diff --git "a/unsloth_whisper-large-v3/TextDecoder.mlmodelc/model.mil" "b/unsloth_whisper-large-v3/TextDecoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/unsloth_whisper-large-v3/TextDecoder.mlmodelc/model.mil" @@ -0,0 +1,4824 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"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_80_axis_0 = const()[name = tensor("op_80_axis_0"), val = tensor(0)]; + tensor var_80_batch_dims_0 = const()[name = tensor("op_80_batch_dims_0"), val = tensor(0)]; + tensor embed_tokens_weight_to_fp16 = const()[name = tensor("embed_tokens_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor var_80_cast_fp16 = gather(axis = var_80_axis_0, batch_dims = var_80_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor("op_80_cast_fp16")]; + tensor var_84_axis_0 = const()[name = tensor("op_84_axis_0"), val = tensor(0)]; + tensor var_84_batch_dims_0 = const()[name = tensor("op_84_batch_dims_0"), val = tensor(0)]; + tensor embed_positions_weight_to_fp16 = const()[name = tensor("embed_positions_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132777088)))]; + tensor var_84_cast_fp16 = gather(axis = var_84_axis_0, batch_dims = var_84_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor("op_84_cast_fp16")]; + tensor hidden_states_1_cast_fp16 = add(x = var_80_cast_fp16, y = var_84_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; + tensor var_98_axes_0 = const()[name = tensor("op_98_axes_0"), val = tensor([2])]; + tensor var_98_cast_fp16 = expand_dims(axes = var_98_axes_0, x = hidden_states_1_cast_fp16)[name = tensor("op_98_cast_fp16")]; + tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([3])]; + tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_98_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280])]; + tensor var_103_axis_0 = const()[name = tensor("op_103_axis_0"), val = tensor(1)]; + tensor var_103_cast_fp16_0, tensor var_103_cast_fp16_1, tensor var_103_cast_fp16_2, tensor var_103_cast_fp16_3, tensor var_103_cast_fp16_4, tensor var_103_cast_fp16_5, tensor var_103_cast_fp16_6, tensor var_103_cast_fp16_7, tensor var_103_cast_fp16_8, tensor var_103_cast_fp16_9, tensor var_103_cast_fp16_10, tensor var_103_cast_fp16_11, tensor var_103_cast_fp16_12, tensor var_103_cast_fp16_13, tensor var_103_cast_fp16_14, tensor var_103_cast_fp16_15, tensor var_103_cast_fp16_16, tensor var_103_cast_fp16_17, tensor var_103_cast_fp16_18, tensor var_103_cast_fp16_19, tensor var_103_cast_fp16_20, tensor var_103_cast_fp16_21, tensor var_103_cast_fp16_22, tensor var_103_cast_fp16_23, tensor var_103_cast_fp16_24, tensor var_103_cast_fp16_25, tensor var_103_cast_fp16_26, tensor var_103_cast_fp16_27, tensor var_103_cast_fp16_28, tensor var_103_cast_fp16_29, tensor var_103_cast_fp16_30, tensor var_103_cast_fp16_31 = split(axis = var_103_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor("op_103_cast_fp16")]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280])]; + tensor var_138_axis_0 = const()[name = tensor("op_138_axis_0"), val = tensor(1)]; + tensor var_138_cast_fp16_0, tensor var_138_cast_fp16_1, tensor var_138_cast_fp16_2, tensor var_138_cast_fp16_3, tensor var_138_cast_fp16_4, tensor var_138_cast_fp16_5, tensor var_138_cast_fp16_6, tensor var_138_cast_fp16_7, tensor var_138_cast_fp16_8, tensor var_138_cast_fp16_9, tensor var_138_cast_fp16_10, tensor var_138_cast_fp16_11, tensor var_138_cast_fp16_12, tensor var_138_cast_fp16_13, tensor var_138_cast_fp16_14, tensor var_138_cast_fp16_15, tensor var_138_cast_fp16_16, tensor var_138_cast_fp16_17, tensor var_138_cast_fp16_18, tensor var_138_cast_fp16_19, tensor var_138_cast_fp16_20, tensor var_138_cast_fp16_21, tensor var_138_cast_fp16_22, tensor var_138_cast_fp16_23, tensor var_138_cast_fp16_24, tensor var_138_cast_fp16_25, tensor var_138_cast_fp16_26, tensor var_138_cast_fp16_27, tensor var_138_cast_fp16_28, tensor var_138_cast_fp16_29, tensor var_138_cast_fp16_30, tensor var_138_cast_fp16_31 = split(axis = var_138_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor("op_138_cast_fp16")]; + tensor var_176 = const()[name = tensor("op_176"), val = tensor(3)]; + tensor out_1_axes_0 = const()[name = tensor("out_1_axes_0"), val = tensor([1])]; + tensor var_201_to_fp16 = const()[name = tensor("op_201_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_201_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(133924032)))]; + tensor obj_1_variance_0_to_fp16 = const()[name = tensor("obj_1_variance_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133926656)))]; + tensor obj_1_gamma_0_to_fp16 = const()[name = tensor("obj_1_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133929280)))]; + tensor obj_1_beta_0_to_fp16 = const()[name = tensor("obj_1_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133931904)))]; + tensor obj_1_epsilon_0_to_fp16 = const()[name = tensor("obj_1_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor("obj_1_cast_fp16")]; + tensor 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(133934528)))]; + tensor layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137211392)))]; + tensor query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = 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(137214016)))]; + 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(140490880)))]; + tensor layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143767744)))]; + tensor current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = 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_236_axes_0 = const()[name = tensor("op_236_axes_0"), val = tensor([1])]; + tensor var_236_cast_fp16 = expand_dims(axes = var_236_axes_0, x = kv_cache_update_mask)[name = tensor("op_236_cast_fp16")]; + tensor var_237_axes_0 = const()[name = tensor("op_237_axes_0"), val = tensor([2])]; + tensor var_237_cast_fp16 = expand_dims(axes = var_237_axes_0, x = var_236_cast_fp16)[name = tensor("op_237_cast_fp16")]; + tensor var_177_to_fp16 = const()[name = tensor("op_177_to_fp16"), val = tensor(0x1p+0)]; + tensor var_239_cast_fp16 = sub(x = var_177_to_fp16, y = var_237_cast_fp16)[name = tensor("op_239_cast_fp16")]; + tensor var_240_cast_fp16 = mul(x = var_103_cast_fp16_0, y = var_239_cast_fp16)[name = tensor("op_240_cast_fp16")]; + tensor var_241_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_241_cast_fp16")]; + tensor key_1_cast_fp16 = add(x = var_240_cast_fp16, y = var_241_cast_fp16)[name = tensor("key_1_cast_fp16")]; + tensor var_244_cast_fp16 = mul(x = var_138_cast_fp16_0, y = var_239_cast_fp16)[name = tensor("op_244_cast_fp16")]; + tensor var_245_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_245_cast_fp16")]; + tensor value_1_cast_fp16 = add(x = var_244_cast_fp16, y = var_245_cast_fp16)[name = tensor("value_1_cast_fp16")]; + tensor var_249 = const()[name = tensor("op_249"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_1_cast_fp16 = reshape(shape = var_249, x = query_1_cast_fp16)[name = tensor("mh_q_1_cast_fp16")]; + tensor var_251_to_fp16 = const()[name = tensor("op_251_to_fp16"), val = tensor(0x1p-3)]; + tensor var_252_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_251_to_fp16)[name = tensor("op_252_cast_fp16")]; + tensor var_255 = const()[name = tensor("op_255"), val = tensor([1, 20, 64, 448])]; + tensor var_256_cast_fp16 = reshape(shape = var_255, x = key_1_cast_fp16)[name = tensor("op_256_cast_fp16")]; + tensor mh_w_1_transpose_x_0 = const()[name = tensor("mh_w_1_transpose_x_0"), val = tensor(true)]; + tensor mh_w_1_transpose_y_0 = const()[name = tensor("mh_w_1_transpose_y_0"), val = tensor(false)]; + tensor mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_252_cast_fp16, y = var_256_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; + tensor var_260_axes_0 = const()[name = tensor("op_260_axes_0"), val = tensor([1])]; + tensor var_260_cast_fp16 = expand_dims(axes = var_260_axes_0, x = decoder_key_padding_mask)[name = tensor("op_260_cast_fp16")]; + tensor var_261_axes_0 = const()[name = tensor("op_261_axes_0"), val = tensor([2])]; + tensor var_261_cast_fp16 = expand_dims(axes = var_261_axes_0, x = var_260_cast_fp16)[name = tensor("op_261_cast_fp16")]; + tensor mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; + tensor var_264_cast_fp16 = softmax(axis = var_176, x = mh_w_3_cast_fp16)[name = tensor("op_264_cast_fp16")]; + tensor var_265 = const()[name = tensor("op_265"), val = tensor([1, 20, 64, 448])]; + tensor var_266_cast_fp16 = reshape(shape = var_265, x = value_1_cast_fp16)[name = tensor("op_266_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_266_cast_fp16, y = var_264_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_269 = const()[name = tensor("op_269"), val = tensor([1, 1280, 1, 1])]; + tensor input_1_cast_fp16 = reshape(shape = var_269, 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(143770368)))]; + tensor layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147047232)))]; + tensor obj_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = 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_291_to_fp16 = const()[name = tensor("op_291_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_291_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(147049856)))]; + tensor obj_9_beta_0_to_fp16 = const()[name = tensor("obj_9_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147052480)))]; + tensor obj_9_epsilon_0_to_fp16 = const()[name = tensor("obj_9_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor("obj_9_cast_fp16")]; + tensor 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(147055104)))]; + tensor layers_0_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150331968)))]; + tensor query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = 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(150334592)))]; + 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(153611456)))]; + tensor layers_0_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156888320)))]; + tensor value_3_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_bias_to_fp16, dilations = 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_327 = const()[name = tensor("op_327"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_3_cast_fp16 = reshape(shape = var_327, x = query_3_cast_fp16)[name = tensor("mh_q_3_cast_fp16")]; + tensor var_329_to_fp16 = const()[name = tensor("op_329_to_fp16"), val = tensor(0x1p-3)]; + tensor var_330_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_329_to_fp16)[name = tensor("op_330_cast_fp16")]; + tensor var_333 = const()[name = tensor("op_333"), val = tensor([1, 20, 64, 1500])]; + tensor var_334_cast_fp16 = reshape(shape = var_333, x = key_3_cast_fp16)[name = tensor("op_334_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_330_cast_fp16, y = var_334_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; + tensor obj_13_cast_fp16 = softmax(axis = var_176, x = mh_w_5_cast_fp16)[name = tensor("obj_13_cast_fp16")]; + tensor var_338 = const()[name = tensor("op_338"), val = tensor([1, 20, 64, 1500])]; + tensor var_339_cast_fp16 = reshape(shape = var_338, x = value_3_cast_fp16)[name = tensor("op_339_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_339_cast_fp16, y = obj_13_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_342 = const()[name = tensor("op_342"), val = tensor([1, 1280, 1, 1])]; + tensor input_3_cast_fp16 = reshape(shape = var_342, 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(156890944)))]; + tensor layers_0_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160167808)))]; + tensor obj_11_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = 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_360_to_fp16 = const()[name = tensor("op_360_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_360_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(160170432)))]; + tensor input_5_beta_0_to_fp16 = const()[name = tensor("input_5_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160173056)))]; + tensor input_5_epsilon_0_to_fp16 = const()[name = tensor("input_5_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = batch_norm(beta = input_5_beta_0_to_fp16, epsilon = input_5_epsilon_0_to_fp16, gamma = input_5_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor 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(160175680)))]; + tensor layers_0_fc1_bias_to_fp16 = const()[name = tensor("layers_0_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173282944)))]; + tensor input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = 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(173293248)))]; + tensor layers_0_fc2_bias_to_fp16 = const()[name = tensor("layers_0_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186400512)))]; + tensor hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = 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_395 = const()[name = tensor("op_395"), val = tensor(3)]; + tensor out_7_axes_0 = const()[name = tensor("out_7_axes_0"), val = tensor([1])]; + tensor var_420_to_fp16 = const()[name = tensor("op_420_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_420_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(186403136)))]; + tensor obj_15_beta_0_to_fp16 = const()[name = tensor("obj_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186405760)))]; + tensor obj_15_epsilon_0_to_fp16 = const()[name = tensor("obj_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor("obj_15_cast_fp16")]; + tensor 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(186408384)))]; + tensor layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189685248)))]; + tensor query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = 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(189687872)))]; + 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(192964736)))]; + tensor layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196241600)))]; + tensor current_value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = 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_459_cast_fp16 = mul(x = var_103_cast_fp16_1, y = var_239_cast_fp16)[name = tensor("op_459_cast_fp16")]; + tensor var_460_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_460_cast_fp16")]; + tensor key_5_cast_fp16 = add(x = var_459_cast_fp16, y = var_460_cast_fp16)[name = tensor("key_5_cast_fp16")]; + tensor var_463_cast_fp16 = mul(x = var_138_cast_fp16_1, y = var_239_cast_fp16)[name = tensor("op_463_cast_fp16")]; + tensor var_464_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_464_cast_fp16")]; + tensor value_5_cast_fp16 = add(x = var_463_cast_fp16, y = var_464_cast_fp16)[name = tensor("value_5_cast_fp16")]; + tensor var_468 = const()[name = tensor("op_468"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_5_cast_fp16 = reshape(shape = var_468, x = query_5_cast_fp16)[name = tensor("mh_q_5_cast_fp16")]; + tensor var_470_to_fp16 = const()[name = tensor("op_470_to_fp16"), val = tensor(0x1p-3)]; + tensor var_471_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_470_to_fp16)[name = tensor("op_471_cast_fp16")]; + tensor var_474 = const()[name = tensor("op_474"), val = tensor([1, 20, 64, 448])]; + tensor var_475_cast_fp16 = reshape(shape = var_474, x = key_5_cast_fp16)[name = tensor("op_475_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_471_cast_fp16, y = var_475_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; + tensor mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; + tensor var_483_cast_fp16 = softmax(axis = var_395, x = mh_w_9_cast_fp16)[name = tensor("op_483_cast_fp16")]; + tensor var_484 = const()[name = tensor("op_484"), val = tensor([1, 20, 64, 448])]; + tensor var_485_cast_fp16 = reshape(shape = var_484, x = value_5_cast_fp16)[name = tensor("op_485_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_485_cast_fp16, y = var_483_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_488 = const()[name = tensor("op_488"), val = tensor([1, 1280, 1, 1])]; + tensor input_11_cast_fp16 = reshape(shape = var_488, 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(196244224)))]; + tensor layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199521088)))]; + tensor obj_21_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = 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_510_to_fp16 = const()[name = tensor("op_510_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_510_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(199523712)))]; + tensor obj_23_beta_0_to_fp16 = const()[name = tensor("obj_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199526336)))]; + tensor obj_23_epsilon_0_to_fp16 = const()[name = tensor("obj_23_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor("obj_23_cast_fp16")]; + tensor 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(199528960)))]; + tensor layers_1_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202805824)))]; + tensor query_7_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = 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(202808448)))]; + 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(206085312)))]; + tensor layers_1_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209362176)))]; + tensor value_7_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_bias_to_fp16, dilations = 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_546 = const()[name = tensor("op_546"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_7_cast_fp16 = reshape(shape = var_546, x = query_7_cast_fp16)[name = tensor("mh_q_7_cast_fp16")]; + tensor var_548_to_fp16 = const()[name = tensor("op_548_to_fp16"), val = tensor(0x1p-3)]; + tensor var_549_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_548_to_fp16)[name = tensor("op_549_cast_fp16")]; + tensor var_552 = const()[name = tensor("op_552"), val = tensor([1, 20, 64, 1500])]; + tensor var_553_cast_fp16 = reshape(shape = var_552, x = key_7_cast_fp16)[name = tensor("op_553_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_549_cast_fp16, y = var_553_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; + tensor obj_27_cast_fp16 = softmax(axis = var_395, x = mh_w_11_cast_fp16)[name = tensor("obj_27_cast_fp16")]; + tensor var_557 = const()[name = tensor("op_557"), val = tensor([1, 20, 64, 1500])]; + tensor var_558_cast_fp16 = reshape(shape = var_557, x = value_7_cast_fp16)[name = tensor("op_558_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_558_cast_fp16, y = obj_27_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_561 = const()[name = tensor("op_561"), val = tensor([1, 1280, 1, 1])]; + tensor input_13_cast_fp16 = reshape(shape = var_561, 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(209364800)))]; + tensor layers_1_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212641664)))]; + tensor obj_25_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = 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_579_to_fp16 = const()[name = tensor("op_579_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_579_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(212644288)))]; + tensor input_15_beta_0_to_fp16 = const()[name = tensor("input_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212646912)))]; + tensor input_15_epsilon_0_to_fp16 = const()[name = tensor("input_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = batch_norm(beta = input_15_beta_0_to_fp16, epsilon = input_15_epsilon_0_to_fp16, gamma = input_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor 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(212649536)))]; + tensor layers_1_fc1_bias_to_fp16 = const()[name = tensor("layers_1_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(225756800)))]; + tensor input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = 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(225767104)))]; + tensor layers_1_fc2_bias_to_fp16 = const()[name = tensor("layers_1_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238874368)))]; + tensor hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = 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_614 = const()[name = tensor("op_614"), val = tensor(3)]; + tensor out_13_axes_0 = const()[name = tensor("out_13_axes_0"), val = tensor([1])]; + tensor var_639_to_fp16 = const()[name = tensor("op_639_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_639_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(238876992)))]; + tensor obj_29_beta_0_to_fp16 = const()[name = tensor("obj_29_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238879616)))]; + tensor obj_29_epsilon_0_to_fp16 = const()[name = tensor("obj_29_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor("obj_29_cast_fp16")]; + tensor 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(238882240)))]; + tensor layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242159104)))]; + tensor query_9_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = 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(242161728)))]; + 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(245438592)))]; + tensor layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248715456)))]; + tensor current_value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = 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_678_cast_fp16 = mul(x = var_103_cast_fp16_2, y = var_239_cast_fp16)[name = tensor("op_678_cast_fp16")]; + tensor var_679_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_679_cast_fp16")]; + tensor key_9_cast_fp16 = add(x = var_678_cast_fp16, y = var_679_cast_fp16)[name = tensor("key_9_cast_fp16")]; + tensor var_682_cast_fp16 = mul(x = var_138_cast_fp16_2, y = var_239_cast_fp16)[name = tensor("op_682_cast_fp16")]; + tensor var_683_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_683_cast_fp16")]; + tensor value_9_cast_fp16 = add(x = var_682_cast_fp16, y = var_683_cast_fp16)[name = tensor("value_9_cast_fp16")]; + tensor var_687 = const()[name = tensor("op_687"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_9_cast_fp16 = reshape(shape = var_687, x = query_9_cast_fp16)[name = tensor("mh_q_9_cast_fp16")]; + tensor var_689_to_fp16 = const()[name = tensor("op_689_to_fp16"), val = tensor(0x1p-3)]; + tensor var_690_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_689_to_fp16)[name = tensor("op_690_cast_fp16")]; + tensor var_693 = const()[name = tensor("op_693"), val = tensor([1, 20, 64, 448])]; + tensor var_694_cast_fp16 = reshape(shape = var_693, x = key_9_cast_fp16)[name = tensor("op_694_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_690_cast_fp16, y = var_694_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; + tensor mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; + tensor var_702_cast_fp16 = softmax(axis = var_614, x = mh_w_15_cast_fp16)[name = tensor("op_702_cast_fp16")]; + tensor var_703 = const()[name = tensor("op_703"), val = tensor([1, 20, 64, 448])]; + tensor var_704_cast_fp16 = reshape(shape = var_703, x = value_9_cast_fp16)[name = tensor("op_704_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_704_cast_fp16, y = var_702_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_707 = const()[name = tensor("op_707"), val = tensor([1, 1280, 1, 1])]; + tensor input_21_cast_fp16 = reshape(shape = var_707, 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(248718080)))]; + tensor layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251994944)))]; + tensor obj_35_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = 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_729_to_fp16 = const()[name = tensor("op_729_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_729_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(251997568)))]; + tensor obj_37_beta_0_to_fp16 = const()[name = tensor("obj_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252000192)))]; + tensor obj_37_epsilon_0_to_fp16 = const()[name = tensor("obj_37_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor("obj_37_cast_fp16")]; + tensor 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(252002816)))]; + tensor layers_2_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255279680)))]; + tensor query_11_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_bias_to_fp16, dilations = 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(255282304)))]; + 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(258559168)))]; + tensor layers_2_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261836032)))]; + tensor value_11_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_bias_to_fp16, dilations = 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_765 = const()[name = tensor("op_765"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_11_cast_fp16 = reshape(shape = var_765, x = query_11_cast_fp16)[name = tensor("mh_q_11_cast_fp16")]; + tensor var_767_to_fp16 = const()[name = tensor("op_767_to_fp16"), val = tensor(0x1p-3)]; + tensor var_768_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_767_to_fp16)[name = tensor("op_768_cast_fp16")]; + tensor var_771 = const()[name = tensor("op_771"), val = tensor([1, 20, 64, 1500])]; + tensor var_772_cast_fp16 = reshape(shape = var_771, x = key_11_cast_fp16)[name = tensor("op_772_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_768_cast_fp16, y = var_772_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; + tensor obj_41_cast_fp16 = softmax(axis = var_614, x = mh_w_17_cast_fp16)[name = tensor("obj_41_cast_fp16")]; + tensor var_776 = const()[name = tensor("op_776"), val = tensor([1, 20, 64, 1500])]; + tensor var_777_cast_fp16 = reshape(shape = var_776, x = value_11_cast_fp16)[name = tensor("op_777_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_777_cast_fp16, y = obj_41_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_780 = const()[name = tensor("op_780"), val = tensor([1, 1280, 1, 1])]; + tensor input_23_cast_fp16 = reshape(shape = var_780, 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(261838656)))]; + tensor layers_2_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265115520)))]; + tensor obj_39_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = 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_798_to_fp16 = const()[name = tensor("op_798_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_798_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(265118144)))]; + tensor input_25_beta_0_to_fp16 = const()[name = tensor("input_25_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265120768)))]; + tensor input_25_epsilon_0_to_fp16 = const()[name = tensor("input_25_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_25_cast_fp16 = batch_norm(beta = input_25_beta_0_to_fp16, epsilon = input_25_epsilon_0_to_fp16, gamma = input_25_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor 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(265123392)))]; + tensor layers_2_fc1_bias_to_fp16 = const()[name = tensor("layers_2_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278230656)))]; + tensor input_27_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = 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(278240960)))]; + tensor layers_2_fc2_bias_to_fp16 = const()[name = tensor("layers_2_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291348224)))]; + tensor hidden_states_7_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = 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_833 = const()[name = tensor("op_833"), val = tensor(3)]; + tensor out_19_axes_0 = const()[name = tensor("out_19_axes_0"), val = tensor([1])]; + tensor var_858_to_fp16 = const()[name = tensor("op_858_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_858_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(291350848)))]; + tensor obj_43_beta_0_to_fp16 = const()[name = tensor("obj_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291353472)))]; + tensor obj_43_epsilon_0_to_fp16 = const()[name = tensor("obj_43_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor("obj_43_cast_fp16")]; + tensor 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(291356096)))]; + tensor layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294632960)))]; + tensor query_13_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = 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(294635584)))]; + 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(297912448)))]; + tensor layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(301189312)))]; + tensor current_value_7_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = 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_897_cast_fp16 = mul(x = var_103_cast_fp16_3, y = var_239_cast_fp16)[name = tensor("op_897_cast_fp16")]; + tensor var_898_cast_fp16 = mul(x = current_key_7_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_898_cast_fp16")]; + tensor key_13_cast_fp16 = add(x = var_897_cast_fp16, y = var_898_cast_fp16)[name = tensor("key_13_cast_fp16")]; + tensor var_901_cast_fp16 = mul(x = var_138_cast_fp16_3, y = var_239_cast_fp16)[name = tensor("op_901_cast_fp16")]; + tensor var_902_cast_fp16 = mul(x = current_value_7_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_902_cast_fp16")]; + tensor value_13_cast_fp16 = add(x = var_901_cast_fp16, y = var_902_cast_fp16)[name = tensor("value_13_cast_fp16")]; + tensor var_906 = const()[name = tensor("op_906"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_13_cast_fp16 = reshape(shape = var_906, x = query_13_cast_fp16)[name = tensor("mh_q_13_cast_fp16")]; + tensor var_908_to_fp16 = const()[name = tensor("op_908_to_fp16"), val = tensor(0x1p-3)]; + tensor var_909_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_908_to_fp16)[name = tensor("op_909_cast_fp16")]; + tensor var_912 = const()[name = tensor("op_912"), val = tensor([1, 20, 64, 448])]; + tensor var_913_cast_fp16 = reshape(shape = var_912, x = key_13_cast_fp16)[name = tensor("op_913_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_909_cast_fp16, y = var_913_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; + tensor mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; + tensor var_921_cast_fp16 = softmax(axis = var_833, x = mh_w_21_cast_fp16)[name = tensor("op_921_cast_fp16")]; + tensor var_922 = const()[name = tensor("op_922"), val = tensor([1, 20, 64, 448])]; + tensor var_923_cast_fp16 = reshape(shape = var_922, x = value_13_cast_fp16)[name = tensor("op_923_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_923_cast_fp16, y = var_921_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_926 = const()[name = tensor("op_926"), val = tensor([1, 1280, 1, 1])]; + tensor input_31_cast_fp16 = reshape(shape = var_926, 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(301191936)))]; + tensor layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(304468800)))]; + tensor obj_49_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = 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_948_to_fp16 = const()[name = tensor("op_948_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_948_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(304471424)))]; + tensor obj_51_beta_0_to_fp16 = const()[name = tensor("obj_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(304474048)))]; + tensor obj_51_epsilon_0_to_fp16 = const()[name = tensor("obj_51_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor("obj_51_cast_fp16")]; + tensor 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(304476672)))]; + tensor layers_3_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307753536)))]; + tensor query_15_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_bias_to_fp16, dilations = 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(307756160)))]; + 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(311033024)))]; + tensor layers_3_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314309888)))]; + tensor value_15_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_bias_to_fp16, dilations = 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_984 = const()[name = tensor("op_984"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_15_cast_fp16 = reshape(shape = var_984, x = query_15_cast_fp16)[name = tensor("mh_q_15_cast_fp16")]; + tensor var_986_to_fp16 = const()[name = tensor("op_986_to_fp16"), val = tensor(0x1p-3)]; + tensor var_987_cast_fp16 = mul(x = mh_q_15_cast_fp16, y = var_986_to_fp16)[name = tensor("op_987_cast_fp16")]; + tensor var_990 = const()[name = tensor("op_990"), val = tensor([1, 20, 64, 1500])]; + tensor var_991_cast_fp16 = reshape(shape = var_990, x = key_15_cast_fp16)[name = tensor("op_991_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_987_cast_fp16, y = var_991_cast_fp16)[name = tensor("mh_w_23_cast_fp16")]; + tensor obj_55_cast_fp16 = softmax(axis = var_833, x = mh_w_23_cast_fp16)[name = tensor("obj_55_cast_fp16")]; + tensor var_995 = const()[name = tensor("op_995"), val = tensor([1, 20, 64, 1500])]; + tensor var_996_cast_fp16 = reshape(shape = var_995, x = value_15_cast_fp16)[name = tensor("op_996_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_996_cast_fp16, y = obj_55_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_999 = const()[name = tensor("op_999"), val = tensor([1, 1280, 1, 1])]; + tensor input_33_cast_fp16 = reshape(shape = var_999, 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(314312512)))]; + tensor layers_3_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317589376)))]; + tensor obj_53_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = 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_1017_to_fp16 = const()[name = tensor("op_1017_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_1017_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(317592000)))]; + tensor input_35_beta_0_to_fp16 = const()[name = tensor("input_35_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317594624)))]; + tensor input_35_epsilon_0_to_fp16 = const()[name = tensor("input_35_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = batch_norm(beta = input_35_beta_0_to_fp16, epsilon = input_35_epsilon_0_to_fp16, gamma = input_35_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor 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(317597248)))]; + tensor layers_3_fc1_bias_to_fp16 = const()[name = tensor("layers_3_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(330704512)))]; + tensor input_37_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = 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(330714816)))]; + tensor layers_3_fc2_bias_to_fp16 = const()[name = tensor("layers_3_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343822080)))]; + tensor hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = 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_1052 = const()[name = tensor("op_1052"), val = tensor(3)]; + tensor out_25_axes_0 = const()[name = tensor("out_25_axes_0"), val = tensor([1])]; + tensor var_1077_to_fp16 = const()[name = tensor("op_1077_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_1077_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(343824704)))]; + tensor obj_57_beta_0_to_fp16 = const()[name = tensor("obj_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343827328)))]; + tensor obj_57_epsilon_0_to_fp16 = const()[name = tensor("obj_57_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_57_cast_fp16 = batch_norm(beta = obj_57_beta_0_to_fp16, epsilon = obj_57_epsilon_0_to_fp16, gamma = obj_57_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor("obj_57_cast_fp16")]; + tensor 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(343829952)))]; + tensor layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347106816)))]; + tensor query_17_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_bias_to_fp16, dilations = 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(347109440)))]; + 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(350386304)))]; + tensor layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353663168)))]; + tensor current_value_9_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_bias_to_fp16, dilations = 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_1116_cast_fp16 = mul(x = var_103_cast_fp16_4, y = var_239_cast_fp16)[name = tensor("op_1116_cast_fp16")]; + tensor var_1117_cast_fp16 = mul(x = current_key_9_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_1117_cast_fp16")]; + tensor key_17_cast_fp16 = add(x = var_1116_cast_fp16, y = var_1117_cast_fp16)[name = tensor("key_17_cast_fp16")]; + tensor var_1120_cast_fp16 = mul(x = var_138_cast_fp16_4, y = var_239_cast_fp16)[name = tensor("op_1120_cast_fp16")]; + tensor var_1121_cast_fp16 = mul(x = current_value_9_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_1121_cast_fp16")]; + tensor value_17_cast_fp16 = add(x = var_1120_cast_fp16, y = var_1121_cast_fp16)[name = tensor("value_17_cast_fp16")]; + tensor var_1125 = const()[name = tensor("op_1125"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_17_cast_fp16 = reshape(shape = var_1125, x = query_17_cast_fp16)[name = tensor("mh_q_17_cast_fp16")]; + tensor var_1127_to_fp16 = const()[name = tensor("op_1127_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1128_cast_fp16 = mul(x = mh_q_17_cast_fp16, y = var_1127_to_fp16)[name = tensor("op_1128_cast_fp16")]; + tensor var_1131 = const()[name = tensor("op_1131"), val = tensor([1, 20, 64, 448])]; + tensor var_1132_cast_fp16 = reshape(shape = var_1131, x = key_17_cast_fp16)[name = tensor("op_1132_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_1128_cast_fp16, y = var_1132_cast_fp16)[name = tensor("mh_w_25_cast_fp16")]; + tensor mh_w_27_cast_fp16 = add(x = mh_w_25_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_27_cast_fp16")]; + tensor var_1140_cast_fp16 = softmax(axis = var_1052, x = mh_w_27_cast_fp16)[name = tensor("op_1140_cast_fp16")]; + tensor var_1141 = const()[name = tensor("op_1141"), val = tensor([1, 20, 64, 448])]; + tensor var_1142_cast_fp16 = reshape(shape = var_1141, x = value_17_cast_fp16)[name = tensor("op_1142_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_1142_cast_fp16, y = var_1140_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([1, 1280, 1, 1])]; + tensor input_41_cast_fp16 = reshape(shape = var_1145, 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(353665792)))]; + tensor layers_4_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356942656)))]; + tensor obj_63_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_bias_to_fp16, dilations = 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_1167_to_fp16 = const()[name = tensor("op_1167_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_1167_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(356945280)))]; + tensor obj_65_beta_0_to_fp16 = const()[name = tensor("obj_65_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356947904)))]; + tensor obj_65_epsilon_0_to_fp16 = const()[name = tensor("obj_65_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_65_cast_fp16 = batch_norm(beta = obj_65_beta_0_to_fp16, epsilon = obj_65_epsilon_0_to_fp16, gamma = obj_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor("obj_65_cast_fp16")]; + tensor 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(356950528)))]; + tensor layers_4_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360227392)))]; + tensor query_19_cast_fp16 = conv(bias = layers_4_encoder_attn_q_proj_bias_to_fp16, dilations = 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(360230016)))]; + 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(363506880)))]; + tensor layers_4_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366783744)))]; + tensor value_19_cast_fp16 = conv(bias = layers_4_encoder_attn_v_proj_bias_to_fp16, dilations = 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_1203 = const()[name = tensor("op_1203"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_19_cast_fp16 = reshape(shape = var_1203, x = query_19_cast_fp16)[name = tensor("mh_q_19_cast_fp16")]; + tensor var_1205_to_fp16 = const()[name = tensor("op_1205_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1206_cast_fp16 = mul(x = mh_q_19_cast_fp16, y = var_1205_to_fp16)[name = tensor("op_1206_cast_fp16")]; + tensor var_1209 = const()[name = tensor("op_1209"), val = tensor([1, 20, 64, 1500])]; + tensor var_1210_cast_fp16 = reshape(shape = var_1209, x = key_19_cast_fp16)[name = tensor("op_1210_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_1206_cast_fp16, y = var_1210_cast_fp16)[name = tensor("mh_w_29_cast_fp16")]; + tensor obj_69_cast_fp16 = softmax(axis = var_1052, x = mh_w_29_cast_fp16)[name = tensor("obj_69_cast_fp16")]; + tensor var_1214 = const()[name = tensor("op_1214"), val = tensor([1, 20, 64, 1500])]; + tensor var_1215_cast_fp16 = reshape(shape = var_1214, x = value_19_cast_fp16)[name = tensor("op_1215_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_1215_cast_fp16, y = obj_69_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_1218 = const()[name = tensor("op_1218"), val = tensor([1, 1280, 1, 1])]; + tensor input_43_cast_fp16 = reshape(shape = var_1218, 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(366786368)))]; + tensor layers_4_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370063232)))]; + tensor obj_67_cast_fp16 = conv(bias = layers_4_encoder_attn_o_proj_bias_to_fp16, dilations = 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_1236_to_fp16 = const()[name = tensor("op_1236_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_1236_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(370065856)))]; + tensor input_45_beta_0_to_fp16 = const()[name = tensor("input_45_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370068480)))]; + tensor input_45_epsilon_0_to_fp16 = const()[name = tensor("input_45_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = batch_norm(beta = input_45_beta_0_to_fp16, epsilon = input_45_epsilon_0_to_fp16, gamma = input_45_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor 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(370071104)))]; + tensor layers_4_fc1_bias_to_fp16 = const()[name = tensor("layers_4_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383178368)))]; + tensor input_47_cast_fp16 = conv(bias = layers_4_fc1_bias_to_fp16, dilations = 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(383188672)))]; + tensor layers_4_fc2_bias_to_fp16 = const()[name = tensor("layers_4_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396295936)))]; + tensor hidden_states_11_cast_fp16 = conv(bias = layers_4_fc2_bias_to_fp16, dilations = 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_1271 = const()[name = tensor("op_1271"), val = tensor(3)]; + tensor out_31_axes_0 = const()[name = tensor("out_31_axes_0"), val = tensor([1])]; + tensor var_1296_to_fp16 = const()[name = tensor("op_1296_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1296_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(396298560)))]; + tensor obj_71_beta_0_to_fp16 = const()[name = tensor("obj_71_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396301184)))]; + tensor obj_71_epsilon_0_to_fp16 = const()[name = tensor("obj_71_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_71_cast_fp16 = batch_norm(beta = obj_71_beta_0_to_fp16, epsilon = obj_71_epsilon_0_to_fp16, gamma = obj_71_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor("obj_71_cast_fp16")]; + tensor 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(396303808)))]; + tensor layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(399580672)))]; + tensor query_21_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_bias_to_fp16, dilations = query_21_dilations_0, groups = query_21_groups_0, pad = query_21_pad_0, pad_type = query_21_pad_type_0, strides = query_21_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("query_21_cast_fp16")]; + tensor current_key_11_pad_type_0 = const()[name = tensor("current_key_11_pad_type_0"), val = tensor("valid")]; + tensor current_key_11_strides_0 = const()[name = tensor("current_key_11_strides_0"), val = tensor([1, 1])]; + tensor current_key_11_pad_0 = const()[name = tensor("current_key_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_11_dilations_0 = const()[name = tensor("current_key_11_dilations_0"), val = tensor([1, 1])]; + tensor current_key_11_groups_0 = const()[name = tensor("current_key_11_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(399583296)))]; + tensor current_key_11_cast_fp16 = conv(dilations = current_key_11_dilations_0, groups = current_key_11_groups_0, pad = current_key_11_pad_0, pad_type = current_key_11_pad_type_0, strides = current_key_11_strides_0, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("current_key_11_cast_fp16")]; + tensor current_value_11_pad_type_0 = const()[name = tensor("current_value_11_pad_type_0"), val = tensor("valid")]; + tensor current_value_11_strides_0 = const()[name = tensor("current_value_11_strides_0"), val = tensor([1, 1])]; + tensor current_value_11_pad_0 = const()[name = tensor("current_value_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_11_dilations_0 = const()[name = tensor("current_value_11_dilations_0"), val = tensor([1, 1])]; + tensor current_value_11_groups_0 = const()[name = tensor("current_value_11_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402860160)))]; + tensor layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(406137024)))]; + tensor current_value_11_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_bias_to_fp16, dilations = current_value_11_dilations_0, groups = current_value_11_groups_0, pad = current_value_11_pad_0, pad_type = current_value_11_pad_type_0, strides = current_value_11_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("current_value_11_cast_fp16")]; + tensor var_1335_cast_fp16 = mul(x = var_103_cast_fp16_5, y = var_239_cast_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor var_1336_cast_fp16 = mul(x = current_key_11_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_1336_cast_fp16")]; + tensor key_21_cast_fp16 = add(x = var_1335_cast_fp16, y = var_1336_cast_fp16)[name = tensor("key_21_cast_fp16")]; + tensor var_1339_cast_fp16 = mul(x = var_138_cast_fp16_5, y = var_239_cast_fp16)[name = tensor("op_1339_cast_fp16")]; + tensor var_1340_cast_fp16 = mul(x = current_value_11_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_1340_cast_fp16")]; + tensor value_21_cast_fp16 = add(x = var_1339_cast_fp16, y = var_1340_cast_fp16)[name = tensor("value_21_cast_fp16")]; + tensor var_1344 = const()[name = tensor("op_1344"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_21_cast_fp16 = reshape(shape = var_1344, x = query_21_cast_fp16)[name = tensor("mh_q_21_cast_fp16")]; + tensor var_1346_to_fp16 = const()[name = tensor("op_1346_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1347_cast_fp16 = mul(x = mh_q_21_cast_fp16, y = var_1346_to_fp16)[name = tensor("op_1347_cast_fp16")]; + tensor var_1350 = const()[name = tensor("op_1350"), val = tensor([1, 20, 64, 448])]; + tensor var_1351_cast_fp16 = reshape(shape = var_1350, x = key_21_cast_fp16)[name = tensor("op_1351_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_1347_cast_fp16, y = var_1351_cast_fp16)[name = tensor("mh_w_31_cast_fp16")]; + tensor mh_w_33_cast_fp16 = add(x = mh_w_31_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_33_cast_fp16")]; + tensor var_1359_cast_fp16 = softmax(axis = var_1271, x = mh_w_33_cast_fp16)[name = tensor("op_1359_cast_fp16")]; + tensor var_1360 = const()[name = tensor("op_1360"), val = tensor([1, 20, 64, 448])]; + tensor var_1361_cast_fp16 = reshape(shape = var_1360, x = value_21_cast_fp16)[name = tensor("op_1361_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_1361_cast_fp16, y = var_1359_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_1364 = const()[name = tensor("op_1364"), val = tensor([1, 1280, 1, 1])]; + tensor input_51_cast_fp16 = reshape(shape = var_1364, 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(406139648)))]; + tensor layers_5_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409416512)))]; + tensor obj_77_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_bias_to_fp16, dilations = 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_1386_to_fp16 = const()[name = tensor("op_1386_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1386_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(409419136)))]; + tensor obj_79_beta_0_to_fp16 = const()[name = tensor("obj_79_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409421760)))]; + tensor obj_79_epsilon_0_to_fp16 = const()[name = tensor("obj_79_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_79_cast_fp16 = batch_norm(beta = obj_79_beta_0_to_fp16, epsilon = obj_79_epsilon_0_to_fp16, gamma = obj_79_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor("obj_79_cast_fp16")]; + tensor query_23_pad_type_0 = const()[name = tensor("query_23_pad_type_0"), val = tensor("valid")]; + tensor query_23_strides_0 = const()[name = tensor("query_23_strides_0"), val = tensor([1, 1])]; + tensor query_23_pad_0 = const()[name = tensor("query_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_23_dilations_0 = const()[name = tensor("query_23_dilations_0"), val = tensor([1, 1])]; + tensor query_23_groups_0 = const()[name = tensor("query_23_groups_0"), val = tensor(1)]; + tensor layers_5_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409424384)))]; + tensor layers_5_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412701248)))]; + tensor query_23_cast_fp16 = conv(bias = layers_5_encoder_attn_q_proj_bias_to_fp16, dilations = query_23_dilations_0, groups = query_23_groups_0, pad = query_23_pad_0, pad_type = query_23_pad_type_0, strides = query_23_strides_0, weight = layers_5_encoder_attn_q_proj_weight_to_fp16, x = obj_79_cast_fp16)[name = tensor("query_23_cast_fp16")]; + tensor key_23_pad_type_0 = const()[name = tensor("key_23_pad_type_0"), val = tensor("valid")]; + tensor key_23_strides_0 = const()[name = tensor("key_23_strides_0"), val = tensor([1, 1])]; + tensor key_23_pad_0 = const()[name = tensor("key_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_23_dilations_0 = const()[name = tensor("key_23_dilations_0"), val = tensor([1, 1])]; + tensor key_23_groups_0 = const()[name = tensor("key_23_groups_0"), val = tensor(1)]; + tensor layers_5_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412703872)))]; + tensor key_23_cast_fp16 = conv(dilations = key_23_dilations_0, groups = key_23_groups_0, pad = key_23_pad_0, pad_type = key_23_pad_type_0, strides = key_23_strides_0, weight = layers_5_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_23_cast_fp16")]; + tensor value_23_pad_type_0 = const()[name = tensor("value_23_pad_type_0"), val = tensor("valid")]; + tensor value_23_strides_0 = const()[name = tensor("value_23_strides_0"), val = tensor([1, 1])]; + tensor value_23_pad_0 = const()[name = tensor("value_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_23_dilations_0 = const()[name = tensor("value_23_dilations_0"), val = tensor([1, 1])]; + tensor value_23_groups_0 = const()[name = tensor("value_23_groups_0"), val = tensor(1)]; + tensor layers_5_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(415980736)))]; + tensor layers_5_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419257600)))]; + tensor value_23_cast_fp16 = conv(bias = layers_5_encoder_attn_v_proj_bias_to_fp16, dilations = value_23_dilations_0, groups = value_23_groups_0, pad = value_23_pad_0, pad_type = value_23_pad_type_0, strides = value_23_strides_0, weight = layers_5_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_23_cast_fp16")]; + tensor var_1422 = const()[name = tensor("op_1422"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_23_cast_fp16 = reshape(shape = var_1422, x = query_23_cast_fp16)[name = tensor("mh_q_23_cast_fp16")]; + tensor var_1424_to_fp16 = const()[name = tensor("op_1424_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1425_cast_fp16 = mul(x = mh_q_23_cast_fp16, y = var_1424_to_fp16)[name = tensor("op_1425_cast_fp16")]; + tensor var_1428 = const()[name = tensor("op_1428"), val = tensor([1, 20, 64, 1500])]; + tensor var_1429_cast_fp16 = reshape(shape = var_1428, x = key_23_cast_fp16)[name = tensor("op_1429_cast_fp16")]; + tensor mh_w_35_transpose_x_0 = const()[name = tensor("mh_w_35_transpose_x_0"), val = tensor(true)]; + tensor mh_w_35_transpose_y_0 = const()[name = tensor("mh_w_35_transpose_y_0"), val = tensor(false)]; + tensor mh_w_35_cast_fp16 = matmul(transpose_x = mh_w_35_transpose_x_0, transpose_y = mh_w_35_transpose_y_0, x = var_1425_cast_fp16, y = var_1429_cast_fp16)[name = tensor("mh_w_35_cast_fp16")]; + tensor obj_83_cast_fp16 = softmax(axis = var_1271, x = mh_w_35_cast_fp16)[name = tensor("obj_83_cast_fp16")]; + tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([1, 20, 64, 1500])]; + tensor var_1434_cast_fp16 = reshape(shape = var_1433, x = value_23_cast_fp16)[name = tensor("op_1434_cast_fp16")]; + tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; + tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(true)]; + tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = var_1434_cast_fp16, y = obj_83_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_1437 = const()[name = tensor("op_1437"), val = tensor([1, 1280, 1, 1])]; + tensor input_53_cast_fp16 = reshape(shape = var_1437, x = attn_23_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor obj_81_pad_type_0 = const()[name = tensor("obj_81_pad_type_0"), val = tensor("valid")]; + tensor obj_81_strides_0 = const()[name = tensor("obj_81_strides_0"), val = tensor([1, 1])]; + tensor obj_81_pad_0 = const()[name = tensor("obj_81_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_81_dilations_0 = const()[name = tensor("obj_81_dilations_0"), val = tensor([1, 1])]; + tensor obj_81_groups_0 = const()[name = tensor("obj_81_groups_0"), val = tensor(1)]; + tensor layers_5_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419260224)))]; + tensor layers_5_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422537088)))]; + tensor obj_81_cast_fp16 = conv(bias = layers_5_encoder_attn_o_proj_bias_to_fp16, dilations = 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_1455_to_fp16 = const()[name = tensor("op_1455_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1455_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(422539712)))]; + tensor input_55_beta_0_to_fp16 = const()[name = tensor("input_55_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422542336)))]; + tensor input_55_epsilon_0_to_fp16 = const()[name = tensor("input_55_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_55_cast_fp16 = batch_norm(beta = input_55_beta_0_to_fp16, epsilon = input_55_epsilon_0_to_fp16, gamma = input_55_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor 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(422544960)))]; + tensor layers_5_fc1_bias_to_fp16 = const()[name = tensor("layers_5_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435652224)))]; + tensor input_57_cast_fp16 = conv(bias = layers_5_fc1_bias_to_fp16, dilations = input_57_dilations_0, groups = input_57_groups_0, pad = input_57_pad_0, pad_type = input_57_pad_type_0, strides = input_57_strides_0, weight = layers_5_fc1_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor hidden_states_13_pad_type_0 = const()[name = tensor("hidden_states_13_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_13_strides_0 = const()[name = tensor("hidden_states_13_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_13_pad_0 = const()[name = tensor("hidden_states_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_13_dilations_0 = const()[name = tensor("hidden_states_13_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_13_groups_0 = const()[name = tensor("hidden_states_13_groups_0"), val = tensor(1)]; + tensor layers_5_fc2_weight_to_fp16 = const()[name = tensor("layers_5_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435662528)))]; + tensor layers_5_fc2_bias_to_fp16 = const()[name = tensor("layers_5_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448769792)))]; + tensor hidden_states_13_cast_fp16 = conv(bias = layers_5_fc2_bias_to_fp16, dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_5_fc2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("hidden_states_13_cast_fp16")]; + tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor("inputs_37_cast_fp16")]; + tensor var_1490 = const()[name = tensor("op_1490"), val = tensor(3)]; + tensor out_37_axes_0 = const()[name = tensor("out_37_axes_0"), val = tensor([1])]; + tensor var_1515_to_fp16 = const()[name = tensor("op_1515_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_1515_to_fp16, x = inputs_37_cast_fp16)[name = tensor("out_37_cast_fp16")]; + tensor obj_85_gamma_0_to_fp16 = const()[name = tensor("obj_85_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448772416)))]; + tensor obj_85_beta_0_to_fp16 = const()[name = tensor("obj_85_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448775040)))]; + tensor obj_85_epsilon_0_to_fp16 = const()[name = tensor("obj_85_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_85_cast_fp16 = batch_norm(beta = obj_85_beta_0_to_fp16, epsilon = obj_85_epsilon_0_to_fp16, gamma = obj_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_37_cast_fp16)[name = tensor("obj_85_cast_fp16")]; + tensor query_25_pad_type_0 = const()[name = tensor("query_25_pad_type_0"), val = tensor("valid")]; + tensor query_25_strides_0 = const()[name = tensor("query_25_strides_0"), val = tensor([1, 1])]; + tensor query_25_pad_0 = const()[name = tensor("query_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_25_dilations_0 = const()[name = tensor("query_25_dilations_0"), val = tensor([1, 1])]; + tensor query_25_groups_0 = const()[name = tensor("query_25_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448777664)))]; + tensor layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452054528)))]; + tensor query_25_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_bias_to_fp16, dilations = query_25_dilations_0, groups = query_25_groups_0, pad = query_25_pad_0, pad_type = query_25_pad_type_0, strides = query_25_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("query_25_cast_fp16")]; + tensor current_key_13_pad_type_0 = const()[name = tensor("current_key_13_pad_type_0"), val = tensor("valid")]; + tensor current_key_13_strides_0 = const()[name = tensor("current_key_13_strides_0"), val = tensor([1, 1])]; + tensor current_key_13_pad_0 = const()[name = tensor("current_key_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_13_dilations_0 = const()[name = tensor("current_key_13_dilations_0"), val = tensor([1, 1])]; + tensor current_key_13_groups_0 = const()[name = tensor("current_key_13_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452057152)))]; + tensor current_key_13_cast_fp16 = conv(dilations = current_key_13_dilations_0, groups = current_key_13_groups_0, pad = current_key_13_pad_0, pad_type = current_key_13_pad_type_0, strides = current_key_13_strides_0, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("current_key_13_cast_fp16")]; + tensor current_value_13_pad_type_0 = const()[name = tensor("current_value_13_pad_type_0"), val = tensor("valid")]; + tensor current_value_13_strides_0 = const()[name = tensor("current_value_13_strides_0"), val = tensor([1, 1])]; + tensor current_value_13_pad_0 = const()[name = tensor("current_value_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_13_dilations_0 = const()[name = tensor("current_value_13_dilations_0"), val = tensor([1, 1])]; + tensor current_value_13_groups_0 = const()[name = tensor("current_value_13_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455334016)))]; + tensor layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(458610880)))]; + tensor current_value_13_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_bias_to_fp16, dilations = current_value_13_dilations_0, groups = current_value_13_groups_0, pad = current_value_13_pad_0, pad_type = current_value_13_pad_type_0, strides = current_value_13_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("current_value_13_cast_fp16")]; + tensor var_1554_cast_fp16 = mul(x = var_103_cast_fp16_6, y = var_239_cast_fp16)[name = tensor("op_1554_cast_fp16")]; + tensor var_1555_cast_fp16 = mul(x = current_key_13_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_1555_cast_fp16")]; + tensor key_25_cast_fp16 = add(x = var_1554_cast_fp16, y = var_1555_cast_fp16)[name = tensor("key_25_cast_fp16")]; + tensor var_1558_cast_fp16 = mul(x = var_138_cast_fp16_6, y = var_239_cast_fp16)[name = tensor("op_1558_cast_fp16")]; + tensor var_1559_cast_fp16 = mul(x = current_value_13_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_1559_cast_fp16")]; + tensor value_25_cast_fp16 = add(x = var_1558_cast_fp16, y = var_1559_cast_fp16)[name = tensor("value_25_cast_fp16")]; + tensor var_1563 = const()[name = tensor("op_1563"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_25_cast_fp16 = reshape(shape = var_1563, x = query_25_cast_fp16)[name = tensor("mh_q_25_cast_fp16")]; + tensor var_1565_to_fp16 = const()[name = tensor("op_1565_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1566_cast_fp16 = mul(x = mh_q_25_cast_fp16, y = var_1565_to_fp16)[name = tensor("op_1566_cast_fp16")]; + tensor var_1569 = const()[name = tensor("op_1569"), val = tensor([1, 20, 64, 448])]; + tensor var_1570_cast_fp16 = reshape(shape = var_1569, x = key_25_cast_fp16)[name = tensor("op_1570_cast_fp16")]; + tensor mh_w_37_transpose_x_0 = const()[name = tensor("mh_w_37_transpose_x_0"), val = tensor(true)]; + tensor mh_w_37_transpose_y_0 = const()[name = tensor("mh_w_37_transpose_y_0"), val = tensor(false)]; + tensor mh_w_37_cast_fp16 = matmul(transpose_x = mh_w_37_transpose_x_0, transpose_y = mh_w_37_transpose_y_0, x = var_1566_cast_fp16, y = var_1570_cast_fp16)[name = tensor("mh_w_37_cast_fp16")]; + tensor mh_w_39_cast_fp16 = add(x = mh_w_37_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_39_cast_fp16")]; + tensor var_1578_cast_fp16 = softmax(axis = var_1490, x = mh_w_39_cast_fp16)[name = tensor("op_1578_cast_fp16")]; + tensor var_1579 = const()[name = tensor("op_1579"), val = tensor([1, 20, 64, 448])]; + tensor var_1580_cast_fp16 = reshape(shape = var_1579, x = value_25_cast_fp16)[name = tensor("op_1580_cast_fp16")]; + tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; + tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(true)]; + tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = var_1580_cast_fp16, y = var_1578_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_1583 = const()[name = tensor("op_1583"), val = tensor([1, 1280, 1, 1])]; + tensor input_61_cast_fp16 = reshape(shape = var_1583, x = attn_25_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor obj_91_pad_type_0 = const()[name = tensor("obj_91_pad_type_0"), val = tensor("valid")]; + tensor obj_91_strides_0 = const()[name = tensor("obj_91_strides_0"), val = tensor([1, 1])]; + tensor obj_91_pad_0 = const()[name = tensor("obj_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_91_dilations_0 = const()[name = tensor("obj_91_dilations_0"), val = tensor([1, 1])]; + tensor obj_91_groups_0 = const()[name = tensor("obj_91_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(458613504)))]; + tensor layers_6_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461890368)))]; + tensor obj_91_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_bias_to_fp16, dilations = obj_91_dilations_0, groups = obj_91_groups_0, pad = obj_91_pad_0, pad_type = obj_91_pad_type_0, strides = obj_91_strides_0, weight = layers_6_self_attn_o_proj_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("obj_91_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_91_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; + tensor out_39_axes_0 = const()[name = tensor("out_39_axes_0"), val = tensor([1])]; + tensor var_1605_to_fp16 = const()[name = tensor("op_1605_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_1605_to_fp16, x = inputs_39_cast_fp16)[name = tensor("out_39_cast_fp16")]; + tensor obj_93_gamma_0_to_fp16 = const()[name = tensor("obj_93_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461892992)))]; + tensor obj_93_beta_0_to_fp16 = const()[name = tensor("obj_93_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461895616)))]; + tensor obj_93_epsilon_0_to_fp16 = const()[name = tensor("obj_93_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_93_cast_fp16 = batch_norm(beta = obj_93_beta_0_to_fp16, epsilon = obj_93_epsilon_0_to_fp16, gamma = obj_93_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_39_cast_fp16)[name = tensor("obj_93_cast_fp16")]; + tensor query_27_pad_type_0 = const()[name = tensor("query_27_pad_type_0"), val = tensor("valid")]; + tensor query_27_strides_0 = const()[name = tensor("query_27_strides_0"), val = tensor([1, 1])]; + tensor query_27_pad_0 = const()[name = tensor("query_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_27_dilations_0 = const()[name = tensor("query_27_dilations_0"), val = tensor([1, 1])]; + tensor query_27_groups_0 = const()[name = tensor("query_27_groups_0"), val = tensor(1)]; + tensor layers_6_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461898240)))]; + tensor layers_6_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465175104)))]; + tensor query_27_cast_fp16 = conv(bias = layers_6_encoder_attn_q_proj_bias_to_fp16, dilations = query_27_dilations_0, groups = query_27_groups_0, pad = query_27_pad_0, pad_type = query_27_pad_type_0, strides = query_27_strides_0, weight = layers_6_encoder_attn_q_proj_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor("query_27_cast_fp16")]; + tensor key_27_pad_type_0 = const()[name = tensor("key_27_pad_type_0"), val = tensor("valid")]; + tensor key_27_strides_0 = const()[name = tensor("key_27_strides_0"), val = tensor([1, 1])]; + tensor key_27_pad_0 = const()[name = tensor("key_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_27_dilations_0 = const()[name = tensor("key_27_dilations_0"), val = tensor([1, 1])]; + tensor key_27_groups_0 = const()[name = tensor("key_27_groups_0"), val = tensor(1)]; + tensor layers_6_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465177728)))]; + tensor key_27_cast_fp16 = conv(dilations = key_27_dilations_0, groups = key_27_groups_0, pad = key_27_pad_0, pad_type = key_27_pad_type_0, strides = key_27_strides_0, weight = layers_6_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_27_cast_fp16")]; + tensor value_27_pad_type_0 = const()[name = tensor("value_27_pad_type_0"), val = tensor("valid")]; + tensor value_27_strides_0 = const()[name = tensor("value_27_strides_0"), val = tensor([1, 1])]; + tensor value_27_pad_0 = const()[name = tensor("value_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_27_dilations_0 = const()[name = tensor("value_27_dilations_0"), val = tensor([1, 1])]; + tensor value_27_groups_0 = const()[name = tensor("value_27_groups_0"), val = tensor(1)]; + tensor layers_6_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468454592)))]; + tensor layers_6_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471731456)))]; + tensor value_27_cast_fp16 = conv(bias = layers_6_encoder_attn_v_proj_bias_to_fp16, dilations = value_27_dilations_0, groups = value_27_groups_0, pad = value_27_pad_0, pad_type = value_27_pad_type_0, strides = value_27_strides_0, weight = layers_6_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_27_cast_fp16")]; + tensor var_1641 = const()[name = tensor("op_1641"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_27_cast_fp16 = reshape(shape = var_1641, x = query_27_cast_fp16)[name = tensor("mh_q_27_cast_fp16")]; + tensor var_1643_to_fp16 = const()[name = tensor("op_1643_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1644_cast_fp16 = mul(x = mh_q_27_cast_fp16, y = var_1643_to_fp16)[name = tensor("op_1644_cast_fp16")]; + tensor var_1647 = const()[name = tensor("op_1647"), val = tensor([1, 20, 64, 1500])]; + tensor var_1648_cast_fp16 = reshape(shape = var_1647, x = key_27_cast_fp16)[name = tensor("op_1648_cast_fp16")]; + tensor mh_w_41_transpose_x_0 = const()[name = tensor("mh_w_41_transpose_x_0"), val = tensor(true)]; + tensor mh_w_41_transpose_y_0 = const()[name = tensor("mh_w_41_transpose_y_0"), val = tensor(false)]; + tensor mh_w_41_cast_fp16 = matmul(transpose_x = mh_w_41_transpose_x_0, transpose_y = mh_w_41_transpose_y_0, x = var_1644_cast_fp16, y = var_1648_cast_fp16)[name = tensor("mh_w_41_cast_fp16")]; + tensor obj_97_cast_fp16 = softmax(axis = var_1490, x = mh_w_41_cast_fp16)[name = tensor("obj_97_cast_fp16")]; + tensor var_1652 = const()[name = tensor("op_1652"), val = tensor([1, 20, 64, 1500])]; + tensor var_1653_cast_fp16 = reshape(shape = var_1652, x = value_27_cast_fp16)[name = tensor("op_1653_cast_fp16")]; + tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; + tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(true)]; + tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = var_1653_cast_fp16, y = obj_97_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_1656 = const()[name = tensor("op_1656"), val = tensor([1, 1280, 1, 1])]; + tensor input_63_cast_fp16 = reshape(shape = var_1656, x = attn_27_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor obj_95_pad_type_0 = const()[name = tensor("obj_95_pad_type_0"), val = tensor("valid")]; + tensor obj_95_strides_0 = const()[name = tensor("obj_95_strides_0"), val = tensor([1, 1])]; + tensor obj_95_pad_0 = const()[name = tensor("obj_95_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_95_dilations_0 = const()[name = tensor("obj_95_dilations_0"), val = tensor([1, 1])]; + tensor obj_95_groups_0 = const()[name = tensor("obj_95_groups_0"), val = tensor(1)]; + tensor layers_6_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471734080)))]; + tensor layers_6_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(475010944)))]; + tensor obj_95_cast_fp16 = conv(bias = layers_6_encoder_attn_o_proj_bias_to_fp16, dilations = obj_95_dilations_0, groups = obj_95_groups_0, pad = obj_95_pad_0, pad_type = obj_95_pad_type_0, strides = obj_95_strides_0, weight = layers_6_encoder_attn_o_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("obj_95_cast_fp16")]; + tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = obj_95_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; + tensor out_41_axes_0 = const()[name = tensor("out_41_axes_0"), val = tensor([1])]; + tensor var_1674_to_fp16 = const()[name = tensor("op_1674_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_1674_to_fp16, x = inputs_41_cast_fp16)[name = tensor("out_41_cast_fp16")]; + tensor input_65_gamma_0_to_fp16 = const()[name = tensor("input_65_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(475013568)))]; + tensor input_65_beta_0_to_fp16 = const()[name = tensor("input_65_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(475016192)))]; + tensor input_65_epsilon_0_to_fp16 = const()[name = tensor("input_65_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = batch_norm(beta = input_65_beta_0_to_fp16, epsilon = input_65_epsilon_0_to_fp16, gamma = input_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_41_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor input_67_pad_type_0 = const()[name = tensor("input_67_pad_type_0"), val = tensor("valid")]; + tensor input_67_strides_0 = const()[name = tensor("input_67_strides_0"), val = tensor([1, 1])]; + tensor input_67_pad_0 = const()[name = tensor("input_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_67_dilations_0 = const()[name = tensor("input_67_dilations_0"), val = tensor([1, 1])]; + tensor input_67_groups_0 = const()[name = tensor("input_67_groups_0"), val = tensor(1)]; + tensor layers_6_fc1_weight_to_fp16 = const()[name = tensor("layers_6_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(475018816)))]; + tensor layers_6_fc1_bias_to_fp16 = const()[name = tensor("layers_6_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488126080)))]; + tensor input_67_cast_fp16 = conv(bias = layers_6_fc1_bias_to_fp16, dilations = input_67_dilations_0, groups = input_67_groups_0, pad = input_67_pad_0, pad_type = input_67_pad_type_0, strides = input_67_strides_0, weight = layers_6_fc1_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_mode_0 = const()[name = tensor("input_69_mode_0"), val = tensor("EXACT")]; + tensor input_69_cast_fp16 = gelu(mode = input_69_mode_0, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor hidden_states_15_pad_type_0 = const()[name = tensor("hidden_states_15_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_15_strides_0 = const()[name = tensor("hidden_states_15_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_15_pad_0 = const()[name = tensor("hidden_states_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_15_dilations_0 = const()[name = tensor("hidden_states_15_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_15_groups_0 = const()[name = tensor("hidden_states_15_groups_0"), val = tensor(1)]; + tensor layers_6_fc2_weight_to_fp16 = const()[name = tensor("layers_6_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488136384)))]; + tensor layers_6_fc2_bias_to_fp16 = const()[name = tensor("layers_6_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(501243648)))]; + tensor hidden_states_15_cast_fp16 = conv(bias = layers_6_fc2_bias_to_fp16, dilations = hidden_states_15_dilations_0, groups = hidden_states_15_groups_0, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = hidden_states_15_strides_0, weight = layers_6_fc2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("hidden_states_15_cast_fp16")]; + tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = hidden_states_15_cast_fp16)[name = tensor("inputs_43_cast_fp16")]; + tensor var_1709 = const()[name = tensor("op_1709"), val = tensor(3)]; + tensor out_43_axes_0 = const()[name = tensor("out_43_axes_0"), val = tensor([1])]; + tensor var_1734_to_fp16 = const()[name = tensor("op_1734_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_1734_to_fp16, x = inputs_43_cast_fp16)[name = tensor("out_43_cast_fp16")]; + tensor obj_99_gamma_0_to_fp16 = const()[name = tensor("obj_99_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(501246272)))]; + tensor obj_99_beta_0_to_fp16 = const()[name = tensor("obj_99_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(501248896)))]; + tensor obj_99_epsilon_0_to_fp16 = const()[name = tensor("obj_99_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_99_cast_fp16 = batch_norm(beta = obj_99_beta_0_to_fp16, epsilon = obj_99_epsilon_0_to_fp16, gamma = obj_99_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_43_cast_fp16)[name = tensor("obj_99_cast_fp16")]; + tensor query_29_pad_type_0 = const()[name = tensor("query_29_pad_type_0"), val = tensor("valid")]; + tensor query_29_strides_0 = const()[name = tensor("query_29_strides_0"), val = tensor([1, 1])]; + tensor query_29_pad_0 = const()[name = tensor("query_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_29_dilations_0 = const()[name = tensor("query_29_dilations_0"), val = tensor([1, 1])]; + tensor query_29_groups_0 = const()[name = tensor("query_29_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(501251520)))]; + tensor layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504528384)))]; + tensor query_29_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_bias_to_fp16, dilations = query_29_dilations_0, groups = query_29_groups_0, pad = query_29_pad_0, pad_type = query_29_pad_type_0, strides = query_29_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("query_29_cast_fp16")]; + tensor current_key_15_pad_type_0 = const()[name = tensor("current_key_15_pad_type_0"), val = tensor("valid")]; + tensor current_key_15_strides_0 = const()[name = tensor("current_key_15_strides_0"), val = tensor([1, 1])]; + tensor current_key_15_pad_0 = const()[name = tensor("current_key_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_15_dilations_0 = const()[name = tensor("current_key_15_dilations_0"), val = tensor([1, 1])]; + tensor current_key_15_groups_0 = const()[name = tensor("current_key_15_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504531008)))]; + tensor current_key_15_cast_fp16 = conv(dilations = current_key_15_dilations_0, groups = current_key_15_groups_0, pad = current_key_15_pad_0, pad_type = current_key_15_pad_type_0, strides = current_key_15_strides_0, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("current_key_15_cast_fp16")]; + tensor current_value_15_pad_type_0 = const()[name = tensor("current_value_15_pad_type_0"), val = tensor("valid")]; + tensor current_value_15_strides_0 = const()[name = tensor("current_value_15_strides_0"), val = tensor([1, 1])]; + tensor current_value_15_pad_0 = const()[name = tensor("current_value_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_15_dilations_0 = const()[name = tensor("current_value_15_dilations_0"), val = tensor([1, 1])]; + tensor current_value_15_groups_0 = const()[name = tensor("current_value_15_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507807872)))]; + tensor layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511084736)))]; + tensor current_value_15_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_bias_to_fp16, dilations = current_value_15_dilations_0, groups = current_value_15_groups_0, pad = current_value_15_pad_0, pad_type = current_value_15_pad_type_0, strides = current_value_15_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("current_value_15_cast_fp16")]; + tensor var_1773_cast_fp16 = mul(x = var_103_cast_fp16_7, y = var_239_cast_fp16)[name = tensor("op_1773_cast_fp16")]; + tensor var_1774_cast_fp16 = mul(x = current_key_15_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_1774_cast_fp16")]; + tensor key_29_cast_fp16 = add(x = var_1773_cast_fp16, y = var_1774_cast_fp16)[name = tensor("key_29_cast_fp16")]; + tensor var_1777_cast_fp16 = mul(x = var_138_cast_fp16_7, y = var_239_cast_fp16)[name = tensor("op_1777_cast_fp16")]; + tensor var_1778_cast_fp16 = mul(x = current_value_15_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_1778_cast_fp16")]; + tensor value_29_cast_fp16 = add(x = var_1777_cast_fp16, y = var_1778_cast_fp16)[name = tensor("value_29_cast_fp16")]; + tensor var_1782 = const()[name = tensor("op_1782"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_29_cast_fp16 = reshape(shape = var_1782, x = query_29_cast_fp16)[name = tensor("mh_q_29_cast_fp16")]; + tensor var_1784_to_fp16 = const()[name = tensor("op_1784_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1785_cast_fp16 = mul(x = mh_q_29_cast_fp16, y = var_1784_to_fp16)[name = tensor("op_1785_cast_fp16")]; + tensor var_1788 = const()[name = tensor("op_1788"), val = tensor([1, 20, 64, 448])]; + tensor var_1789_cast_fp16 = reshape(shape = var_1788, x = key_29_cast_fp16)[name = tensor("op_1789_cast_fp16")]; + tensor mh_w_43_transpose_x_0 = const()[name = tensor("mh_w_43_transpose_x_0"), val = tensor(true)]; + tensor mh_w_43_transpose_y_0 = const()[name = tensor("mh_w_43_transpose_y_0"), val = tensor(false)]; + tensor mh_w_43_cast_fp16 = matmul(transpose_x = mh_w_43_transpose_x_0, transpose_y = mh_w_43_transpose_y_0, x = var_1785_cast_fp16, y = var_1789_cast_fp16)[name = tensor("mh_w_43_cast_fp16")]; + tensor mh_w_45_cast_fp16 = add(x = mh_w_43_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_45_cast_fp16")]; + tensor var_1797_cast_fp16 = softmax(axis = var_1709, x = mh_w_45_cast_fp16)[name = tensor("op_1797_cast_fp16")]; + tensor var_1798 = const()[name = tensor("op_1798"), val = tensor([1, 20, 64, 448])]; + tensor var_1799_cast_fp16 = reshape(shape = var_1798, x = value_29_cast_fp16)[name = tensor("op_1799_cast_fp16")]; + tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; + tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(true)]; + tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = var_1799_cast_fp16, y = var_1797_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_1802 = const()[name = tensor("op_1802"), val = tensor([1, 1280, 1, 1])]; + tensor input_71_cast_fp16 = reshape(shape = var_1802, x = attn_29_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor obj_105_pad_type_0 = const()[name = tensor("obj_105_pad_type_0"), val = tensor("valid")]; + tensor obj_105_strides_0 = const()[name = tensor("obj_105_strides_0"), val = tensor([1, 1])]; + tensor obj_105_pad_0 = const()[name = tensor("obj_105_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_105_dilations_0 = const()[name = tensor("obj_105_dilations_0"), val = tensor([1, 1])]; + tensor obj_105_groups_0 = const()[name = tensor("obj_105_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511087360)))]; + tensor layers_7_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514364224)))]; + tensor obj_105_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_bias_to_fp16, dilations = obj_105_dilations_0, groups = obj_105_groups_0, pad = obj_105_pad_0, pad_type = obj_105_pad_type_0, strides = obj_105_strides_0, weight = layers_7_self_attn_o_proj_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("obj_105_cast_fp16")]; + tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = obj_105_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; + tensor out_45_axes_0 = const()[name = tensor("out_45_axes_0"), val = tensor([1])]; + tensor var_1824_to_fp16 = const()[name = tensor("op_1824_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_1824_to_fp16, x = inputs_45_cast_fp16)[name = tensor("out_45_cast_fp16")]; + tensor obj_107_gamma_0_to_fp16 = const()[name = tensor("obj_107_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514366848)))]; + tensor obj_107_beta_0_to_fp16 = const()[name = tensor("obj_107_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514369472)))]; + tensor obj_107_epsilon_0_to_fp16 = const()[name = tensor("obj_107_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_107_cast_fp16 = batch_norm(beta = obj_107_beta_0_to_fp16, epsilon = obj_107_epsilon_0_to_fp16, gamma = obj_107_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor("obj_107_cast_fp16")]; + tensor query_31_pad_type_0 = const()[name = tensor("query_31_pad_type_0"), val = tensor("valid")]; + tensor query_31_strides_0 = const()[name = tensor("query_31_strides_0"), val = tensor([1, 1])]; + tensor query_31_pad_0 = const()[name = tensor("query_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_31_dilations_0 = const()[name = tensor("query_31_dilations_0"), val = tensor([1, 1])]; + tensor query_31_groups_0 = const()[name = tensor("query_31_groups_0"), val = tensor(1)]; + tensor layers_7_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514372096)))]; + tensor layers_7_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517648960)))]; + tensor query_31_cast_fp16 = conv(bias = layers_7_encoder_attn_q_proj_bias_to_fp16, dilations = query_31_dilations_0, groups = query_31_groups_0, pad = query_31_pad_0, pad_type = query_31_pad_type_0, strides = query_31_strides_0, weight = layers_7_encoder_attn_q_proj_weight_to_fp16, x = obj_107_cast_fp16)[name = tensor("query_31_cast_fp16")]; + tensor key_31_pad_type_0 = const()[name = tensor("key_31_pad_type_0"), val = tensor("valid")]; + tensor key_31_strides_0 = const()[name = tensor("key_31_strides_0"), val = tensor([1, 1])]; + tensor key_31_pad_0 = const()[name = tensor("key_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_31_dilations_0 = const()[name = tensor("key_31_dilations_0"), val = tensor([1, 1])]; + tensor key_31_groups_0 = const()[name = tensor("key_31_groups_0"), val = tensor(1)]; + tensor layers_7_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517651584)))]; + tensor key_31_cast_fp16 = conv(dilations = key_31_dilations_0, groups = key_31_groups_0, pad = key_31_pad_0, pad_type = key_31_pad_type_0, strides = key_31_strides_0, weight = layers_7_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_31_cast_fp16")]; + tensor value_31_pad_type_0 = const()[name = tensor("value_31_pad_type_0"), val = tensor("valid")]; + tensor value_31_strides_0 = const()[name = tensor("value_31_strides_0"), val = tensor([1, 1])]; + tensor value_31_pad_0 = const()[name = tensor("value_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_31_dilations_0 = const()[name = tensor("value_31_dilations_0"), val = tensor([1, 1])]; + tensor value_31_groups_0 = const()[name = tensor("value_31_groups_0"), val = tensor(1)]; + tensor layers_7_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520928448)))]; + tensor layers_7_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524205312)))]; + tensor value_31_cast_fp16 = conv(bias = layers_7_encoder_attn_v_proj_bias_to_fp16, dilations = value_31_dilations_0, groups = value_31_groups_0, pad = value_31_pad_0, pad_type = value_31_pad_type_0, strides = value_31_strides_0, weight = layers_7_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_31_cast_fp16")]; + tensor var_1860 = const()[name = tensor("op_1860"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_31_cast_fp16 = reshape(shape = var_1860, x = query_31_cast_fp16)[name = tensor("mh_q_31_cast_fp16")]; + tensor var_1862_to_fp16 = const()[name = tensor("op_1862_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1863_cast_fp16 = mul(x = mh_q_31_cast_fp16, y = var_1862_to_fp16)[name = tensor("op_1863_cast_fp16")]; + tensor var_1866 = const()[name = tensor("op_1866"), val = tensor([1, 20, 64, 1500])]; + tensor var_1867_cast_fp16 = reshape(shape = var_1866, x = key_31_cast_fp16)[name = tensor("op_1867_cast_fp16")]; + tensor mh_w_47_transpose_x_0 = const()[name = tensor("mh_w_47_transpose_x_0"), val = tensor(true)]; + tensor mh_w_47_transpose_y_0 = const()[name = tensor("mh_w_47_transpose_y_0"), val = tensor(false)]; + tensor mh_w_47_cast_fp16 = matmul(transpose_x = mh_w_47_transpose_x_0, transpose_y = mh_w_47_transpose_y_0, x = var_1863_cast_fp16, y = var_1867_cast_fp16)[name = tensor("mh_w_47_cast_fp16")]; + tensor obj_111_cast_fp16 = softmax(axis = var_1709, x = mh_w_47_cast_fp16)[name = tensor("obj_111_cast_fp16")]; + tensor var_1871 = const()[name = tensor("op_1871"), val = tensor([1, 20, 64, 1500])]; + tensor var_1872_cast_fp16 = reshape(shape = var_1871, x = value_31_cast_fp16)[name = tensor("op_1872_cast_fp16")]; + tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; + tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(true)]; + tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = var_1872_cast_fp16, y = obj_111_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_1875 = const()[name = tensor("op_1875"), val = tensor([1, 1280, 1, 1])]; + tensor input_73_cast_fp16 = reshape(shape = var_1875, x = attn_31_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor obj_109_pad_type_0 = const()[name = tensor("obj_109_pad_type_0"), val = tensor("valid")]; + tensor obj_109_strides_0 = const()[name = tensor("obj_109_strides_0"), val = tensor([1, 1])]; + tensor obj_109_pad_0 = const()[name = tensor("obj_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_109_dilations_0 = const()[name = tensor("obj_109_dilations_0"), val = tensor([1, 1])]; + tensor obj_109_groups_0 = const()[name = tensor("obj_109_groups_0"), val = tensor(1)]; + tensor layers_7_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524207936)))]; + tensor layers_7_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527484800)))]; + tensor obj_109_cast_fp16 = conv(bias = layers_7_encoder_attn_o_proj_bias_to_fp16, dilations = obj_109_dilations_0, groups = obj_109_groups_0, pad = obj_109_pad_0, pad_type = obj_109_pad_type_0, strides = obj_109_strides_0, weight = layers_7_encoder_attn_o_proj_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("obj_109_cast_fp16")]; + tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_109_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; + tensor out_47_axes_0 = const()[name = tensor("out_47_axes_0"), val = tensor([1])]; + tensor var_1896_to_fp16 = const()[name = tensor("op_1896_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_1896_to_fp16, x = inputs_47_cast_fp16)[name = tensor("out_47_cast_fp16")]; + tensor input_75_gamma_0_to_fp16 = const()[name = tensor("input_75_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527487424)))]; + tensor input_75_beta_0_to_fp16 = const()[name = tensor("input_75_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527490048)))]; + tensor input_75_epsilon_0_to_fp16 = const()[name = tensor("input_75_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_75_cast_fp16 = batch_norm(beta = input_75_beta_0_to_fp16, epsilon = input_75_epsilon_0_to_fp16, gamma = input_75_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_47_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_77_pad_type_0 = const()[name = tensor("input_77_pad_type_0"), val = tensor("valid")]; + tensor input_77_strides_0 = const()[name = tensor("input_77_strides_0"), val = tensor([1, 1])]; + tensor input_77_pad_0 = const()[name = tensor("input_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_77_dilations_0 = const()[name = tensor("input_77_dilations_0"), val = tensor([1, 1])]; + tensor input_77_groups_0 = const()[name = tensor("input_77_groups_0"), val = tensor(1)]; + tensor layers_7_fc1_weight_to_fp16 = const()[name = tensor("layers_7_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527492672)))]; + tensor layers_7_fc1_bias_to_fp16 = const()[name = tensor("layers_7_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540599936)))]; + tensor input_77_cast_fp16 = conv(bias = layers_7_fc1_bias_to_fp16, dilations = input_77_dilations_0, groups = input_77_groups_0, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = input_77_strides_0, weight = layers_7_fc1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_mode_0 = const()[name = tensor("input_79_mode_0"), val = tensor("EXACT")]; + tensor input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor hidden_states_17_pad_type_0 = const()[name = tensor("hidden_states_17_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_17_strides_0 = const()[name = tensor("hidden_states_17_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_17_pad_0 = const()[name = tensor("hidden_states_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_17_dilations_0 = const()[name = tensor("hidden_states_17_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_17_groups_0 = const()[name = tensor("hidden_states_17_groups_0"), val = tensor(1)]; + tensor layers_7_fc2_weight_to_fp16 = const()[name = tensor("layers_7_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540610240)))]; + tensor layers_7_fc2_bias_to_fp16 = const()[name = tensor("layers_7_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553717504)))]; + tensor hidden_states_17_cast_fp16 = conv(bias = layers_7_fc2_bias_to_fp16, dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_7_fc2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("hidden_states_17_cast_fp16")]; + tensor inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = hidden_states_17_cast_fp16)[name = tensor("inputs_49_cast_fp16")]; + tensor var_1932 = const()[name = tensor("op_1932"), val = tensor(3)]; + tensor out_49_axes_0 = const()[name = tensor("out_49_axes_0"), val = tensor([1])]; + tensor var_1957_to_fp16 = const()[name = tensor("op_1957_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_1957_to_fp16, x = inputs_49_cast_fp16)[name = tensor("out_49_cast_fp16")]; + tensor obj_113_gamma_0_to_fp16 = const()[name = tensor("obj_113_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553720128)))]; + tensor obj_113_beta_0_to_fp16 = const()[name = tensor("obj_113_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553722752)))]; + tensor obj_113_epsilon_0_to_fp16 = const()[name = tensor("obj_113_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_113_cast_fp16 = batch_norm(beta = obj_113_beta_0_to_fp16, epsilon = obj_113_epsilon_0_to_fp16, gamma = obj_113_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_49_cast_fp16)[name = tensor("obj_113_cast_fp16")]; + tensor query_33_pad_type_0 = const()[name = tensor("query_33_pad_type_0"), val = tensor("valid")]; + tensor query_33_strides_0 = const()[name = tensor("query_33_strides_0"), val = tensor([1, 1])]; + tensor query_33_pad_0 = const()[name = tensor("query_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_33_dilations_0 = const()[name = tensor("query_33_dilations_0"), val = tensor([1, 1])]; + tensor query_33_groups_0 = const()[name = tensor("query_33_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553725376)))]; + tensor layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(557002240)))]; + tensor query_33_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_bias_to_fp16, dilations = query_33_dilations_0, groups = query_33_groups_0, pad = query_33_pad_0, pad_type = query_33_pad_type_0, strides = query_33_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("query_33_cast_fp16")]; + tensor current_key_17_pad_type_0 = const()[name = tensor("current_key_17_pad_type_0"), val = tensor("valid")]; + tensor current_key_17_strides_0 = const()[name = tensor("current_key_17_strides_0"), val = tensor([1, 1])]; + tensor current_key_17_pad_0 = const()[name = tensor("current_key_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_17_dilations_0 = const()[name = tensor("current_key_17_dilations_0"), val = tensor([1, 1])]; + tensor current_key_17_groups_0 = const()[name = tensor("current_key_17_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(557004864)))]; + tensor current_key_17_cast_fp16 = conv(dilations = current_key_17_dilations_0, groups = current_key_17_groups_0, pad = current_key_17_pad_0, pad_type = current_key_17_pad_type_0, strides = current_key_17_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("current_key_17_cast_fp16")]; + tensor current_value_17_pad_type_0 = const()[name = tensor("current_value_17_pad_type_0"), val = tensor("valid")]; + tensor current_value_17_strides_0 = const()[name = tensor("current_value_17_strides_0"), val = tensor([1, 1])]; + tensor current_value_17_pad_0 = const()[name = tensor("current_value_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_17_dilations_0 = const()[name = tensor("current_value_17_dilations_0"), val = tensor([1, 1])]; + tensor current_value_17_groups_0 = const()[name = tensor("current_value_17_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560281728)))]; + tensor layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563558592)))]; + tensor current_value_17_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_bias_to_fp16, dilations = current_value_17_dilations_0, groups = current_value_17_groups_0, pad = current_value_17_pad_0, pad_type = current_value_17_pad_type_0, strides = current_value_17_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("current_value_17_cast_fp16")]; + tensor var_1996_cast_fp16 = mul(x = var_103_cast_fp16_8, y = var_239_cast_fp16)[name = tensor("op_1996_cast_fp16")]; + tensor var_1997_cast_fp16 = mul(x = current_key_17_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_1997_cast_fp16")]; + tensor key_33_cast_fp16 = add(x = var_1996_cast_fp16, y = var_1997_cast_fp16)[name = tensor("key_33_cast_fp16")]; + tensor var_2000_cast_fp16 = mul(x = var_138_cast_fp16_8, y = var_239_cast_fp16)[name = tensor("op_2000_cast_fp16")]; + tensor var_2001_cast_fp16 = mul(x = current_value_17_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_2001_cast_fp16")]; + tensor value_33_cast_fp16 = add(x = var_2000_cast_fp16, y = var_2001_cast_fp16)[name = tensor("value_33_cast_fp16")]; + tensor var_2005 = const()[name = tensor("op_2005"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_33_cast_fp16 = reshape(shape = var_2005, x = query_33_cast_fp16)[name = tensor("mh_q_33_cast_fp16")]; + tensor var_2007_to_fp16 = const()[name = tensor("op_2007_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2008_cast_fp16 = mul(x = mh_q_33_cast_fp16, y = var_2007_to_fp16)[name = tensor("op_2008_cast_fp16")]; + tensor var_2011 = const()[name = tensor("op_2011"), val = tensor([1, 20, 64, 448])]; + tensor var_2012_cast_fp16 = reshape(shape = var_2011, x = key_33_cast_fp16)[name = tensor("op_2012_cast_fp16")]; + tensor mh_w_49_transpose_x_0 = const()[name = tensor("mh_w_49_transpose_x_0"), val = tensor(true)]; + tensor mh_w_49_transpose_y_0 = const()[name = tensor("mh_w_49_transpose_y_0"), val = tensor(false)]; + tensor mh_w_49_cast_fp16 = matmul(transpose_x = mh_w_49_transpose_x_0, transpose_y = mh_w_49_transpose_y_0, x = var_2008_cast_fp16, y = var_2012_cast_fp16)[name = tensor("mh_w_49_cast_fp16")]; + tensor mh_w_51_cast_fp16 = add(x = mh_w_49_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_51_cast_fp16")]; + tensor var_2020_cast_fp16 = softmax(axis = var_1932, x = mh_w_51_cast_fp16)[name = tensor("op_2020_cast_fp16")]; + tensor var_2021 = const()[name = tensor("op_2021"), val = tensor([1, 20, 64, 448])]; + tensor var_2022_cast_fp16 = reshape(shape = var_2021, x = value_33_cast_fp16)[name = tensor("op_2022_cast_fp16")]; + tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; + tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(true)]; + tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = var_2022_cast_fp16, y = var_2020_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_2025 = const()[name = tensor("op_2025"), val = tensor([1, 1280, 1, 1])]; + tensor input_81_cast_fp16 = reshape(shape = var_2025, x = attn_33_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor obj_119_pad_type_0 = const()[name = tensor("obj_119_pad_type_0"), val = tensor("valid")]; + tensor obj_119_strides_0 = const()[name = tensor("obj_119_strides_0"), val = tensor([1, 1])]; + tensor obj_119_pad_0 = const()[name = tensor("obj_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_119_dilations_0 = const()[name = tensor("obj_119_dilations_0"), val = tensor([1, 1])]; + tensor obj_119_groups_0 = const()[name = tensor("obj_119_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563561216)))]; + tensor layers_8_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566838080)))]; + tensor obj_119_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_bias_to_fp16, dilations = obj_119_dilations_0, groups = obj_119_groups_0, pad = obj_119_pad_0, pad_type = obj_119_pad_type_0, strides = obj_119_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("obj_119_cast_fp16")]; + tensor inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_119_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; + tensor out_51_axes_0 = const()[name = tensor("out_51_axes_0"), val = tensor([1])]; + tensor var_2047_to_fp16 = const()[name = tensor("op_2047_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_2047_to_fp16, x = inputs_51_cast_fp16)[name = tensor("out_51_cast_fp16")]; + tensor obj_121_gamma_0_to_fp16 = const()[name = tensor("obj_121_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566840704)))]; + tensor obj_121_beta_0_to_fp16 = const()[name = tensor("obj_121_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566843328)))]; + tensor obj_121_epsilon_0_to_fp16 = const()[name = tensor("obj_121_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_121_cast_fp16 = batch_norm(beta = obj_121_beta_0_to_fp16, epsilon = obj_121_epsilon_0_to_fp16, gamma = obj_121_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_51_cast_fp16)[name = tensor("obj_121_cast_fp16")]; + tensor query_35_pad_type_0 = const()[name = tensor("query_35_pad_type_0"), val = tensor("valid")]; + tensor query_35_strides_0 = const()[name = tensor("query_35_strides_0"), val = tensor([1, 1])]; + tensor query_35_pad_0 = const()[name = tensor("query_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_35_dilations_0 = const()[name = tensor("query_35_dilations_0"), val = tensor([1, 1])]; + tensor query_35_groups_0 = const()[name = tensor("query_35_groups_0"), val = tensor(1)]; + tensor layers_8_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566845952)))]; + tensor layers_8_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570122816)))]; + tensor query_35_cast_fp16 = conv(bias = layers_8_encoder_attn_q_proj_bias_to_fp16, dilations = query_35_dilations_0, groups = query_35_groups_0, pad = query_35_pad_0, pad_type = query_35_pad_type_0, strides = query_35_strides_0, weight = layers_8_encoder_attn_q_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("query_35_cast_fp16")]; + tensor key_35_pad_type_0 = const()[name = tensor("key_35_pad_type_0"), val = tensor("valid")]; + tensor key_35_strides_0 = const()[name = tensor("key_35_strides_0"), val = tensor([1, 1])]; + tensor key_35_pad_0 = const()[name = tensor("key_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_35_dilations_0 = const()[name = tensor("key_35_dilations_0"), val = tensor([1, 1])]; + tensor key_35_groups_0 = const()[name = tensor("key_35_groups_0"), val = tensor(1)]; + tensor layers_8_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570125440)))]; + tensor key_35_cast_fp16 = conv(dilations = key_35_dilations_0, groups = key_35_groups_0, pad = key_35_pad_0, pad_type = key_35_pad_type_0, strides = key_35_strides_0, weight = layers_8_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_35_cast_fp16")]; + tensor value_35_pad_type_0 = const()[name = tensor("value_35_pad_type_0"), val = tensor("valid")]; + tensor value_35_strides_0 = const()[name = tensor("value_35_strides_0"), val = tensor([1, 1])]; + tensor value_35_pad_0 = const()[name = tensor("value_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_35_dilations_0 = const()[name = tensor("value_35_dilations_0"), val = tensor([1, 1])]; + tensor value_35_groups_0 = const()[name = tensor("value_35_groups_0"), val = tensor(1)]; + tensor layers_8_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(573402304)))]; + tensor layers_8_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(576679168)))]; + tensor value_35_cast_fp16 = conv(bias = layers_8_encoder_attn_v_proj_bias_to_fp16, dilations = value_35_dilations_0, groups = value_35_groups_0, pad = value_35_pad_0, pad_type = value_35_pad_type_0, strides = value_35_strides_0, weight = layers_8_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_35_cast_fp16")]; + tensor var_2083 = const()[name = tensor("op_2083"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_35_cast_fp16 = reshape(shape = var_2083, x = query_35_cast_fp16)[name = tensor("mh_q_35_cast_fp16")]; + tensor var_2085_to_fp16 = const()[name = tensor("op_2085_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2086_cast_fp16 = mul(x = mh_q_35_cast_fp16, y = var_2085_to_fp16)[name = tensor("op_2086_cast_fp16")]; + tensor var_2089 = const()[name = tensor("op_2089"), val = tensor([1, 20, 64, 1500])]; + tensor var_2090_cast_fp16 = reshape(shape = var_2089, x = key_35_cast_fp16)[name = tensor("op_2090_cast_fp16")]; + tensor mh_w_53_transpose_x_0 = const()[name = tensor("mh_w_53_transpose_x_0"), val = tensor(true)]; + tensor mh_w_53_transpose_y_0 = const()[name = tensor("mh_w_53_transpose_y_0"), val = tensor(false)]; + tensor mh_w_53_cast_fp16 = matmul(transpose_x = mh_w_53_transpose_x_0, transpose_y = mh_w_53_transpose_y_0, x = var_2086_cast_fp16, y = var_2090_cast_fp16)[name = tensor("mh_w_53_cast_fp16")]; + tensor obj_125_cast_fp16 = softmax(axis = var_1932, x = mh_w_53_cast_fp16)[name = tensor("obj_125_cast_fp16")]; + tensor var_2094 = const()[name = tensor("op_2094"), val = tensor([1, 20, 64, 1500])]; + tensor var_2095_cast_fp16 = reshape(shape = var_2094, x = value_35_cast_fp16)[name = tensor("op_2095_cast_fp16")]; + tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; + tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(true)]; + tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = var_2095_cast_fp16, y = obj_125_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_2098 = const()[name = tensor("op_2098"), val = tensor([1, 1280, 1, 1])]; + tensor input_83_cast_fp16 = reshape(shape = var_2098, x = attn_35_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor obj_123_pad_type_0 = const()[name = tensor("obj_123_pad_type_0"), val = tensor("valid")]; + tensor obj_123_strides_0 = const()[name = tensor("obj_123_strides_0"), val = tensor([1, 1])]; + tensor obj_123_pad_0 = const()[name = tensor("obj_123_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_123_dilations_0 = const()[name = tensor("obj_123_dilations_0"), val = tensor([1, 1])]; + tensor obj_123_groups_0 = const()[name = tensor("obj_123_groups_0"), val = tensor(1)]; + tensor layers_8_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(576681792)))]; + tensor layers_8_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579958656)))]; + tensor obj_123_cast_fp16 = conv(bias = layers_8_encoder_attn_o_proj_bias_to_fp16, dilations = obj_123_dilations_0, groups = obj_123_groups_0, pad = obj_123_pad_0, pad_type = obj_123_pad_type_0, strides = obj_123_strides_0, weight = layers_8_encoder_attn_o_proj_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("obj_123_cast_fp16")]; + tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = obj_123_cast_fp16)[name = tensor("inputs_53_cast_fp16")]; + tensor out_53_axes_0 = const()[name = tensor("out_53_axes_0"), val = tensor([1])]; + tensor var_2116_to_fp16 = const()[name = tensor("op_2116_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_2116_to_fp16, x = inputs_53_cast_fp16)[name = tensor("out_53_cast_fp16")]; + tensor input_85_gamma_0_to_fp16 = const()[name = tensor("input_85_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579961280)))]; + tensor input_85_beta_0_to_fp16 = const()[name = tensor("input_85_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579963904)))]; + tensor input_85_epsilon_0_to_fp16 = const()[name = tensor("input_85_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_85_cast_fp16 = batch_norm(beta = input_85_beta_0_to_fp16, epsilon = input_85_epsilon_0_to_fp16, gamma = input_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_53_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor input_87_pad_type_0 = const()[name = tensor("input_87_pad_type_0"), val = tensor("valid")]; + tensor input_87_strides_0 = const()[name = tensor("input_87_strides_0"), val = tensor([1, 1])]; + tensor input_87_pad_0 = const()[name = tensor("input_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_87_dilations_0 = const()[name = tensor("input_87_dilations_0"), val = tensor([1, 1])]; + tensor input_87_groups_0 = const()[name = tensor("input_87_groups_0"), val = tensor(1)]; + tensor layers_8_fc1_weight_to_fp16 = const()[name = tensor("layers_8_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579966528)))]; + tensor layers_8_fc1_bias_to_fp16 = const()[name = tensor("layers_8_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(593073792)))]; + tensor input_87_cast_fp16 = conv(bias = layers_8_fc1_bias_to_fp16, dilations = input_87_dilations_0, groups = input_87_groups_0, pad = input_87_pad_0, pad_type = input_87_pad_type_0, strides = input_87_strides_0, weight = layers_8_fc1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_mode_0 = const()[name = tensor("input_89_mode_0"), val = tensor("EXACT")]; + tensor input_89_cast_fp16 = gelu(mode = input_89_mode_0, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor hidden_states_19_pad_type_0 = const()[name = tensor("hidden_states_19_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_19_strides_0 = const()[name = tensor("hidden_states_19_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_19_pad_0 = const()[name = tensor("hidden_states_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_19_dilations_0 = const()[name = tensor("hidden_states_19_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_19_groups_0 = const()[name = tensor("hidden_states_19_groups_0"), val = tensor(1)]; + tensor layers_8_fc2_weight_to_fp16 = const()[name = tensor("layers_8_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(593084096)))]; + tensor layers_8_fc2_bias_to_fp16 = const()[name = tensor("layers_8_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(606191360)))]; + tensor hidden_states_19_cast_fp16 = conv(bias = layers_8_fc2_bias_to_fp16, dilations = hidden_states_19_dilations_0, groups = hidden_states_19_groups_0, pad = hidden_states_19_pad_0, pad_type = hidden_states_19_pad_type_0, strides = hidden_states_19_strides_0, weight = layers_8_fc2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("hidden_states_19_cast_fp16")]; + tensor inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = hidden_states_19_cast_fp16)[name = tensor("inputs_55_cast_fp16")]; + tensor var_2151 = const()[name = tensor("op_2151"), val = tensor(3)]; + tensor out_55_axes_0 = const()[name = tensor("out_55_axes_0"), val = tensor([1])]; + tensor var_2176_to_fp16 = const()[name = tensor("op_2176_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_2176_to_fp16, x = inputs_55_cast_fp16)[name = tensor("out_55_cast_fp16")]; + tensor obj_127_gamma_0_to_fp16 = const()[name = tensor("obj_127_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(606193984)))]; + tensor obj_127_beta_0_to_fp16 = const()[name = tensor("obj_127_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(606196608)))]; + tensor obj_127_epsilon_0_to_fp16 = const()[name = tensor("obj_127_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_127_cast_fp16 = batch_norm(beta = obj_127_beta_0_to_fp16, epsilon = obj_127_epsilon_0_to_fp16, gamma = obj_127_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_55_cast_fp16)[name = tensor("obj_127_cast_fp16")]; + tensor query_37_pad_type_0 = const()[name = tensor("query_37_pad_type_0"), val = tensor("valid")]; + tensor query_37_strides_0 = const()[name = tensor("query_37_strides_0"), val = tensor([1, 1])]; + tensor query_37_pad_0 = const()[name = tensor("query_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_37_dilations_0 = const()[name = tensor("query_37_dilations_0"), val = tensor([1, 1])]; + tensor query_37_groups_0 = const()[name = tensor("query_37_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(606199232)))]; + tensor layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(609476096)))]; + tensor query_37_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_bias_to_fp16, dilations = query_37_dilations_0, groups = query_37_groups_0, pad = query_37_pad_0, pad_type = query_37_pad_type_0, strides = query_37_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("query_37_cast_fp16")]; + tensor current_key_19_pad_type_0 = const()[name = tensor("current_key_19_pad_type_0"), val = tensor("valid")]; + tensor current_key_19_strides_0 = const()[name = tensor("current_key_19_strides_0"), val = tensor([1, 1])]; + tensor current_key_19_pad_0 = const()[name = tensor("current_key_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_19_dilations_0 = const()[name = tensor("current_key_19_dilations_0"), val = tensor([1, 1])]; + tensor current_key_19_groups_0 = const()[name = tensor("current_key_19_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(609478720)))]; + tensor current_key_19_cast_fp16 = conv(dilations = current_key_19_dilations_0, groups = current_key_19_groups_0, pad = current_key_19_pad_0, pad_type = current_key_19_pad_type_0, strides = current_key_19_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("current_key_19_cast_fp16")]; + tensor current_value_19_pad_type_0 = const()[name = tensor("current_value_19_pad_type_0"), val = tensor("valid")]; + tensor current_value_19_strides_0 = const()[name = tensor("current_value_19_strides_0"), val = tensor([1, 1])]; + tensor current_value_19_pad_0 = const()[name = tensor("current_value_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_19_dilations_0 = const()[name = tensor("current_value_19_dilations_0"), val = tensor([1, 1])]; + tensor current_value_19_groups_0 = const()[name = tensor("current_value_19_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(612755584)))]; + tensor layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(616032448)))]; + tensor current_value_19_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_bias_to_fp16, dilations = current_value_19_dilations_0, groups = current_value_19_groups_0, pad = current_value_19_pad_0, pad_type = current_value_19_pad_type_0, strides = current_value_19_strides_0, weight = layers_9_self_attn_v_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("current_value_19_cast_fp16")]; + tensor var_2215_cast_fp16 = mul(x = var_103_cast_fp16_9, y = var_239_cast_fp16)[name = tensor("op_2215_cast_fp16")]; + tensor var_2216_cast_fp16 = mul(x = current_key_19_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_2216_cast_fp16")]; + tensor key_37_cast_fp16 = add(x = var_2215_cast_fp16, y = var_2216_cast_fp16)[name = tensor("key_37_cast_fp16")]; + tensor var_2219_cast_fp16 = mul(x = var_138_cast_fp16_9, y = var_239_cast_fp16)[name = tensor("op_2219_cast_fp16")]; + tensor var_2220_cast_fp16 = mul(x = current_value_19_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_2220_cast_fp16")]; + tensor value_37_cast_fp16 = add(x = var_2219_cast_fp16, y = var_2220_cast_fp16)[name = tensor("value_37_cast_fp16")]; + tensor var_2224 = const()[name = tensor("op_2224"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_37_cast_fp16 = reshape(shape = var_2224, x = query_37_cast_fp16)[name = tensor("mh_q_37_cast_fp16")]; + tensor var_2226_to_fp16 = const()[name = tensor("op_2226_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2227_cast_fp16 = mul(x = mh_q_37_cast_fp16, y = var_2226_to_fp16)[name = tensor("op_2227_cast_fp16")]; + tensor var_2230 = const()[name = tensor("op_2230"), val = tensor([1, 20, 64, 448])]; + tensor var_2231_cast_fp16 = reshape(shape = var_2230, x = key_37_cast_fp16)[name = tensor("op_2231_cast_fp16")]; + tensor mh_w_55_transpose_x_0 = const()[name = tensor("mh_w_55_transpose_x_0"), val = tensor(true)]; + tensor mh_w_55_transpose_y_0 = const()[name = tensor("mh_w_55_transpose_y_0"), val = tensor(false)]; + tensor mh_w_55_cast_fp16 = matmul(transpose_x = mh_w_55_transpose_x_0, transpose_y = mh_w_55_transpose_y_0, x = var_2227_cast_fp16, y = var_2231_cast_fp16)[name = tensor("mh_w_55_cast_fp16")]; + tensor mh_w_57_cast_fp16 = add(x = mh_w_55_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_57_cast_fp16")]; + tensor var_2239_cast_fp16 = softmax(axis = var_2151, x = mh_w_57_cast_fp16)[name = tensor("op_2239_cast_fp16")]; + tensor var_2240 = const()[name = tensor("op_2240"), val = tensor([1, 20, 64, 448])]; + tensor var_2241_cast_fp16 = reshape(shape = var_2240, x = value_37_cast_fp16)[name = tensor("op_2241_cast_fp16")]; + tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; + tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(true)]; + tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = var_2241_cast_fp16, y = var_2239_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_2244 = const()[name = tensor("op_2244"), val = tensor([1, 1280, 1, 1])]; + tensor input_91_cast_fp16 = reshape(shape = var_2244, x = attn_37_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor obj_133_pad_type_0 = const()[name = tensor("obj_133_pad_type_0"), val = tensor("valid")]; + tensor obj_133_strides_0 = const()[name = tensor("obj_133_strides_0"), val = tensor([1, 1])]; + tensor obj_133_pad_0 = const()[name = tensor("obj_133_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_133_dilations_0 = const()[name = tensor("obj_133_dilations_0"), val = tensor([1, 1])]; + tensor obj_133_groups_0 = const()[name = tensor("obj_133_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(616035072)))]; + tensor layers_9_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(619311936)))]; + tensor obj_133_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_bias_to_fp16, dilations = obj_133_dilations_0, groups = obj_133_groups_0, pad = obj_133_pad_0, pad_type = obj_133_pad_type_0, strides = obj_133_strides_0, weight = layers_9_self_attn_o_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("obj_133_cast_fp16")]; + tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = obj_133_cast_fp16)[name = tensor("inputs_57_cast_fp16")]; + tensor out_57_axes_0 = const()[name = tensor("out_57_axes_0"), val = tensor([1])]; + tensor var_2266_to_fp16 = const()[name = tensor("op_2266_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_57_cast_fp16 = layer_norm(axes = out_57_axes_0, epsilon = var_2266_to_fp16, x = inputs_57_cast_fp16)[name = tensor("out_57_cast_fp16")]; + tensor obj_135_gamma_0_to_fp16 = const()[name = tensor("obj_135_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(619314560)))]; + tensor obj_135_beta_0_to_fp16 = const()[name = tensor("obj_135_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(619317184)))]; + tensor obj_135_epsilon_0_to_fp16 = const()[name = tensor("obj_135_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_135_cast_fp16 = batch_norm(beta = obj_135_beta_0_to_fp16, epsilon = obj_135_epsilon_0_to_fp16, gamma = obj_135_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_57_cast_fp16)[name = tensor("obj_135_cast_fp16")]; + tensor query_39_pad_type_0 = const()[name = tensor("query_39_pad_type_0"), val = tensor("valid")]; + tensor query_39_strides_0 = const()[name = tensor("query_39_strides_0"), val = tensor([1, 1])]; + tensor query_39_pad_0 = const()[name = tensor("query_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_39_dilations_0 = const()[name = tensor("query_39_dilations_0"), val = tensor([1, 1])]; + tensor query_39_groups_0 = const()[name = tensor("query_39_groups_0"), val = tensor(1)]; + tensor layers_9_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(619319808)))]; + tensor layers_9_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(622596672)))]; + tensor query_39_cast_fp16 = conv(bias = layers_9_encoder_attn_q_proj_bias_to_fp16, dilations = query_39_dilations_0, groups = query_39_groups_0, pad = query_39_pad_0, pad_type = query_39_pad_type_0, strides = query_39_strides_0, weight = layers_9_encoder_attn_q_proj_weight_to_fp16, x = obj_135_cast_fp16)[name = tensor("query_39_cast_fp16")]; + tensor key_39_pad_type_0 = const()[name = tensor("key_39_pad_type_0"), val = tensor("valid")]; + tensor key_39_strides_0 = const()[name = tensor("key_39_strides_0"), val = tensor([1, 1])]; + tensor key_39_pad_0 = const()[name = tensor("key_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_39_dilations_0 = const()[name = tensor("key_39_dilations_0"), val = tensor([1, 1])]; + tensor key_39_groups_0 = const()[name = tensor("key_39_groups_0"), val = tensor(1)]; + tensor layers_9_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(622599296)))]; + tensor key_39_cast_fp16 = conv(dilations = key_39_dilations_0, groups = key_39_groups_0, pad = key_39_pad_0, pad_type = key_39_pad_type_0, strides = key_39_strides_0, weight = layers_9_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_39_cast_fp16")]; + tensor value_39_pad_type_0 = const()[name = tensor("value_39_pad_type_0"), val = tensor("valid")]; + tensor value_39_strides_0 = const()[name = tensor("value_39_strides_0"), val = tensor([1, 1])]; + tensor value_39_pad_0 = const()[name = tensor("value_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_39_dilations_0 = const()[name = tensor("value_39_dilations_0"), val = tensor([1, 1])]; + tensor value_39_groups_0 = const()[name = tensor("value_39_groups_0"), val = tensor(1)]; + tensor layers_9_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(625876160)))]; + tensor layers_9_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(629153024)))]; + tensor value_39_cast_fp16 = conv(bias = layers_9_encoder_attn_v_proj_bias_to_fp16, dilations = value_39_dilations_0, groups = value_39_groups_0, pad = value_39_pad_0, pad_type = value_39_pad_type_0, strides = value_39_strides_0, weight = layers_9_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_39_cast_fp16")]; + tensor var_2302 = const()[name = tensor("op_2302"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_39_cast_fp16 = reshape(shape = var_2302, x = query_39_cast_fp16)[name = tensor("mh_q_39_cast_fp16")]; + tensor var_2304_to_fp16 = const()[name = tensor("op_2304_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2305_cast_fp16 = mul(x = mh_q_39_cast_fp16, y = var_2304_to_fp16)[name = tensor("op_2305_cast_fp16")]; + tensor var_2308 = const()[name = tensor("op_2308"), val = tensor([1, 20, 64, 1500])]; + tensor var_2309_cast_fp16 = reshape(shape = var_2308, x = key_39_cast_fp16)[name = tensor("op_2309_cast_fp16")]; + tensor mh_w_59_transpose_x_0 = const()[name = tensor("mh_w_59_transpose_x_0"), val = tensor(true)]; + tensor mh_w_59_transpose_y_0 = const()[name = tensor("mh_w_59_transpose_y_0"), val = tensor(false)]; + tensor mh_w_59_cast_fp16 = matmul(transpose_x = mh_w_59_transpose_x_0, transpose_y = mh_w_59_transpose_y_0, x = var_2305_cast_fp16, y = var_2309_cast_fp16)[name = tensor("mh_w_59_cast_fp16")]; + tensor obj_139_cast_fp16 = softmax(axis = var_2151, x = mh_w_59_cast_fp16)[name = tensor("obj_139_cast_fp16")]; + tensor var_2313 = const()[name = tensor("op_2313"), val = tensor([1, 20, 64, 1500])]; + tensor var_2314_cast_fp16 = reshape(shape = var_2313, x = value_39_cast_fp16)[name = tensor("op_2314_cast_fp16")]; + tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; + tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(true)]; + tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = var_2314_cast_fp16, y = obj_139_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_2317 = const()[name = tensor("op_2317"), val = tensor([1, 1280, 1, 1])]; + tensor input_93_cast_fp16 = reshape(shape = var_2317, x = attn_39_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor obj_137_pad_type_0 = const()[name = tensor("obj_137_pad_type_0"), val = tensor("valid")]; + tensor obj_137_strides_0 = const()[name = tensor("obj_137_strides_0"), val = tensor([1, 1])]; + tensor obj_137_pad_0 = const()[name = tensor("obj_137_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_137_dilations_0 = const()[name = tensor("obj_137_dilations_0"), val = tensor([1, 1])]; + tensor obj_137_groups_0 = const()[name = tensor("obj_137_groups_0"), val = tensor(1)]; + tensor layers_9_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(629155648)))]; + tensor layers_9_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(632432512)))]; + tensor obj_137_cast_fp16 = conv(bias = layers_9_encoder_attn_o_proj_bias_to_fp16, dilations = obj_137_dilations_0, groups = obj_137_groups_0, pad = obj_137_pad_0, pad_type = obj_137_pad_type_0, strides = obj_137_strides_0, weight = layers_9_encoder_attn_o_proj_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("obj_137_cast_fp16")]; + tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_137_cast_fp16)[name = tensor("inputs_59_cast_fp16")]; + tensor out_59_axes_0 = const()[name = tensor("out_59_axes_0"), val = tensor([1])]; + tensor var_2335_to_fp16 = const()[name = tensor("op_2335_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_59_cast_fp16 = layer_norm(axes = out_59_axes_0, epsilon = var_2335_to_fp16, x = inputs_59_cast_fp16)[name = tensor("out_59_cast_fp16")]; + tensor input_95_gamma_0_to_fp16 = const()[name = tensor("input_95_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(632435136)))]; + tensor input_95_beta_0_to_fp16 = const()[name = tensor("input_95_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(632437760)))]; + tensor input_95_epsilon_0_to_fp16 = const()[name = tensor("input_95_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_95_cast_fp16 = batch_norm(beta = input_95_beta_0_to_fp16, epsilon = input_95_epsilon_0_to_fp16, gamma = input_95_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_59_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor input_97_pad_type_0 = const()[name = tensor("input_97_pad_type_0"), val = tensor("valid")]; + tensor input_97_strides_0 = const()[name = tensor("input_97_strides_0"), val = tensor([1, 1])]; + tensor input_97_pad_0 = const()[name = tensor("input_97_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_97_dilations_0 = const()[name = tensor("input_97_dilations_0"), val = tensor([1, 1])]; + tensor input_97_groups_0 = const()[name = tensor("input_97_groups_0"), val = tensor(1)]; + tensor layers_9_fc1_weight_to_fp16 = const()[name = tensor("layers_9_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(632440384)))]; + tensor layers_9_fc1_bias_to_fp16 = const()[name = tensor("layers_9_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(645547648)))]; + tensor input_97_cast_fp16 = conv(bias = layers_9_fc1_bias_to_fp16, dilations = input_97_dilations_0, groups = input_97_groups_0, pad = input_97_pad_0, pad_type = input_97_pad_type_0, strides = input_97_strides_0, weight = layers_9_fc1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_mode_0 = const()[name = tensor("input_99_mode_0"), val = tensor("EXACT")]; + tensor input_99_cast_fp16 = gelu(mode = input_99_mode_0, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor hidden_states_21_pad_type_0 = const()[name = tensor("hidden_states_21_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_21_strides_0 = const()[name = tensor("hidden_states_21_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_21_pad_0 = const()[name = tensor("hidden_states_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_21_dilations_0 = const()[name = tensor("hidden_states_21_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_21_groups_0 = const()[name = tensor("hidden_states_21_groups_0"), val = tensor(1)]; + tensor layers_9_fc2_weight_to_fp16 = const()[name = tensor("layers_9_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(645557952)))]; + tensor layers_9_fc2_bias_to_fp16 = const()[name = tensor("layers_9_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(658665216)))]; + tensor hidden_states_21_cast_fp16 = conv(bias = layers_9_fc2_bias_to_fp16, dilations = hidden_states_21_dilations_0, groups = hidden_states_21_groups_0, pad = hidden_states_21_pad_0, pad_type = hidden_states_21_pad_type_0, strides = hidden_states_21_strides_0, weight = layers_9_fc2_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("hidden_states_21_cast_fp16")]; + tensor inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = hidden_states_21_cast_fp16)[name = tensor("inputs_61_cast_fp16")]; + tensor var_2370 = const()[name = tensor("op_2370"), val = tensor(3)]; + tensor out_61_axes_0 = const()[name = tensor("out_61_axes_0"), val = tensor([1])]; + tensor var_2395_to_fp16 = const()[name = tensor("op_2395_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_61_cast_fp16 = layer_norm(axes = out_61_axes_0, epsilon = var_2395_to_fp16, x = inputs_61_cast_fp16)[name = tensor("out_61_cast_fp16")]; + tensor obj_141_gamma_0_to_fp16 = const()[name = tensor("obj_141_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(658667840)))]; + tensor obj_141_beta_0_to_fp16 = const()[name = tensor("obj_141_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(658670464)))]; + tensor obj_141_epsilon_0_to_fp16 = const()[name = tensor("obj_141_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_141_cast_fp16 = batch_norm(beta = obj_141_beta_0_to_fp16, epsilon = obj_141_epsilon_0_to_fp16, gamma = obj_141_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_61_cast_fp16)[name = tensor("obj_141_cast_fp16")]; + tensor query_41_pad_type_0 = const()[name = tensor("query_41_pad_type_0"), val = tensor("valid")]; + tensor query_41_strides_0 = const()[name = tensor("query_41_strides_0"), val = tensor([1, 1])]; + tensor query_41_pad_0 = const()[name = tensor("query_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_41_dilations_0 = const()[name = tensor("query_41_dilations_0"), val = tensor([1, 1])]; + tensor query_41_groups_0 = const()[name = tensor("query_41_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(658673088)))]; + tensor layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(661949952)))]; + tensor query_41_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_bias_to_fp16, dilations = query_41_dilations_0, groups = query_41_groups_0, pad = query_41_pad_0, pad_type = query_41_pad_type_0, strides = query_41_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("query_41_cast_fp16")]; + tensor current_key_21_pad_type_0 = const()[name = tensor("current_key_21_pad_type_0"), val = tensor("valid")]; + tensor current_key_21_strides_0 = const()[name = tensor("current_key_21_strides_0"), val = tensor([1, 1])]; + tensor current_key_21_pad_0 = const()[name = tensor("current_key_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_21_dilations_0 = const()[name = tensor("current_key_21_dilations_0"), val = tensor([1, 1])]; + tensor current_key_21_groups_0 = const()[name = tensor("current_key_21_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(661952576)))]; + tensor current_key_21_cast_fp16 = conv(dilations = current_key_21_dilations_0, groups = current_key_21_groups_0, pad = current_key_21_pad_0, pad_type = current_key_21_pad_type_0, strides = current_key_21_strides_0, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("current_key_21_cast_fp16")]; + tensor current_value_21_pad_type_0 = const()[name = tensor("current_value_21_pad_type_0"), val = tensor("valid")]; + tensor current_value_21_strides_0 = const()[name = tensor("current_value_21_strides_0"), val = tensor([1, 1])]; + tensor current_value_21_pad_0 = const()[name = tensor("current_value_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_21_dilations_0 = const()[name = tensor("current_value_21_dilations_0"), val = tensor([1, 1])]; + tensor current_value_21_groups_0 = const()[name = tensor("current_value_21_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(665229440)))]; + tensor layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(668506304)))]; + tensor current_value_21_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_bias_to_fp16, dilations = current_value_21_dilations_0, groups = current_value_21_groups_0, pad = current_value_21_pad_0, pad_type = current_value_21_pad_type_0, strides = current_value_21_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("current_value_21_cast_fp16")]; + tensor var_2434_cast_fp16 = mul(x = var_103_cast_fp16_10, y = var_239_cast_fp16)[name = tensor("op_2434_cast_fp16")]; + tensor var_2435_cast_fp16 = mul(x = current_key_21_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor key_41_cast_fp16 = add(x = var_2434_cast_fp16, y = var_2435_cast_fp16)[name = tensor("key_41_cast_fp16")]; + tensor var_2438_cast_fp16 = mul(x = var_138_cast_fp16_10, y = var_239_cast_fp16)[name = tensor("op_2438_cast_fp16")]; + tensor var_2439_cast_fp16 = mul(x = current_value_21_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_2439_cast_fp16")]; + tensor value_41_cast_fp16 = add(x = var_2438_cast_fp16, y = var_2439_cast_fp16)[name = tensor("value_41_cast_fp16")]; + tensor var_2443 = const()[name = tensor("op_2443"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_41_cast_fp16 = reshape(shape = var_2443, x = query_41_cast_fp16)[name = tensor("mh_q_41_cast_fp16")]; + tensor var_2445_to_fp16 = const()[name = tensor("op_2445_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2446_cast_fp16 = mul(x = mh_q_41_cast_fp16, y = var_2445_to_fp16)[name = tensor("op_2446_cast_fp16")]; + tensor var_2449 = const()[name = tensor("op_2449"), val = tensor([1, 20, 64, 448])]; + tensor var_2450_cast_fp16 = reshape(shape = var_2449, x = key_41_cast_fp16)[name = tensor("op_2450_cast_fp16")]; + tensor mh_w_61_transpose_x_0 = const()[name = tensor("mh_w_61_transpose_x_0"), val = tensor(true)]; + tensor mh_w_61_transpose_y_0 = const()[name = tensor("mh_w_61_transpose_y_0"), val = tensor(false)]; + tensor mh_w_61_cast_fp16 = matmul(transpose_x = mh_w_61_transpose_x_0, transpose_y = mh_w_61_transpose_y_0, x = var_2446_cast_fp16, y = var_2450_cast_fp16)[name = tensor("mh_w_61_cast_fp16")]; + tensor mh_w_63_cast_fp16 = add(x = mh_w_61_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_63_cast_fp16")]; + tensor var_2458_cast_fp16 = softmax(axis = var_2370, x = mh_w_63_cast_fp16)[name = tensor("op_2458_cast_fp16")]; + tensor var_2459 = const()[name = tensor("op_2459"), val = tensor([1, 20, 64, 448])]; + tensor var_2460_cast_fp16 = reshape(shape = var_2459, x = value_41_cast_fp16)[name = tensor("op_2460_cast_fp16")]; + tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; + tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(true)]; + tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = var_2460_cast_fp16, y = var_2458_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_2463 = const()[name = tensor("op_2463"), val = tensor([1, 1280, 1, 1])]; + tensor input_101_cast_fp16 = reshape(shape = var_2463, x = attn_41_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor obj_147_pad_type_0 = const()[name = tensor("obj_147_pad_type_0"), val = tensor("valid")]; + tensor obj_147_strides_0 = const()[name = tensor("obj_147_strides_0"), val = tensor([1, 1])]; + tensor obj_147_pad_0 = const()[name = tensor("obj_147_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_147_dilations_0 = const()[name = tensor("obj_147_dilations_0"), val = tensor([1, 1])]; + tensor obj_147_groups_0 = const()[name = tensor("obj_147_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(668508928)))]; + tensor layers_10_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(671785792)))]; + tensor obj_147_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_bias_to_fp16, dilations = obj_147_dilations_0, groups = obj_147_groups_0, pad = obj_147_pad_0, pad_type = obj_147_pad_type_0, strides = obj_147_strides_0, weight = layers_10_self_attn_o_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("obj_147_cast_fp16")]; + tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_147_cast_fp16)[name = tensor("inputs_63_cast_fp16")]; + tensor out_63_axes_0 = const()[name = tensor("out_63_axes_0"), val = tensor([1])]; + tensor var_2485_to_fp16 = const()[name = tensor("op_2485_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_63_cast_fp16 = layer_norm(axes = out_63_axes_0, epsilon = var_2485_to_fp16, x = inputs_63_cast_fp16)[name = tensor("out_63_cast_fp16")]; + tensor obj_149_gamma_0_to_fp16 = const()[name = tensor("obj_149_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(671788416)))]; + tensor obj_149_beta_0_to_fp16 = const()[name = tensor("obj_149_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(671791040)))]; + tensor obj_149_epsilon_0_to_fp16 = const()[name = tensor("obj_149_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_149_cast_fp16 = batch_norm(beta = obj_149_beta_0_to_fp16, epsilon = obj_149_epsilon_0_to_fp16, gamma = obj_149_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_63_cast_fp16)[name = tensor("obj_149_cast_fp16")]; + tensor query_43_pad_type_0 = const()[name = tensor("query_43_pad_type_0"), val = tensor("valid")]; + tensor query_43_strides_0 = const()[name = tensor("query_43_strides_0"), val = tensor([1, 1])]; + tensor query_43_pad_0 = const()[name = tensor("query_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_43_dilations_0 = const()[name = tensor("query_43_dilations_0"), val = tensor([1, 1])]; + tensor query_43_groups_0 = const()[name = tensor("query_43_groups_0"), val = tensor(1)]; + tensor layers_10_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(671793664)))]; + tensor layers_10_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(675070528)))]; + tensor query_43_cast_fp16 = conv(bias = layers_10_encoder_attn_q_proj_bias_to_fp16, dilations = query_43_dilations_0, groups = query_43_groups_0, pad = query_43_pad_0, pad_type = query_43_pad_type_0, strides = query_43_strides_0, weight = layers_10_encoder_attn_q_proj_weight_to_fp16, x = obj_149_cast_fp16)[name = tensor("query_43_cast_fp16")]; + tensor key_43_pad_type_0 = const()[name = tensor("key_43_pad_type_0"), val = tensor("valid")]; + tensor key_43_strides_0 = const()[name = tensor("key_43_strides_0"), val = tensor([1, 1])]; + tensor key_43_pad_0 = const()[name = tensor("key_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_43_dilations_0 = const()[name = tensor("key_43_dilations_0"), val = tensor([1, 1])]; + tensor key_43_groups_0 = const()[name = tensor("key_43_groups_0"), val = tensor(1)]; + tensor layers_10_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(675073152)))]; + tensor key_43_cast_fp16 = conv(dilations = key_43_dilations_0, groups = key_43_groups_0, pad = key_43_pad_0, pad_type = key_43_pad_type_0, strides = key_43_strides_0, weight = layers_10_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_43_cast_fp16")]; + tensor value_43_pad_type_0 = const()[name = tensor("value_43_pad_type_0"), val = tensor("valid")]; + tensor value_43_strides_0 = const()[name = tensor("value_43_strides_0"), val = tensor([1, 1])]; + tensor value_43_pad_0 = const()[name = tensor("value_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_43_dilations_0 = const()[name = tensor("value_43_dilations_0"), val = tensor([1, 1])]; + tensor value_43_groups_0 = const()[name = tensor("value_43_groups_0"), val = tensor(1)]; + tensor layers_10_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(678350016)))]; + tensor layers_10_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(681626880)))]; + tensor value_43_cast_fp16 = conv(bias = layers_10_encoder_attn_v_proj_bias_to_fp16, dilations = value_43_dilations_0, groups = value_43_groups_0, pad = value_43_pad_0, pad_type = value_43_pad_type_0, strides = value_43_strides_0, weight = layers_10_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_43_cast_fp16")]; + tensor var_2521 = const()[name = tensor("op_2521"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_43_cast_fp16 = reshape(shape = var_2521, x = query_43_cast_fp16)[name = tensor("mh_q_43_cast_fp16")]; + tensor var_2523_to_fp16 = const()[name = tensor("op_2523_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2524_cast_fp16 = mul(x = mh_q_43_cast_fp16, y = var_2523_to_fp16)[name = tensor("op_2524_cast_fp16")]; + tensor var_2527 = const()[name = tensor("op_2527"), val = tensor([1, 20, 64, 1500])]; + tensor var_2528_cast_fp16 = reshape(shape = var_2527, x = key_43_cast_fp16)[name = tensor("op_2528_cast_fp16")]; + tensor mh_w_65_transpose_x_0 = const()[name = tensor("mh_w_65_transpose_x_0"), val = tensor(true)]; + tensor mh_w_65_transpose_y_0 = const()[name = tensor("mh_w_65_transpose_y_0"), val = tensor(false)]; + tensor mh_w_65_cast_fp16 = matmul(transpose_x = mh_w_65_transpose_x_0, transpose_y = mh_w_65_transpose_y_0, x = var_2524_cast_fp16, y = var_2528_cast_fp16)[name = tensor("mh_w_65_cast_fp16")]; + tensor obj_153_cast_fp16 = softmax(axis = var_2370, x = mh_w_65_cast_fp16)[name = tensor("obj_153_cast_fp16")]; + tensor var_2532 = const()[name = tensor("op_2532"), val = tensor([1, 20, 64, 1500])]; + tensor var_2533_cast_fp16 = reshape(shape = var_2532, x = value_43_cast_fp16)[name = tensor("op_2533_cast_fp16")]; + tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; + tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(true)]; + tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = var_2533_cast_fp16, y = obj_153_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_2536 = const()[name = tensor("op_2536"), val = tensor([1, 1280, 1, 1])]; + tensor input_103_cast_fp16 = reshape(shape = var_2536, x = attn_43_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor obj_151_pad_type_0 = const()[name = tensor("obj_151_pad_type_0"), val = tensor("valid")]; + tensor obj_151_strides_0 = const()[name = tensor("obj_151_strides_0"), val = tensor([1, 1])]; + tensor obj_151_pad_0 = const()[name = tensor("obj_151_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_151_dilations_0 = const()[name = tensor("obj_151_dilations_0"), val = tensor([1, 1])]; + tensor obj_151_groups_0 = const()[name = tensor("obj_151_groups_0"), val = tensor(1)]; + tensor layers_10_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(681629504)))]; + tensor layers_10_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(684906368)))]; + tensor obj_151_cast_fp16 = conv(bias = layers_10_encoder_attn_o_proj_bias_to_fp16, dilations = obj_151_dilations_0, groups = obj_151_groups_0, pad = obj_151_pad_0, pad_type = obj_151_pad_type_0, strides = obj_151_strides_0, weight = layers_10_encoder_attn_o_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("obj_151_cast_fp16")]; + tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = obj_151_cast_fp16)[name = tensor("inputs_65_cast_fp16")]; + tensor out_65_axes_0 = const()[name = tensor("out_65_axes_0"), val = tensor([1])]; + tensor var_2557_to_fp16 = const()[name = tensor("op_2557_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_65_cast_fp16 = layer_norm(axes = out_65_axes_0, epsilon = var_2557_to_fp16, x = inputs_65_cast_fp16)[name = tensor("out_65_cast_fp16")]; + tensor input_105_gamma_0_to_fp16 = const()[name = tensor("input_105_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(684908992)))]; + tensor input_105_beta_0_to_fp16 = const()[name = tensor("input_105_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(684911616)))]; + tensor input_105_epsilon_0_to_fp16 = const()[name = tensor("input_105_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_105_cast_fp16 = batch_norm(beta = input_105_beta_0_to_fp16, epsilon = input_105_epsilon_0_to_fp16, gamma = input_105_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_65_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor input_107_pad_type_0 = const()[name = tensor("input_107_pad_type_0"), val = tensor("valid")]; + tensor input_107_strides_0 = const()[name = tensor("input_107_strides_0"), val = tensor([1, 1])]; + tensor input_107_pad_0 = const()[name = tensor("input_107_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_107_dilations_0 = const()[name = tensor("input_107_dilations_0"), val = tensor([1, 1])]; + tensor input_107_groups_0 = const()[name = tensor("input_107_groups_0"), val = tensor(1)]; + tensor layers_10_fc1_weight_to_fp16 = const()[name = tensor("layers_10_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(684914240)))]; + tensor layers_10_fc1_bias_to_fp16 = const()[name = tensor("layers_10_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(698021504)))]; + tensor input_107_cast_fp16 = conv(bias = layers_10_fc1_bias_to_fp16, dilations = input_107_dilations_0, groups = input_107_groups_0, pad = input_107_pad_0, pad_type = input_107_pad_type_0, strides = input_107_strides_0, weight = layers_10_fc1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_mode_0 = const()[name = tensor("input_109_mode_0"), val = tensor("EXACT")]; + tensor input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor hidden_states_23_pad_type_0 = const()[name = tensor("hidden_states_23_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_23_strides_0 = const()[name = tensor("hidden_states_23_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_23_pad_0 = const()[name = tensor("hidden_states_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_23_dilations_0 = const()[name = tensor("hidden_states_23_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_23_groups_0 = const()[name = tensor("hidden_states_23_groups_0"), val = tensor(1)]; + tensor layers_10_fc2_weight_to_fp16 = const()[name = tensor("layers_10_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(698031808)))]; + tensor layers_10_fc2_bias_to_fp16 = const()[name = tensor("layers_10_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(711139072)))]; + tensor hidden_states_23_cast_fp16 = conv(bias = layers_10_fc2_bias_to_fp16, dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_10_fc2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("hidden_states_23_cast_fp16")]; + tensor inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = hidden_states_23_cast_fp16)[name = tensor("inputs_67_cast_fp16")]; + tensor var_2593 = const()[name = tensor("op_2593"), val = tensor(3)]; + tensor out_67_axes_0 = const()[name = tensor("out_67_axes_0"), val = tensor([1])]; + tensor var_2618_to_fp16 = const()[name = tensor("op_2618_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_67_cast_fp16 = layer_norm(axes = out_67_axes_0, epsilon = var_2618_to_fp16, x = inputs_67_cast_fp16)[name = tensor("out_67_cast_fp16")]; + tensor obj_155_gamma_0_to_fp16 = const()[name = tensor("obj_155_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(711141696)))]; + tensor obj_155_beta_0_to_fp16 = const()[name = tensor("obj_155_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(711144320)))]; + tensor obj_155_epsilon_0_to_fp16 = const()[name = tensor("obj_155_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_155_cast_fp16 = batch_norm(beta = obj_155_beta_0_to_fp16, epsilon = obj_155_epsilon_0_to_fp16, gamma = obj_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_67_cast_fp16)[name = tensor("obj_155_cast_fp16")]; + tensor query_45_pad_type_0 = const()[name = tensor("query_45_pad_type_0"), val = tensor("valid")]; + tensor query_45_strides_0 = const()[name = tensor("query_45_strides_0"), val = tensor([1, 1])]; + tensor query_45_pad_0 = const()[name = tensor("query_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_45_dilations_0 = const()[name = tensor("query_45_dilations_0"), val = tensor([1, 1])]; + tensor query_45_groups_0 = const()[name = tensor("query_45_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(711146944)))]; + tensor layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(714423808)))]; + tensor query_45_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_bias_to_fp16, dilations = query_45_dilations_0, groups = query_45_groups_0, pad = query_45_pad_0, pad_type = query_45_pad_type_0, strides = query_45_strides_0, weight = layers_11_self_attn_q_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("query_45_cast_fp16")]; + tensor current_key_23_pad_type_0 = const()[name = tensor("current_key_23_pad_type_0"), val = tensor("valid")]; + tensor current_key_23_strides_0 = const()[name = tensor("current_key_23_strides_0"), val = tensor([1, 1])]; + tensor current_key_23_pad_0 = const()[name = tensor("current_key_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_23_dilations_0 = const()[name = tensor("current_key_23_dilations_0"), val = tensor([1, 1])]; + tensor current_key_23_groups_0 = const()[name = tensor("current_key_23_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(714426432)))]; + tensor current_key_23_cast_fp16 = conv(dilations = current_key_23_dilations_0, groups = current_key_23_groups_0, pad = current_key_23_pad_0, pad_type = current_key_23_pad_type_0, strides = current_key_23_strides_0, weight = layers_11_self_attn_k_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("current_key_23_cast_fp16")]; + tensor current_value_23_pad_type_0 = const()[name = tensor("current_value_23_pad_type_0"), val = tensor("valid")]; + tensor current_value_23_strides_0 = const()[name = tensor("current_value_23_strides_0"), val = tensor([1, 1])]; + tensor current_value_23_pad_0 = const()[name = tensor("current_value_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_23_dilations_0 = const()[name = tensor("current_value_23_dilations_0"), val = tensor([1, 1])]; + tensor current_value_23_groups_0 = const()[name = tensor("current_value_23_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(717703296)))]; + tensor layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(720980160)))]; + tensor current_value_23_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_bias_to_fp16, dilations = current_value_23_dilations_0, groups = current_value_23_groups_0, pad = current_value_23_pad_0, pad_type = current_value_23_pad_type_0, strides = current_value_23_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("current_value_23_cast_fp16")]; + tensor var_2657_cast_fp16 = mul(x = var_103_cast_fp16_11, y = var_239_cast_fp16)[name = tensor("op_2657_cast_fp16")]; + tensor var_2658_cast_fp16 = mul(x = current_key_23_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_2658_cast_fp16")]; + tensor key_45_cast_fp16 = add(x = var_2657_cast_fp16, y = var_2658_cast_fp16)[name = tensor("key_45_cast_fp16")]; + tensor var_2661_cast_fp16 = mul(x = var_138_cast_fp16_11, y = var_239_cast_fp16)[name = tensor("op_2661_cast_fp16")]; + tensor var_2662_cast_fp16 = mul(x = current_value_23_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_2662_cast_fp16")]; + tensor value_45_cast_fp16 = add(x = var_2661_cast_fp16, y = var_2662_cast_fp16)[name = tensor("value_45_cast_fp16")]; + tensor var_2666 = const()[name = tensor("op_2666"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_45_cast_fp16 = reshape(shape = var_2666, x = query_45_cast_fp16)[name = tensor("mh_q_45_cast_fp16")]; + tensor var_2668_to_fp16 = const()[name = tensor("op_2668_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2669_cast_fp16 = mul(x = mh_q_45_cast_fp16, y = var_2668_to_fp16)[name = tensor("op_2669_cast_fp16")]; + tensor var_2672 = const()[name = tensor("op_2672"), val = tensor([1, 20, 64, 448])]; + tensor var_2673_cast_fp16 = reshape(shape = var_2672, x = key_45_cast_fp16)[name = tensor("op_2673_cast_fp16")]; + tensor mh_w_67_transpose_x_0 = const()[name = tensor("mh_w_67_transpose_x_0"), val = tensor(true)]; + tensor mh_w_67_transpose_y_0 = const()[name = tensor("mh_w_67_transpose_y_0"), val = tensor(false)]; + tensor mh_w_67_cast_fp16 = matmul(transpose_x = mh_w_67_transpose_x_0, transpose_y = mh_w_67_transpose_y_0, x = var_2669_cast_fp16, y = var_2673_cast_fp16)[name = tensor("mh_w_67_cast_fp16")]; + tensor mh_w_69_cast_fp16 = add(x = mh_w_67_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_69_cast_fp16")]; + tensor var_2681_cast_fp16 = softmax(axis = var_2593, x = mh_w_69_cast_fp16)[name = tensor("op_2681_cast_fp16")]; + tensor var_2682 = const()[name = tensor("op_2682"), val = tensor([1, 20, 64, 448])]; + tensor var_2683_cast_fp16 = reshape(shape = var_2682, x = value_45_cast_fp16)[name = tensor("op_2683_cast_fp16")]; + tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; + tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(true)]; + tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = var_2683_cast_fp16, y = var_2681_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_2686 = const()[name = tensor("op_2686"), val = tensor([1, 1280, 1, 1])]; + tensor input_111_cast_fp16 = reshape(shape = var_2686, x = attn_45_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor obj_161_pad_type_0 = const()[name = tensor("obj_161_pad_type_0"), val = tensor("valid")]; + tensor obj_161_strides_0 = const()[name = tensor("obj_161_strides_0"), val = tensor([1, 1])]; + tensor obj_161_pad_0 = const()[name = tensor("obj_161_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_161_dilations_0 = const()[name = tensor("obj_161_dilations_0"), val = tensor([1, 1])]; + tensor obj_161_groups_0 = const()[name = tensor("obj_161_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(720982784)))]; + tensor layers_11_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(724259648)))]; + tensor obj_161_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_bias_to_fp16, dilations = obj_161_dilations_0, groups = obj_161_groups_0, pad = obj_161_pad_0, pad_type = obj_161_pad_type_0, strides = obj_161_strides_0, weight = layers_11_self_attn_o_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("obj_161_cast_fp16")]; + tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = obj_161_cast_fp16)[name = tensor("inputs_69_cast_fp16")]; + tensor out_69_axes_0 = const()[name = tensor("out_69_axes_0"), val = tensor([1])]; + tensor var_2708_to_fp16 = const()[name = tensor("op_2708_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_69_cast_fp16 = layer_norm(axes = out_69_axes_0, epsilon = var_2708_to_fp16, x = inputs_69_cast_fp16)[name = tensor("out_69_cast_fp16")]; + tensor obj_163_gamma_0_to_fp16 = const()[name = tensor("obj_163_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(724262272)))]; + tensor obj_163_beta_0_to_fp16 = const()[name = tensor("obj_163_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(724264896)))]; + tensor obj_163_epsilon_0_to_fp16 = const()[name = tensor("obj_163_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_163_cast_fp16 = batch_norm(beta = obj_163_beta_0_to_fp16, epsilon = obj_163_epsilon_0_to_fp16, gamma = obj_163_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_69_cast_fp16)[name = tensor("obj_163_cast_fp16")]; + tensor query_47_pad_type_0 = const()[name = tensor("query_47_pad_type_0"), val = tensor("valid")]; + tensor query_47_strides_0 = const()[name = tensor("query_47_strides_0"), val = tensor([1, 1])]; + tensor query_47_pad_0 = const()[name = tensor("query_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_47_dilations_0 = const()[name = tensor("query_47_dilations_0"), val = tensor([1, 1])]; + tensor query_47_groups_0 = const()[name = tensor("query_47_groups_0"), val = tensor(1)]; + tensor layers_11_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(724267520)))]; + tensor layers_11_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(727544384)))]; + tensor query_47_cast_fp16 = conv(bias = layers_11_encoder_attn_q_proj_bias_to_fp16, dilations = query_47_dilations_0, groups = query_47_groups_0, pad = query_47_pad_0, pad_type = query_47_pad_type_0, strides = query_47_strides_0, weight = layers_11_encoder_attn_q_proj_weight_to_fp16, x = obj_163_cast_fp16)[name = tensor("query_47_cast_fp16")]; + tensor key_47_pad_type_0 = const()[name = tensor("key_47_pad_type_0"), val = tensor("valid")]; + tensor key_47_strides_0 = const()[name = tensor("key_47_strides_0"), val = tensor([1, 1])]; + tensor key_47_pad_0 = const()[name = tensor("key_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_47_dilations_0 = const()[name = tensor("key_47_dilations_0"), val = tensor([1, 1])]; + tensor key_47_groups_0 = const()[name = tensor("key_47_groups_0"), val = tensor(1)]; + tensor layers_11_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(727547008)))]; + tensor key_47_cast_fp16 = conv(dilations = key_47_dilations_0, groups = key_47_groups_0, pad = key_47_pad_0, pad_type = key_47_pad_type_0, strides = key_47_strides_0, weight = layers_11_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_47_cast_fp16")]; + tensor value_47_pad_type_0 = const()[name = tensor("value_47_pad_type_0"), val = tensor("valid")]; + tensor value_47_strides_0 = const()[name = tensor("value_47_strides_0"), val = tensor([1, 1])]; + tensor value_47_pad_0 = const()[name = tensor("value_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_47_dilations_0 = const()[name = tensor("value_47_dilations_0"), val = tensor([1, 1])]; + tensor value_47_groups_0 = const()[name = tensor("value_47_groups_0"), val = tensor(1)]; + tensor layers_11_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(730823872)))]; + tensor layers_11_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(734100736)))]; + tensor value_47_cast_fp16 = conv(bias = layers_11_encoder_attn_v_proj_bias_to_fp16, dilations = value_47_dilations_0, groups = value_47_groups_0, pad = value_47_pad_0, pad_type = value_47_pad_type_0, strides = value_47_strides_0, weight = layers_11_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_47_cast_fp16")]; + tensor var_2744 = const()[name = tensor("op_2744"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_47_cast_fp16 = reshape(shape = var_2744, x = query_47_cast_fp16)[name = tensor("mh_q_47_cast_fp16")]; + tensor var_2746_to_fp16 = const()[name = tensor("op_2746_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2747_cast_fp16 = mul(x = mh_q_47_cast_fp16, y = var_2746_to_fp16)[name = tensor("op_2747_cast_fp16")]; + tensor var_2750 = const()[name = tensor("op_2750"), val = tensor([1, 20, 64, 1500])]; + tensor var_2751_cast_fp16 = reshape(shape = var_2750, x = key_47_cast_fp16)[name = tensor("op_2751_cast_fp16")]; + tensor mh_w_71_transpose_x_0 = const()[name = tensor("mh_w_71_transpose_x_0"), val = tensor(true)]; + tensor mh_w_71_transpose_y_0 = const()[name = tensor("mh_w_71_transpose_y_0"), val = tensor(false)]; + tensor mh_w_71_cast_fp16 = matmul(transpose_x = mh_w_71_transpose_x_0, transpose_y = mh_w_71_transpose_y_0, x = var_2747_cast_fp16, y = var_2751_cast_fp16)[name = tensor("mh_w_71_cast_fp16")]; + tensor obj_167_cast_fp16 = softmax(axis = var_2593, x = mh_w_71_cast_fp16)[name = tensor("obj_167_cast_fp16")]; + tensor var_2755 = const()[name = tensor("op_2755"), val = tensor([1, 20, 64, 1500])]; + tensor var_2756_cast_fp16 = reshape(shape = var_2755, x = value_47_cast_fp16)[name = tensor("op_2756_cast_fp16")]; + tensor attn_47_transpose_x_0 = const()[name = tensor("attn_47_transpose_x_0"), val = tensor(false)]; + tensor attn_47_transpose_y_0 = const()[name = tensor("attn_47_transpose_y_0"), val = tensor(true)]; + tensor attn_47_cast_fp16 = matmul(transpose_x = attn_47_transpose_x_0, transpose_y = attn_47_transpose_y_0, x = var_2756_cast_fp16, y = obj_167_cast_fp16)[name = tensor("attn_47_cast_fp16")]; + tensor var_2759 = const()[name = tensor("op_2759"), val = tensor([1, 1280, 1, 1])]; + tensor input_113_cast_fp16 = reshape(shape = var_2759, x = attn_47_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor obj_165_pad_type_0 = const()[name = tensor("obj_165_pad_type_0"), val = tensor("valid")]; + tensor obj_165_strides_0 = const()[name = tensor("obj_165_strides_0"), val = tensor([1, 1])]; + tensor obj_165_pad_0 = const()[name = tensor("obj_165_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_165_dilations_0 = const()[name = tensor("obj_165_dilations_0"), val = tensor([1, 1])]; + tensor obj_165_groups_0 = const()[name = tensor("obj_165_groups_0"), val = tensor(1)]; + tensor layers_11_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(734103360)))]; + tensor layers_11_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(737380224)))]; + tensor obj_165_cast_fp16 = conv(bias = layers_11_encoder_attn_o_proj_bias_to_fp16, dilations = obj_165_dilations_0, groups = obj_165_groups_0, pad = obj_165_pad_0, pad_type = obj_165_pad_type_0, strides = obj_165_strides_0, weight = layers_11_encoder_attn_o_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("obj_165_cast_fp16")]; + tensor inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_165_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; + tensor out_71_axes_0 = const()[name = tensor("out_71_axes_0"), val = tensor([1])]; + tensor var_2777_to_fp16 = const()[name = tensor("op_2777_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_71_cast_fp16 = layer_norm(axes = out_71_axes_0, epsilon = var_2777_to_fp16, x = inputs_71_cast_fp16)[name = tensor("out_71_cast_fp16")]; + tensor input_115_gamma_0_to_fp16 = const()[name = tensor("input_115_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(737382848)))]; + tensor input_115_beta_0_to_fp16 = const()[name = tensor("input_115_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(737385472)))]; + tensor input_115_epsilon_0_to_fp16 = const()[name = tensor("input_115_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_115_cast_fp16 = batch_norm(beta = input_115_beta_0_to_fp16, epsilon = input_115_epsilon_0_to_fp16, gamma = input_115_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_71_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor input_117_pad_type_0 = const()[name = tensor("input_117_pad_type_0"), val = tensor("valid")]; + tensor input_117_strides_0 = const()[name = tensor("input_117_strides_0"), val = tensor([1, 1])]; + tensor input_117_pad_0 = const()[name = tensor("input_117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_117_dilations_0 = const()[name = tensor("input_117_dilations_0"), val = tensor([1, 1])]; + tensor input_117_groups_0 = const()[name = tensor("input_117_groups_0"), val = tensor(1)]; + tensor layers_11_fc1_weight_to_fp16 = const()[name = tensor("layers_11_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(737388096)))]; + tensor layers_11_fc1_bias_to_fp16 = const()[name = tensor("layers_11_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(750495360)))]; + tensor input_117_cast_fp16 = conv(bias = layers_11_fc1_bias_to_fp16, dilations = input_117_dilations_0, groups = input_117_groups_0, pad = input_117_pad_0, pad_type = input_117_pad_type_0, strides = input_117_strides_0, weight = layers_11_fc1_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_mode_0 = const()[name = tensor("input_119_mode_0"), val = tensor("EXACT")]; + tensor input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor hidden_states_25_pad_type_0 = const()[name = tensor("hidden_states_25_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_25_strides_0 = const()[name = tensor("hidden_states_25_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_25_pad_0 = const()[name = tensor("hidden_states_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_25_dilations_0 = const()[name = tensor("hidden_states_25_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_25_groups_0 = const()[name = tensor("hidden_states_25_groups_0"), val = tensor(1)]; + tensor layers_11_fc2_weight_to_fp16 = const()[name = tensor("layers_11_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(750505664)))]; + tensor layers_11_fc2_bias_to_fp16 = const()[name = tensor("layers_11_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(763612928)))]; + tensor hidden_states_25_cast_fp16 = conv(bias = layers_11_fc2_bias_to_fp16, dilations = hidden_states_25_dilations_0, groups = hidden_states_25_groups_0, pad = hidden_states_25_pad_0, pad_type = hidden_states_25_pad_type_0, strides = hidden_states_25_strides_0, weight = layers_11_fc2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("hidden_states_25_cast_fp16")]; + tensor inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = hidden_states_25_cast_fp16)[name = tensor("inputs_73_cast_fp16")]; + tensor var_2812 = const()[name = tensor("op_2812"), val = tensor(3)]; + tensor out_73_axes_0 = const()[name = tensor("out_73_axes_0"), val = tensor([1])]; + tensor var_2837_to_fp16 = const()[name = tensor("op_2837_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_73_cast_fp16 = layer_norm(axes = out_73_axes_0, epsilon = var_2837_to_fp16, x = inputs_73_cast_fp16)[name = tensor("out_73_cast_fp16")]; + tensor obj_169_gamma_0_to_fp16 = const()[name = tensor("obj_169_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(763615552)))]; + tensor obj_169_beta_0_to_fp16 = const()[name = tensor("obj_169_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(763618176)))]; + tensor obj_169_epsilon_0_to_fp16 = const()[name = tensor("obj_169_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_169_cast_fp16 = batch_norm(beta = obj_169_beta_0_to_fp16, epsilon = obj_169_epsilon_0_to_fp16, gamma = obj_169_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_73_cast_fp16)[name = tensor("obj_169_cast_fp16")]; + tensor query_49_pad_type_0 = const()[name = tensor("query_49_pad_type_0"), val = tensor("valid")]; + tensor query_49_strides_0 = const()[name = tensor("query_49_strides_0"), val = tensor([1, 1])]; + tensor query_49_pad_0 = const()[name = tensor("query_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_49_dilations_0 = const()[name = tensor("query_49_dilations_0"), val = tensor([1, 1])]; + tensor query_49_groups_0 = const()[name = tensor("query_49_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(763620800)))]; + tensor layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(766897664)))]; + tensor query_49_cast_fp16 = conv(bias = layers_12_self_attn_q_proj_bias_to_fp16, dilations = query_49_dilations_0, groups = query_49_groups_0, pad = query_49_pad_0, pad_type = query_49_pad_type_0, strides = query_49_strides_0, weight = layers_12_self_attn_q_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("query_49_cast_fp16")]; + tensor current_key_25_pad_type_0 = const()[name = tensor("current_key_25_pad_type_0"), val = tensor("valid")]; + tensor current_key_25_strides_0 = const()[name = tensor("current_key_25_strides_0"), val = tensor([1, 1])]; + tensor current_key_25_pad_0 = const()[name = tensor("current_key_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_25_dilations_0 = const()[name = tensor("current_key_25_dilations_0"), val = tensor([1, 1])]; + tensor current_key_25_groups_0 = const()[name = tensor("current_key_25_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(766900288)))]; + tensor current_key_25_cast_fp16 = conv(dilations = current_key_25_dilations_0, groups = current_key_25_groups_0, pad = current_key_25_pad_0, pad_type = current_key_25_pad_type_0, strides = current_key_25_strides_0, weight = layers_12_self_attn_k_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("current_key_25_cast_fp16")]; + tensor current_value_25_pad_type_0 = const()[name = tensor("current_value_25_pad_type_0"), val = tensor("valid")]; + tensor current_value_25_strides_0 = const()[name = tensor("current_value_25_strides_0"), val = tensor([1, 1])]; + tensor current_value_25_pad_0 = const()[name = tensor("current_value_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_25_dilations_0 = const()[name = tensor("current_value_25_dilations_0"), val = tensor([1, 1])]; + tensor current_value_25_groups_0 = const()[name = tensor("current_value_25_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(770177152)))]; + tensor layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(773454016)))]; + tensor current_value_25_cast_fp16 = conv(bias = layers_12_self_attn_v_proj_bias_to_fp16, dilations = current_value_25_dilations_0, groups = current_value_25_groups_0, pad = current_value_25_pad_0, pad_type = current_value_25_pad_type_0, strides = current_value_25_strides_0, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("current_value_25_cast_fp16")]; + tensor var_2876_cast_fp16 = mul(x = var_103_cast_fp16_12, y = var_239_cast_fp16)[name = tensor("op_2876_cast_fp16")]; + tensor var_2877_cast_fp16 = mul(x = current_key_25_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_2877_cast_fp16")]; + tensor key_49_cast_fp16 = add(x = var_2876_cast_fp16, y = var_2877_cast_fp16)[name = tensor("key_49_cast_fp16")]; + tensor var_2880_cast_fp16 = mul(x = var_138_cast_fp16_12, y = var_239_cast_fp16)[name = tensor("op_2880_cast_fp16")]; + tensor var_2881_cast_fp16 = mul(x = current_value_25_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_2881_cast_fp16")]; + tensor value_49_cast_fp16 = add(x = var_2880_cast_fp16, y = var_2881_cast_fp16)[name = tensor("value_49_cast_fp16")]; + tensor var_2885 = const()[name = tensor("op_2885"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_49_cast_fp16 = reshape(shape = var_2885, x = query_49_cast_fp16)[name = tensor("mh_q_49_cast_fp16")]; + tensor var_2887_to_fp16 = const()[name = tensor("op_2887_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2888_cast_fp16 = mul(x = mh_q_49_cast_fp16, y = var_2887_to_fp16)[name = tensor("op_2888_cast_fp16")]; + tensor var_2891 = const()[name = tensor("op_2891"), val = tensor([1, 20, 64, 448])]; + tensor var_2892_cast_fp16 = reshape(shape = var_2891, x = key_49_cast_fp16)[name = tensor("op_2892_cast_fp16")]; + tensor mh_w_73_transpose_x_0 = const()[name = tensor("mh_w_73_transpose_x_0"), val = tensor(true)]; + tensor mh_w_73_transpose_y_0 = const()[name = tensor("mh_w_73_transpose_y_0"), val = tensor(false)]; + tensor mh_w_73_cast_fp16 = matmul(transpose_x = mh_w_73_transpose_x_0, transpose_y = mh_w_73_transpose_y_0, x = var_2888_cast_fp16, y = var_2892_cast_fp16)[name = tensor("mh_w_73_cast_fp16")]; + tensor mh_w_75_cast_fp16 = add(x = mh_w_73_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_75_cast_fp16")]; + tensor var_2900_cast_fp16 = softmax(axis = var_2812, x = mh_w_75_cast_fp16)[name = tensor("op_2900_cast_fp16")]; + tensor var_2901 = const()[name = tensor("op_2901"), val = tensor([1, 20, 64, 448])]; + tensor var_2902_cast_fp16 = reshape(shape = var_2901, x = value_49_cast_fp16)[name = tensor("op_2902_cast_fp16")]; + tensor attn_49_transpose_x_0 = const()[name = tensor("attn_49_transpose_x_0"), val = tensor(false)]; + tensor attn_49_transpose_y_0 = const()[name = tensor("attn_49_transpose_y_0"), val = tensor(true)]; + tensor attn_49_cast_fp16 = matmul(transpose_x = attn_49_transpose_x_0, transpose_y = attn_49_transpose_y_0, x = var_2902_cast_fp16, y = var_2900_cast_fp16)[name = tensor("attn_49_cast_fp16")]; + tensor var_2905 = const()[name = tensor("op_2905"), val = tensor([1, 1280, 1, 1])]; + tensor input_121_cast_fp16 = reshape(shape = var_2905, x = attn_49_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor obj_175_pad_type_0 = const()[name = tensor("obj_175_pad_type_0"), val = tensor("valid")]; + tensor obj_175_strides_0 = const()[name = tensor("obj_175_strides_0"), val = tensor([1, 1])]; + tensor obj_175_pad_0 = const()[name = tensor("obj_175_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_175_dilations_0 = const()[name = tensor("obj_175_dilations_0"), val = tensor([1, 1])]; + tensor obj_175_groups_0 = const()[name = tensor("obj_175_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(773456640)))]; + tensor layers_12_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(776733504)))]; + tensor obj_175_cast_fp16 = conv(bias = layers_12_self_attn_o_proj_bias_to_fp16, dilations = obj_175_dilations_0, groups = obj_175_groups_0, pad = obj_175_pad_0, pad_type = obj_175_pad_type_0, strides = obj_175_strides_0, weight = layers_12_self_attn_o_proj_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("obj_175_cast_fp16")]; + tensor inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = obj_175_cast_fp16)[name = tensor("inputs_75_cast_fp16")]; + tensor out_75_axes_0 = const()[name = tensor("out_75_axes_0"), val = tensor([1])]; + tensor var_2927_to_fp16 = const()[name = tensor("op_2927_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_75_cast_fp16 = layer_norm(axes = out_75_axes_0, epsilon = var_2927_to_fp16, x = inputs_75_cast_fp16)[name = tensor("out_75_cast_fp16")]; + tensor obj_177_gamma_0_to_fp16 = const()[name = tensor("obj_177_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(776736128)))]; + tensor obj_177_beta_0_to_fp16 = const()[name = tensor("obj_177_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(776738752)))]; + tensor obj_177_epsilon_0_to_fp16 = const()[name = tensor("obj_177_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_177_cast_fp16 = batch_norm(beta = obj_177_beta_0_to_fp16, epsilon = obj_177_epsilon_0_to_fp16, gamma = obj_177_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_75_cast_fp16)[name = tensor("obj_177_cast_fp16")]; + tensor query_51_pad_type_0 = const()[name = tensor("query_51_pad_type_0"), val = tensor("valid")]; + tensor query_51_strides_0 = const()[name = tensor("query_51_strides_0"), val = tensor([1, 1])]; + tensor query_51_pad_0 = const()[name = tensor("query_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_51_dilations_0 = const()[name = tensor("query_51_dilations_0"), val = tensor([1, 1])]; + tensor query_51_groups_0 = const()[name = tensor("query_51_groups_0"), val = tensor(1)]; + tensor layers_12_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(776741376)))]; + tensor layers_12_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(780018240)))]; + tensor query_51_cast_fp16 = conv(bias = layers_12_encoder_attn_q_proj_bias_to_fp16, dilations = query_51_dilations_0, groups = query_51_groups_0, pad = query_51_pad_0, pad_type = query_51_pad_type_0, strides = query_51_strides_0, weight = layers_12_encoder_attn_q_proj_weight_to_fp16, x = obj_177_cast_fp16)[name = tensor("query_51_cast_fp16")]; + tensor key_51_pad_type_0 = const()[name = tensor("key_51_pad_type_0"), val = tensor("valid")]; + tensor key_51_strides_0 = const()[name = tensor("key_51_strides_0"), val = tensor([1, 1])]; + tensor key_51_pad_0 = const()[name = tensor("key_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_51_dilations_0 = const()[name = tensor("key_51_dilations_0"), val = tensor([1, 1])]; + tensor key_51_groups_0 = const()[name = tensor("key_51_groups_0"), val = tensor(1)]; + tensor layers_12_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(780020864)))]; + tensor key_51_cast_fp16 = conv(dilations = key_51_dilations_0, groups = key_51_groups_0, pad = key_51_pad_0, pad_type = key_51_pad_type_0, strides = key_51_strides_0, weight = layers_12_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_51_cast_fp16")]; + tensor value_51_pad_type_0 = const()[name = tensor("value_51_pad_type_0"), val = tensor("valid")]; + tensor value_51_strides_0 = const()[name = tensor("value_51_strides_0"), val = tensor([1, 1])]; + tensor value_51_pad_0 = const()[name = tensor("value_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_51_dilations_0 = const()[name = tensor("value_51_dilations_0"), val = tensor([1, 1])]; + tensor value_51_groups_0 = const()[name = tensor("value_51_groups_0"), val = tensor(1)]; + tensor layers_12_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(783297728)))]; + tensor layers_12_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(786574592)))]; + tensor value_51_cast_fp16 = conv(bias = layers_12_encoder_attn_v_proj_bias_to_fp16, dilations = value_51_dilations_0, groups = value_51_groups_0, pad = value_51_pad_0, pad_type = value_51_pad_type_0, strides = value_51_strides_0, weight = layers_12_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_51_cast_fp16")]; + tensor var_2963 = const()[name = tensor("op_2963"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_51_cast_fp16 = reshape(shape = var_2963, x = query_51_cast_fp16)[name = tensor("mh_q_51_cast_fp16")]; + tensor var_2965_to_fp16 = const()[name = tensor("op_2965_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2966_cast_fp16 = mul(x = mh_q_51_cast_fp16, y = var_2965_to_fp16)[name = tensor("op_2966_cast_fp16")]; + tensor var_2969 = const()[name = tensor("op_2969"), val = tensor([1, 20, 64, 1500])]; + tensor var_2970_cast_fp16 = reshape(shape = var_2969, x = key_51_cast_fp16)[name = tensor("op_2970_cast_fp16")]; + tensor mh_w_77_transpose_x_0 = const()[name = tensor("mh_w_77_transpose_x_0"), val = tensor(true)]; + tensor mh_w_77_transpose_y_0 = const()[name = tensor("mh_w_77_transpose_y_0"), val = tensor(false)]; + tensor mh_w_77_cast_fp16 = matmul(transpose_x = mh_w_77_transpose_x_0, transpose_y = mh_w_77_transpose_y_0, x = var_2966_cast_fp16, y = var_2970_cast_fp16)[name = tensor("mh_w_77_cast_fp16")]; + tensor obj_181_cast_fp16 = softmax(axis = var_2812, x = mh_w_77_cast_fp16)[name = tensor("obj_181_cast_fp16")]; + tensor var_2974 = const()[name = tensor("op_2974"), val = tensor([1, 20, 64, 1500])]; + tensor var_2975_cast_fp16 = reshape(shape = var_2974, x = value_51_cast_fp16)[name = tensor("op_2975_cast_fp16")]; + tensor attn_51_transpose_x_0 = const()[name = tensor("attn_51_transpose_x_0"), val = tensor(false)]; + tensor attn_51_transpose_y_0 = const()[name = tensor("attn_51_transpose_y_0"), val = tensor(true)]; + tensor attn_51_cast_fp16 = matmul(transpose_x = attn_51_transpose_x_0, transpose_y = attn_51_transpose_y_0, x = var_2975_cast_fp16, y = obj_181_cast_fp16)[name = tensor("attn_51_cast_fp16")]; + tensor var_2978 = const()[name = tensor("op_2978"), val = tensor([1, 1280, 1, 1])]; + tensor input_123_cast_fp16 = reshape(shape = var_2978, x = attn_51_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor obj_179_pad_type_0 = const()[name = tensor("obj_179_pad_type_0"), val = tensor("valid")]; + tensor obj_179_strides_0 = const()[name = tensor("obj_179_strides_0"), val = tensor([1, 1])]; + tensor obj_179_pad_0 = const()[name = tensor("obj_179_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_179_dilations_0 = const()[name = tensor("obj_179_dilations_0"), val = tensor([1, 1])]; + tensor obj_179_groups_0 = const()[name = tensor("obj_179_groups_0"), val = tensor(1)]; + tensor layers_12_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(786577216)))]; + tensor layers_12_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(789854080)))]; + tensor obj_179_cast_fp16 = conv(bias = layers_12_encoder_attn_o_proj_bias_to_fp16, dilations = obj_179_dilations_0, groups = obj_179_groups_0, pad = obj_179_pad_0, pad_type = obj_179_pad_type_0, strides = obj_179_strides_0, weight = layers_12_encoder_attn_o_proj_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("obj_179_cast_fp16")]; + tensor inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = obj_179_cast_fp16)[name = tensor("inputs_77_cast_fp16")]; + tensor out_77_axes_0 = const()[name = tensor("out_77_axes_0"), val = tensor([1])]; + tensor var_2999_to_fp16 = const()[name = tensor("op_2999_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_77_cast_fp16 = layer_norm(axes = out_77_axes_0, epsilon = var_2999_to_fp16, x = inputs_77_cast_fp16)[name = tensor("out_77_cast_fp16")]; + tensor input_125_gamma_0_to_fp16 = const()[name = tensor("input_125_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(789856704)))]; + tensor input_125_beta_0_to_fp16 = const()[name = tensor("input_125_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(789859328)))]; + tensor input_125_epsilon_0_to_fp16 = const()[name = tensor("input_125_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_125_cast_fp16 = batch_norm(beta = input_125_beta_0_to_fp16, epsilon = input_125_epsilon_0_to_fp16, gamma = input_125_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_77_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor input_127_pad_type_0 = const()[name = tensor("input_127_pad_type_0"), val = tensor("valid")]; + tensor input_127_strides_0 = const()[name = tensor("input_127_strides_0"), val = tensor([1, 1])]; + tensor input_127_pad_0 = const()[name = tensor("input_127_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_127_dilations_0 = const()[name = tensor("input_127_dilations_0"), val = tensor([1, 1])]; + tensor input_127_groups_0 = const()[name = tensor("input_127_groups_0"), val = tensor(1)]; + tensor layers_12_fc1_weight_to_fp16 = const()[name = tensor("layers_12_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(789861952)))]; + tensor layers_12_fc1_bias_to_fp16 = const()[name = tensor("layers_12_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(802969216)))]; + tensor input_127_cast_fp16 = conv(bias = layers_12_fc1_bias_to_fp16, dilations = input_127_dilations_0, groups = input_127_groups_0, pad = input_127_pad_0, pad_type = input_127_pad_type_0, strides = input_127_strides_0, weight = layers_12_fc1_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_mode_0 = const()[name = tensor("input_129_mode_0"), val = tensor("EXACT")]; + tensor input_129_cast_fp16 = gelu(mode = input_129_mode_0, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor hidden_states_27_pad_type_0 = const()[name = tensor("hidden_states_27_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_27_strides_0 = const()[name = tensor("hidden_states_27_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_27_pad_0 = const()[name = tensor("hidden_states_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_27_dilations_0 = const()[name = tensor("hidden_states_27_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_27_groups_0 = const()[name = tensor("hidden_states_27_groups_0"), val = tensor(1)]; + tensor layers_12_fc2_weight_to_fp16 = const()[name = tensor("layers_12_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(802979520)))]; + tensor layers_12_fc2_bias_to_fp16 = const()[name = tensor("layers_12_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(816086784)))]; + tensor hidden_states_27_cast_fp16 = conv(bias = layers_12_fc2_bias_to_fp16, dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_12_fc2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("hidden_states_27_cast_fp16")]; + tensor inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = hidden_states_27_cast_fp16)[name = tensor("inputs_79_cast_fp16")]; + tensor var_3035 = const()[name = tensor("op_3035"), val = tensor(3)]; + tensor out_79_axes_0 = const()[name = tensor("out_79_axes_0"), val = tensor([1])]; + tensor var_3060_to_fp16 = const()[name = tensor("op_3060_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_79_cast_fp16 = layer_norm(axes = out_79_axes_0, epsilon = var_3060_to_fp16, x = inputs_79_cast_fp16)[name = tensor("out_79_cast_fp16")]; + tensor obj_183_gamma_0_to_fp16 = const()[name = tensor("obj_183_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(816089408)))]; + tensor obj_183_beta_0_to_fp16 = const()[name = tensor("obj_183_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(816092032)))]; + tensor obj_183_epsilon_0_to_fp16 = const()[name = tensor("obj_183_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_183_cast_fp16 = batch_norm(beta = obj_183_beta_0_to_fp16, epsilon = obj_183_epsilon_0_to_fp16, gamma = obj_183_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_79_cast_fp16)[name = tensor("obj_183_cast_fp16")]; + tensor query_53_pad_type_0 = const()[name = tensor("query_53_pad_type_0"), val = tensor("valid")]; + tensor query_53_strides_0 = const()[name = tensor("query_53_strides_0"), val = tensor([1, 1])]; + tensor query_53_pad_0 = const()[name = tensor("query_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_53_dilations_0 = const()[name = tensor("query_53_dilations_0"), val = tensor([1, 1])]; + tensor query_53_groups_0 = const()[name = tensor("query_53_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(816094656)))]; + tensor layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(819371520)))]; + tensor query_53_cast_fp16 = conv(bias = layers_13_self_attn_q_proj_bias_to_fp16, dilations = query_53_dilations_0, groups = query_53_groups_0, pad = query_53_pad_0, pad_type = query_53_pad_type_0, strides = query_53_strides_0, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor("query_53_cast_fp16")]; + tensor current_key_27_pad_type_0 = const()[name = tensor("current_key_27_pad_type_0"), val = tensor("valid")]; + tensor current_key_27_strides_0 = const()[name = tensor("current_key_27_strides_0"), val = tensor([1, 1])]; + tensor current_key_27_pad_0 = const()[name = tensor("current_key_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_27_dilations_0 = const()[name = tensor("current_key_27_dilations_0"), val = tensor([1, 1])]; + tensor current_key_27_groups_0 = const()[name = tensor("current_key_27_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(819374144)))]; + tensor current_key_27_cast_fp16 = conv(dilations = current_key_27_dilations_0, groups = current_key_27_groups_0, pad = current_key_27_pad_0, pad_type = current_key_27_pad_type_0, strides = current_key_27_strides_0, weight = layers_13_self_attn_k_proj_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor("current_key_27_cast_fp16")]; + tensor current_value_27_pad_type_0 = const()[name = tensor("current_value_27_pad_type_0"), val = tensor("valid")]; + tensor current_value_27_strides_0 = const()[name = tensor("current_value_27_strides_0"), val = tensor([1, 1])]; + tensor current_value_27_pad_0 = const()[name = tensor("current_value_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_27_dilations_0 = const()[name = tensor("current_value_27_dilations_0"), val = tensor([1, 1])]; + tensor current_value_27_groups_0 = const()[name = tensor("current_value_27_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(822651008)))]; + tensor layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(825927872)))]; + tensor current_value_27_cast_fp16 = conv(bias = layers_13_self_attn_v_proj_bias_to_fp16, dilations = current_value_27_dilations_0, groups = current_value_27_groups_0, pad = current_value_27_pad_0, pad_type = current_value_27_pad_type_0, strides = current_value_27_strides_0, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor("current_value_27_cast_fp16")]; + tensor var_3099_cast_fp16 = mul(x = var_103_cast_fp16_13, y = var_239_cast_fp16)[name = tensor("op_3099_cast_fp16")]; + tensor var_3100_cast_fp16 = mul(x = current_key_27_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_3100_cast_fp16")]; + tensor key_53_cast_fp16 = add(x = var_3099_cast_fp16, y = var_3100_cast_fp16)[name = tensor("key_53_cast_fp16")]; + tensor var_3103_cast_fp16 = mul(x = var_138_cast_fp16_13, y = var_239_cast_fp16)[name = tensor("op_3103_cast_fp16")]; + tensor var_3104_cast_fp16 = mul(x = current_value_27_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_3104_cast_fp16")]; + tensor value_53_cast_fp16 = add(x = var_3103_cast_fp16, y = var_3104_cast_fp16)[name = tensor("value_53_cast_fp16")]; + tensor var_3108 = const()[name = tensor("op_3108"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_53_cast_fp16 = reshape(shape = var_3108, x = query_53_cast_fp16)[name = tensor("mh_q_53_cast_fp16")]; + tensor var_3110_to_fp16 = const()[name = tensor("op_3110_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3111_cast_fp16 = mul(x = mh_q_53_cast_fp16, y = var_3110_to_fp16)[name = tensor("op_3111_cast_fp16")]; + tensor var_3114 = const()[name = tensor("op_3114"), val = tensor([1, 20, 64, 448])]; + tensor var_3115_cast_fp16 = reshape(shape = var_3114, x = key_53_cast_fp16)[name = tensor("op_3115_cast_fp16")]; + tensor mh_w_79_transpose_x_0 = const()[name = tensor("mh_w_79_transpose_x_0"), val = tensor(true)]; + tensor mh_w_79_transpose_y_0 = const()[name = tensor("mh_w_79_transpose_y_0"), val = tensor(false)]; + tensor mh_w_79_cast_fp16 = matmul(transpose_x = mh_w_79_transpose_x_0, transpose_y = mh_w_79_transpose_y_0, x = var_3111_cast_fp16, y = var_3115_cast_fp16)[name = tensor("mh_w_79_cast_fp16")]; + tensor mh_w_81_cast_fp16 = add(x = mh_w_79_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_81_cast_fp16")]; + tensor var_3123_cast_fp16 = softmax(axis = var_3035, x = mh_w_81_cast_fp16)[name = tensor("op_3123_cast_fp16")]; + tensor var_3124 = const()[name = tensor("op_3124"), val = tensor([1, 20, 64, 448])]; + tensor var_3125_cast_fp16 = reshape(shape = var_3124, x = value_53_cast_fp16)[name = tensor("op_3125_cast_fp16")]; + tensor attn_53_transpose_x_0 = const()[name = tensor("attn_53_transpose_x_0"), val = tensor(false)]; + tensor attn_53_transpose_y_0 = const()[name = tensor("attn_53_transpose_y_0"), val = tensor(true)]; + tensor attn_53_cast_fp16 = matmul(transpose_x = attn_53_transpose_x_0, transpose_y = attn_53_transpose_y_0, x = var_3125_cast_fp16, y = var_3123_cast_fp16)[name = tensor("attn_53_cast_fp16")]; + tensor var_3128 = const()[name = tensor("op_3128"), val = tensor([1, 1280, 1, 1])]; + tensor input_131_cast_fp16 = reshape(shape = var_3128, x = attn_53_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor obj_189_pad_type_0 = const()[name = tensor("obj_189_pad_type_0"), val = tensor("valid")]; + tensor obj_189_strides_0 = const()[name = tensor("obj_189_strides_0"), val = tensor([1, 1])]; + tensor obj_189_pad_0 = const()[name = tensor("obj_189_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_189_dilations_0 = const()[name = tensor("obj_189_dilations_0"), val = tensor([1, 1])]; + tensor obj_189_groups_0 = const()[name = tensor("obj_189_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(825930496)))]; + tensor layers_13_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(829207360)))]; + tensor obj_189_cast_fp16 = conv(bias = layers_13_self_attn_o_proj_bias_to_fp16, dilations = obj_189_dilations_0, groups = obj_189_groups_0, pad = obj_189_pad_0, pad_type = obj_189_pad_type_0, strides = obj_189_strides_0, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("obj_189_cast_fp16")]; + tensor inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = obj_189_cast_fp16)[name = tensor("inputs_81_cast_fp16")]; + tensor out_81_axes_0 = const()[name = tensor("out_81_axes_0"), val = tensor([1])]; + tensor var_3150_to_fp16 = const()[name = tensor("op_3150_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_81_cast_fp16 = layer_norm(axes = out_81_axes_0, epsilon = var_3150_to_fp16, x = inputs_81_cast_fp16)[name = tensor("out_81_cast_fp16")]; + tensor obj_191_gamma_0_to_fp16 = const()[name = tensor("obj_191_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(829209984)))]; + tensor obj_191_beta_0_to_fp16 = const()[name = tensor("obj_191_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(829212608)))]; + tensor obj_191_epsilon_0_to_fp16 = const()[name = tensor("obj_191_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_191_cast_fp16 = batch_norm(beta = obj_191_beta_0_to_fp16, epsilon = obj_191_epsilon_0_to_fp16, gamma = obj_191_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_81_cast_fp16)[name = tensor("obj_191_cast_fp16")]; + tensor query_55_pad_type_0 = const()[name = tensor("query_55_pad_type_0"), val = tensor("valid")]; + tensor query_55_strides_0 = const()[name = tensor("query_55_strides_0"), val = tensor([1, 1])]; + tensor query_55_pad_0 = const()[name = tensor("query_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_55_dilations_0 = const()[name = tensor("query_55_dilations_0"), val = tensor([1, 1])]; + tensor query_55_groups_0 = const()[name = tensor("query_55_groups_0"), val = tensor(1)]; + tensor layers_13_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(829215232)))]; + tensor layers_13_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(832492096)))]; + tensor query_55_cast_fp16 = conv(bias = layers_13_encoder_attn_q_proj_bias_to_fp16, dilations = query_55_dilations_0, groups = query_55_groups_0, pad = query_55_pad_0, pad_type = query_55_pad_type_0, strides = query_55_strides_0, weight = layers_13_encoder_attn_q_proj_weight_to_fp16, x = obj_191_cast_fp16)[name = tensor("query_55_cast_fp16")]; + tensor key_55_pad_type_0 = const()[name = tensor("key_55_pad_type_0"), val = tensor("valid")]; + tensor key_55_strides_0 = const()[name = tensor("key_55_strides_0"), val = tensor([1, 1])]; + tensor key_55_pad_0 = const()[name = tensor("key_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_55_dilations_0 = const()[name = tensor("key_55_dilations_0"), val = tensor([1, 1])]; + tensor key_55_groups_0 = const()[name = tensor("key_55_groups_0"), val = tensor(1)]; + tensor layers_13_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(832494720)))]; + tensor key_55_cast_fp16 = conv(dilations = key_55_dilations_0, groups = key_55_groups_0, pad = key_55_pad_0, pad_type = key_55_pad_type_0, strides = key_55_strides_0, weight = layers_13_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_55_cast_fp16")]; + tensor value_55_pad_type_0 = const()[name = tensor("value_55_pad_type_0"), val = tensor("valid")]; + tensor value_55_strides_0 = const()[name = tensor("value_55_strides_0"), val = tensor([1, 1])]; + tensor value_55_pad_0 = const()[name = tensor("value_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_55_dilations_0 = const()[name = tensor("value_55_dilations_0"), val = tensor([1, 1])]; + tensor value_55_groups_0 = const()[name = tensor("value_55_groups_0"), val = tensor(1)]; + tensor layers_13_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(835771584)))]; + tensor layers_13_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(839048448)))]; + tensor value_55_cast_fp16 = conv(bias = layers_13_encoder_attn_v_proj_bias_to_fp16, dilations = value_55_dilations_0, groups = value_55_groups_0, pad = value_55_pad_0, pad_type = value_55_pad_type_0, strides = value_55_strides_0, weight = layers_13_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_55_cast_fp16")]; + tensor var_3186 = const()[name = tensor("op_3186"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_55_cast_fp16 = reshape(shape = var_3186, x = query_55_cast_fp16)[name = tensor("mh_q_55_cast_fp16")]; + tensor var_3188_to_fp16 = const()[name = tensor("op_3188_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3189_cast_fp16 = mul(x = mh_q_55_cast_fp16, y = var_3188_to_fp16)[name = tensor("op_3189_cast_fp16")]; + tensor var_3192 = const()[name = tensor("op_3192"), val = tensor([1, 20, 64, 1500])]; + tensor var_3193_cast_fp16 = reshape(shape = var_3192, x = key_55_cast_fp16)[name = tensor("op_3193_cast_fp16")]; + tensor mh_w_83_transpose_x_0 = const()[name = tensor("mh_w_83_transpose_x_0"), val = tensor(true)]; + tensor mh_w_83_transpose_y_0 = const()[name = tensor("mh_w_83_transpose_y_0"), val = tensor(false)]; + tensor mh_w_83_cast_fp16 = matmul(transpose_x = mh_w_83_transpose_x_0, transpose_y = mh_w_83_transpose_y_0, x = var_3189_cast_fp16, y = var_3193_cast_fp16)[name = tensor("mh_w_83_cast_fp16")]; + tensor obj_195_cast_fp16 = softmax(axis = var_3035, x = mh_w_83_cast_fp16)[name = tensor("obj_195_cast_fp16")]; + tensor var_3197 = const()[name = tensor("op_3197"), val = tensor([1, 20, 64, 1500])]; + tensor var_3198_cast_fp16 = reshape(shape = var_3197, x = value_55_cast_fp16)[name = tensor("op_3198_cast_fp16")]; + tensor attn_55_transpose_x_0 = const()[name = tensor("attn_55_transpose_x_0"), val = tensor(false)]; + tensor attn_55_transpose_y_0 = const()[name = tensor("attn_55_transpose_y_0"), val = tensor(true)]; + tensor attn_55_cast_fp16 = matmul(transpose_x = attn_55_transpose_x_0, transpose_y = attn_55_transpose_y_0, x = var_3198_cast_fp16, y = obj_195_cast_fp16)[name = tensor("attn_55_cast_fp16")]; + tensor var_3201 = const()[name = tensor("op_3201"), val = tensor([1, 1280, 1, 1])]; + tensor input_133_cast_fp16 = reshape(shape = var_3201, x = attn_55_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor obj_193_pad_type_0 = const()[name = tensor("obj_193_pad_type_0"), val = tensor("valid")]; + tensor obj_193_strides_0 = const()[name = tensor("obj_193_strides_0"), val = tensor([1, 1])]; + tensor obj_193_pad_0 = const()[name = tensor("obj_193_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_193_dilations_0 = const()[name = tensor("obj_193_dilations_0"), val = tensor([1, 1])]; + tensor obj_193_groups_0 = const()[name = tensor("obj_193_groups_0"), val = tensor(1)]; + tensor layers_13_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(839051072)))]; + tensor layers_13_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(842327936)))]; + tensor obj_193_cast_fp16 = conv(bias = layers_13_encoder_attn_o_proj_bias_to_fp16, dilations = obj_193_dilations_0, groups = obj_193_groups_0, pad = obj_193_pad_0, pad_type = obj_193_pad_type_0, strides = obj_193_strides_0, weight = layers_13_encoder_attn_o_proj_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("obj_193_cast_fp16")]; + tensor inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = obj_193_cast_fp16)[name = tensor("inputs_83_cast_fp16")]; + tensor out_83_axes_0 = const()[name = tensor("out_83_axes_0"), val = tensor([1])]; + tensor var_3222_to_fp16 = const()[name = tensor("op_3222_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_83_cast_fp16 = layer_norm(axes = out_83_axes_0, epsilon = var_3222_to_fp16, x = inputs_83_cast_fp16)[name = tensor("out_83_cast_fp16")]; + tensor input_135_gamma_0_to_fp16 = const()[name = tensor("input_135_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(842330560)))]; + tensor input_135_beta_0_to_fp16 = const()[name = tensor("input_135_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(842333184)))]; + tensor input_135_epsilon_0_to_fp16 = const()[name = tensor("input_135_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_135_cast_fp16 = batch_norm(beta = input_135_beta_0_to_fp16, epsilon = input_135_epsilon_0_to_fp16, gamma = input_135_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_83_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor input_137_pad_type_0 = const()[name = tensor("input_137_pad_type_0"), val = tensor("valid")]; + tensor input_137_strides_0 = const()[name = tensor("input_137_strides_0"), val = tensor([1, 1])]; + tensor input_137_pad_0 = const()[name = tensor("input_137_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_137_dilations_0 = const()[name = tensor("input_137_dilations_0"), val = tensor([1, 1])]; + tensor input_137_groups_0 = const()[name = tensor("input_137_groups_0"), val = tensor(1)]; + tensor layers_13_fc1_weight_to_fp16 = const()[name = tensor("layers_13_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(842335808)))]; + tensor layers_13_fc1_bias_to_fp16 = const()[name = tensor("layers_13_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(855443072)))]; + tensor input_137_cast_fp16 = conv(bias = layers_13_fc1_bias_to_fp16, dilations = input_137_dilations_0, groups = input_137_groups_0, pad = input_137_pad_0, pad_type = input_137_pad_type_0, strides = input_137_strides_0, weight = layers_13_fc1_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor input_139_mode_0 = const()[name = tensor("input_139_mode_0"), val = tensor("EXACT")]; + tensor input_139_cast_fp16 = gelu(mode = input_139_mode_0, x = input_137_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor hidden_states_29_pad_type_0 = const()[name = tensor("hidden_states_29_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_29_strides_0 = const()[name = tensor("hidden_states_29_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_29_pad_0 = const()[name = tensor("hidden_states_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_29_dilations_0 = const()[name = tensor("hidden_states_29_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_29_groups_0 = const()[name = tensor("hidden_states_29_groups_0"), val = tensor(1)]; + tensor layers_13_fc2_weight_to_fp16 = const()[name = tensor("layers_13_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(855453376)))]; + tensor layers_13_fc2_bias_to_fp16 = const()[name = tensor("layers_13_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(868560640)))]; + tensor hidden_states_29_cast_fp16 = conv(bias = layers_13_fc2_bias_to_fp16, dilations = hidden_states_29_dilations_0, groups = hidden_states_29_groups_0, pad = hidden_states_29_pad_0, pad_type = hidden_states_29_pad_type_0, strides = hidden_states_29_strides_0, weight = layers_13_fc2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("hidden_states_29_cast_fp16")]; + tensor inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = hidden_states_29_cast_fp16)[name = tensor("inputs_85_cast_fp16")]; + tensor var_3258 = const()[name = tensor("op_3258"), val = tensor(3)]; + tensor out_85_axes_0 = const()[name = tensor("out_85_axes_0"), val = tensor([1])]; + tensor var_3283_to_fp16 = const()[name = tensor("op_3283_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_85_cast_fp16 = layer_norm(axes = out_85_axes_0, epsilon = var_3283_to_fp16, x = inputs_85_cast_fp16)[name = tensor("out_85_cast_fp16")]; + tensor obj_197_gamma_0_to_fp16 = const()[name = tensor("obj_197_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(868563264)))]; + tensor obj_197_beta_0_to_fp16 = const()[name = tensor("obj_197_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(868565888)))]; + tensor obj_197_epsilon_0_to_fp16 = const()[name = tensor("obj_197_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_197_cast_fp16 = batch_norm(beta = obj_197_beta_0_to_fp16, epsilon = obj_197_epsilon_0_to_fp16, gamma = obj_197_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_85_cast_fp16)[name = tensor("obj_197_cast_fp16")]; + tensor query_57_pad_type_0 = const()[name = tensor("query_57_pad_type_0"), val = tensor("valid")]; + tensor query_57_strides_0 = const()[name = tensor("query_57_strides_0"), val = tensor([1, 1])]; + tensor query_57_pad_0 = const()[name = tensor("query_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_57_dilations_0 = const()[name = tensor("query_57_dilations_0"), val = tensor([1, 1])]; + tensor query_57_groups_0 = const()[name = tensor("query_57_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(868568512)))]; + tensor layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(871845376)))]; + tensor query_57_cast_fp16 = conv(bias = layers_14_self_attn_q_proj_bias_to_fp16, dilations = query_57_dilations_0, groups = query_57_groups_0, pad = query_57_pad_0, pad_type = query_57_pad_type_0, strides = query_57_strides_0, weight = layers_14_self_attn_q_proj_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor("query_57_cast_fp16")]; + tensor current_key_29_pad_type_0 = const()[name = tensor("current_key_29_pad_type_0"), val = tensor("valid")]; + tensor current_key_29_strides_0 = const()[name = tensor("current_key_29_strides_0"), val = tensor([1, 1])]; + tensor current_key_29_pad_0 = const()[name = tensor("current_key_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_29_dilations_0 = const()[name = tensor("current_key_29_dilations_0"), val = tensor([1, 1])]; + tensor current_key_29_groups_0 = const()[name = tensor("current_key_29_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(871848000)))]; + tensor current_key_29_cast_fp16 = conv(dilations = current_key_29_dilations_0, groups = current_key_29_groups_0, pad = current_key_29_pad_0, pad_type = current_key_29_pad_type_0, strides = current_key_29_strides_0, weight = layers_14_self_attn_k_proj_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor("current_key_29_cast_fp16")]; + tensor current_value_29_pad_type_0 = const()[name = tensor("current_value_29_pad_type_0"), val = tensor("valid")]; + tensor current_value_29_strides_0 = const()[name = tensor("current_value_29_strides_0"), val = tensor([1, 1])]; + tensor current_value_29_pad_0 = const()[name = tensor("current_value_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_29_dilations_0 = const()[name = tensor("current_value_29_dilations_0"), val = tensor([1, 1])]; + tensor current_value_29_groups_0 = const()[name = tensor("current_value_29_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(875124864)))]; + tensor layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(878401728)))]; + tensor current_value_29_cast_fp16 = conv(bias = layers_14_self_attn_v_proj_bias_to_fp16, dilations = current_value_29_dilations_0, groups = current_value_29_groups_0, pad = current_value_29_pad_0, pad_type = current_value_29_pad_type_0, strides = current_value_29_strides_0, weight = layers_14_self_attn_v_proj_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor("current_value_29_cast_fp16")]; + tensor var_3322_cast_fp16 = mul(x = var_103_cast_fp16_14, y = var_239_cast_fp16)[name = tensor("op_3322_cast_fp16")]; + tensor var_3323_cast_fp16 = mul(x = current_key_29_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_3323_cast_fp16")]; + tensor key_57_cast_fp16 = add(x = var_3322_cast_fp16, y = var_3323_cast_fp16)[name = tensor("key_57_cast_fp16")]; + tensor var_3326_cast_fp16 = mul(x = var_138_cast_fp16_14, y = var_239_cast_fp16)[name = tensor("op_3326_cast_fp16")]; + tensor var_3327_cast_fp16 = mul(x = current_value_29_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_3327_cast_fp16")]; + tensor value_57_cast_fp16 = add(x = var_3326_cast_fp16, y = var_3327_cast_fp16)[name = tensor("value_57_cast_fp16")]; + tensor var_3331 = const()[name = tensor("op_3331"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_57_cast_fp16 = reshape(shape = var_3331, x = query_57_cast_fp16)[name = tensor("mh_q_57_cast_fp16")]; + tensor var_3333_to_fp16 = const()[name = tensor("op_3333_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3334_cast_fp16 = mul(x = mh_q_57_cast_fp16, y = var_3333_to_fp16)[name = tensor("op_3334_cast_fp16")]; + tensor var_3337 = const()[name = tensor("op_3337"), val = tensor([1, 20, 64, 448])]; + tensor var_3338_cast_fp16 = reshape(shape = var_3337, x = key_57_cast_fp16)[name = tensor("op_3338_cast_fp16")]; + tensor mh_w_85_transpose_x_0 = const()[name = tensor("mh_w_85_transpose_x_0"), val = tensor(true)]; + tensor mh_w_85_transpose_y_0 = const()[name = tensor("mh_w_85_transpose_y_0"), val = tensor(false)]; + tensor mh_w_85_cast_fp16 = matmul(transpose_x = mh_w_85_transpose_x_0, transpose_y = mh_w_85_transpose_y_0, x = var_3334_cast_fp16, y = var_3338_cast_fp16)[name = tensor("mh_w_85_cast_fp16")]; + tensor mh_w_87_cast_fp16 = add(x = mh_w_85_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_87_cast_fp16")]; + tensor var_3346_cast_fp16 = softmax(axis = var_3258, x = mh_w_87_cast_fp16)[name = tensor("op_3346_cast_fp16")]; + tensor var_3347 = const()[name = tensor("op_3347"), val = tensor([1, 20, 64, 448])]; + tensor var_3348_cast_fp16 = reshape(shape = var_3347, x = value_57_cast_fp16)[name = tensor("op_3348_cast_fp16")]; + tensor attn_57_transpose_x_0 = const()[name = tensor("attn_57_transpose_x_0"), val = tensor(false)]; + tensor attn_57_transpose_y_0 = const()[name = tensor("attn_57_transpose_y_0"), val = tensor(true)]; + tensor attn_57_cast_fp16 = matmul(transpose_x = attn_57_transpose_x_0, transpose_y = attn_57_transpose_y_0, x = var_3348_cast_fp16, y = var_3346_cast_fp16)[name = tensor("attn_57_cast_fp16")]; + tensor var_3351 = const()[name = tensor("op_3351"), val = tensor([1, 1280, 1, 1])]; + tensor input_141_cast_fp16 = reshape(shape = var_3351, x = attn_57_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor obj_203_pad_type_0 = const()[name = tensor("obj_203_pad_type_0"), val = tensor("valid")]; + tensor obj_203_strides_0 = const()[name = tensor("obj_203_strides_0"), val = tensor([1, 1])]; + tensor obj_203_pad_0 = const()[name = tensor("obj_203_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_203_dilations_0 = const()[name = tensor("obj_203_dilations_0"), val = tensor([1, 1])]; + tensor obj_203_groups_0 = const()[name = tensor("obj_203_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(878404352)))]; + tensor layers_14_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(881681216)))]; + tensor obj_203_cast_fp16 = conv(bias = layers_14_self_attn_o_proj_bias_to_fp16, dilations = obj_203_dilations_0, groups = obj_203_groups_0, pad = obj_203_pad_0, pad_type = obj_203_pad_type_0, strides = obj_203_strides_0, weight = layers_14_self_attn_o_proj_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("obj_203_cast_fp16")]; + tensor inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = obj_203_cast_fp16)[name = tensor("inputs_87_cast_fp16")]; + tensor out_87_axes_0 = const()[name = tensor("out_87_axes_0"), val = tensor([1])]; + tensor var_3373_to_fp16 = const()[name = tensor("op_3373_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_87_cast_fp16 = layer_norm(axes = out_87_axes_0, epsilon = var_3373_to_fp16, x = inputs_87_cast_fp16)[name = tensor("out_87_cast_fp16")]; + tensor obj_205_gamma_0_to_fp16 = const()[name = tensor("obj_205_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(881683840)))]; + tensor obj_205_beta_0_to_fp16 = const()[name = tensor("obj_205_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(881686464)))]; + tensor obj_205_epsilon_0_to_fp16 = const()[name = tensor("obj_205_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_205_cast_fp16 = batch_norm(beta = obj_205_beta_0_to_fp16, epsilon = obj_205_epsilon_0_to_fp16, gamma = obj_205_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_87_cast_fp16)[name = tensor("obj_205_cast_fp16")]; + tensor query_59_pad_type_0 = const()[name = tensor("query_59_pad_type_0"), val = tensor("valid")]; + tensor query_59_strides_0 = const()[name = tensor("query_59_strides_0"), val = tensor([1, 1])]; + tensor query_59_pad_0 = const()[name = tensor("query_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_59_dilations_0 = const()[name = tensor("query_59_dilations_0"), val = tensor([1, 1])]; + tensor query_59_groups_0 = const()[name = tensor("query_59_groups_0"), val = tensor(1)]; + tensor layers_14_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(881689088)))]; + tensor layers_14_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(884965952)))]; + tensor query_59_cast_fp16 = conv(bias = layers_14_encoder_attn_q_proj_bias_to_fp16, dilations = query_59_dilations_0, groups = query_59_groups_0, pad = query_59_pad_0, pad_type = query_59_pad_type_0, strides = query_59_strides_0, weight = layers_14_encoder_attn_q_proj_weight_to_fp16, x = obj_205_cast_fp16)[name = tensor("query_59_cast_fp16")]; + tensor key_59_pad_type_0 = const()[name = tensor("key_59_pad_type_0"), val = tensor("valid")]; + tensor key_59_strides_0 = const()[name = tensor("key_59_strides_0"), val = tensor([1, 1])]; + tensor key_59_pad_0 = const()[name = tensor("key_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_59_dilations_0 = const()[name = tensor("key_59_dilations_0"), val = tensor([1, 1])]; + tensor key_59_groups_0 = const()[name = tensor("key_59_groups_0"), val = tensor(1)]; + tensor layers_14_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(884968576)))]; + tensor key_59_cast_fp16 = conv(dilations = key_59_dilations_0, groups = key_59_groups_0, pad = key_59_pad_0, pad_type = key_59_pad_type_0, strides = key_59_strides_0, weight = layers_14_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_59_cast_fp16")]; + tensor value_59_pad_type_0 = const()[name = tensor("value_59_pad_type_0"), val = tensor("valid")]; + tensor value_59_strides_0 = const()[name = tensor("value_59_strides_0"), val = tensor([1, 1])]; + tensor value_59_pad_0 = const()[name = tensor("value_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_59_dilations_0 = const()[name = tensor("value_59_dilations_0"), val = tensor([1, 1])]; + tensor value_59_groups_0 = const()[name = tensor("value_59_groups_0"), val = tensor(1)]; + tensor layers_14_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(888245440)))]; + tensor layers_14_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(891522304)))]; + tensor value_59_cast_fp16 = conv(bias = layers_14_encoder_attn_v_proj_bias_to_fp16, dilations = value_59_dilations_0, groups = value_59_groups_0, pad = value_59_pad_0, pad_type = value_59_pad_type_0, strides = value_59_strides_0, weight = layers_14_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_59_cast_fp16")]; + tensor var_3409 = const()[name = tensor("op_3409"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_59_cast_fp16 = reshape(shape = var_3409, x = query_59_cast_fp16)[name = tensor("mh_q_59_cast_fp16")]; + tensor var_3411_to_fp16 = const()[name = tensor("op_3411_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3412_cast_fp16 = mul(x = mh_q_59_cast_fp16, y = var_3411_to_fp16)[name = tensor("op_3412_cast_fp16")]; + tensor var_3415 = const()[name = tensor("op_3415"), val = tensor([1, 20, 64, 1500])]; + tensor var_3416_cast_fp16 = reshape(shape = var_3415, x = key_59_cast_fp16)[name = tensor("op_3416_cast_fp16")]; + tensor mh_w_89_transpose_x_0 = const()[name = tensor("mh_w_89_transpose_x_0"), val = tensor(true)]; + tensor mh_w_89_transpose_y_0 = const()[name = tensor("mh_w_89_transpose_y_0"), val = tensor(false)]; + tensor mh_w_89_cast_fp16 = matmul(transpose_x = mh_w_89_transpose_x_0, transpose_y = mh_w_89_transpose_y_0, x = var_3412_cast_fp16, y = var_3416_cast_fp16)[name = tensor("mh_w_89_cast_fp16")]; + tensor obj_209_cast_fp16 = softmax(axis = var_3258, x = mh_w_89_cast_fp16)[name = tensor("obj_209_cast_fp16")]; + tensor var_3420 = const()[name = tensor("op_3420"), val = tensor([1, 20, 64, 1500])]; + tensor var_3421_cast_fp16 = reshape(shape = var_3420, x = value_59_cast_fp16)[name = tensor("op_3421_cast_fp16")]; + tensor attn_59_transpose_x_0 = const()[name = tensor("attn_59_transpose_x_0"), val = tensor(false)]; + tensor attn_59_transpose_y_0 = const()[name = tensor("attn_59_transpose_y_0"), val = tensor(true)]; + tensor attn_59_cast_fp16 = matmul(transpose_x = attn_59_transpose_x_0, transpose_y = attn_59_transpose_y_0, x = var_3421_cast_fp16, y = obj_209_cast_fp16)[name = tensor("attn_59_cast_fp16")]; + tensor var_3424 = const()[name = tensor("op_3424"), val = tensor([1, 1280, 1, 1])]; + tensor input_143_cast_fp16 = reshape(shape = var_3424, x = attn_59_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor obj_207_pad_type_0 = const()[name = tensor("obj_207_pad_type_0"), val = tensor("valid")]; + tensor obj_207_strides_0 = const()[name = tensor("obj_207_strides_0"), val = tensor([1, 1])]; + tensor obj_207_pad_0 = const()[name = tensor("obj_207_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_207_dilations_0 = const()[name = tensor("obj_207_dilations_0"), val = tensor([1, 1])]; + tensor obj_207_groups_0 = const()[name = tensor("obj_207_groups_0"), val = tensor(1)]; + tensor layers_14_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(891524928)))]; + tensor layers_14_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(894801792)))]; + tensor obj_207_cast_fp16 = conv(bias = layers_14_encoder_attn_o_proj_bias_to_fp16, dilations = obj_207_dilations_0, groups = obj_207_groups_0, pad = obj_207_pad_0, pad_type = obj_207_pad_type_0, strides = obj_207_strides_0, weight = layers_14_encoder_attn_o_proj_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("obj_207_cast_fp16")]; + tensor inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = obj_207_cast_fp16)[name = tensor("inputs_89_cast_fp16")]; + tensor out_89_axes_0 = const()[name = tensor("out_89_axes_0"), val = tensor([1])]; + tensor var_3442_to_fp16 = const()[name = tensor("op_3442_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_89_cast_fp16 = layer_norm(axes = out_89_axes_0, epsilon = var_3442_to_fp16, x = inputs_89_cast_fp16)[name = tensor("out_89_cast_fp16")]; + tensor input_145_gamma_0_to_fp16 = const()[name = tensor("input_145_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(894804416)))]; + tensor input_145_beta_0_to_fp16 = const()[name = tensor("input_145_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(894807040)))]; + tensor input_145_epsilon_0_to_fp16 = const()[name = tensor("input_145_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_145_cast_fp16 = batch_norm(beta = input_145_beta_0_to_fp16, epsilon = input_145_epsilon_0_to_fp16, gamma = input_145_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_89_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor input_147_pad_type_0 = const()[name = tensor("input_147_pad_type_0"), val = tensor("valid")]; + tensor input_147_strides_0 = const()[name = tensor("input_147_strides_0"), val = tensor([1, 1])]; + tensor input_147_pad_0 = const()[name = tensor("input_147_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_147_dilations_0 = const()[name = tensor("input_147_dilations_0"), val = tensor([1, 1])]; + tensor input_147_groups_0 = const()[name = tensor("input_147_groups_0"), val = tensor(1)]; + tensor layers_14_fc1_weight_to_fp16 = const()[name = tensor("layers_14_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(894809664)))]; + tensor layers_14_fc1_bias_to_fp16 = const()[name = tensor("layers_14_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(907916928)))]; + tensor input_147_cast_fp16 = conv(bias = layers_14_fc1_bias_to_fp16, dilations = input_147_dilations_0, groups = input_147_groups_0, pad = input_147_pad_0, pad_type = input_147_pad_type_0, strides = input_147_strides_0, weight = layers_14_fc1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_mode_0 = const()[name = tensor("input_149_mode_0"), val = tensor("EXACT")]; + tensor input_149_cast_fp16 = gelu(mode = input_149_mode_0, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor hidden_states_31_pad_type_0 = const()[name = tensor("hidden_states_31_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_31_strides_0 = const()[name = tensor("hidden_states_31_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_31_pad_0 = const()[name = tensor("hidden_states_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_31_dilations_0 = const()[name = tensor("hidden_states_31_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_31_groups_0 = const()[name = tensor("hidden_states_31_groups_0"), val = tensor(1)]; + tensor layers_14_fc2_weight_to_fp16 = const()[name = tensor("layers_14_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(907927232)))]; + tensor layers_14_fc2_bias_to_fp16 = const()[name = tensor("layers_14_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(921034496)))]; + tensor hidden_states_31_cast_fp16 = conv(bias = layers_14_fc2_bias_to_fp16, dilations = hidden_states_31_dilations_0, groups = hidden_states_31_groups_0, pad = hidden_states_31_pad_0, pad_type = hidden_states_31_pad_type_0, strides = hidden_states_31_strides_0, weight = layers_14_fc2_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("hidden_states_31_cast_fp16")]; + tensor inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = hidden_states_31_cast_fp16)[name = tensor("inputs_91_cast_fp16")]; + tensor var_3477 = const()[name = tensor("op_3477"), val = tensor(3)]; + tensor out_91_axes_0 = const()[name = tensor("out_91_axes_0"), val = tensor([1])]; + tensor var_3502_to_fp16 = const()[name = tensor("op_3502_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_91_cast_fp16 = layer_norm(axes = out_91_axes_0, epsilon = var_3502_to_fp16, x = inputs_91_cast_fp16)[name = tensor("out_91_cast_fp16")]; + tensor obj_211_gamma_0_to_fp16 = const()[name = tensor("obj_211_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(921037120)))]; + tensor obj_211_beta_0_to_fp16 = const()[name = tensor("obj_211_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(921039744)))]; + tensor obj_211_epsilon_0_to_fp16 = const()[name = tensor("obj_211_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_211_cast_fp16 = batch_norm(beta = obj_211_beta_0_to_fp16, epsilon = obj_211_epsilon_0_to_fp16, gamma = obj_211_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_91_cast_fp16)[name = tensor("obj_211_cast_fp16")]; + tensor query_61_pad_type_0 = const()[name = tensor("query_61_pad_type_0"), val = tensor("valid")]; + tensor query_61_strides_0 = const()[name = tensor("query_61_strides_0"), val = tensor([1, 1])]; + tensor query_61_pad_0 = const()[name = tensor("query_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_61_dilations_0 = const()[name = tensor("query_61_dilations_0"), val = tensor([1, 1])]; + tensor query_61_groups_0 = const()[name = tensor("query_61_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(921042368)))]; + tensor layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(924319232)))]; + tensor query_61_cast_fp16 = conv(bias = layers_15_self_attn_q_proj_bias_to_fp16, dilations = query_61_dilations_0, groups = query_61_groups_0, pad = query_61_pad_0, pad_type = query_61_pad_type_0, strides = query_61_strides_0, weight = layers_15_self_attn_q_proj_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor("query_61_cast_fp16")]; + tensor current_key_31_pad_type_0 = const()[name = tensor("current_key_31_pad_type_0"), val = tensor("valid")]; + tensor current_key_31_strides_0 = const()[name = tensor("current_key_31_strides_0"), val = tensor([1, 1])]; + tensor current_key_31_pad_0 = const()[name = tensor("current_key_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_31_dilations_0 = const()[name = tensor("current_key_31_dilations_0"), val = tensor([1, 1])]; + tensor current_key_31_groups_0 = const()[name = tensor("current_key_31_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(924321856)))]; + tensor current_key_31_cast_fp16 = conv(dilations = current_key_31_dilations_0, groups = current_key_31_groups_0, pad = current_key_31_pad_0, pad_type = current_key_31_pad_type_0, strides = current_key_31_strides_0, weight = layers_15_self_attn_k_proj_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor("current_key_31_cast_fp16")]; + tensor current_value_31_pad_type_0 = const()[name = tensor("current_value_31_pad_type_0"), val = tensor("valid")]; + tensor current_value_31_strides_0 = const()[name = tensor("current_value_31_strides_0"), val = tensor([1, 1])]; + tensor current_value_31_pad_0 = const()[name = tensor("current_value_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_31_dilations_0 = const()[name = tensor("current_value_31_dilations_0"), val = tensor([1, 1])]; + tensor current_value_31_groups_0 = const()[name = tensor("current_value_31_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(927598720)))]; + tensor layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(930875584)))]; + tensor current_value_31_cast_fp16 = conv(bias = layers_15_self_attn_v_proj_bias_to_fp16, dilations = current_value_31_dilations_0, groups = current_value_31_groups_0, pad = current_value_31_pad_0, pad_type = current_value_31_pad_type_0, strides = current_value_31_strides_0, weight = layers_15_self_attn_v_proj_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor("current_value_31_cast_fp16")]; + tensor var_3541_cast_fp16 = mul(x = var_103_cast_fp16_15, y = var_239_cast_fp16)[name = tensor("op_3541_cast_fp16")]; + tensor var_3542_cast_fp16 = mul(x = current_key_31_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_3542_cast_fp16")]; + tensor key_61_cast_fp16 = add(x = var_3541_cast_fp16, y = var_3542_cast_fp16)[name = tensor("key_61_cast_fp16")]; + tensor var_3545_cast_fp16 = mul(x = var_138_cast_fp16_15, y = var_239_cast_fp16)[name = tensor("op_3545_cast_fp16")]; + tensor var_3546_cast_fp16 = mul(x = current_value_31_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_3546_cast_fp16")]; + tensor value_61_cast_fp16 = add(x = var_3545_cast_fp16, y = var_3546_cast_fp16)[name = tensor("value_61_cast_fp16")]; + tensor var_3550 = const()[name = tensor("op_3550"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_61_cast_fp16 = reshape(shape = var_3550, x = query_61_cast_fp16)[name = tensor("mh_q_61_cast_fp16")]; + tensor var_3552_to_fp16 = const()[name = tensor("op_3552_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3553_cast_fp16 = mul(x = mh_q_61_cast_fp16, y = var_3552_to_fp16)[name = tensor("op_3553_cast_fp16")]; + tensor var_3556 = const()[name = tensor("op_3556"), val = tensor([1, 20, 64, 448])]; + tensor var_3557_cast_fp16 = reshape(shape = var_3556, x = key_61_cast_fp16)[name = tensor("op_3557_cast_fp16")]; + tensor mh_w_91_transpose_x_0 = const()[name = tensor("mh_w_91_transpose_x_0"), val = tensor(true)]; + tensor mh_w_91_transpose_y_0 = const()[name = tensor("mh_w_91_transpose_y_0"), val = tensor(false)]; + tensor mh_w_91_cast_fp16 = matmul(transpose_x = mh_w_91_transpose_x_0, transpose_y = mh_w_91_transpose_y_0, x = var_3553_cast_fp16, y = var_3557_cast_fp16)[name = tensor("mh_w_91_cast_fp16")]; + tensor mh_w_93_cast_fp16 = add(x = mh_w_91_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_93_cast_fp16")]; + tensor var_3565_cast_fp16 = softmax(axis = var_3477, x = mh_w_93_cast_fp16)[name = tensor("op_3565_cast_fp16")]; + tensor var_3566 = const()[name = tensor("op_3566"), val = tensor([1, 20, 64, 448])]; + tensor var_3567_cast_fp16 = reshape(shape = var_3566, x = value_61_cast_fp16)[name = tensor("op_3567_cast_fp16")]; + tensor attn_61_transpose_x_0 = const()[name = tensor("attn_61_transpose_x_0"), val = tensor(false)]; + tensor attn_61_transpose_y_0 = const()[name = tensor("attn_61_transpose_y_0"), val = tensor(true)]; + tensor attn_61_cast_fp16 = matmul(transpose_x = attn_61_transpose_x_0, transpose_y = attn_61_transpose_y_0, x = var_3567_cast_fp16, y = var_3565_cast_fp16)[name = tensor("attn_61_cast_fp16")]; + tensor var_3570 = const()[name = tensor("op_3570"), val = tensor([1, 1280, 1, 1])]; + tensor input_151_cast_fp16 = reshape(shape = var_3570, x = attn_61_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor obj_217_pad_type_0 = const()[name = tensor("obj_217_pad_type_0"), val = tensor("valid")]; + tensor obj_217_strides_0 = const()[name = tensor("obj_217_strides_0"), val = tensor([1, 1])]; + tensor obj_217_pad_0 = const()[name = tensor("obj_217_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_217_dilations_0 = const()[name = tensor("obj_217_dilations_0"), val = tensor([1, 1])]; + tensor obj_217_groups_0 = const()[name = tensor("obj_217_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(930878208)))]; + tensor layers_15_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(934155072)))]; + tensor obj_217_cast_fp16 = conv(bias = layers_15_self_attn_o_proj_bias_to_fp16, dilations = obj_217_dilations_0, groups = obj_217_groups_0, pad = obj_217_pad_0, pad_type = obj_217_pad_type_0, strides = obj_217_strides_0, weight = layers_15_self_attn_o_proj_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("obj_217_cast_fp16")]; + tensor inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = obj_217_cast_fp16)[name = tensor("inputs_93_cast_fp16")]; + tensor out_93_axes_0 = const()[name = tensor("out_93_axes_0"), val = tensor([1])]; + tensor var_3592_to_fp16 = const()[name = tensor("op_3592_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_93_cast_fp16 = layer_norm(axes = out_93_axes_0, epsilon = var_3592_to_fp16, x = inputs_93_cast_fp16)[name = tensor("out_93_cast_fp16")]; + tensor obj_219_gamma_0_to_fp16 = const()[name = tensor("obj_219_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(934157696)))]; + tensor obj_219_beta_0_to_fp16 = const()[name = tensor("obj_219_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(934160320)))]; + tensor obj_219_epsilon_0_to_fp16 = const()[name = tensor("obj_219_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_219_cast_fp16 = batch_norm(beta = obj_219_beta_0_to_fp16, epsilon = obj_219_epsilon_0_to_fp16, gamma = obj_219_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_93_cast_fp16)[name = tensor("obj_219_cast_fp16")]; + tensor query_63_pad_type_0 = const()[name = tensor("query_63_pad_type_0"), val = tensor("valid")]; + tensor query_63_strides_0 = const()[name = tensor("query_63_strides_0"), val = tensor([1, 1])]; + tensor query_63_pad_0 = const()[name = tensor("query_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_63_dilations_0 = const()[name = tensor("query_63_dilations_0"), val = tensor([1, 1])]; + tensor query_63_groups_0 = const()[name = tensor("query_63_groups_0"), val = tensor(1)]; + tensor layers_15_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(934162944)))]; + tensor layers_15_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(937439808)))]; + tensor query_63_cast_fp16 = conv(bias = layers_15_encoder_attn_q_proj_bias_to_fp16, dilations = query_63_dilations_0, groups = query_63_groups_0, pad = query_63_pad_0, pad_type = query_63_pad_type_0, strides = query_63_strides_0, weight = layers_15_encoder_attn_q_proj_weight_to_fp16, x = obj_219_cast_fp16)[name = tensor("query_63_cast_fp16")]; + tensor key_63_pad_type_0 = const()[name = tensor("key_63_pad_type_0"), val = tensor("valid")]; + tensor key_63_strides_0 = const()[name = tensor("key_63_strides_0"), val = tensor([1, 1])]; + tensor key_63_pad_0 = const()[name = tensor("key_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_63_dilations_0 = const()[name = tensor("key_63_dilations_0"), val = tensor([1, 1])]; + tensor key_63_groups_0 = const()[name = tensor("key_63_groups_0"), val = tensor(1)]; + tensor layers_15_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(937442432)))]; + tensor key_63_cast_fp16 = conv(dilations = key_63_dilations_0, groups = key_63_groups_0, pad = key_63_pad_0, pad_type = key_63_pad_type_0, strides = key_63_strides_0, weight = layers_15_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_63_cast_fp16")]; + tensor value_63_pad_type_0 = const()[name = tensor("value_63_pad_type_0"), val = tensor("valid")]; + tensor value_63_strides_0 = const()[name = tensor("value_63_strides_0"), val = tensor([1, 1])]; + tensor value_63_pad_0 = const()[name = tensor("value_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_63_dilations_0 = const()[name = tensor("value_63_dilations_0"), val = tensor([1, 1])]; + tensor value_63_groups_0 = const()[name = tensor("value_63_groups_0"), val = tensor(1)]; + tensor layers_15_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(940719296)))]; + tensor layers_15_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(943996160)))]; + tensor value_63_cast_fp16 = conv(bias = layers_15_encoder_attn_v_proj_bias_to_fp16, dilations = value_63_dilations_0, groups = value_63_groups_0, pad = value_63_pad_0, pad_type = value_63_pad_type_0, strides = value_63_strides_0, weight = layers_15_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_63_cast_fp16")]; + tensor var_3628 = const()[name = tensor("op_3628"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_63_cast_fp16 = reshape(shape = var_3628, x = query_63_cast_fp16)[name = tensor("mh_q_63_cast_fp16")]; + tensor var_3630_to_fp16 = const()[name = tensor("op_3630_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3631_cast_fp16 = mul(x = mh_q_63_cast_fp16, y = var_3630_to_fp16)[name = tensor("op_3631_cast_fp16")]; + tensor var_3634 = const()[name = tensor("op_3634"), val = tensor([1, 20, 64, 1500])]; + tensor var_3635_cast_fp16 = reshape(shape = var_3634, x = key_63_cast_fp16)[name = tensor("op_3635_cast_fp16")]; + tensor mh_w_95_transpose_x_0 = const()[name = tensor("mh_w_95_transpose_x_0"), val = tensor(true)]; + tensor mh_w_95_transpose_y_0 = const()[name = tensor("mh_w_95_transpose_y_0"), val = tensor(false)]; + tensor mh_w_95_cast_fp16 = matmul(transpose_x = mh_w_95_transpose_x_0, transpose_y = mh_w_95_transpose_y_0, x = var_3631_cast_fp16, y = var_3635_cast_fp16)[name = tensor("mh_w_95_cast_fp16")]; + tensor obj_223_cast_fp16 = softmax(axis = var_3477, x = mh_w_95_cast_fp16)[name = tensor("obj_223_cast_fp16")]; + tensor var_3639 = const()[name = tensor("op_3639"), val = tensor([1, 20, 64, 1500])]; + tensor var_3640_cast_fp16 = reshape(shape = var_3639, x = value_63_cast_fp16)[name = tensor("op_3640_cast_fp16")]; + tensor attn_63_transpose_x_0 = const()[name = tensor("attn_63_transpose_x_0"), val = tensor(false)]; + tensor attn_63_transpose_y_0 = const()[name = tensor("attn_63_transpose_y_0"), val = tensor(true)]; + tensor attn_63_cast_fp16 = matmul(transpose_x = attn_63_transpose_x_0, transpose_y = attn_63_transpose_y_0, x = var_3640_cast_fp16, y = obj_223_cast_fp16)[name = tensor("attn_63_cast_fp16")]; + tensor var_3643 = const()[name = tensor("op_3643"), val = tensor([1, 1280, 1, 1])]; + tensor input_153_cast_fp16 = reshape(shape = var_3643, x = attn_63_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor obj_221_pad_type_0 = const()[name = tensor("obj_221_pad_type_0"), val = tensor("valid")]; + tensor obj_221_strides_0 = const()[name = tensor("obj_221_strides_0"), val = tensor([1, 1])]; + tensor obj_221_pad_0 = const()[name = tensor("obj_221_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_221_dilations_0 = const()[name = tensor("obj_221_dilations_0"), val = tensor([1, 1])]; + tensor obj_221_groups_0 = const()[name = tensor("obj_221_groups_0"), val = tensor(1)]; + tensor layers_15_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(943998784)))]; + tensor layers_15_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(947275648)))]; + tensor obj_221_cast_fp16 = conv(bias = layers_15_encoder_attn_o_proj_bias_to_fp16, dilations = obj_221_dilations_0, groups = obj_221_groups_0, pad = obj_221_pad_0, pad_type = obj_221_pad_type_0, strides = obj_221_strides_0, weight = layers_15_encoder_attn_o_proj_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("obj_221_cast_fp16")]; + tensor inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = obj_221_cast_fp16)[name = tensor("inputs_95_cast_fp16")]; + tensor out_95_axes_0 = const()[name = tensor("out_95_axes_0"), val = tensor([1])]; + tensor var_3661_to_fp16 = const()[name = tensor("op_3661_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_95_cast_fp16 = layer_norm(axes = out_95_axes_0, epsilon = var_3661_to_fp16, x = inputs_95_cast_fp16)[name = tensor("out_95_cast_fp16")]; + tensor input_155_gamma_0_to_fp16 = const()[name = tensor("input_155_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(947278272)))]; + tensor input_155_beta_0_to_fp16 = const()[name = tensor("input_155_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(947280896)))]; + tensor input_155_epsilon_0_to_fp16 = const()[name = tensor("input_155_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_155_cast_fp16 = batch_norm(beta = input_155_beta_0_to_fp16, epsilon = input_155_epsilon_0_to_fp16, gamma = input_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_95_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor input_157_pad_type_0 = const()[name = tensor("input_157_pad_type_0"), val = tensor("valid")]; + tensor input_157_strides_0 = const()[name = tensor("input_157_strides_0"), val = tensor([1, 1])]; + tensor input_157_pad_0 = const()[name = tensor("input_157_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_157_dilations_0 = const()[name = tensor("input_157_dilations_0"), val = tensor([1, 1])]; + tensor input_157_groups_0 = const()[name = tensor("input_157_groups_0"), val = tensor(1)]; + tensor layers_15_fc1_weight_to_fp16 = const()[name = tensor("layers_15_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(947283520)))]; + tensor