diff --git "a/base/TextDecoder.mlmodelc/model.mil" "b/base/TextDecoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/base/TextDecoder.mlmodelc/model.mil" @@ -0,0 +1,1010 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3405.2.1"}, {"coremlc-version", "3404.23.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + 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_28_axis_0 = const()[name = tensor("op_28_axis_0"), val = tensor(0)]; + tensor var_28_batch_dims_0 = const()[name = tensor("op_28_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_28_cast_fp16 = gather(axis = var_28_axis_0, batch_dims = var_28_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor("op_28_cast_fp16")]; + tensor var_32_axis_0 = const()[name = tensor("op_32_axis_0"), val = tensor(0)]; + tensor var_32_batch_dims_0 = const()[name = tensor("op_32_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(53109888)))]; + tensor var_32_cast_fp16 = gather(axis = var_32_axis_0, batch_dims = var_32_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor("op_32_cast_fp16")]; + tensor hidden_states_1_cast_fp16 = add(x = var_28_cast_fp16, y = var_32_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; + tensor var_46_axes_0 = const()[name = tensor("op_46_axes_0"), val = tensor([2])]; + tensor var_46_cast_fp16 = expand_dims(axes = var_46_axes_0, x = hidden_states_1_cast_fp16)[name = tensor("op_46_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_46_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([512, 512, 512, 512, 512, 512])]; + 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, tensor var_51_cast_fp16_3, tensor var_51_cast_fp16_4, tensor var_51_cast_fp16_5 = split(axis = var_51_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor("op_51_cast_fp16")]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([512, 512, 512, 512, 512, 512])]; + tensor var_60_axis_0 = const()[name = tensor("op_60_axis_0"), val = tensor(1)]; + tensor var_60_cast_fp16_0, tensor var_60_cast_fp16_1, tensor var_60_cast_fp16_2, tensor var_60_cast_fp16_3, tensor var_60_cast_fp16_4, tensor var_60_cast_fp16_5 = split(axis = var_60_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor("op_60_cast_fp16")]; + tensor var_72 = const()[name = tensor("op_72"), val = tensor(3)]; + tensor out_1_axes_0 = const()[name = tensor("out_1_axes_0"), val = tensor([1])]; + tensor var_97_to_fp16 = const()[name = tensor("op_97_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_97_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(53568704)))]; + 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(53569792)))]; + 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(53570880)))]; + 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(53571968)))]; + 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(53573056)))]; + 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(54097408)))]; + 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(54098496)))]; + 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(54622848)))]; + 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(55147200)))]; + 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_132_axes_0 = const()[name = tensor("op_132_axes_0"), val = tensor([1])]; + tensor var_132_cast_fp16 = expand_dims(axes = var_132_axes_0, x = kv_cache_update_mask)[name = tensor("op_132_cast_fp16")]; + tensor var_133_axes_0 = const()[name = tensor("op_133_axes_0"), val = tensor([2])]; + tensor var_133_cast_fp16 = expand_dims(axes = var_133_axes_0, x = var_132_cast_fp16)[name = tensor("op_133_cast_fp16")]; + tensor var_73_to_fp16 = const()[name = tensor("op_73_to_fp16"), val = tensor(0x1p+0)]; + tensor var_135_cast_fp16 = sub(x = var_73_to_fp16, y = var_133_cast_fp16)[name = tensor("op_135_cast_fp16")]; + tensor var_136_cast_fp16 = mul(x = var_51_cast_fp16_0, y = var_135_cast_fp16)[name = tensor("op_136_cast_fp16")]; + tensor var_137_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_133_cast_fp16)[name = tensor("op_137_cast_fp16")]; + tensor key_1_cast_fp16 = add(x = var_136_cast_fp16, y = var_137_cast_fp16)[name = tensor("key_1_cast_fp16")]; + tensor var_140_cast_fp16 = mul(x = var_60_cast_fp16_0, y = var_135_cast_fp16)[name = tensor("op_140_cast_fp16")]; + tensor var_141_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_133_cast_fp16)[name = tensor("op_141_cast_fp16")]; + tensor value_1_cast_fp16 = add(x = var_140_cast_fp16, y = var_141_cast_fp16)[name = tensor("value_1_cast_fp16")]; + tensor var_145 = const()[name = tensor("op_145"), val = tensor([1, 8, 64, 1])]; + tensor mh_q_1_cast_fp16 = reshape(shape = var_145, x = query_1_cast_fp16)[name = tensor("mh_q_1_cast_fp16")]; + tensor var_147_to_fp16 = const()[name = tensor("op_147_to_fp16"), val = tensor(0x1p-3)]; + tensor var_148_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_147_to_fp16)[name = tensor("op_148_cast_fp16")]; + tensor var_151 = const()[name = tensor("op_151"), val = tensor([1, 8, 64, 448])]; + tensor var_152_cast_fp16 = reshape(shape = var_151, x = key_1_cast_fp16)[name = tensor("op_152_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_148_cast_fp16, y = var_152_cast_fp16)[name = tensor("mh_w_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 = decoder_key_padding_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 mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_157_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; + tensor var_160_cast_fp16 = softmax(axis = var_72, x = mh_w_3_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor var_161 = const()[name = tensor("op_161"), val = tensor([1, 8, 64, 448])]; + tensor var_162_cast_fp16 = reshape(shape = var_161, x = value_1_cast_fp16)[name = tensor("op_162_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_162_cast_fp16, y = var_160_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_165 = const()[name = tensor("op_165"), val = tensor([1, 512, 1, 1])]; + tensor input_1_cast_fp16 = reshape(shape = var_165, 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(55148288)))]; + 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(55672640)))]; + 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_187_to_fp16 = const()[name = tensor("op_187_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_187_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(55673728)))]; + 