, strides = query_45_strides_0, weight = layers_11_self_attn_q_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("query_45_cast_fp16")]; + tensor current_key_pad_type_0 = const()[name = tensor("current_key_pad_type_0"), val = tensor("valid")]; + tensor current_key_strides_0 = const()[name = tensor("current_key_strides_0"), val = tensor([1, 1])]; + tensor current_key_pad_0 = const()[name = tensor("current_key_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_dilations_0 = const()[name = tensor("current_key_dilations_0"), val = tensor([1, 1])]; + tensor current_key_groups_0 = const()[name = tensor("current_key_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289462848)))]; + tensor current_key_cast_fp16 = conv(dilations = current_key_dilations_0, groups = current_key_groups_0, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = current_key_strides_0, weight = layers_11_self_attn_k_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("current_key_cast_fp16")]; + tensor current_value_pad_type_0 = const()[name = tensor("current_value_pad_type_0"), val = tensor("valid")]; + tensor current_value_strides_0 = const()[name = tensor("current_value_strides_0"), val = tensor([1, 1])]; + tensor current_value_pad_0 = const()[name = tensor("current_value_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_dilations_0 = const()[name = tensor("current_value_dilations_0"), val = tensor([1, 1])]; + tensor current_value_groups_0 = const()[name = tensor("current_value_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290642560)))]; + tensor layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291822272)))]; + tensor current_value_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_bias_to_fp16, dilations = current_value_dilations_0, groups = current_value_groups_0, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = current_value_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("current_value_cast_fp16")]; + tensor var_2585_cast_fp16 = mul(x = var_63_cast_fp16_11, y = var_159_cast_fp16)[name = tensor("op_2585_cast_fp16")]; + tensor var_2586_cast_fp16 = mul(x = current_key_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_2586_cast_fp16")]; + tensor key_45_cast_fp16 = add(x = var_2585_cast_fp16, y = var_2586_cast_fp16)[name = tensor("key_45_cast_fp16")]; + tensor var_2589_cast_fp16 = mul(x = var_78_cast_fp16_11, y = var_159_cast_fp16)[name = tensor("op_2589_cast_fp16")]; + tensor var_2590_cast_fp16 = mul(x = current_value_cast_fp16, y = var_157_cast_fp16)[name = tensor("op_2590_cast_fp16")]; + tensor value_45_cast_fp16 = add(x = var_2589_cast_fp16, y = var_2590_cast_fp16)[name = tensor("value_45_cast_fp16")]; + tensor var_2594 = const()[name = tensor("op_2594"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_45_cast_fp16 = reshape(shape = var_2594, x = query_45_cast_fp16)[name = tensor("mh_q_45_cast_fp16")]; + tensor var_2596_to_fp16 = const()[name = tensor("op_2596_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2597_cast_fp16 = mul(x = mh_q_45_cast_fp16, y = var_2596_to_fp16)[name = tensor("op_2597_cast_fp16")]; + tensor var_2600 = const()[name = tensor("op_2600"), val = tensor([1, 12, 64, 448])]; + tensor var_2601_cast_fp16 = reshape(shape = var_2600, x = key_45_cast_fp16)[name = tensor("op_2601_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_2597_cast_fp16, y = var_2601_cast_fp16)[name = tensor("mh_w_67_cast_fp16")]; + tensor mh_w_69_cast_fp16 = add(x = mh_w_67_cast_fp16, y = var_181_cast_fp16)[name = tensor("mh_w_69_cast_fp16")]; + tensor var_2609_cast_fp16 = softmax(axis = var_2521, x = mh_w_69_cast_fp16)[name = tensor("op_2609_cast_fp16")]; + tensor var_2610 = const()[name = tensor("op_2610"), val = tensor([1, 12, 64, 448])]; + tensor var_2611_cast_fp16 = reshape(shape = var_2610, x = value_45_cast_fp16)[name = tensor("op_2611_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_2611_cast_fp16, y = var_2609_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_2614 = const()[name = tensor("op_2614"), val = tensor([1, 768, 1, 1])]; + tensor input_111_cast_fp16 = reshape(shape = var_2614, x = attn_45_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor obj_161_pad_type_0 = const()[name = tensor("obj_161_pad_type_0"), val = tensor("valid")]; + tensor obj_161_strides_0 = const()[name = tensor("obj_161_strides_0"), val = tensor([1, 1])]; + tensor obj_161_pad_0 = const()[name = tensor("obj_161_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_161_dilations_0 = const()[name = tensor("obj_161_dilations_0"), val = tensor([1, 1])]; + tensor obj_161_groups_0 = const()[name = tensor("obj_161_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291823872)))]; + tensor layers_11_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293003584)))]; + tensor obj_161_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_bias_to_fp16, dilations = obj_161_dilations_0, groups = obj_161_groups_0, pad = obj_161_pad_0, pad_type = obj_161_pad_type_0, strides = obj_161_strides_0, weight = layers_11_self_attn_o_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("obj_161_cast_fp16")]; + tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = obj_161_cast_fp16)[name = tensor("inputs_69_cast_fp16")]; + tensor out_69_axes_0 = const()[name = tensor("out_69_axes_0"), val = tensor([1])]; + tensor var_2636_to_fp16 = const()[name = tensor("op_2636_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_69_cast_fp16 = layer_norm(axes = out_69_axes_0, epsilon = var_2636_to_fp16, x = inputs_69_cast_fp16)[name = tensor("out_69_cast_fp16")]; + tensor obj_163_gamma_0_to_fp16 = const()[name = tensor("obj_163_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293005184)))]; + tensor obj_163_beta_0_to_fp16 = const()[name = tensor("obj_163_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293006784)))]; + tensor obj_163_epsilon_0_to_fp16 = const()[name = tensor("obj_163_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_163_cast_fp16 = batch_norm(beta = obj_163_beta_0_to_fp16, epsilon = obj_163_epsilon_0_to_fp16, gamma = obj_163_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_69_cast_fp16)[name = tensor("obj_163_cast_fp16")]; + tensor query_pad_type_0 = const()[name = tensor("query_pad_type_0"), val = tensor("valid")]; + tensor query_strides_0 = const()[name = tensor("query_strides_0"), val = tensor([1, 1])]; + tensor query_pad_0 = const()[name = tensor("query_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_dilations_0 = const()[name = tensor("query_dilations_0"), val = tensor([1, 1])]; + tensor query_groups_0 = const()[name = tensor("query_groups_0"), val = tensor(1)]; + tensor layers_11_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293008384)))]; + tensor layers_11_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294188096)))]; + tensor query_cast_fp16 = conv(bias = layers_11_encoder_attn_q_proj_bias_to_fp16, dilations = query_dilations_0, groups = query_groups_0, pad = query_pad_0, pad_type = query_pad_type_0, strides = query_strides_0, weight = layers_11_encoder_attn_q_proj_weight_to_fp16, x = obj_163_cast_fp16)[name = tensor("query_cast_fp16")]; + tensor key_pad_type_0 = const()[name = tensor("key_pad_type_0"), val = tensor("valid")]; + tensor key_strides_0 = const()[name = tensor("key_strides_0"), val = tensor([1, 1])]; + tensor key_pad_0 = const()[name = tensor("key_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_dilations_0 = const()[name = tensor("key_dilations_0"), val = tensor([1, 1])]; + tensor key_groups_0 = const()[name = tensor("key_groups_0"), val = tensor(1)]; + tensor layers_11_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294189696)))]; + tensor key_cast_fp16 = conv(dilations = key_dilations_0, groups = key_groups_0, pad = key_pad_0, pad_type = key_pad_type_0, strides = key_strides_0, weight = layers_11_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_cast_fp16")]; + tensor value_pad_type_0 = const()[name = tensor("value_pad_type_0"), val = tensor("valid")]; + tensor value_strides_0 = const()[name = tensor("value_strides_0"), val = tensor([1, 1])]; + tensor value_pad_0 = const()[name = tensor("value_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_dilations_0 = const()[name = tensor("value_dilations_0"), val = tensor([1, 1])]; + tensor value_groups_0 = const()[name = tensor("value_groups_0"), val = tensor(1)]; + tensor layers_11_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295369408)))]; + tensor layers_11_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296549120)))]; + tensor value_cast_fp16 = conv(bias = layers_11_encoder_attn_v_proj_bias_to_fp16, dilations = value_dilations_0, groups = value_groups_0, pad = value_pad_0, pad_type = value_pad_type_0, strides = value_strides_0, weight = layers_11_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_cast_fp16")]; + tensor var_2672 = const()[name = tensor("op_2672"), val = tensor([1, 12, 64, 1])]; + tensor mh_q_cast_fp16 = reshape(shape = var_2672, x = query_cast_fp16)[name = tensor("mh_q_cast_fp16")]; + tensor var_2674_to_fp16 = const()[name = tensor("op_2674_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2675_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_2674_to_fp16)[name = tensor("op_2675_cast_fp16")]; + tensor var_2678 = const()[name = tensor("op_2678"), val = tensor([1, 12, 64, 1500])]; + tensor var_2679_cast_fp16 = reshape(shape = var_2678, x = key_cast_fp16)[name = tensor("op_2679_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_2675_cast_fp16, y = var_2679_cast_fp16)[name = tensor("mh_w_cast_fp16")]; + tensor obj_167_cast_fp16 = softmax(axis = var_2521, x = mh_w_cast_fp16)[name = tensor("obj_167_cast_fp16")]; + tensor var_2683 = const()[name = tensor("op_2683"), val = tensor([1, 12, 64, 1500])]; + tensor var_2684_cast_fp16 = reshape(shape = var_2683, x = value_cast_fp16)[name = tensor("op_2684_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_2684_cast_fp16, y = obj_167_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_2687 = const()[name = tensor("op_2687"), val = tensor([1, 768, 1, 1])]; + tensor input_113_cast_fp16 = reshape(shape = var_2687, x = attn_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor obj_165_pad_type_0 = const()[name = tensor("obj_165_pad_type_0"), val = tensor("valid")]; + tensor obj_165_strides_0 = const()[name = tensor("obj_165_strides_0"), val = tensor([1, 1])]; + tensor obj_165_pad_0 = const()[name = tensor("obj_165_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_165_dilations_0 = const()[name = tensor("obj_165_dilations_0"), val = tensor([1, 1])]; + tensor obj_165_groups_0 = const()[name = tensor("obj_165_groups_0"), val = tensor(1)]; + tensor layers_11_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296550720)))]; + tensor layers_11_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297730432)))]; + tensor obj_165_cast_fp16 = conv(bias = layers_11_encoder_attn_o_proj_bias_to_fp16, dilations = obj_165_dilations_0, groups = obj_165_groups_0, pad = obj_165_pad_0, pad_type = obj_165_pad_type_0, strides = obj_165_strides_0, weight = layers_11_encoder_attn_o_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("obj_165_cast_fp16")]; + tensor inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_165_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; + tensor out_71_axes_0 = const()[name = tensor("out_71_axes_0"), val = tensor([1])]; + tensor var_2705_to_fp16 = const()[name = tensor("op_2705_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_71_cast_fp16 = layer_norm(axes = out_71_axes_0, epsilon = var_2705_to_fp16, x = inputs_71_cast_fp16)[name = tensor("out_71_cast_fp16")]; + tensor input_115_gamma_0_to_fp16 = const()[name = tensor("input_115_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297732032)))]; + tensor input_115_beta_0_to_fp16 = const()[name = tensor("input_115_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297733632)))]; + tensor input_115_epsilon_0_to_fp16 = const()[name = tensor("input_115_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_115_cast_fp16 = batch_norm(beta = input_115_beta_0_to_fp16, epsilon = input_115_epsilon_0_to_fp16, gamma = input_115_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_71_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor input_117_pad_type_0 = const()[name = tensor("input_117_pad_type_0"), val = tensor("valid")]; + tensor input_117_strides_0 = const()[name = tensor("input_117_strides_0"), val = tensor([1, 1])]; + tensor input_117_pad_0 = const()[name = tensor("input_117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_117_dilations_0 = const()[name = tensor("input_117_dilations_0"), val = tensor([1, 1])]; + tensor input_117_groups_0 = const()[name = tensor("input_117_groups_0"), val = tensor(1)]; + tensor layers_11_fc1_weight_to_fp16 = const()[name = tensor("layers_11_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297735232)))]; + tensor layers_11_fc1_bias_to_fp16 = const()[name = tensor("layers_11_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302453888)))]; + tensor input_117_cast_fp16 = conv(bias = layers_11_fc1_bias_to_fp16, dilations = input_117_dilations_0, groups = input_117_groups_0, pad = input_117_pad_0, pad_type = input_117_pad_type_0, strides = input_117_strides_0, weight = layers_11_fc1_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_117_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor hidden_states_25_pad_type_0 = const()[name = tensor("hidden_states_25_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_25_strides_0 = const()[name = tensor("hidden_states_25_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_25_pad_0 = const()[name = tensor("hidden_states_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_25_dilations_0 = const()[name = tensor("hidden_states_25_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_25_groups_0 = const()[name = tensor("hidden_states_25_groups_0"), val = tensor(1)]; + tensor layers_11_fc2_weight_to_fp16 = const()[name = tensor("layers_11_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302460096)))]; + tensor layers_11_fc2_bias_to_fp16 = const()[name = tensor("layers_11_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307178752)))]; + tensor hidden_states_25_cast_fp16 = conv(bias = layers_11_fc2_bias_to_fp16, dilations = hidden_states_25_dilations_0, groups = hidden_states_25_groups_0, pad = hidden_states_25_pad_0, pad_type = hidden_states_25_pad_type_0, strides = hidden_states_25_strides_0, weight = layers_11_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor("hidden_states_25_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_71_cast_fp16, y = hidden_states_25_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor out_axes_0 = const()[name = tensor("out_axes_0"), val = tensor([1])]; + tensor var_2747_to_fp16 = const()[name = tensor("op_2747_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_2747_to_fp16, x = inputs_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(307180352)))]; + 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(307181952)))]; + 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_2758_axes_0 = const()[name = tensor("op_2758_axes_0"), val = tensor([2])]; + tensor var_2758_cast_fp16 = squeeze(axes = var_2758_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_2758_cast_fp16")]; + tensor var_2761_perm_0 = const()[name = tensor("op_2761_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(307183552)))]; + tensor var_2761_cast_fp16 = transpose(perm = var_2761_perm_0, x = var_2758_cast_fp16)[name = tensor("transpose_0")]; + tensor logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_2761_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_2765 = const()[name = tensor("op_2765"), val = tensor(1)]; + tensor obj_171_interleave_0 = const()[name = tensor("obj_171_interleave_0"), val = tensor(false)]; + tensor key_cache_updates = concat(axis = var_2765, interleave = obj_171_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_cast_fp16))[name = tensor("obj_171_cast_fp16")]; + tensor var_2768 = const()[name = tensor("op_2768"), val = tensor(1)]; + tensor obj_173_interleave_0 = const()[name = tensor("obj_173_interleave_0"), val = tensor(false)]; + tensor value_cache_updates = concat(axis = var_2768, interleave = obj_173_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_cast_fp16))[name = tensor("obj_173_cast_fp16")]; + tensor var_2779_begin_0 = const()[name = tensor("op_2779_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_2779_end_0 = const()[name = tensor("op_2779_end_0"), val = tensor([1, 4, 1, 1500])]; + tensor var_2779_end_mask_0 = const()[name = tensor("op_2779_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2779_cast_fp16 = slice_by_index(begin = var_2779_begin_0, end = var_2779_end_0, end_mask = var_2779_end_mask_0, x = obj_83_cast_fp16)[name = tensor("op_2779_cast_fp16")]; + tensor var_2782_begin_0 = const()[name = tensor("op_2782_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2782_end_0 = const()[name = tensor("op_2782_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_2782_end_mask_0 = const()[name = tensor("op_2782_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_2782_squeeze_mask_0 = const()[name = tensor("op_2782_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_2782_cast_fp16 = slice_by_index(begin = var_2782_begin_0, end = var_2782_end_0, end_mask = var_2782_end_mask_0, squeeze_mask = var_2782_squeeze_mask_0, x = var_2779_cast_fp16)[name = tensor("op_2782_cast_fp16")]; + tensor var_2797_begin_0 = const()[name = tensor("op_2797_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_2797_end_0 = const()[name = tensor("op_2797_end_0"), val = tensor([1, 10, 1, 1500])]; + tensor var_2797_end_mask_0 = const()[name = tensor("op_2797_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2797_cast_fp16 = slice_by_index(begin = var_2797_begin_0, end = var_2797_end_0, end_mask = var_2797_end_mask_0, x = obj_83_cast_fp16)[name = tensor("op_2797_cast_fp16")]; + tensor var_2800_begin_0 = const()[name = tensor("op_2800_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2800_end_0 = const()[name = tensor("op_2800_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_2800_end_mask_0 = const()[name = tensor("op_2800_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_2800_squeeze_mask_0 = const()[name = tensor("op_2800_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_2800_cast_fp16 = slice_by_index(begin = var_2800_begin_0, end = var_2800_end_0, end_mask = var_2800_end_mask_0, squeeze_mask = var_2800_squeeze_mask_0, x = var_2797_cast_fp16)[name = tensor("op_2800_cast_fp16")]; + tensor var_2815_begin_0 = const()[name = tensor("op_2815_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2815_end_0 = const()[name = tensor("op_2815_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_2815_end_mask_0 = const()[name = tensor("op_2815_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2815_cast_fp16 = slice_by_index(begin = var_2815_begin_0, end = var_2815_end_0, end_mask = var_2815_end_mask_0, x = obj_125_cast_fp16)[name = tensor("op_2815_cast_fp16")]; + tensor var_2818_begin_0 = const()[name = tensor("op_2818_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2818_end_0 = const()[name = tensor("op_2818_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_2818_end_mask_0 = const()[name = tensor("op_2818_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_2818_squeeze_mask_0 = const()[name = tensor("op_2818_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_2818_cast_fp16 = slice_by_index(begin = var_2818_begin_0, end = var_2818_end_0, end_mask = var_2818_end_mask_0, squeeze_mask = var_2818_squeeze_mask_0, x = var_2815_cast_fp16)[name = tensor("op_2818_cast_fp16")]; + tensor var_2833_begin_0 = const()[name = tensor("op_2833_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_2833_end_0 = const()[name = tensor("op_2833_end_0"), val = tensor([1, 5, 1, 1500])]; + tensor var_2833_end_mask_0 = const()[name = tensor("op_2833_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2833_cast_fp16 = slice_by_index(begin = var_2833_begin_0, end = var_2833_end_0, end_mask = var_2833_end_mask_0, x = obj_125_cast_fp16)[name = tensor("op_2833_cast_fp16")]; + tensor var_2836_begin_0 = const()[name = tensor("op_2836_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2836_end_0 = const()[name = tensor("op_2836_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_2836_end_mask_0 = const()[name = tensor("op_2836_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_2836_squeeze_mask_0 = const()[name = tensor("op_2836_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_2836_cast_fp16 = slice_by_index(begin = var_2836_begin_0, end = var_2836_end_0, end_mask = var_2836_end_mask_0, squeeze_mask = var_2836_squeeze_mask_0, x = var_2833_cast_fp16)[name = tensor("op_2836_cast_fp16")]; + tensor var_2851_begin_0 = const()[name = tensor("op_2851_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_2851_end_0 = const()[name = tensor("op_2851_end_0"), val = tensor([1, 8, 1, 1500])]; + tensor var_2851_end_mask_0 = const()[name = tensor("op_2851_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2851_cast_fp16 = slice_by_index(begin = var_2851_begin_0, end = var_2851_end_0, end_mask = var_2851_end_mask_0, x = obj_125_cast_fp16)[name = tensor("op_2851_cast_fp16")]; + tensor var_2854_begin_0 = const()[name = tensor("op_2854_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2854_end_0 = const()[name = tensor("op_2854_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_2854_end_mask_0 = const()[name = tensor("op_2854_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_2854_squeeze_mask_0 = const()[name = tensor("op_2854_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_2854_cast_fp16 = slice_by_index(begin = var_2854_begin_0, end = var_2854_end_0, end_mask = var_2854_end_mask_0, squeeze_mask = var_2854_squeeze_mask_0, x = var_2851_cast_fp16)[name = tensor("op_2854_cast_fp16")]; + tensor var_2869_begin_0 = const()[name = tensor("op_2869_begin_0"), val = tensor([0, 8, 0, 0])]; + tensor var_2869_end_0 = const()[name = tensor("op_2869_end_0"), val = tensor([1, 9, 1, 1500])]; + tensor var_2869_end_mask_0 = const()[name = tensor("op_2869_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2869_cast_fp16 = slice_by_index(begin = var_2869_begin_0, end = var_2869_end_0, end_mask = var_2869_end_mask_0, x = obj_125_cast_fp16)[name = tensor("op_2869_cast_fp16")]; + tensor var_2872_begin_0 = const()[name = tensor("op_2872_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2872_end_0 = const()[name = tensor("op_2872_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_2872_end_mask_0 = const()[name = tensor("op_2872_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_2872_squeeze_mask_0 = const()[name = tensor("op_2872_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_2872_cast_fp16 = slice_by_index(begin = var_2872_begin_0, end = var_2872_end_0, end_mask = var_2872_end_mask_0, squeeze_mask = var_2872_squeeze_mask_0, x = var_2869_cast_fp16)[name = tensor("op_2872_cast_fp16")]; + tensor var_2887_begin_0 = const()[name = tensor("op_2887_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2887_end_0 = const()[name = tensor("op_2887_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_2887_end_mask_0 = const()[name = tensor("op_2887_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2887_cast_fp16 = slice_by_index(begin = var_2887_begin_0, end = var_2887_end_0, end_mask = var_2887_end_mask_0, x = obj_139_cast_fp16)[name = tensor("op_2887_cast_fp16")]; + tensor var_2890_begin_0 = const()[name = tensor("op_2890_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2890_end_0 = const()[name = tensor("op_2890_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_2890_end_mask_0 = const()[name = tensor("op_2890_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_2890_squeeze_mask_0 = const()[name = tensor("op_2890_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_2890_cast_fp16 = slice_by_index(begin = var_2890_begin_0, end = var_2890_end_0, end_mask = var_2890_end_mask_0, squeeze_mask = var_2890_squeeze_mask_0, x = var_2887_cast_fp16)[name = tensor("op_2890_cast_fp16")]; + tensor var_2905_begin_0 = const()[name = tensor("op_2905_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_2905_end_0 = const()[name = tensor("op_2905_end_0"), val = tensor([1, 8, 1, 1500])]; + tensor