program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.5.0"}, {"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_22_axis_0 = const()[name = tensor("op_22_axis_0"), val = tensor(0)]; tensor var_22_batch_dims_0 = const()[name = tensor("op_22_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_22_cast_fp16 = gather(axis = var_22_axis_0, batch_dims = var_22_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor("op_22_cast_fp16")]; tensor var_26_axis_0 = const()[name = tensor("op_26_axis_0"), val = tensor(0)]; tensor var_26_batch_dims_0 = const()[name = tensor("op_26_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_26_cast_fp16 = gather(axis = var_26_axis_0, batch_dims = var_26_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor("op_26_cast_fp16")]; tensor hidden_states_1_cast_fp16 = add(x = var_22_cast_fp16, y = var_26_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; tensor var_40_axes_0 = const()[name = tensor("op_40_axes_0"), val = tensor([2])]; tensor var_40_cast_fp16 = expand_dims(axes = var_40_axes_0, x = hidden_states_1_cast_fp16)[name = tensor("op_40_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_40_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([768, 768, 768])]; tensor var_45_axis_0 = const()[name = tensor("op_45_axis_0"), val = tensor(1)]; tensor var_45_cast_fp16_0, tensor var_45_cast_fp16_1, tensor var_45_cast_fp16_2 = split(axis = var_45_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor("op_45_cast_fp16")]; tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([768, 768, 768])]; tensor var_51_axis_0 = const()[name = tensor("op_51_axis_0"), val = tensor(1)]; tensor var_51_cast_fp16_0, tensor var_51_cast_fp16_1, tensor var_51_cast_fp16_2 = split(axis = var_51_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor("op_51_cast_fp16")]; tensor var_60 = const()[name = tensor("op_60"), val = tensor(3)]; tensor out_1_axes_0 = const()[name = tensor("out_1_axes_0"), val = tensor([1])]; tensor var_85_to_fp16 = const()[name = tensor("op_85_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_85_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_120_axes_0 = const()[name = tensor("op_120_axes_0"), val = tensor([1])]; tensor var_120_cast_fp16 = expand_dims(axes = var_120_axes_0, x = kv_cache_update_mask)[name = tensor("op_120_cast_fp16")]; tensor var_121_axes_0 = const()[name = tensor("op_121_axes_0"), val = tensor([2])]; tensor var_121_cast_fp16 = expand_dims(axes = var_121_axes_0, x = var_120_cast_fp16)[name = tensor("op_121_cast_fp16")]; tensor var_61_to_fp16 = const()[name = tensor("op_61_to_fp16"), val = tensor(0x1p+0)]; tensor var_123_cast_fp16 = sub(x = var_61_to_fp16, y = var_121_cast_fp16)[name = tensor("op_123_cast_fp16")]; tensor var_124_cast_fp16 = mul(x = var_45_cast_fp16_0, y = var_123_cast_fp16)[name = tensor("op_124_cast_fp16")]; tensor var_125_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_121_cast_fp16)[name = tensor("op_125_cast_fp16")]; tensor key_1_cast_fp16 = add(x = var_124_cast_fp16, y = var_125_cast_fp16)[name = tensor("key_1_cast_fp16")]; tensor var_128_cast_fp16 = mul(x = var_51_cast_fp16_0, y = var_123_cast_fp16)[name = tensor("op_128_cast_fp16")]; tensor var_129_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_121_cast_fp16)[name = tensor("op_129_cast_fp16")]; tensor value_1_cast_fp16 = add(x = var_128_cast_fp16, y = var_129_cast_fp16)[name = tensor("value_1_cast_fp16")]; tensor var_133 = const()[name = tensor("op_133"), val = tensor([1, 12, 64, 1])]; tensor mh_q_1_cast_fp16 = reshape(shape = var_133, x = query_1_cast_fp16)[name = tensor("mh_q_1_cast_fp16")]; tensor var_135_to_fp16 = const()[name = tensor("op_135_to_fp16"), val = tensor(0x1p-3)]; tensor var_136_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_135_to_fp16)[name = tensor("op_136_cast_fp16")]; tensor var_139 = const()[name = tensor("op_139"), val = tensor([1, 12, 64, 448])]; tensor var_140_cast_fp16 = reshape(shape = var_139, x = key_1_cast_fp16)[name = tensor("op_140_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_136_cast_fp16, y = var_140_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; tensor var_144_axes_0 = const()[name = tensor("op_144_axes_0"), val = tensor([1])]; tensor var_144_cast_fp16 = expand_dims(axes = var_144_axes_0, x = decoder_key_padding_mask)[name = tensor("op_144_cast_fp16")]; tensor var_145_axes_0 = const()[name = tensor("op_145_axes_0"), val = tensor([2])]; tensor var_145_cast_fp16 = expand_dims(axes = var_145_axes_0, x = var_144_cast_fp16)[name = tensor("op_145_cast_fp16")]; tensor mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_145_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; tensor var_148_cast_fp16 = softmax(axis = var_60, x = mh_w_3_cast_fp16)[name = tensor("op_148_cast_fp16")]; tensor var_149 = const()[name = tensor("op_149"), val = tensor([1, 12, 64, 448])]; tensor var_150_cast_fp16 = reshape(shape = var_149, x = value_1_cast_fp16)[name = tensor("op_150_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_150_cast_fp16, y = var_148_cast_fp16)[name = tensor("attn_1_cast_fp16")]; tensor var_153 = const()[name = tensor("op_153"), val = tensor([1, 768, 1, 1])]; tensor input_1_cast_fp16 = reshape(shape = var_153, 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_175_to_fp16 = const()[name = tensor("op_175_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_175_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_211 = const()[name = tensor("op_211"), val = tensor([1, 12, 64, 1])]; tensor mh_q_3_cast_fp16 = reshape(shape = var_211, x = query_3_cast_fp16)[name = tensor("mh_q_3_cast_fp16")]; tensor var_213_to_fp16 = const()[name = tensor("op_213_to_fp16"), val = tensor(0x1p-3)]; tensor var_214_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_213_to_fp16)[name = tensor("op_214_cast_fp16")]; tensor var_217 = const()[name = tensor("op_217"), val = tensor([1, 12, 64, 1500])]; tensor var_218_cast_fp16 = reshape(shape = var_217, x = key_3_cast_fp16)[name = tensor("op_218_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_214_cast_fp16, y = var_218_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; tensor var_221_cast_fp16 = softmax(axis = var_60, x = mh_w_5_cast_fp16)[name = tensor("op_221_cast_fp16")]; tensor var_222 = const()[name = tensor("op_222"), val = tensor([1, 12, 64, 1500])]; tensor var_223_cast_fp16 = reshape(shape = var_222, x = value_3_cast_fp16)[name = tensor("op_223_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_223_cast_fp16, y = var_221_cast_fp16)[name = tensor("attn_3_cast_fp16")]; tensor var_226 = const()[name = tensor("op_226"), val = tensor([1, 768, 1, 1])]; tensor input_3_cast_fp16 = reshape(shape = var_226, 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_244_to_fp16 = const()[name = tensor("op_244_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_244_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_279 = const()[name = tensor("op_279"), val = tensor(3)]; tensor out_7_axes_0 = const()[name = tensor("out_7_axes_0"), val = tensor([1])]; tensor var_304_to_fp16 = const()[name = tensor("op_304_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_304_to_fp16, x = inputs_7_cast_fp16)[name = tensor("out_7_cast_fp16")]; tensor obj_13_gamma_0_to_fp16 = const()[name = tensor("obj_13_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99258176)))]; tensor obj_13_beta_0_to_fp16 = const()[name = tensor("obj_13_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99259776)))]; tensor obj_13_epsilon_0_to_fp16 = const()[name = tensor("obj_13_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_13_cast_fp16 = batch_norm(beta = obj_13_beta_0_to_fp16, epsilon = obj_13_epsilon_0_to_fp16, gamma = obj_13_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor("obj_13_cast_fp16")]; tensor 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_13_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_13_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_13_cast_fp16)[name = tensor("current_value_3_cast_fp16")]; tensor