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(55674816)))]; + 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(55675904)))]; + 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(56200256)))]; + 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(56201344)))]; + 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(56725696)))]; + 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(57250048)))]; + 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_223 = const()[name = tensor("op_223"), val = tensor([1, 8, 64, 1])]; + tensor mh_q_3_cast_fp16 = reshape(shape = var_223, x = query_3_cast_fp16)[name = tensor("mh_q_3_cast_fp16")]; + tensor var_225_to_fp16 = const()[name = tensor("op_225_to_fp16"), val = tensor(0x1p-3)]; + tensor var_226_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_225_to_fp16)[name = tensor("op_226_cast_fp16")]; + tensor var_229 = const()[name = tensor("op_229"), val = tensor([1, 8, 64, 1500])]; + tensor var_230_cast_fp16 = reshape(shape = var_229, x = key_3_cast_fp16)[name = tensor("op_230_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_226_cast_fp16, y = var_230_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; + tensor obj_13_cast_fp16 = softmax(axis = var_72, x = mh_w_5_cast_fp16)[name = tensor("obj_13_cast_fp16")]; + tensor var_234 = const()[name = tensor("op_234"), val = tensor([1, 8, 64, 1500])]; + tensor var_235_cast_fp16 = reshape(shape = var_234, x = value_3_cast_fp16)[name = tensor("op_235_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_235_cast_fp16, y = obj_13_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_238 = const()[name = tensor("op_238"), val = tensor([1, 512, 1, 1])]; + tensor input_3_cast_fp16 = reshape(shape = var_238, 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(57251136)))]; + 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(57775488)))]; + 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_256_to_fp16 = const()[name = tensor("op_256_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_256_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(57776576)))]; + 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(57777664)))]; + 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(57778752)))]; + 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(59875968)))]; + 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(59880128)))]; + 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(61977344)))]; + 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_291 = const()[name = tensor("op_291"), val = tensor(3)]; + tensor out_7_axes_0 = const()[name = tensor("out_7_axes_0"), val = tensor([1])]; + tensor var_316_to_fp16 = const()[name = tensor("op_316_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_316_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(61978432)))]; + 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(61979520)))]; + 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(61980608)))]; + 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(62504960)))]; + 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(62506048)))]; + 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(63030400)))]; + 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(63554752)))]; + 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_355_cast_fp16 = mul(x = var_51_cast_fp16_1, y = var_135_cast_fp16)[name = tensor("op_355_cast_fp16")]; + tensor var_356_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_133_cast_fp16)[name = tensor("op_356_cast_fp16")]; + tensor key_5_cast_fp16 = add(x = var_355_cast_fp16, y = var_356_cast_fp16)[name = tensor("key_5_cast_fp16")]; + tensor var_359_cast_fp16 = mul(x = var_60_cast_fp16_1, y = var_135_cast_fp16)[name = tensor("op_359_cast_fp16")]; + tensor var_360_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_133_cast_fp16)[name = tensor("op_360_cast_fp16")]; + tensor value_5_cast_fp16 = add(x = var_359_cast_fp16, y = var_360_cast_fp16)[name = tensor("value_5_cast_fp16")]; + tensor var_364 = const()[name = tensor("op_364"), val = tensor([1, 8, 64, 1])]; + tensor mh_q_5_cast_fp16 = reshape(shape = var_364, x = query_5_cast_fp16)[name = tensor("mh_q_5_cast_fp16")]; + tensor var_366_to_fp16 = const()[name = tensor("op_366_to_fp16"), val = tensor(0x1p-3)]; + tensor var_367_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_366_to_fp16)[name = tensor("op_367_cast_fp16")]; + tensor var_370 = const()[name = tensor("op_370"), val = tensor([1, 8, 64, 448])]; + tensor var_371_cast_fp16 = reshape(shape = var_370, x = key_5_cast_fp16)[name = tensor("op_371_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_367_cast_fp16, y = var_371_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; + tensor mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_157_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; + tensor var_379_cast_fp16 = softmax(axis = var_291, x = mh_w_9_cast_fp16)[name = tensor("op_379_cast_fp16")]; + tensor var_380 = const()[name = tensor("op_380"), val = tensor([1, 8, 64, 448])]; + tensor var_381_cast_fp16 = reshape(shape = var_380, x = value_5_cast_fp16)[name = tensor("op_381_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_381_cast_fp16, y = var_379_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_384 = const()[name = tensor("op_384"), val = tensor([1, 512, 1, 1])]; + tensor input_11_cast_fp16 = reshape(shape = var_384, 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(63555840)))]; + 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(64080192)))]; + 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_406_to_fp16 = const()[name = tensor("op_406_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_406_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(64081280)))]; + 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(64082368)))]; + 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(64083456)))]; + 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(64607808)))]; + 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(64608896)))]; + 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(65133248)))]; + 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(65657600)))]; + 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_442 = const()[name = tensor("op_442"), val = tensor([1, 8, 64, 1])]; + tensor mh_q_7_cast_fp16 = reshape(shape = var_442, x = query_7_cast_fp16)[name = tensor("mh_q_7_cast_fp16")]; + tensor var_444_to_fp16 = const()[name = tensor("op_444_to_fp16"), val = tensor(0x1p-3)]; + tensor var_445_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_444_to_fp16)[name = tensor("op_445_cast_fp16")]; + tensor var_448 = const()[name = tensor("op_448"), val = tensor([1, 8, 64, 1500])]; + tensor var_449_cast_fp16 = reshape(shape = var_448, x = key_7_cast_fp16)[name = tensor("op_449_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_445_cast_fp16, y = var_449_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; + tensor obj_27_cast_fp16 = softmax(axis = var_291, x = mh_w_11_cast_fp16)[name = tensor("obj_27_cast_fp16")]; + tensor var_453 = const()[name = tensor("op_453"), val = tensor([1, 8, 64, 1500])]; + tensor var_454_cast_fp16 = reshape(shape = var_453, x = value_7_cast_fp16)[name = tensor("op_454_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_454_cast_fp16, y = obj_27_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_457 = const()[name = tensor("op_457"), val = tensor([1, 512, 1, 1])]; + tensor input_13_cast_fp16 = reshape(shape = var_457, 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(65658688)))]; + 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(66183040)))]; + 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_475_to_fp16 = const()[name = tensor("op_475_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_475_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(66184128)))]; + 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(66185216)))]; + 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(66186304)))]; + 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(68283520)))]; + 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(68287680)))]; + 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(70384896)))]; + 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_510 = const()[name = tensor("op_510"), val = tensor(3)]; + tensor out_13_axes_0 = const()[name = tensor("out_13_axes_0"), val = tensor([1])]; + tensor var_535_to_fp16 = const()[name = tensor("op_535_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_535_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(70385984)))]; + 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(70387072)))]; + 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(70388160)))]; + 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(70912512)))]; + 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(70913600)))]; + 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(71437952)))]; + 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(71962304)))]; + 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_574_cast_fp16 = mul(x = var_51_cast_fp16_2, y = var_135_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor var_575_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_133_cast_fp16)[name = tensor("op_575_cast_fp16")]; + tensor key_9_cast_fp16 = add(x = var_574_cast_fp16, y = var_575_cast_fp16)[name = tensor("key_9_cast_fp16")]; + tensor var_578_cast_fp16 = mul(x = var_60_cast_fp16_2, y = var_135_cast_fp16)[name = tensor("op_578_cast_fp16")]; + tensor var_579_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_133_cast_fp16)[name = tensor("op_579_cast_fp16")]; + tensor value_9_cast_fp16 = add(x = var_578_cast_fp16, y = var_579_cast_fp16)[name = tensor("value_9_cast_fp16")]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 8, 64, 1])]; + tensor mh_q_9_cast_fp16 = reshape(shape = var_583, x = query_9_cast_fp16)[name = tensor("mh_q_9_cast_fp16")]; + tensor var_585_to_fp16 = const()[name = tensor("op_585_to_fp16"), val = tensor(0x1p-3)]; + tensor var_586_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_585_to_fp16)[name = tensor("op_586_cast_fp16")]; + tensor var_589 = const()[name = tensor("op_589"), val = tensor([1, 8, 64, 448])]; + tensor var_590_cast_fp16 = reshape(shape = var_589, x = key_9_cast_fp16)[name = tensor("op_590_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_586_cast_fp16, y = var_590_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; + tensor mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_157_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; + tensor var_598_cast_fp16 = softmax(axis = var_510, x = mh_w_15_cast_fp16)[name = tensor("op_598_cast_fp16")]; + tensor var_599 = const()[name = tensor("op_599"), val = tensor([1, 8, 64, 448])]; + tensor var_600_cast_fp16 = reshape(shape = var_599, x = value_9_cast_fp16)[name = tensor("op_600_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_600_cast_fp16, y = var_598_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_603 = const()[name = tensor("op_603"), val = tensor([1, 512, 1, 1])]; + tensor input_21_cast_fp16 = reshape(shape = var_603, 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(71963392)))]; + 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(72487744)))]; + 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_625_to_fp16 = const()[name = tensor("op_625_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_625_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(72488832)))]; + 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(72489920)))]; + 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(72491008)))]; + 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(73015360)))]; + 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(73016448)))]; + 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(73540800)))]; + 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(74065152)))]; + 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_661 = const()[name = tensor("op_661"), val = tensor([1, 8, 64, 1])]; + tensor mh_q_11_cast_fp16 = reshape(shape = var_661, x = query_11_cast_fp16)[name = tensor("mh_q_11_cast_fp16")]; + tensor var_663_to_fp16 = const()[name = tensor("op_663_to_fp16"), val = tensor(0x1p-3)]; + tensor var_664_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_663_to_fp16)[name = tensor("op_664_cast_fp16")]; + tensor var_667 = const()[name = tensor("op_667"), val = tensor([1, 8, 64, 1500])]; + tensor var_668_cast_fp16 = reshape(shape = var_667, x = key_11_cast_fp16)[name = tensor("op_668_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_664_cast_fp16, y = var_668_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; + tensor obj_41_cast_fp16 = softmax(axis = var_510, x = mh_w_17_cast_fp16)[name = tensor("obj_41_cast_fp16")]; + tensor var_672 = const()[name = tensor("op_672"), val = tensor([1, 8, 64, 1500])]; + tensor var_673_cast_fp16 = reshape(shape = var_672, x = value_11_cast_fp16)[name = tensor("op_673_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_673_cast_fp16, y = obj_41_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_676 = const()[name = tensor("op_676"), val = tensor([1, 512, 1, 1])]; + tensor input_23_cast_fp16 = reshape(shape = var_676, 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(74066240)))]; + 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(74590592)))]; + 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_694_to_fp16 = const()[name = tensor("op_694_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_694_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(74591680)))]; + 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(74592768)))]; + 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(74593856)))]; + 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(76691072)))]; + 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(76695232)))]; + 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(78792448)))]; + 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_729 = const()[name = tensor("op_729"), val = tensor(3)]; + tensor out_19_axes_0 = const()[name = tensor("out_19_axes_0"), val = tensor([1])]; + tensor var_754_to_fp16 = const()[name = tensor("op_754_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_754_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(78793536)))]; + 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(78794624)))]; + 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(78795712)))]; + 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(79320064)))]; + 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(79321152)))]; + 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(79845504)))]; + 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(80369856)))]; + 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_793_cast_fp16 = mul(x = var_51_cast_fp16_3, y = var_135_cast_fp16)[name = tensor("op_793_cast_fp16")]; + tensor var_794_cast_fp16 = mul(x = current_key_7_cast_fp16, y = var_133_cast_fp16)[name = tensor("op_794_cast_fp16")]; + tensor key_13_cast_fp16 = add(x = var_793_cast_fp16, y = var_794_cast_fp16)[name = tensor("key_13_cast_fp16")]; + tensor var_797_cast_fp16 = mul(x = var_60_cast_fp16_3, y = var_135_cast_fp16)[name = tensor("op_797_cast_fp16")]; + tensor var_798_cast_fp16 = mul(x = current_value_7_cast_fp16, y = var_133_cast_fp16)[name = tensor("op_798_cast_fp16")]; + tensor value_13_cast_fp16 = add(x = var_797_cast_fp16, y = var_798_cast_fp16)[name = tensor("value_13_cast_fp16")]; + tensor var_802 = const()[name = tensor("op_802"), val = tensor([1, 8, 64, 1])]; + tensor mh_q_13_cast_fp16 = reshape(shape = var_802, x = query_13_cast_fp16)[name = tensor("mh_q_13_cast_fp16")]; + tensor var_804_to_fp16 = const()[name = tensor("op_804_to_fp16"), val = tensor(0x1p-3)]; + tensor var_805_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_804_to_fp16)[name = tensor("op_805_cast_fp16")]; + tensor var_808 = const()[name = tensor("op_808"), val = tensor([1, 8, 64, 448])]; + tensor var_809_cast_fp16 = reshape(shape = var_808, x = key_13_cast_fp16)[name = tensor("op_809_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_805_cast_fp16, y = var_809_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; + tensor mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_157_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; + tensor var_817_cast_fp16 = softmax(axis = var_729, x = mh_w_21_cast_fp16)[name = tensor("op_817_cast_fp16")]; + tensor var_818 = const()[name = tensor("op_818"), val = tensor([1, 8, 64, 448])]; + tensor var_819_cast_fp16 = reshape(shape = var_818, x = value_13_cast_fp16)[name = tensor("op_819_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_819_cast_fp16, y = var_817_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_822 = const()[name = tensor("op_822"), val = tensor([1, 512, 1, 1])]; + tensor input_31_cast_fp16 = reshape(shape = var_822, 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(80370944)))]; + 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(80895296)))]; + 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_844_to_fp16 = const()[name = tensor("op_844_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_844_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(80896384)))]; + 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(80897472)))]; + 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(80898560)))]; + 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(81422912)))]; + 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(81424000)))]; + 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(81948352)))]; + 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(82472704)))]; + 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_880 = const()[name = tensor("op_880"), val = tensor([1, 8, 64, 1])]; + tensor mh_q_15_cast_fp16 = reshape(shape = var_880, x = query_15_cast_fp16)[name = tensor("mh_q_15_cast_fp16")]; + tensor var_882_to_fp16 = const()[name = tensor("op_882_to_fp16"), val = tensor(0x1p-3)]; + tensor var_883_cast_fp16 = mul(x = mh_q_15_cast_fp16, y = var_882_to_fp16)[name = tensor("op_883_cast_fp16")]; + tensor var_886 = const()[name = tensor("op_886"), val = tensor([1, 8, 64, 1500])]; + tensor var_887_cast_fp16 = reshape(shape = var_886, x = key_15_cast_fp16)[name = tensor("op_887_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_883_cast_fp16, y = var_887_cast_fp16)[name = tensor("mh_w_23_cast_fp16")]; + tensor obj_55_cast_fp16 = softmax(axis = var_729, x = mh_w_23_cast_fp16)[name = tensor("obj_55_cast_fp16")]; + tensor var_891 = const()[name = tensor("op_891"), val = tensor([1, 8, 64, 1500])]; + tensor var_892_cast_fp16 = reshape(shape = var_891, x = value_15_cast_fp16)[name = tensor("op_892_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_892_cast_fp16, y = obj_55_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_895 = const()[name = tensor("op_895"), val = tensor([1, 512, 1, 1])]; + tensor input_33_cast_fp16 = reshape(shape = var_895, 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(82473792)))]; + 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(82998144)))]; + 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_916_to_fp16 = const()[name = tensor("op_916_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_916_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(82999232)))]; + 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(83000320)))]; + 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(83001408)))]; + 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(85098624)))]; + 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(85102784)))]; + 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(87200000)))]; + 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_952 = const()[name = tensor("op_952"), val = tensor(3)]; + tensor out_25_axes_0 = const()[name = tensor("out_25_axes_0"), val = tensor([1])]; + tensor var_977_to_fp16 = const()[name = tensor("op_977_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_977_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(87201088)))]; + 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(87202176)))]; + 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(87203264)))]; + 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(87727616)))]; + 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(87728704)))]; + 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(88253056)))]; + 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(88777408)))]; + 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_1016_cast_fp16 = mul(x = var_51_cast_fp16_4, y = var_135_cast_fp16)[name = tensor("op_1016_cast_fp16")]; + tensor var_1017_cast_fp16 = mul(x = current_key_9_cast_fp16, y = var_133_cast_fp16)[name = tensor("op_1017_cast_fp16")]; + tensor key_17_cast_fp16 = add(x = var_1016_cast_fp16, y = var_1017_cast_fp16)[name = tensor("key_17_cast_fp16")]; + tensor var_1020_cast_fp16 = mul(x = var_60_cast_fp16_4, y = var_135_cast_fp16)[name = tensor("op_1020_cast_fp16")]; + tensor var_1021_cast_fp16 = mul(x = current_value_9_cast_fp16, y = var_133_cast_fp16)[name = tensor("op_1021_cast_fp16")]; + tensor value_17_cast_fp16 = add(x = var_1020_cast_fp16, y = var_1021_cast_fp16)[name = tensor("value_17_cast_fp16")]; + tensor var_1025 = const()[name = tensor("op_1025"), val = tensor([1, 8, 64, 1])]; + tensor mh_q_17_cast_fp16 = reshape(shape = var_1025, x = query_17_cast_fp16)[name = tensor("mh_q_17_cast_fp16")]; + tensor var_1027_to_fp16 = const()[name = tensor("op_1027_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1028_cast_fp16 = mul(x = mh_q_17_cast_fp16, y = var_1027_to_fp16)[name = tensor("op_1028_cast_fp16")]; + tensor var_1031 = const()[name = tensor("op_1031"), val = tensor([1, 8, 64, 448])]; + tensor var_1032_cast_fp16 = reshape(shape = var_1031, x = key_17_cast_fp16)[name = tensor("op_1032_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_1028_cast_fp16, y = var_1032_cast_fp16)[name = tensor("mh_w_25_cast_fp16")]; + tensor mh_w_27_cast_fp16 = add(x = mh_w_25_cast_fp16, y = var_157_cast_fp16)[name = tensor("mh_w_27_cast_fp16")]; + tensor var_1040_cast_fp16 = softmax(axis = var_952, x = mh_w_27_cast_fp16)[name = tensor("op_1040_cast_fp16")]; + tensor var_1041 = const()[name = tensor("op_1041"), val = tensor([1, 8, 64, 448])]; + tensor var_1042_cast_fp16 = reshape(shape = var_1041, x = value_17_cast_fp16)[name = tensor("op_1042_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_1042_cast_fp16, y = var_1040_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_1045 = const()[name = tensor("op_1045"), val = tensor([1, 512, 1, 1])]; + tensor input_41_cast_fp16 = reshape(shape = var_1045, 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(88778496)))]; + 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(89302848)))]; + 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_1067_to_fp16 = const()[name = tensor("op_1067_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_1067_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(89303936)))]; + 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(89305024)))]; + 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(89306112)))]; + 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(89830464)))]; + 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(89831552)))]; + 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(90355904)))]; + 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(90880256)))]; + 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_1103 = const()[name = tensor("op_1103"), val = tensor([1, 8, 64, 1])]; + tensor mh_q_19_cast_fp16 = reshape(shape = var_1103, x = query_19_cast_fp16)[name = tensor("mh_q_19_cast_fp16")]; + tensor var_1105_to_fp16 = const()[name = tensor("op_1105_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1106_cast_fp16 = mul(x = mh_q_19_cast_fp16, y = var_1105_to_fp16)[name = tensor("op_1106_cast_fp16")]; + tensor var_1109 = const()[name = tensor("op_1109"), val = tensor([1, 8, 64, 1500])]; + tensor var_1110_cast_fp16 = reshape(shape = var_1109, x = key_19_cast_fp16)[name = tensor("op_1110_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_1106_cast_fp16, y = var_1110_cast_fp16)[name = tensor("mh_w_29_cast_fp16")]; + tensor obj_69_cast_fp16 = softmax(axis = var_952, x = mh_w_29_cast_fp16)[name = tensor("obj_69_cast_fp16")]; + tensor var_1114 = const()[name = tensor("op_1114"), val = tensor([1, 8, 64, 1500])]; + tensor var_1115_cast_fp16 = reshape(shape = var_1114, x = value_19_cast_fp16)[name = tensor("op_1115_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_1115_cast_fp16, y = obj_69_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_1118 = const()[name = tensor("op_1118"), val = tensor([1, 512, 1, 1])]; + tensor input_43_cast_fp16 = reshape(shape = var_1118, 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(90881344)))]; + 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(91405696)))]; + 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_1139_to_fp16 = const()[name = tensor("op_1139_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_1139_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(91406784)))]; + 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(91407872)))]; + 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(91408960)))]; + 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(93506176)))]; + 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(93510336)))]; + 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(95607552)))]; + 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_1175 = const()[name = tensor("op_1175"), val = tensor(3)]; + tensor out_31_axes_0 = const()[name = tensor("out_31_axes_0"), val = tensor([1])]; + tensor var_1200_to_fp16 = const()[name = tensor("op_1200_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1200_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(95608640)))]; + 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(95609728)))]; + 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(95610816)))]; + 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(96135168)))]; + 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_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_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(96136256)))]; + 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_5_self_attn_k_proj_weight_to_fp16, x = obj_71_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_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(96660608)))]; + 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(97184960)))]; + tensor current_value_cast_fp16 = conv(bias = layers_5_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_5_self_attn_v_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("current_value_cast_fp16")]; + tensor var_1239_cast_fp16 = mul(x = var_51_cast_fp16_5, y = var_135_cast_fp16)[name = tensor("op_1239_cast_fp16")]; + tensor var_1240_cast_fp16 = mul(x = current_key_cast_fp16, y = var_133_cast_fp16)[name = tensor("op_1240_cast_fp16")]; + tensor key_21_cast_fp16 = add(x = var_1239_cast_fp16, y = var_1240_cast_fp16)[name = tensor("key_21_cast_fp16")]; + tensor var_1243_cast_fp16 = mul(x = var_60_cast_fp16_5, y = var_135_cast_fp16)[name = tensor("op_1243_cast_fp16")]; + tensor var_1244_cast_fp16 = mul(x = current_value_cast_fp16, y = var_133_cast_fp16)[name = tensor("op_1244_cast_fp16")]; + tensor value_21_cast_fp16 = add(x = var_1243_cast_fp16, y = var_1244_cast_fp16)[name = tensor("value_21_cast_fp16")]; + tensor var_1248 = const()[name = tensor("op_1248"), val = tensor([1, 8, 64, 1])]; + tensor mh_q_21_cast_fp16 = reshape(shape = var_1248, x = query_21_cast_fp16)[name = tensor("mh_q_21_cast_fp16")]; + tensor var_1250_to_fp16 = const()[name = tensor("op_1250_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1251_cast_fp16 = mul(x = mh_q_21_cast_fp16, y = var_1250_to_fp16)[name = tensor("op_1251_cast_fp16")]; + tensor var_1254 = const()[name = tensor("op_1254"), val = tensor([1, 8, 64, 448])]; + tensor var_1255_cast_fp16 = reshape(shape = var_1254, x = key_21_cast_fp16)[name = tensor("op_1255_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_1251_cast_fp16, y = var_1255_cast_fp16)[name = tensor("mh_w_31_cast_fp16")]; + tensor mh_w_33_cast_fp16 = add(x = mh_w_31_cast_fp16, y = var_157_cast_fp16)[name = tensor("mh_w_33_cast_fp16")]; + tensor var_1263_cast_fp16 = softmax(axis = var_1175, x = mh_w_33_cast_fp16)[name = tensor("op_1263_cast_fp16")]; + tensor var_1264 = const()[name = tensor("op_1264"), val = tensor([1, 8, 64, 448])]; + tensor var_1265_cast_fp16 = reshape(shape = var_1264, x = value_21_cast_fp16)[name = tensor("op_1265_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_1265_cast_fp16, y = var_1263_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_1268 = const()[name = tensor("op_1268"), val = tensor([1, 512, 1, 1])]; + tensor input_51_cast_fp16 = reshape(shape = var_1268, 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(97186048)))]; + 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(97710400)))]; + 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_1290_to_fp16 = const()[name = tensor("op_1290_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1290_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(97711488)))]; + 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(97712576)))]; + 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_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_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(97713664)))]; + 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(98238016)))]; + tensor query_cast_fp16 = conv(bias = layers_5_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_5_encoder_attn_q_proj_weight_to_fp16, x = obj_79_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_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(98239104)))]; + 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_5_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_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(98763456)))]; + 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(99287808)))]; + tensor value_cast_fp16 = conv(bias = layers_5_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_5_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_cast_fp16")]; + tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([1, 8, 64, 1])]; + tensor mh_q_cast_fp16 = reshape(shape = var_1326, x = query_cast_fp16)[name = tensor("mh_q_cast_fp16")]; + tensor var_1328_to_fp16 = const()[name = tensor("op_1328_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1329_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_1328_to_fp16)[name = tensor("op_1329_cast_fp16")]; + tensor var_1332 = const()[name = tensor("op_1332"), val = tensor([1, 8, 64, 1500])]; + tensor var_1333_cast_fp16 = reshape(shape = var_1332, x = key_cast_fp16)[name = tensor("op_1333_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_1329_cast_fp16, y = var_1333_cast_fp16)[name = tensor("mh_w_cast_fp16")]; + tensor obj_83_cast_fp16 = softmax(axis = var_1175, x = mh_w_cast_fp16)[name = tensor("obj_83_cast_fp16")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([1, 8, 64, 1500])]; + tensor var_1338_cast_fp16 = reshape(shape = var_1337, x = value_cast_fp16)[name = tensor("op_1338_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_1338_cast_fp16, y = obj_83_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_1341 = const()[name = tensor("op_1341"), val = tensor([1, 512, 1, 1])]; + tensor input_53_cast_fp16 = reshape(shape = var_1341, x = attn_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(99288896)))]; + 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(99813248)))]; + 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_1362_to_fp16 = const()[name = tensor("op_1362_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1362_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(99814336)))]; + 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(99815424)))]; + 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(99816512)))]; + 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(101913728)))]; + 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_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_57_cast_fp16)[name = tensor("input_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(101917888)))]; + 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(104015104)))]; + 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_cast_fp16)[name = tensor("hidden_states_13_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor out_axes_0 = const()[name = tensor("out_axes_0"), val = tensor([1])]; + tensor var_1405_to_fp16 = const()[name = tensor("op_1405_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_1405_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(104016192)))]; + 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(104017280)))]; + 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_1416_axes_0 = const()[name = tensor("op_1416_axes_0"), val = tensor([2])]; + tensor var_1416_cast_fp16 = squeeze(axes = var_1416_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_1416_cast_fp16")]; + tensor var_1419_perm_0 = const()[name = tensor("op_1419_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(104018368)))]; + tensor var_1419_cast_fp16 = transpose(perm = var_1419_perm_0, x = var_1416_cast_fp16)[name = tensor("transpose_0")]; + tensor logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_1419_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_1423 = const()[name = tensor("op_1423"), val = tensor(1)]; + tensor obj_87_interleave_0 = const()[name = tensor("obj_87_interleave_0"), val = tensor(false)]; + tensor key_cache_updates = concat(axis = var_1423, interleave = obj_87_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_cast_fp16))[name = tensor("obj_87_cast_fp16")]; + tensor var_1426 = const()[name = tensor("op_1426"), val = tensor(1)]; + tensor obj_89_interleave_0 = const()[name = tensor("obj_89_interleave_0"), val = tensor(false)]; + tensor value_cache_updates = concat(axis = var_1426, interleave = obj_89_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_cast_fp16))[name = tensor("obj_89_cast_fp16")]; + tensor var_1437_begin_0 = const()[name = tensor("op_1437_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_1437_end_0 = const()[name = tensor("op_1437_end_0"), val = tensor([1, 2, 1, 1500])]; + tensor var_1437_end_mask_0 = const()[name = tensor("op_1437_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1437_cast_fp16 = slice_by_index(begin = var_1437_begin_0, end = var_1437_end_0, end_mask = var_1437_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1437_cast_fp16")]; + tensor var_1440_begin_0 = const()[name = tensor("op_1440_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1440_end_0 = const()[name = tensor("op_1440_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1440_end_mask_0 = const()[name = tensor("op_1440_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1440_squeeze_mask_0 = const()[name = tensor("op_1440_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1440_cast_fp16 = slice_by_index(begin = var_1440_begin_0, end = var_1440_end_0, end_mask = var_1440_end_mask_0, squeeze_mask = var_1440_squeeze_mask_0, x = var_1437_cast_fp16)[name = tensor("op_1440_cast_fp16")]; + tensor var_1455_begin_0 = const()[name = tensor("op_1455_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_1455_end_0 = const()[name = tensor("op_1455_end_0"), val = tensor([1, 3, 1, 1500])]; + tensor var_1455_end_mask_0 = const()[name = tensor("op_1455_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1455_cast_fp16 = slice_by_index(begin = var_1455_begin_0, end = var_1455_end_0, end_mask = var_1455_end_mask_0, x = obj_69_cast_fp16)[name = tensor("op_1455_cast_fp16")]; + tensor var_1458_begin_0 = const()[name = tensor("op_1458_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1458_end_0 = const()[name = tensor("op_1458_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1458_end_mask_0 = const()[name = tensor("op_1458_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1458_squeeze_mask_0 = const()[name = tensor("op_1458_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1458_cast_fp16 = slice_by_index(begin = var_1458_begin_0, end = var_1458_end_0, end_mask = var_1458_end_mask_0, squeeze_mask = var_1458_squeeze_mask_0, x = var_1455_cast_fp16)[name = tensor("op_1458_cast_fp16")]; + tensor var_1473_begin_0 = const()[name = tensor("op_1473_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_1473_end_0 = const()[name = tensor("op_1473_end_0"), val = tensor([1, 4, 1, 1500])]; + tensor var_1473_end_mask_0 = const()[name = tensor("op_1473_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1473_cast_fp16 = slice_by_index(begin = var_1473_begin_0, end = var_1473_end_0, end_mask = var_1473_end_mask_0, x = obj_69_cast_fp16)[name = tensor("op_1473_cast_fp16")]; + tensor var_1476_begin_0 = const()[name = tensor("op_1476_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1476_end_0 = const()[name = tensor("op_1476_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1476_end_mask_0 = const()[name = tensor("op_1476_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1476_squeeze_mask_0 = const()[name = tensor("op_1476_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1476_cast_fp16 = slice_by_index(begin = var_1476_begin_0, end = var_1476_end_0, end_mask = var_1476_end_mask_0, squeeze_mask = var_1476_squeeze_mask_0, x = var_1473_cast_fp16)[name = tensor("op_1476_cast_fp16")]; + tensor var_1491_begin_0 = const()[name = tensor("op_1491_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_1491_end_0 = const()[name = tensor("op_1491_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1491_end_mask_0 = const()[name = tensor("op_1491_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1491_cast_fp16 = slice_by_index(begin = var_1491_begin_0, end = var_1491_end_0, end_mask = var_1491_end_mask_0, x = obj_69_cast_fp16)[name = tensor("op_1491_cast_fp16")]; + tensor var_1494_begin_0 = const()[name = tensor("op_1494_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1494_end_0 = const()[name = tensor("op_1494_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1494_end_mask_0 = const()[name = tensor("op_1494_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1494_squeeze_mask_0 = const()[name = tensor("op_1494_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1494_cast_fp16 = slice_by_index(begin = var_1494_begin_0, end = var_1494_end_0, end_mask = var_1494_end_mask_0, squeeze_mask = var_1494_squeeze_mask_0, x = var_1491_cast_fp16)[name = tensor("op_1494_cast_fp16")]; + tensor var_1509_begin_0 = const()[name = tensor("op_1509_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_1509_end_0 = const()[name = tensor("op_1509_end_0"), val = tensor([1, 2, 1, 1500])]; + tensor var_1509_end_mask_0 = const()[name = tensor("op_1509_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1509_cast_fp16 = slice_by_index(begin = var_1509_begin_0, end = var_1509_end_0, end_mask = var_1509_end_mask_0, x = obj_83_cast_fp16)[name = tensor("op_1509_cast_fp16")]; + tensor var_1512_begin_0 = const()[name = tensor("op_1512_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1512_end_0 = const()[name = tensor("op_1512_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1512_end_mask_0 = const()[name = tensor("op_1512_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1512_squeeze_mask_0 = const()[name = tensor("op_1512_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1512_cast_fp16 = slice_by_index(begin = var_1512_begin_0, end = var_1512_end_0, end_mask = var_1512_end_mask_0, squeeze_mask = var_1512_squeeze_mask_0, x = var_1509_cast_fp16)[name = tensor("op_1512_cast_fp16")]; + tensor var_1527_begin_0 = const()[name = tensor("op_1527_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_1527_end_0 = const()[name = tensor("op_1527_end_0"), val = tensor([1, 3, 1, 1500])]; + tensor var_1527_end_mask_0 = const()[name = tensor("op_1527_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1527_cast_fp16 = slice_by_index(begin = var_1527_begin_0, end = var_1527_end_0, end_mask = var_1527_end_mask_0, x = obj_83_cast_fp16)[name = tensor("op_1527_cast_fp16")]; + tensor var_1530_begin_0 = const()[name = tensor("op_1530_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1530_end_0 = const()[name = tensor("op_1530_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1530_end_mask_0 = const()[name = tensor("op_1530_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1530_squeeze_mask_0 = const()[name = tensor("op_1530_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1530_cast_fp16 = slice_by_index(begin = var_1530_begin_0, end = var_1530_end_0, end_mask = var_1530_end_mask_0, squeeze_mask = var_1530_squeeze_mask_0, x = var_1527_cast_fp16)[name = tensor("op_1530_cast_fp16")]; + tensor var_1545_begin_0 = const()[name = tensor("op_1545_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_1545_end_0 = const()[name = tensor("op_1545_end_0"), val = tensor([1, 5, 1, 1500])]; + tensor var_1545_end_mask_0 = const()[name = tensor("op_1545_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1545_cast_fp16 = slice_by_index(begin = var_1545_begin_0, end = var_1545_end_0, end_mask = var_1545_end_mask_0, x = obj_83_cast_fp16)[name = tensor("op_1545_cast_fp16")]; + tensor var_1548_begin_0 = const()[name = tensor("op_1548_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1548_end_0 = const()[name = tensor("op_1548_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1548_end_mask_0 = const()[name = tensor("op_1548_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1548_squeeze_mask_0 = const()[name = tensor("op_1548_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1548_cast_fp16 = slice_by_index(begin = var_1548_begin_0, end = var_1548_end_0, end_mask = var_1548_end_mask_0, squeeze_mask = var_1548_squeeze_mask_0, x = var_1545_cast_fp16)[name = tensor("op_1548_cast_fp16")]; + tensor var_1563_begin_0 = const()[name = tensor("op_1563_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_1563_end_0 = const()[name = tensor("op_1563_end_0"), val = tensor([1, 7, 1, 1500])]; + tensor var_1563_end_mask_0 = const()[name = tensor("op_1563_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1563_cast_fp16 = slice_by_index(begin = var_1563_begin_0, end = var_1563_end_0, end_mask = var_1563_end_mask_0, x = obj_83_cast_fp16)[name = tensor("op_1563_cast_fp16")]; + tensor var_1566_begin_0 = const()[name = tensor("op_1566_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1566_end_0 = const()[name = tensor("op_1566_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1566_end_mask_0 = const()[name = tensor("op_1566_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1566_squeeze_mask_0 = const()[name = tensor("op_1566_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1566_cast_fp16 = slice_by_index(begin = var_1566_begin_0, end = var_1566_end_0, end_mask = var_1566_end_mask_0, squeeze_mask = var_1566_squeeze_mask_0, x = var_1563_cast_fp16)[name = tensor("op_1566_cast_fp16")]; + tensor var_1573 = const()[name = tensor("op_1573"), val = tensor(1)]; + tensor var_1574_interleave_0 = const()[name = tensor("op_1574_interleave_0"), val = tensor(false)]; + tensor var_1574_cast_fp16 = concat(axis = var_1573, interleave = var_1574_interleave_0, values = (var_1440_cast_fp16, var_1458_cast_fp16, var_1476_cast_fp16, var_1494_cast_fp16, var_1512_cast_fp16, var_1530_cast_fp16, var_1548_cast_fp16, var_1566_cast_fp16))[name = tensor("op_1574_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_1574_cast_fp16)[name = tensor("obj_cast_fp16")]; + } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); +} \ No newline at end of file