var_2905_end_mask_0 = const()[name = tensor("op_2905_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2905_cast_fp16 = slice_by_index(begin = var_2905_begin_0, end = var_2905_end_0, end_mask = var_2905_end_mask_0, x = obj_139_cast_fp16)[name = tensor("op_2905_cast_fp16")]; + tensor var_2908_begin_0 = const()[name = tensor("op_2908_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2908_end_0 = const()[name = tensor("op_2908_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_2908_end_mask_0 = const()[name = tensor("op_2908_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_2908_squeeze_mask_0 = const()[name = tensor("op_2908_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_2908_cast_fp16 = slice_by_index(begin = var_2908_begin_0, end = var_2908_end_0, end_mask = var_2908_end_mask_0, squeeze_mask = var_2908_squeeze_mask_0, x = var_2905_cast_fp16)[name = tensor("op_2908_cast_fp16")]; + tensor var_2923_begin_0 = const()[name = tensor("op_2923_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_2923_end_0 = const()[name = tensor("op_2923_end_0"), val = tensor([1, 10, 1, 1500])]; + tensor var_2923_end_mask_0 = const()[name = tensor("op_2923_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2923_cast_fp16 = slice_by_index(begin = var_2923_begin_0, end = var_2923_end_0, end_mask = var_2923_end_mask_0, x = obj_139_cast_fp16)[name = tensor("op_2923_cast_fp16")]; + tensor var_2926_begin_0 = const()[name = tensor("op_2926_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2926_end_0 = const()[name = tensor("op_2926_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_2926_end_mask_0 = const()[name = tensor("op_2926_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_2926_squeeze_mask_0 = const()[name = tensor("op_2926_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_2926_cast_fp16 = slice_by_index(begin = var_2926_begin_0, end = var_2926_end_0, end_mask = var_2926_end_mask_0, squeeze_mask = var_2926_squeeze_mask_0, x = var_2923_cast_fp16)[name = tensor("op_2926_cast_fp16")]; + tensor var_2941_begin_0 = const()[name = tensor("op_2941_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_2941_end_0 = const()[name = tensor("op_2941_end_0"), val = tensor([1, 6, 1, 1500])]; + tensor var_2941_end_mask_0 = const()[name = tensor("op_2941_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_2941_cast_fp16 = slice_by_index(begin = var_2941_begin_0, end = var_2941_end_0, end_mask = var_2941_end_mask_0, x = obj_153_cast_fp16)[name = tensor("op_2941_cast_fp16")]; + tensor var_2944_begin_0 = const()[name = tensor("op_2944_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2944_end_0 = const()[name = tensor("op_2944_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_2944_end_mask_0 = const()[name = tensor("op_2944_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_2944_squeeze_mask_0 = const()[name = tensor("op_2944_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_2944_cast_fp16 = slice_by_index(begin = var_2944_begin_0, end = var_2944_end_0, end_mask = var_2944_end_mask_0, squeeze_mask = var_2944_squeeze_mask_0, x = var_2941_cast_fp16)[name = tensor("op_2944_cast_fp16")]; + tensor var_2951 = const()[name = tensor("op_2951"), val = tensor(1)]; + tensor var_2952_interleave_0 = const()[name = tensor("op_2952_interleave_0"), val = tensor(false)]; + tensor var_2952_cast_fp16 = concat(axis = var_2951, interleave = var_2952_interleave_0, values = (var_2782_cast_fp16, var_2800_cast_fp16, var_2818_cast_fp16, var_2836_cast_fp16, var_2854_cast_fp16, var_2872_cast_fp16, var_2890_cast_fp16, var_2908_cast_fp16, var_2926_cast_fp16, var_2944_cast_fp16))[name = tensor("op_2952_cast_fp16")]; + tensor obj_axes_0 = const()[name = tensor("obj_axes_0"), val = tensor([1])]; + tensor obj_keep_dims_0 = const()[name = tensor("obj_keep_dims_0"), val = tensor(false)]; + tensor alignment_heads_weights = reduce_mean(axes = obj_axes_0, keep_dims = obj_keep_dims_0, x = var_2952_cast_fp16)[name = tensor("obj_cast_fp16")]; + } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); +} \ No newline at end of file