var_343_cast_fp16 = mul(x = var_45_cast_fp16_1, y = var_123_cast_fp16)[name = tensor("op_343_cast_fp16")]; tensor var_344_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_121_cast_fp16)[name = tensor("op_344_cast_fp16")]; tensor key_5_cast_fp16 = add(x = var_343_cast_fp16, y = var_344_cast_fp16)[name = tensor("key_5_cast_fp16")]; tensor var_347_cast_fp16 = mul(x = var_51_cast_fp16_1, y = var_123_cast_fp16)[name = tensor("op_347_cast_fp16")]; tensor var_348_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_121_cast_fp16)[name = tensor("op_348_cast_fp16")]; tensor value_5_cast_fp16 = add(x = var_347_cast_fp16, y = var_348_cast_fp16)[name = tensor("value_5_cast_fp16")]; tensor var_352 = const()[name = tensor("op_352"), val = tensor([1, 12, 64, 1])]; tensor mh_q_5_cast_fp16 = reshape(shape = var_352, x = query_5_cast_fp16)[name = tensor("mh_q_5_cast_fp16")]; tensor var_354_to_fp16 = const()[name = tensor("op_354_to_fp16"), val = tensor(0x1p-3)]; tensor var_355_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_354_to_fp16)[name = tensor("op_355_cast_fp16")]; tensor var_358 = const()[name = tensor("op_358"), val = tensor([1, 12, 64, 448])]; tensor var_359_cast_fp16 = reshape(shape = var_358, x = key_5_cast_fp16)[name = tensor("op_359_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_355_cast_fp16, y = var_359_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; tensor mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_145_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; tensor var_367_cast_fp16 = softmax(axis = var_279, x = mh_w_9_cast_fp16)[name = tensor("op_367_cast_fp16")]; tensor var_368 = const()[name = tensor("op_368"), val = tensor([1, 12, 64, 448])]; tensor var_369_cast_fp16 = reshape(shape = var_368, x = value_5_cast_fp16)[name = tensor("op_369_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_369_cast_fp16, y = var_367_cast_fp16)[name = tensor("attn_5_cast_fp16")]; tensor var_372 = const()[name = tensor("op_372"), val = tensor([1, 768, 1, 1])]; tensor input_11_cast_fp16 = reshape(shape = var_372, x = attn_5_cast_fp16)[name = tensor("input_11_cast_fp16")]; tensor obj_19_pad_type_0 = const()[name = tensor("obj_19_pad_type_0"), val = tensor("valid")]; tensor obj_19_strides_0 = const()[name = tensor("obj_19_strides_0"), val = tensor([1, 1])]; tensor obj_19_pad_0 = const()[name = tensor("obj_19_pad_0"), val = tensor([0, 0, 0, 0])]; tensor obj_19_dilations_0 = const()[name = tensor("obj_19_dilations_0"), val = tensor([1, 1])]; tensor obj_19_groups_0 = const()[name = tensor("obj_19_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_19_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = obj_19_dilations_0, groups = obj_19_groups_0, pad = obj_19_pad_0, pad_type = obj_19_pad_type_0, strides = obj_19_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("obj_19_cast_fp16")]; tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_19_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([1])]; tensor var_394_to_fp16 = const()[name = tensor("op_394_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_394_to_fp16, x = inputs_9_cast_fp16)[name = tensor("out_9_cast_fp16")]; tensor obj_21_gamma_0_to_fp16 = const()[name = tensor("obj_21_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103985024)))]; tensor obj_21_beta_0_to_fp16 = const()[name = tensor("obj_21_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103986624)))]; tensor obj_21_epsilon_0_to_fp16 = const()[name = tensor("obj_21_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_21_cast_fp16 = batch_norm(beta = obj_21_beta_0_to_fp16, epsilon = obj_21_epsilon_0_to_fp16, gamma = obj_21_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor("obj_21_cast_fp16")]; tensor 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_21_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_430 = const()[name = tensor("op_430"), val = tensor([1, 12, 64, 1])]; tensor mh_q_7_cast_fp16 = reshape(shape = var_430, x = query_7_cast_fp16)[name = tensor("mh_q_7_cast_fp16")]; tensor var_432_to_fp16 = const()[name = tensor("op_432_to_fp16"), val = tensor(0x1p-3)]; tensor var_433_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_432_to_fp16)[name = tensor("op_433_cast_fp16")]; tensor var_436 = const()[name = tensor("op_436"), val = tensor([1, 12, 64, 1500])]; tensor var_437_cast_fp16 = reshape(shape = var_436, x = key_7_cast_fp16)[name = tensor("op_437_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_433_cast_fp16, y = var_437_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; tensor var_440_cast_fp16 = softmax(axis = var_279, x = mh_w_11_cast_fp16)[name = tensor("op_440_cast_fp16")]; tensor var_441 = const()[name = tensor("op_441"), val = tensor([1, 12, 64, 1500])]; tensor var_442_cast_fp16 = reshape(shape = var_441, x = value_7_cast_fp16)[name = tensor("op_442_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_442_cast_fp16, y = var_440_cast_fp16)[name = tensor("attn_7_cast_fp16")]; tensor var_445 = const()[name = tensor("op_445"), val = tensor([1, 768, 1, 1])]; tensor input_13_cast_fp16 = reshape(shape = var_445, x = attn_7_cast_fp16)[name = tensor("input_13_cast_fp16")]; tensor obj_23_pad_type_0 = const()[name = tensor("obj_23_pad_type_0"), val = tensor("valid")]; tensor obj_23_strides_0 = const()[name = tensor("obj_23_strides_0"), val = tensor([1, 1])]; tensor obj_23_pad_0 = const()[name = tensor("obj_23_pad_0"), val = tensor([0, 0, 0, 0])]; tensor obj_23_dilations_0 = const()[name = tensor("obj_23_dilations_0"), val = tensor([1, 1])]; tensor obj_23_groups_0 = const()[name = tensor("obj_23_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_23_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = obj_23_dilations_0, groups = obj_23_groups_0, pad = obj_23_pad_0, pad_type = obj_23_pad_type_0, strides = obj_23_strides_0, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("obj_23_cast_fp16")]; tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_23_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; tensor out_11_axes_0 = const()[name = tensor("out_11_axes_0"), val = tensor([1])]; tensor var_463_to_fp16 = const()[name = tensor("op_463_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_463_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_498 = const()[name = tensor("op_498"), val = tensor(3)]; tensor out_13_axes_0 = const()[name = tensor("out_13_axes_0"), val = tensor([1])]; tensor var_523_to_fp16 = const()[name = tensor("op_523_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_523_to_fp16, x = inputs_13_cast_fp16)[name = tensor("out_13_cast_fp16")]; tensor obj_25_gamma_0_to_fp16 = const()[name = tensor("obj_25_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118160192)))]; tensor obj_25_beta_0_to_fp16 = const()[name = tensor("obj_25_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118161792)))]; tensor obj_25_epsilon_0_to_fp16 = const()[name = tensor("obj_25_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_25_cast_fp16 = batch_norm(beta = obj_25_beta_0_to_fp16, epsilon = obj_25_epsilon_0_to_fp16, gamma = obj_25_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor("obj_25_cast_fp16")]; tensor 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_25_cast_fp16)[name = tensor("query_9_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_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_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_2_self_attn_k_proj_weight_to_fp16, x = obj_25_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_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_cast_fp16 = conv(bias = layers_2_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_2_self_attn_v_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor("current_value_cast_fp16")]; tensor var_562_cast_fp16 = mul(x = var_45_cast_fp16_2, y = var_123_cast_fp16)[name = tensor("op_562_cast_fp16")]; tensor var_563_cast_fp16 = mul(x = current_key_cast_fp16, y = var_121_cast_fp16)[name = tensor("op_563_cast_fp16")]; tensor key_9_cast_fp16 = add(x = var_562_cast_fp16, y = var_563_cast_fp16)[name = tensor("key_9_cast_fp16")]; tensor var_566_cast_fp16 = mul(x = var_51_cast_fp16_2, y = var_123_cast_fp16)[name = tensor("op_566_cast_fp16")]; tensor var_567_cast_fp16 = mul(x = current_value_cast_fp16, y = var_121_cast_fp16)[name = tensor("op_567_cast_fp16")]; tensor value_9_cast_fp16 = add(x = var_566_cast_fp16, y = var_567_cast_fp16)[name = tensor("value_9_cast_fp16")]; tensor var_571 = const()[name = tensor("op_571"), val = tensor([1, 12, 64, 1])]; tensor mh_q_9_cast_fp16 = reshape(shape = var_571, x = query_9_cast_fp16)[name = tensor("mh_q_9_cast_fp16")]; tensor var_573_to_fp16 = const()[name = tensor("op_573_to_fp16"), val = tensor(0x1p-3)]; tensor var_574_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_573_to_fp16)[name = tensor("op_574_cast_fp16")]; tensor var_577 = const()[name = tensor("op_577"), val = tensor([1, 12, 64, 448])]; tensor var_578_cast_fp16 = reshape(shape = var_577, x = key_9_cast_fp16)[name = tensor("op_578_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_574_cast_fp16, y = var_578_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; tensor mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_145_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; tensor var_586_cast_fp16 = softmax(axis = var_498, x = mh_w_15_cast_fp16)[name = tensor("op_586_cast_fp16")]; tensor var_587 = const()[name = tensor("op_587"), val = tensor([1, 12, 64, 448])]; tensor var_588_cast_fp16 = reshape(shape = var_587, x = value_9_cast_fp16)[name = tensor("op_588_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_588_cast_fp16, y = var_586_cast_fp16)[name = tensor("attn_9_cast_fp16")]; tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 768, 1, 1])]; tensor input_21_cast_fp16 = reshape(shape = var_591, x = attn_9_cast_fp16)[name = tensor("input_21_cast_fp16")]; tensor obj_31_pad_type_0 = const()[name = tensor("obj_31_pad_type_0"), val = tensor("valid")]; tensor obj_31_strides_0 = const()[name = tensor("obj_31_strides_0"), val = tensor([1, 1])]; tensor obj_31_pad_0 = const()[name = tensor("obj_31_pad_0"), val = tensor([0, 0, 0, 0])]; tensor obj_31_dilations_0 = const()[name = tensor("obj_31_dilations_0"), val = tensor([1, 1])]; tensor obj_31_groups_0 = const()[name = tensor("obj_31_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_31_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = obj_31_dilations_0, groups = obj_31_groups_0, pad = obj_31_pad_0, pad_type = obj_31_pad_type_0, strides = obj_31_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("obj_31_cast_fp16")]; tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_31_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([1])]; tensor var_613_to_fp16 = const()[name = tensor("op_613_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_613_to_fp16, x = inputs_15_cast_fp16)[name = tensor("out_15_cast_fp16")]; tensor obj_33_gamma_0_to_fp16 = const()[name = tensor("obj_33_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122887040)))]; tensor obj_33_beta_0_to_fp16 = const()[name = tensor("obj_33_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122888640)))]; tensor obj_33_epsilon_0_to_fp16 = const()[name = tensor("obj_33_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_33_cast_fp16 = batch_norm(beta = obj_33_beta_0_to_fp16, epsilon = obj_33_epsilon_0_to_fp16, gamma = obj_33_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor("obj_33_cast_fp16")]; tensor 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_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_cast_fp16 = conv(bias = layers_2_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_2_encoder_attn_q_proj_weight_to_fp16, x = obj_33_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_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_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_2_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_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_cast_fp16 = conv(bias = layers_2_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_2_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_cast_fp16")]; tensor var_649 = const()[name = tensor("op_649"), val = tensor([1, 12, 64, 1])]; tensor mh_q_cast_fp16 = reshape(shape = var_649, x = query_cast_fp16)[name = tensor("mh_q_cast_fp16")]; tensor var_651_to_fp16 = const()[name = tensor("op_651_to_fp16"), val = tensor(0x1p-3)]; tensor var_652_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_651_to_fp16)[name = tensor("op_652_cast_fp16")]; tensor var_655 = const()[name = tensor("op_655"), val = tensor([1, 12, 64, 1500])]; tensor var_656_cast_fp16 = reshape(shape = var_655, x = key_cast_fp16)[name = tensor("op_656_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_652_cast_fp16, y = var_656_cast_fp16)[name = tensor("mh_w_cast_fp16")]; tensor var_659_cast_fp16 = softmax(axis = var_498, x = mh_w_cast_fp16)[name = tensor("op_659_cast_fp16")]; tensor var_660 = const()[name = tensor("op_660"), val = tensor([1, 12, 64, 1500])]; tensor var_661_cast_fp16 = reshape(shape = var_660, x = value_cast_fp16)[name = tensor("op_661_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_661_cast_fp16, y = var_659_cast_fp16)[name = tensor("attn_cast_fp16")]; tensor var_664 = const()[name = tensor("op_664"), val = tensor([1, 768, 1, 1])]; tensor input_23_cast_fp16 = reshape(shape = var_664, x = attn_cast_fp16)[name = tensor("input_23_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_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_35_cast_fp16 = conv(bias = layers_2_encoder_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_encoder_attn_o_proj_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("obj_35_cast_fp16")]; tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_35_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; tensor out_17_axes_0 = const()[name = tensor("out_17_axes_0"), val = tensor([1])]; tensor var_682_to_fp16 = const()[name = tensor("op_682_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_682_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_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_27_cast_fp16)[name = tensor("input_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_cast_fp16)[name = tensor("hidden_states_7_cast_fp16")]; tensor inputs_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor("inputs_cast_fp16")]; tensor out_axes_0 = const()[name = tensor("out_axes_0"), val = tensor([1])]; tensor var_724_to_fp16 = const()[name = tensor("op_724_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_724_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(137062208)))]; 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(137063808)))]; 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_735_axes_0 = const()[name = tensor("op_735_axes_0"), val = tensor([2])]; tensor var_735_cast_fp16 = squeeze(axes = var_735_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_735_cast_fp16")]; tensor var_738_perm_0 = const()[name = tensor("op_738_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(137065408)))]; tensor var_738_cast_fp16 = transpose(perm = var_738_perm_0, x = var_735_cast_fp16)[name = tensor("transpose_0")]; tensor logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_738_cast_fp16)[name = tensor("linear_0_cast_fp16")]; tensor var_742 = const()[name = tensor("op_742"), val = tensor(1)]; tensor obj_39_interleave_0 = const()[name = tensor("obj_39_interleave_0"), val = tensor(false)]; tensor key_cache_updates = concat(axis = var_742, interleave = obj_39_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_cast_fp16))[name = tensor("obj_39_cast_fp16")]; tensor var_745 = const()[name = tensor("op_745"), val = tensor(1)]; tensor obj_interleave_0 = const()[name = tensor("obj_interleave_0"), val = tensor(false)]; tensor value_cache_updates = concat(axis = var_745, interleave = obj_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_cast_fp16))[name = tensor("obj_cast_fp16")]; } -> (logits, key_cache_updates, value_cache_updates); }