layers_15_fc1_bias_to_fp16 = const()[name = tensor("layers_15_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(960390784)))]; + tensor input_157_cast_fp16 = conv(bias = layers_15_fc1_bias_to_fp16, dilations = input_157_dilations_0, groups = input_157_groups_0, pad = input_157_pad_0, pad_type = input_157_pad_type_0, strides = input_157_strides_0, weight = layers_15_fc1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_mode_0 = const()[name = tensor("input_159_mode_0"), val = tensor("EXACT")]; + tensor input_159_cast_fp16 = gelu(mode = input_159_mode_0, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor hidden_states_33_pad_type_0 = const()[name = tensor("hidden_states_33_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_33_strides_0 = const()[name = tensor("hidden_states_33_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_33_pad_0 = const()[name = tensor("hidden_states_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_33_dilations_0 = const()[name = tensor("hidden_states_33_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_33_groups_0 = const()[name = tensor("hidden_states_33_groups_0"), val = tensor(1)]; + tensor layers_15_fc2_weight_to_fp16 = const()[name = tensor("layers_15_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(960401088)))]; + tensor layers_15_fc2_bias_to_fp16 = const()[name = tensor("layers_15_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(973508352)))]; + tensor hidden_states_33_cast_fp16 = conv(bias = layers_15_fc2_bias_to_fp16, dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_15_fc2_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("hidden_states_33_cast_fp16")]; + tensor inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = hidden_states_33_cast_fp16)[name = tensor("inputs_97_cast_fp16")]; + tensor var_3696 = const()[name = tensor("op_3696"), val = tensor(3)]; + tensor out_97_axes_0 = const()[name = tensor("out_97_axes_0"), val = tensor([1])]; + tensor var_3721_to_fp16 = const()[name = tensor("op_3721_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_97_cast_fp16 = layer_norm(axes = out_97_axes_0, epsilon = var_3721_to_fp16, x = inputs_97_cast_fp16)[name = tensor("out_97_cast_fp16")]; + tensor obj_225_gamma_0_to_fp16 = const()[name = tensor("obj_225_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(973510976)))]; + tensor obj_225_beta_0_to_fp16 = const()[name = tensor("obj_225_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(973513600)))]; + tensor obj_225_epsilon_0_to_fp16 = const()[name = tensor("obj_225_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_225_cast_fp16 = batch_norm(beta = obj_225_beta_0_to_fp16, epsilon = obj_225_epsilon_0_to_fp16, gamma = obj_225_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_97_cast_fp16)[name = tensor("obj_225_cast_fp16")]; + tensor query_65_pad_type_0 = const()[name = tensor("query_65_pad_type_0"), val = tensor("valid")]; + tensor query_65_strides_0 = const()[name = tensor("query_65_strides_0"), val = tensor([1, 1])]; + tensor query_65_pad_0 = const()[name = tensor("query_65_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_65_dilations_0 = const()[name = tensor("query_65_dilations_0"), val = tensor([1, 1])]; + tensor query_65_groups_0 = const()[name = tensor("query_65_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(973516224)))]; + tensor layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(976793088)))]; + tensor query_65_cast_fp16 = conv(bias = layers_16_self_attn_q_proj_bias_to_fp16, dilations = query_65_dilations_0, groups = query_65_groups_0, pad = query_65_pad_0, pad_type = query_65_pad_type_0, strides = query_65_strides_0, weight = layers_16_self_attn_q_proj_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("query_65_cast_fp16")]; + tensor current_key_33_pad_type_0 = const()[name = tensor("current_key_33_pad_type_0"), val = tensor("valid")]; + tensor current_key_33_strides_0 = const()[name = tensor("current_key_33_strides_0"), val = tensor([1, 1])]; + tensor current_key_33_pad_0 = const()[name = tensor("current_key_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_33_dilations_0 = const()[name = tensor("current_key_33_dilations_0"), val = tensor([1, 1])]; + tensor current_key_33_groups_0 = const()[name = tensor("current_key_33_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(976795712)))]; + tensor current_key_33_cast_fp16 = conv(dilations = current_key_33_dilations_0, groups = current_key_33_groups_0, pad = current_key_33_pad_0, pad_type = current_key_33_pad_type_0, strides = current_key_33_strides_0, weight = layers_16_self_attn_k_proj_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("current_key_33_cast_fp16")]; + tensor current_value_33_pad_type_0 = const()[name = tensor("current_value_33_pad_type_0"), val = tensor("valid")]; + tensor current_value_33_strides_0 = const()[name = tensor("current_value_33_strides_0"), val = tensor([1, 1])]; + tensor current_value_33_pad_0 = const()[name = tensor("current_value_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_33_dilations_0 = const()[name = tensor("current_value_33_dilations_0"), val = tensor([1, 1])]; + tensor current_value_33_groups_0 = const()[name = tensor("current_value_33_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(980072576)))]; + tensor layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(983349440)))]; + tensor current_value_33_cast_fp16 = conv(bias = layers_16_self_attn_v_proj_bias_to_fp16, dilations = current_value_33_dilations_0, groups = current_value_33_groups_0, pad = current_value_33_pad_0, pad_type = current_value_33_pad_type_0, strides = current_value_33_strides_0, weight = layers_16_self_attn_v_proj_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("current_value_33_cast_fp16")]; + tensor var_3760_cast_fp16 = mul(x = var_103_cast_fp16_16, y = var_239_cast_fp16)[name = tensor("op_3760_cast_fp16")]; + tensor var_3761_cast_fp16 = mul(x = current_key_33_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_3761_cast_fp16")]; + tensor key_65_cast_fp16 = add(x = var_3760_cast_fp16, y = var_3761_cast_fp16)[name = tensor("key_65_cast_fp16")]; + tensor var_3764_cast_fp16 = mul(x = var_138_cast_fp16_16, y = var_239_cast_fp16)[name = tensor("op_3764_cast_fp16")]; + tensor var_3765_cast_fp16 = mul(x = current_value_33_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_3765_cast_fp16")]; + tensor value_65_cast_fp16 = add(x = var_3764_cast_fp16, y = var_3765_cast_fp16)[name = tensor("value_65_cast_fp16")]; + tensor var_3769 = const()[name = tensor("op_3769"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_65_cast_fp16 = reshape(shape = var_3769, x = query_65_cast_fp16)[name = tensor("mh_q_65_cast_fp16")]; + tensor var_3771_to_fp16 = const()[name = tensor("op_3771_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3772_cast_fp16 = mul(x = mh_q_65_cast_fp16, y = var_3771_to_fp16)[name = tensor("op_3772_cast_fp16")]; + tensor var_3775 = const()[name = tensor("op_3775"), val = tensor([1, 20, 64, 448])]; + tensor var_3776_cast_fp16 = reshape(shape = var_3775, x = key_65_cast_fp16)[name = tensor("op_3776_cast_fp16")]; + tensor mh_w_97_transpose_x_0 = const()[name = tensor("mh_w_97_transpose_x_0"), val = tensor(true)]; + tensor mh_w_97_transpose_y_0 = const()[name = tensor("mh_w_97_transpose_y_0"), val = tensor(false)]; + tensor mh_w_97_cast_fp16 = matmul(transpose_x = mh_w_97_transpose_x_0, transpose_y = mh_w_97_transpose_y_0, x = var_3772_cast_fp16, y = var_3776_cast_fp16)[name = tensor("mh_w_97_cast_fp16")]; + tensor mh_w_99_cast_fp16 = add(x = mh_w_97_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_99_cast_fp16")]; + tensor var_3784_cast_fp16 = softmax(axis = var_3696, x = mh_w_99_cast_fp16)[name = tensor("op_3784_cast_fp16")]; + tensor var_3785 = const()[name = tensor("op_3785"), val = tensor([1, 20, 64, 448])]; + tensor var_3786_cast_fp16 = reshape(shape = var_3785, x = value_65_cast_fp16)[name = tensor("op_3786_cast_fp16")]; + tensor attn_65_transpose_x_0 = const()[name = tensor("attn_65_transpose_x_0"), val = tensor(false)]; + tensor attn_65_transpose_y_0 = const()[name = tensor("attn_65_transpose_y_0"), val = tensor(true)]; + tensor attn_65_cast_fp16 = matmul(transpose_x = attn_65_transpose_x_0, transpose_y = attn_65_transpose_y_0, x = var_3786_cast_fp16, y = var_3784_cast_fp16)[name = tensor("attn_65_cast_fp16")]; + tensor var_3789 = const()[name = tensor("op_3789"), val = tensor([1, 1280, 1, 1])]; + tensor input_161_cast_fp16 = reshape(shape = var_3789, x = attn_65_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor obj_231_pad_type_0 = const()[name = tensor("obj_231_pad_type_0"), val = tensor("valid")]; + tensor obj_231_strides_0 = const()[name = tensor("obj_231_strides_0"), val = tensor([1, 1])]; + tensor obj_231_pad_0 = const()[name = tensor("obj_231_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_231_dilations_0 = const()[name = tensor("obj_231_dilations_0"), val = tensor([1, 1])]; + tensor obj_231_groups_0 = const()[name = tensor("obj_231_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(983352064)))]; + tensor layers_16_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(986628928)))]; + tensor obj_231_cast_fp16 = conv(bias = layers_16_self_attn_o_proj_bias_to_fp16, dilations = obj_231_dilations_0, groups = obj_231_groups_0, pad = obj_231_pad_0, pad_type = obj_231_pad_type_0, strides = obj_231_strides_0, weight = layers_16_self_attn_o_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("obj_231_cast_fp16")]; + tensor inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = obj_231_cast_fp16)[name = tensor("inputs_99_cast_fp16")]; + tensor out_99_axes_0 = const()[name = tensor("out_99_axes_0"), val = tensor([1])]; + tensor var_3811_to_fp16 = const()[name = tensor("op_3811_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_99_cast_fp16 = layer_norm(axes = out_99_axes_0, epsilon = var_3811_to_fp16, x = inputs_99_cast_fp16)[name = tensor("out_99_cast_fp16")]; + tensor obj_233_gamma_0_to_fp16 = const()[name = tensor("obj_233_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(986631552)))]; + tensor obj_233_beta_0_to_fp16 = const()[name = tensor("obj_233_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(986634176)))]; + tensor obj_233_epsilon_0_to_fp16 = const()[name = tensor("obj_233_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_233_cast_fp16 = batch_norm(beta = obj_233_beta_0_to_fp16, epsilon = obj_233_epsilon_0_to_fp16, gamma = obj_233_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_99_cast_fp16)[name = tensor("obj_233_cast_fp16")]; + tensor query_67_pad_type_0 = const()[name = tensor("query_67_pad_type_0"), val = tensor("valid")]; + tensor query_67_strides_0 = const()[name = tensor("query_67_strides_0"), val = tensor([1, 1])]; + tensor query_67_pad_0 = const()[name = tensor("query_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_67_dilations_0 = const()[name = tensor("query_67_dilations_0"), val = tensor([1, 1])]; + tensor query_67_groups_0 = const()[name = tensor("query_67_groups_0"), val = tensor(1)]; + tensor layers_16_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(986636800)))]; + tensor layers_16_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(989913664)))]; + tensor query_67_cast_fp16 = conv(bias = layers_16_encoder_attn_q_proj_bias_to_fp16, dilations = query_67_dilations_0, groups = query_67_groups_0, pad = query_67_pad_0, pad_type = query_67_pad_type_0, strides = query_67_strides_0, weight = layers_16_encoder_attn_q_proj_weight_to_fp16, x = obj_233_cast_fp16)[name = tensor("query_67_cast_fp16")]; + tensor key_67_pad_type_0 = const()[name = tensor("key_67_pad_type_0"), val = tensor("valid")]; + tensor key_67_strides_0 = const()[name = tensor("key_67_strides_0"), val = tensor([1, 1])]; + tensor key_67_pad_0 = const()[name = tensor("key_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_67_dilations_0 = const()[name = tensor("key_67_dilations_0"), val = tensor([1, 1])]; + tensor key_67_groups_0 = const()[name = tensor("key_67_groups_0"), val = tensor(1)]; + tensor layers_16_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(989916288)))]; + tensor key_67_cast_fp16 = conv(dilations = key_67_dilations_0, groups = key_67_groups_0, pad = key_67_pad_0, pad_type = key_67_pad_type_0, strides = key_67_strides_0, weight = layers_16_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_67_cast_fp16")]; + tensor value_67_pad_type_0 = const()[name = tensor("value_67_pad_type_0"), val = tensor("valid")]; + tensor value_67_strides_0 = const()[name = tensor("value_67_strides_0"), val = tensor([1, 1])]; + tensor value_67_pad_0 = const()[name = tensor("value_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_67_dilations_0 = const()[name = tensor("value_67_dilations_0"), val = tensor([1, 1])]; + tensor value_67_groups_0 = const()[name = tensor("value_67_groups_0"), val = tensor(1)]; + tensor layers_16_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(993193152)))]; + tensor layers_16_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(996470016)))]; + tensor value_67_cast_fp16 = conv(bias = layers_16_encoder_attn_v_proj_bias_to_fp16, dilations = value_67_dilations_0, groups = value_67_groups_0, pad = value_67_pad_0, pad_type = value_67_pad_type_0, strides = value_67_strides_0, weight = layers_16_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_67_cast_fp16")]; + tensor var_3847 = const()[name = tensor("op_3847"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_67_cast_fp16 = reshape(shape = var_3847, x = query_67_cast_fp16)[name = tensor("mh_q_67_cast_fp16")]; + tensor var_3849_to_fp16 = const()[name = tensor("op_3849_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3850_cast_fp16 = mul(x = mh_q_67_cast_fp16, y = var_3849_to_fp16)[name = tensor("op_3850_cast_fp16")]; + tensor var_3853 = const()[name = tensor("op_3853"), val = tensor([1, 20, 64, 1500])]; + tensor var_3854_cast_fp16 = reshape(shape = var_3853, x = key_67_cast_fp16)[name = tensor("op_3854_cast_fp16")]; + tensor mh_w_101_transpose_x_0 = const()[name = tensor("mh_w_101_transpose_x_0"), val = tensor(true)]; + tensor mh_w_101_transpose_y_0 = const()[name = tensor("mh_w_101_transpose_y_0"), val = tensor(false)]; + tensor mh_w_101_cast_fp16 = matmul(transpose_x = mh_w_101_transpose_x_0, transpose_y = mh_w_101_transpose_y_0, x = var_3850_cast_fp16, y = var_3854_cast_fp16)[name = tensor("mh_w_101_cast_fp16")]; + tensor obj_237_cast_fp16 = softmax(axis = var_3696, x = mh_w_101_cast_fp16)[name = tensor("obj_237_cast_fp16")]; + tensor var_3858 = const()[name = tensor("op_3858"), val = tensor([1, 20, 64, 1500])]; + tensor var_3859_cast_fp16 = reshape(shape = var_3858, x = value_67_cast_fp16)[name = tensor("op_3859_cast_fp16")]; + tensor attn_67_transpose_x_0 = const()[name = tensor("attn_67_transpose_x_0"), val = tensor(false)]; + tensor attn_67_transpose_y_0 = const()[name = tensor("attn_67_transpose_y_0"), val = tensor(true)]; + tensor attn_67_cast_fp16 = matmul(transpose_x = attn_67_transpose_x_0, transpose_y = attn_67_transpose_y_0, x = var_3859_cast_fp16, y = obj_237_cast_fp16)[name = tensor("attn_67_cast_fp16")]; + tensor var_3862 = const()[name = tensor("op_3862"), val = tensor([1, 1280, 1, 1])]; + tensor input_163_cast_fp16 = reshape(shape = var_3862, x = attn_67_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor obj_235_pad_type_0 = const()[name = tensor("obj_235_pad_type_0"), val = tensor("valid")]; + tensor obj_235_strides_0 = const()[name = tensor("obj_235_strides_0"), val = tensor([1, 1])]; + tensor obj_235_pad_0 = const()[name = tensor("obj_235_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_235_dilations_0 = const()[name = tensor("obj_235_dilations_0"), val = tensor([1, 1])]; + tensor obj_235_groups_0 = const()[name = tensor("obj_235_groups_0"), val = tensor(1)]; + tensor layers_16_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(996472640)))]; + tensor layers_16_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(999749504)))]; + tensor obj_235_cast_fp16 = conv(bias = layers_16_encoder_attn_o_proj_bias_to_fp16, dilations = obj_235_dilations_0, groups = obj_235_groups_0, pad = obj_235_pad_0, pad_type = obj_235_pad_type_0, strides = obj_235_strides_0, weight = layers_16_encoder_attn_o_proj_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("obj_235_cast_fp16")]; + tensor inputs_101_cast_fp16 = add(x = inputs_99_cast_fp16, y = obj_235_cast_fp16)[name = tensor("inputs_101_cast_fp16")]; + tensor out_101_axes_0 = const()[name = tensor("out_101_axes_0"), val = tensor([1])]; + tensor var_3883_to_fp16 = const()[name = tensor("op_3883_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_101_cast_fp16 = layer_norm(axes = out_101_axes_0, epsilon = var_3883_to_fp16, x = inputs_101_cast_fp16)[name = tensor("out_101_cast_fp16")]; + tensor input_165_gamma_0_to_fp16 = const()[name = tensor("input_165_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(999752128)))]; + tensor input_165_beta_0_to_fp16 = const()[name = tensor("input_165_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(999754752)))]; + tensor input_165_epsilon_0_to_fp16 = const()[name = tensor("input_165_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_165_cast_fp16 = batch_norm(beta = input_165_beta_0_to_fp16, epsilon = input_165_epsilon_0_to_fp16, gamma = input_165_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_101_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor input_167_pad_type_0 = const()[name = tensor("input_167_pad_type_0"), val = tensor("valid")]; + tensor input_167_strides_0 = const()[name = tensor("input_167_strides_0"), val = tensor([1, 1])]; + tensor input_167_pad_0 = const()[name = tensor("input_167_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_167_dilations_0 = const()[name = tensor("input_167_dilations_0"), val = tensor([1, 1])]; + tensor input_167_groups_0 = const()[name = tensor("input_167_groups_0"), val = tensor(1)]; + tensor layers_16_fc1_weight_to_fp16 = const()[name = tensor("layers_16_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(999757376)))]; + tensor layers_16_fc1_bias_to_fp16 = const()[name = tensor("layers_16_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1012864640)))]; + tensor input_167_cast_fp16 = conv(bias = layers_16_fc1_bias_to_fp16, dilations = input_167_dilations_0, groups = input_167_groups_0, pad = input_167_pad_0, pad_type = input_167_pad_type_0, strides = input_167_strides_0, weight = layers_16_fc1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor input_169_mode_0 = const()[name = tensor("input_169_mode_0"), val = tensor("EXACT")]; + tensor input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor hidden_states_35_pad_type_0 = const()[name = tensor("hidden_states_35_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_35_strides_0 = const()[name = tensor("hidden_states_35_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_35_pad_0 = const()[name = tensor("hidden_states_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_35_dilations_0 = const()[name = tensor("hidden_states_35_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_35_groups_0 = const()[name = tensor("hidden_states_35_groups_0"), val = tensor(1)]; + tensor layers_16_fc2_weight_to_fp16 = const()[name = tensor("layers_16_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1012874944)))]; + tensor layers_16_fc2_bias_to_fp16 = const()[name = tensor("layers_16_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1025982208)))]; + tensor hidden_states_35_cast_fp16 = conv(bias = layers_16_fc2_bias_to_fp16, dilations = hidden_states_35_dilations_0, groups = hidden_states_35_groups_0, pad = hidden_states_35_pad_0, pad_type = hidden_states_35_pad_type_0, strides = hidden_states_35_strides_0, weight = layers_16_fc2_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("hidden_states_35_cast_fp16")]; + tensor inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = hidden_states_35_cast_fp16)[name = tensor("inputs_103_cast_fp16")]; + tensor var_3919 = const()[name = tensor("op_3919"), val = tensor(3)]; + tensor out_103_axes_0 = const()[name = tensor("out_103_axes_0"), val = tensor([1])]; + tensor var_3944_to_fp16 = const()[name = tensor("op_3944_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_103_cast_fp16 = layer_norm(axes = out_103_axes_0, epsilon = var_3944_to_fp16, x = inputs_103_cast_fp16)[name = tensor("out_103_cast_fp16")]; + tensor obj_239_gamma_0_to_fp16 = const()[name = tensor("obj_239_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1025984832)))]; + tensor obj_239_beta_0_to_fp16 = const()[name = tensor("obj_239_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1025987456)))]; + tensor obj_239_epsilon_0_to_fp16 = const()[name = tensor("obj_239_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_239_cast_fp16 = batch_norm(beta = obj_239_beta_0_to_fp16, epsilon = obj_239_epsilon_0_to_fp16, gamma = obj_239_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_103_cast_fp16)[name = tensor("obj_239_cast_fp16")]; + tensor query_69_pad_type_0 = const()[name = tensor("query_69_pad_type_0"), val = tensor("valid")]; + tensor query_69_strides_0 = const()[name = tensor("query_69_strides_0"), val = tensor([1, 1])]; + tensor query_69_pad_0 = const()[name = tensor("query_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_69_dilations_0 = const()[name = tensor("query_69_dilations_0"), val = tensor([1, 1])]; + tensor query_69_groups_0 = const()[name = tensor("query_69_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1025990080)))]; + tensor layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1029266944)))]; + tensor query_69_cast_fp16 = conv(bias = layers_17_self_attn_q_proj_bias_to_fp16, dilations = query_69_dilations_0, groups = query_69_groups_0, pad = query_69_pad_0, pad_type = query_69_pad_type_0, strides = query_69_strides_0, weight = layers_17_self_attn_q_proj_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor("query_69_cast_fp16")]; + tensor current_key_35_pad_type_0 = const()[name = tensor("current_key_35_pad_type_0"), val = tensor("valid")]; + tensor current_key_35_strides_0 = const()[name = tensor("current_key_35_strides_0"), val = tensor([1, 1])]; + tensor current_key_35_pad_0 = const()[name = tensor("current_key_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_35_dilations_0 = const()[name = tensor("current_key_35_dilations_0"), val = tensor([1, 1])]; + tensor current_key_35_groups_0 = const()[name = tensor("current_key_35_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1029269568)))]; + tensor current_key_35_cast_fp16 = conv(dilations = current_key_35_dilations_0, groups = current_key_35_groups_0, pad = current_key_35_pad_0, pad_type = current_key_35_pad_type_0, strides = current_key_35_strides_0, weight = layers_17_self_attn_k_proj_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor("current_key_35_cast_fp16")]; + tensor current_value_35_pad_type_0 = const()[name = tensor("current_value_35_pad_type_0"), val = tensor("valid")]; + tensor current_value_35_strides_0 = const()[name = tensor("current_value_35_strides_0"), val = tensor([1, 1])]; + tensor current_value_35_pad_0 = const()[name = tensor("current_value_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_35_dilations_0 = const()[name = tensor("current_value_35_dilations_0"), val = tensor([1, 1])]; + tensor current_value_35_groups_0 = const()[name = tensor("current_value_35_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1032546432)))]; + tensor layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1035823296)))]; + tensor current_value_35_cast_fp16 = conv(bias = layers_17_self_attn_v_proj_bias_to_fp16, dilations = current_value_35_dilations_0, groups = current_value_35_groups_0, pad = current_value_35_pad_0, pad_type = current_value_35_pad_type_0, strides = current_value_35_strides_0, weight = layers_17_self_attn_v_proj_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor("current_value_35_cast_fp16")]; + tensor var_3983_cast_fp16 = mul(x = var_103_cast_fp16_17, y = var_239_cast_fp16)[name = tensor("op_3983_cast_fp16")]; + tensor var_3984_cast_fp16 = mul(x = current_key_35_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_3984_cast_fp16")]; + tensor key_69_cast_fp16 = add(x = var_3983_cast_fp16, y = var_3984_cast_fp16)[name = tensor("key_69_cast_fp16")]; + tensor var_3987_cast_fp16 = mul(x = var_138_cast_fp16_17, y = var_239_cast_fp16)[name = tensor("op_3987_cast_fp16")]; + tensor var_3988_cast_fp16 = mul(x = current_value_35_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_3988_cast_fp16")]; + tensor value_69_cast_fp16 = add(x = var_3987_cast_fp16, y = var_3988_cast_fp16)[name = tensor("value_69_cast_fp16")]; + tensor var_3992 = const()[name = tensor("op_3992"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_69_cast_fp16 = reshape(shape = var_3992, x = query_69_cast_fp16)[name = tensor("mh_q_69_cast_fp16")]; + tensor var_3994_to_fp16 = const()[name = tensor("op_3994_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3995_cast_fp16 = mul(x = mh_q_69_cast_fp16, y = var_3994_to_fp16)[name = tensor("op_3995_cast_fp16")]; + tensor var_3998 = const()[name = tensor("op_3998"), val = tensor([1, 20, 64, 448])]; + tensor var_3999_cast_fp16 = reshape(shape = var_3998, x = key_69_cast_fp16)[name = tensor("op_3999_cast_fp16")]; + tensor mh_w_103_transpose_x_0 = const()[name = tensor("mh_w_103_transpose_x_0"), val = tensor(true)]; + tensor mh_w_103_transpose_y_0 = const()[name = tensor("mh_w_103_transpose_y_0"), val = tensor(false)]; + tensor mh_w_103_cast_fp16 = matmul(transpose_x = mh_w_103_transpose_x_0, transpose_y = mh_w_103_transpose_y_0, x = var_3995_cast_fp16, y = var_3999_cast_fp16)[name = tensor("mh_w_103_cast_fp16")]; + tensor mh_w_105_cast_fp16 = add(x = mh_w_103_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_105_cast_fp16")]; + tensor var_4007_cast_fp16 = softmax(axis = var_3919, x = mh_w_105_cast_fp16)[name = tensor("op_4007_cast_fp16")]; + tensor var_4008 = const()[name = tensor("op_4008"), val = tensor([1, 20, 64, 448])]; + tensor var_4009_cast_fp16 = reshape(shape = var_4008, x = value_69_cast_fp16)[name = tensor("op_4009_cast_fp16")]; + tensor attn_69_transpose_x_0 = const()[name = tensor("attn_69_transpose_x_0"), val = tensor(false)]; + tensor attn_69_transpose_y_0 = const()[name = tensor("attn_69_transpose_y_0"), val = tensor(true)]; + tensor attn_69_cast_fp16 = matmul(transpose_x = attn_69_transpose_x_0, transpose_y = attn_69_transpose_y_0, x = var_4009_cast_fp16, y = var_4007_cast_fp16)[name = tensor("attn_69_cast_fp16")]; + tensor var_4012 = const()[name = tensor("op_4012"), val = tensor([1, 1280, 1, 1])]; + tensor input_171_cast_fp16 = reshape(shape = var_4012, x = attn_69_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor obj_245_pad_type_0 = const()[name = tensor("obj_245_pad_type_0"), val = tensor("valid")]; + tensor obj_245_strides_0 = const()[name = tensor("obj_245_strides_0"), val = tensor([1, 1])]; + tensor obj_245_pad_0 = const()[name = tensor("obj_245_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_245_dilations_0 = const()[name = tensor("obj_245_dilations_0"), val = tensor([1, 1])]; + tensor obj_245_groups_0 = const()[name = tensor("obj_245_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1035825920)))]; + tensor layers_17_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1039102784)))]; + tensor obj_245_cast_fp16 = conv(bias = layers_17_self_attn_o_proj_bias_to_fp16, dilations = obj_245_dilations_0, groups = obj_245_groups_0, pad = obj_245_pad_0, pad_type = obj_245_pad_type_0, strides = obj_245_strides_0, weight = layers_17_self_attn_o_proj_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("obj_245_cast_fp16")]; + tensor inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = obj_245_cast_fp16)[name = tensor("inputs_105_cast_fp16")]; + tensor out_105_axes_0 = const()[name = tensor("out_105_axes_0"), val = tensor([1])]; + tensor var_4034_to_fp16 = const()[name = tensor("op_4034_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_105_cast_fp16 = layer_norm(axes = out_105_axes_0, epsilon = var_4034_to_fp16, x = inputs_105_cast_fp16)[name = tensor("out_105_cast_fp16")]; + tensor obj_247_gamma_0_to_fp16 = const()[name = tensor("obj_247_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1039105408)))]; + tensor obj_247_beta_0_to_fp16 = const()[name = tensor("obj_247_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1039108032)))]; + tensor obj_247_epsilon_0_to_fp16 = const()[name = tensor("obj_247_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_247_cast_fp16 = batch_norm(beta = obj_247_beta_0_to_fp16, epsilon = obj_247_epsilon_0_to_fp16, gamma = obj_247_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_105_cast_fp16)[name = tensor("obj_247_cast_fp16")]; + tensor query_71_pad_type_0 = const()[name = tensor("query_71_pad_type_0"), val = tensor("valid")]; + tensor query_71_strides_0 = const()[name = tensor("query_71_strides_0"), val = tensor([1, 1])]; + tensor query_71_pad_0 = const()[name = tensor("query_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_71_dilations_0 = const()[name = tensor("query_71_dilations_0"), val = tensor([1, 1])]; + tensor query_71_groups_0 = const()[name = tensor("query_71_groups_0"), val = tensor(1)]; + tensor layers_17_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1039110656)))]; + tensor layers_17_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1042387520)))]; + tensor query_71_cast_fp16 = conv(bias = layers_17_encoder_attn_q_proj_bias_to_fp16, dilations = query_71_dilations_0, groups = query_71_groups_0, pad = query_71_pad_0, pad_type = query_71_pad_type_0, strides = query_71_strides_0, weight = layers_17_encoder_attn_q_proj_weight_to_fp16, x = obj_247_cast_fp16)[name = tensor("query_71_cast_fp16")]; + tensor key_71_pad_type_0 = const()[name = tensor("key_71_pad_type_0"), val = tensor("valid")]; + tensor key_71_strides_0 = const()[name = tensor("key_71_strides_0"), val = tensor([1, 1])]; + tensor key_71_pad_0 = const()[name = tensor("key_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_71_dilations_0 = const()[name = tensor("key_71_dilations_0"), val = tensor([1, 1])]; + tensor key_71_groups_0 = const()[name = tensor("key_71_groups_0"), val = tensor(1)]; + tensor layers_17_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1042390144)))]; + tensor key_71_cast_fp16 = conv(dilations = key_71_dilations_0, groups = key_71_groups_0, pad = key_71_pad_0, pad_type = key_71_pad_type_0, strides = key_71_strides_0, weight = layers_17_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_71_cast_fp16")]; + tensor value_71_pad_type_0 = const()[name = tensor("value_71_pad_type_0"), val = tensor("valid")]; + tensor value_71_strides_0 = const()[name = tensor("value_71_strides_0"), val = tensor([1, 1])]; + tensor value_71_pad_0 = const()[name = tensor("value_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_71_dilations_0 = const()[name = tensor("value_71_dilations_0"), val = tensor([1, 1])]; + tensor value_71_groups_0 = const()[name = tensor("value_71_groups_0"), val = tensor(1)]; + tensor layers_17_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1045667008)))]; + tensor layers_17_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1048943872)))]; + tensor value_71_cast_fp16 = conv(bias = layers_17_encoder_attn_v_proj_bias_to_fp16, dilations = value_71_dilations_0, groups = value_71_groups_0, pad = value_71_pad_0, pad_type = value_71_pad_type_0, strides = value_71_strides_0, weight = layers_17_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_71_cast_fp16")]; + tensor var_4070 = const()[name = tensor("op_4070"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_71_cast_fp16 = reshape(shape = var_4070, x = query_71_cast_fp16)[name = tensor("mh_q_71_cast_fp16")]; + tensor var_4072_to_fp16 = const()[name = tensor("op_4072_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4073_cast_fp16 = mul(x = mh_q_71_cast_fp16, y = var_4072_to_fp16)[name = tensor("op_4073_cast_fp16")]; + tensor var_4076 = const()[name = tensor("op_4076"), val = tensor([1, 20, 64, 1500])]; + tensor var_4077_cast_fp16 = reshape(shape = var_4076, x = key_71_cast_fp16)[name = tensor("op_4077_cast_fp16")]; + tensor mh_w_107_transpose_x_0 = const()[name = tensor("mh_w_107_transpose_x_0"), val = tensor(true)]; + tensor mh_w_107_transpose_y_0 = const()[name = tensor("mh_w_107_transpose_y_0"), val = tensor(false)]; + tensor mh_w_107_cast_fp16 = matmul(transpose_x = mh_w_107_transpose_x_0, transpose_y = mh_w_107_transpose_y_0, x = var_4073_cast_fp16, y = var_4077_cast_fp16)[name = tensor("mh_w_107_cast_fp16")]; + tensor obj_251_cast_fp16 = softmax(axis = var_3919, x = mh_w_107_cast_fp16)[name = tensor("obj_251_cast_fp16")]; + tensor var_4081 = const()[name = tensor("op_4081"), val = tensor([1, 20, 64, 1500])]; + tensor var_4082_cast_fp16 = reshape(shape = var_4081, x = value_71_cast_fp16)[name = tensor("op_4082_cast_fp16")]; + tensor attn_71_transpose_x_0 = const()[name = tensor("attn_71_transpose_x_0"), val = tensor(false)]; + tensor attn_71_transpose_y_0 = const()[name = tensor("attn_71_transpose_y_0"), val = tensor(true)]; + tensor attn_71_cast_fp16 = matmul(transpose_x = attn_71_transpose_x_0, transpose_y = attn_71_transpose_y_0, x = var_4082_cast_fp16, y = obj_251_cast_fp16)[name = tensor("attn_71_cast_fp16")]; + tensor var_4085 = const()[name = tensor("op_4085"), val = tensor([1, 1280, 1, 1])]; + tensor input_173_cast_fp16 = reshape(shape = var_4085, x = attn_71_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor obj_249_pad_type_0 = const()[name = tensor("obj_249_pad_type_0"), val = tensor("valid")]; + tensor obj_249_strides_0 = const()[name = tensor("obj_249_strides_0"), val = tensor([1, 1])]; + tensor obj_249_pad_0 = const()[name = tensor("obj_249_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_249_dilations_0 = const()[name = tensor("obj_249_dilations_0"), val = tensor([1, 1])]; + tensor obj_249_groups_0 = const()[name = tensor("obj_249_groups_0"), val = tensor(1)]; + tensor layers_17_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1048946496)))]; + tensor layers_17_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1052223360)))]; + tensor obj_249_cast_fp16 = conv(bias = layers_17_encoder_attn_o_proj_bias_to_fp16, dilations = obj_249_dilations_0, groups = obj_249_groups_0, pad = obj_249_pad_0, pad_type = obj_249_pad_type_0, strides = obj_249_strides_0, weight = layers_17_encoder_attn_o_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("obj_249_cast_fp16")]; + tensor inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = obj_249_cast_fp16)[name = tensor("inputs_107_cast_fp16")]; + tensor out_107_axes_0 = const()[name = tensor("out_107_axes_0"), val = tensor([1])]; + tensor var_4106_to_fp16 = const()[name = tensor("op_4106_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_107_cast_fp16 = layer_norm(axes = out_107_axes_0, epsilon = var_4106_to_fp16, x = inputs_107_cast_fp16)[name = tensor("out_107_cast_fp16")]; + tensor input_175_gamma_0_to_fp16 = const()[name = tensor("input_175_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1052225984)))]; + tensor input_175_beta_0_to_fp16 = const()[name = tensor("input_175_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1052228608)))]; + tensor input_175_epsilon_0_to_fp16 = const()[name = tensor("input_175_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_175_cast_fp16 = batch_norm(beta = input_175_beta_0_to_fp16, epsilon = input_175_epsilon_0_to_fp16, gamma = input_175_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_107_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor input_177_pad_type_0 = const()[name = tensor("input_177_pad_type_0"), val = tensor("valid")]; + tensor input_177_strides_0 = const()[name = tensor("input_177_strides_0"), val = tensor([1, 1])]; + tensor input_177_pad_0 = const()[name = tensor("input_177_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_177_dilations_0 = const()[name = tensor("input_177_dilations_0"), val = tensor([1, 1])]; + tensor input_177_groups_0 = const()[name = tensor("input_177_groups_0"), val = tensor(1)]; + tensor layers_17_fc1_weight_to_fp16 = const()[name = tensor("layers_17_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1052231232)))]; + tensor layers_17_fc1_bias_to_fp16 = const()[name = tensor("layers_17_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1065338496)))]; + tensor input_177_cast_fp16 = conv(bias = layers_17_fc1_bias_to_fp16, dilations = input_177_dilations_0, groups = input_177_groups_0, pad = input_177_pad_0, pad_type = input_177_pad_type_0, strides = input_177_strides_0, weight = layers_17_fc1_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_mode_0 = const()[name = tensor("input_179_mode_0"), val = tensor("EXACT")]; + tensor input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor hidden_states_37_pad_type_0 = const()[name = tensor("hidden_states_37_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_37_strides_0 = const()[name = tensor("hidden_states_37_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_37_pad_0 = const()[name = tensor("hidden_states_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_37_dilations_0 = const()[name = tensor("hidden_states_37_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_37_groups_0 = const()[name = tensor("hidden_states_37_groups_0"), val = tensor(1)]; + tensor layers_17_fc2_weight_to_fp16 = const()[name = tensor("layers_17_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1065348800)))]; + tensor layers_17_fc2_bias_to_fp16 = const()[name = tensor("layers_17_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1078456064)))]; + tensor hidden_states_37_cast_fp16 = conv(bias = layers_17_fc2_bias_to_fp16, dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_17_fc2_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("hidden_states_37_cast_fp16")]; + tensor inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = hidden_states_37_cast_fp16)[name = tensor("inputs_109_cast_fp16")]; + tensor var_4142 = const()[name = tensor("op_4142"), val = tensor(3)]; + tensor out_109_axes_0 = const()[name = tensor("out_109_axes_0"), val = tensor([1])]; + tensor var_4167_to_fp16 = const()[name = tensor("op_4167_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_109_cast_fp16 = layer_norm(axes = out_109_axes_0, epsilon = var_4167_to_fp16, x = inputs_109_cast_fp16)[name = tensor("out_109_cast_fp16")]; + tensor obj_253_gamma_0_to_fp16 = const()[name = tensor("obj_253_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1078458688)))]; + tensor obj_253_beta_0_to_fp16 = const()[name = tensor("obj_253_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1078461312)))]; + tensor obj_253_epsilon_0_to_fp16 = const()[name = tensor("obj_253_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_253_cast_fp16 = batch_norm(beta = obj_253_beta_0_to_fp16, epsilon = obj_253_epsilon_0_to_fp16, gamma = obj_253_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_109_cast_fp16)[name = tensor("obj_253_cast_fp16")]; + tensor query_73_pad_type_0 = const()[name = tensor("query_73_pad_type_0"), val = tensor("valid")]; + tensor query_73_strides_0 = const()[name = tensor("query_73_strides_0"), val = tensor([1, 1])]; + tensor query_73_pad_0 = const()[name = tensor("query_73_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_73_dilations_0 = const()[name = tensor("query_73_dilations_0"), val = tensor([1, 1])]; + tensor query_73_groups_0 = const()[name = tensor("query_73_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1078463936)))]; + tensor layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1081740800)))]; + tensor query_73_cast_fp16 = conv(bias = layers_18_self_attn_q_proj_bias_to_fp16, dilations = query_73_dilations_0, groups = query_73_groups_0, pad = query_73_pad_0, pad_type = query_73_pad_type_0, strides = query_73_strides_0, weight = layers_18_self_attn_q_proj_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("query_73_cast_fp16")]; + tensor current_key_37_pad_type_0 = const()[name = tensor("current_key_37_pad_type_0"), val = tensor("valid")]; + tensor current_key_37_strides_0 = const()[name = tensor("current_key_37_strides_0"), val = tensor([1, 1])]; + tensor current_key_37_pad_0 = const()[name = tensor("current_key_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_37_dilations_0 = const()[name = tensor("current_key_37_dilations_0"), val = tensor([1, 1])]; + tensor current_key_37_groups_0 = const()[name = tensor("current_key_37_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1081743424)))]; + tensor current_key_37_cast_fp16 = conv(dilations = current_key_37_dilations_0, groups = current_key_37_groups_0, pad = current_key_37_pad_0, pad_type = current_key_37_pad_type_0, strides = current_key_37_strides_0, weight = layers_18_self_attn_k_proj_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("current_key_37_cast_fp16")]; + tensor current_value_37_pad_type_0 = const()[name = tensor("current_value_37_pad_type_0"), val = tensor("valid")]; + tensor current_value_37_strides_0 = const()[name = tensor("current_value_37_strides_0"), val = tensor([1, 1])]; + tensor current_value_37_pad_0 = const()[name = tensor("current_value_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_37_dilations_0 = const()[name = tensor("current_value_37_dilations_0"), val = tensor([1, 1])]; + tensor current_value_37_groups_0 = const()[name = tensor("current_value_37_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1085020288)))]; + tensor layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1088297152)))]; + tensor current_value_37_cast_fp16 = conv(bias = layers_18_self_attn_v_proj_bias_to_fp16, dilations = current_value_37_dilations_0, groups = current_value_37_groups_0, pad = current_value_37_pad_0, pad_type = current_value_37_pad_type_0, strides = current_value_37_strides_0, weight = layers_18_self_attn_v_proj_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("current_value_37_cast_fp16")]; + tensor var_4206_cast_fp16 = mul(x = var_103_cast_fp16_18, y = var_239_cast_fp16)[name = tensor("op_4206_cast_fp16")]; + tensor var_4207_cast_fp16 = mul(x = current_key_37_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_4207_cast_fp16")]; + tensor key_73_cast_fp16 = add(x = var_4206_cast_fp16, y = var_4207_cast_fp16)[name = tensor("key_73_cast_fp16")]; + tensor var_4210_cast_fp16 = mul(x = var_138_cast_fp16_18, y = var_239_cast_fp16)[name = tensor("op_4210_cast_fp16")]; + tensor var_4211_cast_fp16 = mul(x = current_value_37_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_4211_cast_fp16")]; + tensor value_73_cast_fp16 = add(x = var_4210_cast_fp16, y = var_4211_cast_fp16)[name = tensor("value_73_cast_fp16")]; + tensor var_4215 = const()[name = tensor("op_4215"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_73_cast_fp16 = reshape(shape = var_4215, x = query_73_cast_fp16)[name = tensor("mh_q_73_cast_fp16")]; + tensor var_4217_to_fp16 = const()[name = tensor("op_4217_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4218_cast_fp16 = mul(x = mh_q_73_cast_fp16, y = var_4217_to_fp16)[name = tensor("op_4218_cast_fp16")]; + tensor var_4221 = const()[name = tensor("op_4221"), val = tensor([1, 20, 64, 448])]; + tensor var_4222_cast_fp16 = reshape(shape = var_4221, x = key_73_cast_fp16)[name = tensor("op_4222_cast_fp16")]; + tensor mh_w_109_transpose_x_0 = const()[name = tensor("mh_w_109_transpose_x_0"), val = tensor(true)]; + tensor mh_w_109_transpose_y_0 = const()[name = tensor("mh_w_109_transpose_y_0"), val = tensor(false)]; + tensor mh_w_109_cast_fp16 = matmul(transpose_x = mh_w_109_transpose_x_0, transpose_y = mh_w_109_transpose_y_0, x = var_4218_cast_fp16, y = var_4222_cast_fp16)[name = tensor("mh_w_109_cast_fp16")]; + tensor mh_w_111_cast_fp16 = add(x = mh_w_109_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_111_cast_fp16")]; + tensor var_4230_cast_fp16 = softmax(axis = var_4142, x = mh_w_111_cast_fp16)[name = tensor("op_4230_cast_fp16")]; + tensor var_4231 = const()[name = tensor("op_4231"), val = tensor([1, 20, 64, 448])]; + tensor var_4232_cast_fp16 = reshape(shape = var_4231, x = value_73_cast_fp16)[name = tensor("op_4232_cast_fp16")]; + tensor attn_73_transpose_x_0 = const()[name = tensor("attn_73_transpose_x_0"), val = tensor(false)]; + tensor attn_73_transpose_y_0 = const()[name = tensor("attn_73_transpose_y_0"), val = tensor(true)]; + tensor attn_73_cast_fp16 = matmul(transpose_x = attn_73_transpose_x_0, transpose_y = attn_73_transpose_y_0, x = var_4232_cast_fp16, y = var_4230_cast_fp16)[name = tensor("attn_73_cast_fp16")]; + tensor var_4235 = const()[name = tensor("op_4235"), val = tensor([1, 1280, 1, 1])]; + tensor input_181_cast_fp16 = reshape(shape = var_4235, x = attn_73_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor obj_259_pad_type_0 = const()[name = tensor("obj_259_pad_type_0"), val = tensor("valid")]; + tensor obj_259_strides_0 = const()[name = tensor("obj_259_strides_0"), val = tensor([1, 1])]; + tensor obj_259_pad_0 = const()[name = tensor("obj_259_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_259_dilations_0 = const()[name = tensor("obj_259_dilations_0"), val = tensor([1, 1])]; + tensor obj_259_groups_0 = const()[name = tensor("obj_259_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1088299776)))]; + tensor layers_18_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1091576640)))]; + tensor obj_259_cast_fp16 = conv(bias = layers_18_self_attn_o_proj_bias_to_fp16, dilations = obj_259_dilations_0, groups = obj_259_groups_0, pad = obj_259_pad_0, pad_type = obj_259_pad_type_0, strides = obj_259_strides_0, weight = layers_18_self_attn_o_proj_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("obj_259_cast_fp16")]; + tensor inputs_111_cast_fp16 = add(x = inputs_109_cast_fp16, y = obj_259_cast_fp16)[name = tensor("inputs_111_cast_fp16")]; + tensor out_111_axes_0 = const()[name = tensor("out_111_axes_0"), val = tensor([1])]; + tensor var_4257_to_fp16 = const()[name = tensor("op_4257_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_111_cast_fp16 = layer_norm(axes = out_111_axes_0, epsilon = var_4257_to_fp16, x = inputs_111_cast_fp16)[name = tensor("out_111_cast_fp16")]; + tensor obj_261_gamma_0_to_fp16 = const()[name = tensor("obj_261_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1091579264)))]; + tensor obj_261_beta_0_to_fp16 = const()[name = tensor("obj_261_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1091581888)))]; + tensor obj_261_epsilon_0_to_fp16 = const()[name = tensor("obj_261_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_261_cast_fp16 = batch_norm(beta = obj_261_beta_0_to_fp16, epsilon = obj_261_epsilon_0_to_fp16, gamma = obj_261_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_111_cast_fp16)[name = tensor("obj_261_cast_fp16")]; + tensor query_75_pad_type_0 = const()[name = tensor("query_75_pad_type_0"), val = tensor("valid")]; + tensor query_75_strides_0 = const()[name = tensor("query_75_strides_0"), val = tensor([1, 1])]; + tensor query_75_pad_0 = const()[name = tensor("query_75_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_75_dilations_0 = const()[name = tensor("query_75_dilations_0"), val = tensor([1, 1])]; + tensor query_75_groups_0 = const()[name = tensor("query_75_groups_0"), val = tensor(1)]; + tensor layers_18_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1091584512)))]; + tensor layers_18_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1094861376)))]; + tensor query_75_cast_fp16 = conv(bias = layers_18_encoder_attn_q_proj_bias_to_fp16, dilations = query_75_dilations_0, groups = query_75_groups_0, pad = query_75_pad_0, pad_type = query_75_pad_type_0, strides = query_75_strides_0, weight = layers_18_encoder_attn_q_proj_weight_to_fp16, x = obj_261_cast_fp16)[name = tensor("query_75_cast_fp16")]; + tensor key_75_pad_type_0 = const()[name = tensor("key_75_pad_type_0"), val = tensor("valid")]; + tensor key_75_strides_0 = const()[name = tensor("key_75_strides_0"), val = tensor([1, 1])]; + tensor key_75_pad_0 = const()[name = tensor("key_75_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_75_dilations_0 = const()[name = tensor("key_75_dilations_0"), val = tensor([1, 1])]; + tensor key_75_groups_0 = const()[name = tensor("key_75_groups_0"), val = tensor(1)]; + tensor layers_18_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1094864000)))]; + tensor key_75_cast_fp16 = conv(dilations = key_75_dilations_0, groups = key_75_groups_0, pad = key_75_pad_0, pad_type = key_75_pad_type_0, strides = key_75_strides_0, weight = layers_18_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_75_cast_fp16")]; + tensor value_75_pad_type_0 = const()[name = tensor("value_75_pad_type_0"), val = tensor("valid")]; + tensor value_75_strides_0 = const()[name = tensor("value_75_strides_0"), val = tensor([1, 1])]; + tensor value_75_pad_0 = const()[name = tensor("value_75_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_75_dilations_0 = const()[name = tensor("value_75_dilations_0"), val = tensor([1, 1])]; + tensor value_75_groups_0 = const()[name = tensor("value_75_groups_0"), val = tensor(1)]; + tensor layers_18_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1098140864)))]; + tensor layers_18_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1101417728)))]; + tensor value_75_cast_fp16 = conv(bias = layers_18_encoder_attn_v_proj_bias_to_fp16, dilations = value_75_dilations_0, groups = value_75_groups_0, pad = value_75_pad_0, pad_type = value_75_pad_type_0, strides = value_75_strides_0, weight = layers_18_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_75_cast_fp16")]; + tensor var_4293 = const()[name = tensor("op_4293"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_75_cast_fp16 = reshape(shape = var_4293, x = query_75_cast_fp16)[name = tensor("mh_q_75_cast_fp16")]; + tensor var_4295_to_fp16 = const()[name = tensor("op_4295_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4296_cast_fp16 = mul(x = mh_q_75_cast_fp16, y = var_4295_to_fp16)[name = tensor("op_4296_cast_fp16")]; + tensor var_4299 = const()[name = tensor("op_4299"), val = tensor([1, 20, 64, 1500])]; + tensor var_4300_cast_fp16 = reshape(shape = var_4299, x = key_75_cast_fp16)[name = tensor("op_4300_cast_fp16")]; + tensor mh_w_113_transpose_x_0 = const()[name = tensor("mh_w_113_transpose_x_0"), val = tensor(true)]; + tensor mh_w_113_transpose_y_0 = const()[name = tensor("mh_w_113_transpose_y_0"), val = tensor(false)]; + tensor mh_w_113_cast_fp16 = matmul(transpose_x = mh_w_113_transpose_x_0, transpose_y = mh_w_113_transpose_y_0, x = var_4296_cast_fp16, y = var_4300_cast_fp16)[name = tensor("mh_w_113_cast_fp16")]; + tensor obj_265_cast_fp16 = softmax(axis = var_4142, x = mh_w_113_cast_fp16)[name = tensor("obj_265_cast_fp16")]; + tensor var_4304 = const()[name = tensor("op_4304"), val = tensor([1, 20, 64, 1500])]; + tensor var_4305_cast_fp16 = reshape(shape = var_4304, x = value_75_cast_fp16)[name = tensor("op_4305_cast_fp16")]; + tensor attn_75_transpose_x_0 = const()[name = tensor("attn_75_transpose_x_0"), val = tensor(false)]; + tensor attn_75_transpose_y_0 = const()[name = tensor("attn_75_transpose_y_0"), val = tensor(true)]; + tensor attn_75_cast_fp16 = matmul(transpose_x = attn_75_transpose_x_0, transpose_y = attn_75_transpose_y_0, x = var_4305_cast_fp16, y = obj_265_cast_fp16)[name = tensor("attn_75_cast_fp16")]; + tensor var_4308 = const()[name = tensor("op_4308"), val = tensor([1, 1280, 1, 1])]; + tensor input_183_cast_fp16 = reshape(shape = var_4308, x = attn_75_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor obj_263_pad_type_0 = const()[name = tensor("obj_263_pad_type_0"), val = tensor("valid")]; + tensor obj_263_strides_0 = const()[name = tensor("obj_263_strides_0"), val = tensor([1, 1])]; + tensor obj_263_pad_0 = const()[name = tensor("obj_263_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_263_dilations_0 = const()[name = tensor("obj_263_dilations_0"), val = tensor([1, 1])]; + tensor obj_263_groups_0 = const()[name = tensor("obj_263_groups_0"), val = tensor(1)]; + tensor layers_18_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1101420352)))]; + tensor layers_18_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1104697216)))]; + tensor obj_263_cast_fp16 = conv(bias = layers_18_encoder_attn_o_proj_bias_to_fp16, dilations = obj_263_dilations_0, groups = obj_263_groups_0, pad = obj_263_pad_0, pad_type = obj_263_pad_type_0, strides = obj_263_strides_0, weight = layers_18_encoder_attn_o_proj_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("obj_263_cast_fp16")]; + tensor inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = obj_263_cast_fp16)[name = tensor("inputs_113_cast_fp16")]; + tensor out_113_axes_0 = const()[name = tensor("out_113_axes_0"), val = tensor([1])]; + tensor var_4326_to_fp16 = const()[name = tensor("op_4326_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_113_cast_fp16 = layer_norm(axes = out_113_axes_0, epsilon = var_4326_to_fp16, x = inputs_113_cast_fp16)[name = tensor("out_113_cast_fp16")]; + tensor input_185_gamma_0_to_fp16 = const()[name = tensor("input_185_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1104699840)))]; + tensor input_185_beta_0_to_fp16 = const()[name = tensor("input_185_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1104702464)))]; + tensor input_185_epsilon_0_to_fp16 = const()[name = tensor("input_185_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_185_cast_fp16 = batch_norm(beta = input_185_beta_0_to_fp16, epsilon = input_185_epsilon_0_to_fp16, gamma = input_185_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_113_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor input_187_pad_type_0 = const()[name = tensor("input_187_pad_type_0"), val = tensor("valid")]; + tensor input_187_strides_0 = const()[name = tensor("input_187_strides_0"), val = tensor([1, 1])]; + tensor input_187_pad_0 = const()[name = tensor("input_187_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_187_dilations_0 = const()[name = tensor("input_187_dilations_0"), val = tensor([1, 1])]; + tensor input_187_groups_0 = const()[name = tensor("input_187_groups_0"), val = tensor(1)]; + tensor layers_18_fc1_weight_to_fp16 = const()[name = tensor("layers_18_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1104705088)))]; + tensor layers_18_fc1_bias_to_fp16 = const()[name = tensor("layers_18_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1117812352)))]; + tensor input_187_cast_fp16 = conv(bias = layers_18_fc1_bias_to_fp16, dilations = input_187_dilations_0, groups = input_187_groups_0, pad = input_187_pad_0, pad_type = input_187_pad_type_0, strides = input_187_strides_0, weight = layers_18_fc1_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_mode_0 = const()[name = tensor("input_189_mode_0"), val = tensor("EXACT")]; + tensor input_189_cast_fp16 = gelu(mode = input_189_mode_0, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor hidden_states_39_pad_type_0 = const()[name = tensor("hidden_states_39_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_39_strides_0 = const()[name = tensor("hidden_states_39_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_39_pad_0 = const()[name = tensor("hidden_states_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_39_dilations_0 = const()[name = tensor("hidden_states_39_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_39_groups_0 = const()[name = tensor("hidden_states_39_groups_0"), val = tensor(1)]; + tensor layers_18_fc2_weight_to_fp16 = const()[name = tensor("layers_18_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1117822656)))]; + tensor layers_18_fc2_bias_to_fp16 = const()[name = tensor("layers_18_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1130929920)))]; + tensor hidden_states_39_cast_fp16 = conv(bias = layers_18_fc2_bias_to_fp16, dilations = hidden_states_39_dilations_0, groups = hidden_states_39_groups_0, pad = hidden_states_39_pad_0, pad_type = hidden_states_39_pad_type_0, strides = hidden_states_39_strides_0, weight = layers_18_fc2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("hidden_states_39_cast_fp16")]; + tensor inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = hidden_states_39_cast_fp16)[name = tensor("inputs_115_cast_fp16")]; + tensor var_4361 = const()[name = tensor("op_4361"), val = tensor(3)]; + tensor out_115_axes_0 = const()[name = tensor("out_115_axes_0"), val = tensor([1])]; + tensor var_4386_to_fp16 = const()[name = tensor("op_4386_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_115_cast_fp16 = layer_norm(axes = out_115_axes_0, epsilon = var_4386_to_fp16, x = inputs_115_cast_fp16)[name = tensor("out_115_cast_fp16")]; + tensor obj_267_gamma_0_to_fp16 = const()[name = tensor("obj_267_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1130932544)))]; + tensor obj_267_beta_0_to_fp16 = const()[name = tensor("obj_267_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1130935168)))]; + tensor obj_267_epsilon_0_to_fp16 = const()[name = tensor("obj_267_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_267_cast_fp16 = batch_norm(beta = obj_267_beta_0_to_fp16, epsilon = obj_267_epsilon_0_to_fp16, gamma = obj_267_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_115_cast_fp16)[name = tensor("obj_267_cast_fp16")]; + tensor query_77_pad_type_0 = const()[name = tensor("query_77_pad_type_0"), val = tensor("valid")]; + tensor query_77_strides_0 = const()[name = tensor("query_77_strides_0"), val = tensor([1, 1])]; + tensor query_77_pad_0 = const()[name = tensor("query_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_77_dilations_0 = const()[name = tensor("query_77_dilations_0"), val = tensor([1, 1])]; + tensor query_77_groups_0 = const()[name = tensor("query_77_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1130937792)))]; + tensor layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1134214656)))]; + tensor query_77_cast_fp16 = conv(bias = layers_19_self_attn_q_proj_bias_to_fp16, dilations = query_77_dilations_0, groups = query_77_groups_0, pad = query_77_pad_0, pad_type = query_77_pad_type_0, strides = query_77_strides_0, weight = layers_19_self_attn_q_proj_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor("query_77_cast_fp16")]; + tensor current_key_39_pad_type_0 = const()[name = tensor("current_key_39_pad_type_0"), val = tensor("valid")]; + tensor current_key_39_strides_0 = const()[name = tensor("current_key_39_strides_0"), val = tensor([1, 1])]; + tensor current_key_39_pad_0 = const()[name = tensor("current_key_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_39_dilations_0 = const()[name = tensor("current_key_39_dilations_0"), val = tensor([1, 1])]; + tensor current_key_39_groups_0 = const()[name = tensor("current_key_39_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1134217280)))]; + tensor current_key_39_cast_fp16 = conv(dilations = current_key_39_dilations_0, groups = current_key_39_groups_0, pad = current_key_39_pad_0, pad_type = current_key_39_pad_type_0, strides = current_key_39_strides_0, weight = layers_19_self_attn_k_proj_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor("current_key_39_cast_fp16")]; + tensor current_value_39_pad_type_0 = const()[name = tensor("current_value_39_pad_type_0"), val = tensor("valid")]; + tensor current_value_39_strides_0 = const()[name = tensor("current_value_39_strides_0"), val = tensor([1, 1])]; + tensor current_value_39_pad_0 = const()[name = tensor("current_value_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_39_dilations_0 = const()[name = tensor("current_value_39_dilations_0"), val = tensor([1, 1])]; + tensor current_value_39_groups_0 = const()[name = tensor("current_value_39_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1137494144)))]; + tensor layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1140771008)))]; + tensor current_value_39_cast_fp16 = conv(bias = layers_19_self_attn_v_proj_bias_to_fp16, dilations = current_value_39_dilations_0, groups = current_value_39_groups_0, pad = current_value_39_pad_0, pad_type = current_value_39_pad_type_0, strides = current_value_39_strides_0, weight = layers_19_self_attn_v_proj_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor("current_value_39_cast_fp16")]; + tensor var_4425_cast_fp16 = mul(x = var_103_cast_fp16_19, y = var_239_cast_fp16)[name = tensor("op_4425_cast_fp16")]; + tensor var_4426_cast_fp16 = mul(x = current_key_39_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_4426_cast_fp16")]; + tensor key_77_cast_fp16 = add(x = var_4425_cast_fp16, y = var_4426_cast_fp16)[name = tensor("key_77_cast_fp16")]; + tensor var_4429_cast_fp16 = mul(x = var_138_cast_fp16_19, y = var_239_cast_fp16)[name = tensor("op_4429_cast_fp16")]; + tensor var_4430_cast_fp16 = mul(x = current_value_39_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_4430_cast_fp16")]; + tensor value_77_cast_fp16 = add(x = var_4429_cast_fp16, y = var_4430_cast_fp16)[name = tensor("value_77_cast_fp16")]; + tensor var_4434 = const()[name = tensor("op_4434"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_77_cast_fp16 = reshape(shape = var_4434, x = query_77_cast_fp16)[name = tensor("mh_q_77_cast_fp16")]; + tensor var_4436_to_fp16 = const()[name = tensor("op_4436_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4437_cast_fp16 = mul(x = mh_q_77_cast_fp16, y = var_4436_to_fp16)[name = tensor("op_4437_cast_fp16")]; + tensor var_4440 = const()[name = tensor("op_4440"), val = tensor([1, 20, 64, 448])]; + tensor var_4441_cast_fp16 = reshape(shape = var_4440, x = key_77_cast_fp16)[name = tensor("op_4441_cast_fp16")]; + tensor mh_w_115_transpose_x_0 = const()[name = tensor("mh_w_115_transpose_x_0"), val = tensor(true)]; + tensor mh_w_115_transpose_y_0 = const()[name = tensor("mh_w_115_transpose_y_0"), val = tensor(false)]; + tensor mh_w_115_cast_fp16 = matmul(transpose_x = mh_w_115_transpose_x_0, transpose_y = mh_w_115_transpose_y_0, x = var_4437_cast_fp16, y = var_4441_cast_fp16)[name = tensor("mh_w_115_cast_fp16")]; + tensor mh_w_117_cast_fp16 = add(x = mh_w_115_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_117_cast_fp16")]; + tensor var_4449_cast_fp16 = softmax(axis = var_4361, x = mh_w_117_cast_fp16)[name = tensor("op_4449_cast_fp16")]; + tensor var_4450 = const()[name = tensor("op_4450"), val = tensor([1, 20, 64, 448])]; + tensor var_4451_cast_fp16 = reshape(shape = var_4450, x = value_77_cast_fp16)[name = tensor("op_4451_cast_fp16")]; + tensor attn_77_transpose_x_0 = const()[name = tensor("attn_77_transpose_x_0"), val = tensor(false)]; + tensor attn_77_transpose_y_0 = const()[name = tensor("attn_77_transpose_y_0"), val = tensor(true)]; + tensor attn_77_cast_fp16 = matmul(transpose_x = attn_77_transpose_x_0, transpose_y = attn_77_transpose_y_0, x = var_4451_cast_fp16, y = var_4449_cast_fp16)[name = tensor("attn_77_cast_fp16")]; + tensor var_4454 = const()[name = tensor("op_4454"), val = tensor([1, 1280, 1, 1])]; + tensor input_191_cast_fp16 = reshape(shape = var_4454, x = attn_77_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor obj_273_pad_type_0 = const()[name = tensor("obj_273_pad_type_0"), val = tensor("valid")]; + tensor obj_273_strides_0 = const()[name = tensor("obj_273_strides_0"), val = tensor([1, 1])]; + tensor obj_273_pad_0 = const()[name = tensor("obj_273_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_273_dilations_0 = const()[name = tensor("obj_273_dilations_0"), val = tensor([1, 1])]; + tensor obj_273_groups_0 = const()[name = tensor("obj_273_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1140773632)))]; + tensor layers_19_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1144050496)))]; + tensor obj_273_cast_fp16 = conv(bias = layers_19_self_attn_o_proj_bias_to_fp16, dilations = obj_273_dilations_0, groups = obj_273_groups_0, pad = obj_273_pad_0, pad_type = obj_273_pad_type_0, strides = obj_273_strides_0, weight = layers_19_self_attn_o_proj_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("obj_273_cast_fp16")]; + tensor inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = obj_273_cast_fp16)[name = tensor("inputs_117_cast_fp16")]; + tensor out_117_axes_0 = const()[name = tensor("out_117_axes_0"), val = tensor([1])]; + tensor var_4476_to_fp16 = const()[name = tensor("op_4476_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_117_cast_fp16 = layer_norm(axes = out_117_axes_0, epsilon = var_4476_to_fp16, x = inputs_117_cast_fp16)[name = tensor("out_117_cast_fp16")]; + tensor obj_275_gamma_0_to_fp16 = const()[name = tensor("obj_275_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1144053120)))]; + tensor obj_275_beta_0_to_fp16 = const()[name = tensor("obj_275_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1144055744)))]; + tensor obj_275_epsilon_0_to_fp16 = const()[name = tensor("obj_275_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_275_cast_fp16 = batch_norm(beta = obj_275_beta_0_to_fp16, epsilon = obj_275_epsilon_0_to_fp16, gamma = obj_275_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_117_cast_fp16)[name = tensor("obj_275_cast_fp16")]; + tensor query_79_pad_type_0 = const()[name = tensor("query_79_pad_type_0"), val = tensor("valid")]; + tensor query_79_strides_0 = const()[name = tensor("query_79_strides_0"), val = tensor([1, 1])]; + tensor query_79_pad_0 = const()[name = tensor("query_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_79_dilations_0 = const()[name = tensor("query_79_dilations_0"), val = tensor([1, 1])]; + tensor query_79_groups_0 = const()[name = tensor("query_79_groups_0"), val = tensor(1)]; + tensor layers_19_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1144058368)))]; + tensor layers_19_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1147335232)))]; + tensor query_79_cast_fp16 = conv(bias = layers_19_encoder_attn_q_proj_bias_to_fp16, dilations = query_79_dilations_0, groups = query_79_groups_0, pad = query_79_pad_0, pad_type = query_79_pad_type_0, strides = query_79_strides_0, weight = layers_19_encoder_attn_q_proj_weight_to_fp16, x = obj_275_cast_fp16)[name = tensor("query_79_cast_fp16")]; + tensor key_79_pad_type_0 = const()[name = tensor("key_79_pad_type_0"), val = tensor("valid")]; + tensor key_79_strides_0 = const()[name = tensor("key_79_strides_0"), val = tensor([1, 1])]; + tensor key_79_pad_0 = const()[name = tensor("key_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_79_dilations_0 = const()[name = tensor("key_79_dilations_0"), val = tensor([1, 1])]; + tensor key_79_groups_0 = const()[name = tensor("key_79_groups_0"), val = tensor(1)]; + tensor layers_19_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1147337856)))]; + tensor key_79_cast_fp16 = conv(dilations = key_79_dilations_0, groups = key_79_groups_0, pad = key_79_pad_0, pad_type = key_79_pad_type_0, strides = key_79_strides_0, weight = layers_19_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_79_cast_fp16")]; + tensor value_79_pad_type_0 = const()[name = tensor("value_79_pad_type_0"), val = tensor("valid")]; + tensor value_79_strides_0 = const()[name = tensor("value_79_strides_0"), val = tensor([1, 1])]; + tensor value_79_pad_0 = const()[name = tensor("value_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_79_dilations_0 = const()[name = tensor("value_79_dilations_0"), val = tensor([1, 1])]; + tensor value_79_groups_0 = const()[name = tensor("value_79_groups_0"), val = tensor(1)]; + tensor layers_19_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1150614720)))]; + tensor layers_19_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1153891584)))]; + tensor value_79_cast_fp16 = conv(bias = layers_19_encoder_attn_v_proj_bias_to_fp16, dilations = value_79_dilations_0, groups = value_79_groups_0, pad = value_79_pad_0, pad_type = value_79_pad_type_0, strides = value_79_strides_0, weight = layers_19_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_79_cast_fp16")]; + tensor var_4512 = const()[name = tensor("op_4512"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_79_cast_fp16 = reshape(shape = var_4512, x = query_79_cast_fp16)[name = tensor("mh_q_79_cast_fp16")]; + tensor var_4514_to_fp16 = const()[name = tensor("op_4514_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4515_cast_fp16 = mul(x = mh_q_79_cast_fp16, y = var_4514_to_fp16)[name = tensor("op_4515_cast_fp16")]; + tensor var_4518 = const()[name = tensor("op_4518"), val = tensor([1, 20, 64, 1500])]; + tensor var_4519_cast_fp16 = reshape(shape = var_4518, x = key_79_cast_fp16)[name = tensor("op_4519_cast_fp16")]; + tensor mh_w_119_transpose_x_0 = const()[name = tensor("mh_w_119_transpose_x_0"), val = tensor(true)]; + tensor mh_w_119_transpose_y_0 = const()[name = tensor("mh_w_119_transpose_y_0"), val = tensor(false)]; + tensor mh_w_119_cast_fp16 = matmul(transpose_x = mh_w_119_transpose_x_0, transpose_y = mh_w_119_transpose_y_0, x = var_4515_cast_fp16, y = var_4519_cast_fp16)[name = tensor("mh_w_119_cast_fp16")]; + tensor obj_279_cast_fp16 = softmax(axis = var_4361, x = mh_w_119_cast_fp16)[name = tensor("obj_279_cast_fp16")]; + tensor var_4523 = const()[name = tensor("op_4523"), val = tensor([1, 20, 64, 1500])]; + tensor var_4524_cast_fp16 = reshape(shape = var_4523, x = value_79_cast_fp16)[name = tensor("op_4524_cast_fp16")]; + tensor attn_79_transpose_x_0 = const()[name = tensor("attn_79_transpose_x_0"), val = tensor(false)]; + tensor attn_79_transpose_y_0 = const()[name = tensor("attn_79_transpose_y_0"), val = tensor(true)]; + tensor attn_79_cast_fp16 = matmul(transpose_x = attn_79_transpose_x_0, transpose_y = attn_79_transpose_y_0, x = var_4524_cast_fp16, y = obj_279_cast_fp16)[name = tensor("attn_79_cast_fp16")]; + tensor var_4527 = const()[name = tensor("op_4527"), val = tensor([1, 1280, 1, 1])]; + tensor input_193_cast_fp16 = reshape(shape = var_4527, x = attn_79_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor obj_277_pad_type_0 = const()[name = tensor("obj_277_pad_type_0"), val = tensor("valid")]; + tensor obj_277_strides_0 = const()[name = tensor("obj_277_strides_0"), val = tensor([1, 1])]; + tensor obj_277_pad_0 = const()[name = tensor("obj_277_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_277_dilations_0 = const()[name = tensor("obj_277_dilations_0"), val = tensor([1, 1])]; + tensor obj_277_groups_0 = const()[name = tensor("obj_277_groups_0"), val = tensor(1)]; + tensor layers_19_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1153894208)))]; + tensor layers_19_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1157171072)))]; + tensor obj_277_cast_fp16 = conv(bias = layers_19_encoder_attn_o_proj_bias_to_fp16, dilations = obj_277_dilations_0, groups = obj_277_groups_0, pad = obj_277_pad_0, pad_type = obj_277_pad_type_0, strides = obj_277_strides_0, weight = layers_19_encoder_attn_o_proj_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("obj_277_cast_fp16")]; + tensor inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = obj_277_cast_fp16)[name = tensor("inputs_119_cast_fp16")]; + tensor out_119_axes_0 = const()[name = tensor("out_119_axes_0"), val = tensor([1])]; + tensor var_4548_to_fp16 = const()[name = tensor("op_4548_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_119_cast_fp16 = layer_norm(axes = out_119_axes_0, epsilon = var_4548_to_fp16, x = inputs_119_cast_fp16)[name = tensor("out_119_cast_fp16")]; + tensor input_195_gamma_0_to_fp16 = const()[name = tensor("input_195_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1157173696)))]; + tensor input_195_beta_0_to_fp16 = const()[name = tensor("input_195_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1157176320)))]; + tensor input_195_epsilon_0_to_fp16 = const()[name = tensor("input_195_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_195_cast_fp16 = batch_norm(beta = input_195_beta_0_to_fp16, epsilon = input_195_epsilon_0_to_fp16, gamma = input_195_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_119_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor input_197_pad_type_0 = const()[name = tensor("input_197_pad_type_0"), val = tensor("valid")]; + tensor input_197_strides_0 = const()[name = tensor("input_197_strides_0"), val = tensor([1, 1])]; + tensor input_197_pad_0 = const()[name = tensor("input_197_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_197_dilations_0 = const()[name = tensor("input_197_dilations_0"), val = tensor([1, 1])]; + tensor input_197_groups_0 = const()[name = tensor("input_197_groups_0"), val = tensor(1)]; + tensor layers_19_fc1_weight_to_fp16 = const()[name = tensor("layers_19_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1157178944)))]; + tensor layers_19_fc1_bias_to_fp16 = const()[name = tensor("layers_19_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1170286208)))]; + tensor input_197_cast_fp16 = conv(bias = layers_19_fc1_bias_to_fp16, dilations = input_197_dilations_0, groups = input_197_groups_0, pad = input_197_pad_0, pad_type = input_197_pad_type_0, strides = input_197_strides_0, weight = layers_19_fc1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor input_199_mode_0 = const()[name = tensor("input_199_mode_0"), val = tensor("EXACT")]; + tensor input_199_cast_fp16 = gelu(mode = input_199_mode_0, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor hidden_states_41_pad_type_0 = const()[name = tensor("hidden_states_41_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_41_strides_0 = const()[name = tensor("hidden_states_41_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_41_pad_0 = const()[name = tensor("hidden_states_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_41_dilations_0 = const()[name = tensor("hidden_states_41_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_41_groups_0 = const()[name = tensor("hidden_states_41_groups_0"), val = tensor(1)]; + tensor layers_19_fc2_weight_to_fp16 = const()[name = tensor("layers_19_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1170296512)))]; + tensor layers_19_fc2_bias_to_fp16 = const()[name = tensor("layers_19_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1183403776)))]; + tensor hidden_states_41_cast_fp16 = conv(bias = layers_19_fc2_bias_to_fp16, dilations = hidden_states_41_dilations_0, groups = hidden_states_41_groups_0, pad = hidden_states_41_pad_0, pad_type = hidden_states_41_pad_type_0, strides = hidden_states_41_strides_0, weight = layers_19_fc2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("hidden_states_41_cast_fp16")]; + tensor inputs_121_cast_fp16 = add(x = inputs_119_cast_fp16, y = hidden_states_41_cast_fp16)[name = tensor("inputs_121_cast_fp16")]; + tensor var_4584 = const()[name = tensor("op_4584"), val = tensor(3)]; + tensor out_121_axes_0 = const()[name = tensor("out_121_axes_0"), val = tensor([1])]; + tensor var_4609_to_fp16 = const()[name = tensor("op_4609_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_121_cast_fp16 = layer_norm(axes = out_121_axes_0, epsilon = var_4609_to_fp16, x = inputs_121_cast_fp16)[name = tensor("out_121_cast_fp16")]; + tensor obj_281_gamma_0_to_fp16 = const()[name = tensor("obj_281_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1183406400)))]; + tensor obj_281_beta_0_to_fp16 = const()[name = tensor("obj_281_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1183409024)))]; + tensor obj_281_epsilon_0_to_fp16 = const()[name = tensor("obj_281_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_281_cast_fp16 = batch_norm(beta = obj_281_beta_0_to_fp16, epsilon = obj_281_epsilon_0_to_fp16, gamma = obj_281_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_121_cast_fp16)[name = tensor("obj_281_cast_fp16")]; + tensor query_81_pad_type_0 = const()[name = tensor("query_81_pad_type_0"), val = tensor("valid")]; + tensor query_81_strides_0 = const()[name = tensor("query_81_strides_0"), val = tensor([1, 1])]; + tensor query_81_pad_0 = const()[name = tensor("query_81_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_81_dilations_0 = const()[name = tensor("query_81_dilations_0"), val = tensor([1, 1])]; + tensor query_81_groups_0 = const()[name = tensor("query_81_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1183411648)))]; + tensor layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1186688512)))]; + tensor query_81_cast_fp16 = conv(bias = layers_20_self_attn_q_proj_bias_to_fp16, dilations = query_81_dilations_0, groups = query_81_groups_0, pad = query_81_pad_0, pad_type = query_81_pad_type_0, strides = query_81_strides_0, weight = layers_20_self_attn_q_proj_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor("query_81_cast_fp16")]; + tensor current_key_41_pad_type_0 = const()[name = tensor("current_key_41_pad_type_0"), val = tensor("valid")]; + tensor current_key_41_strides_0 = const()[name = tensor("current_key_41_strides_0"), val = tensor([1, 1])]; + tensor current_key_41_pad_0 = const()[name = tensor("current_key_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_41_dilations_0 = const()[name = tensor("current_key_41_dilations_0"), val = tensor([1, 1])]; + tensor current_key_41_groups_0 = const()[name = tensor("current_key_41_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1186691136)))]; + tensor current_key_41_cast_fp16 = conv(dilations = current_key_41_dilations_0, groups = current_key_41_groups_0, pad = current_key_41_pad_0, pad_type = current_key_41_pad_type_0, strides = current_key_41_strides_0, weight = layers_20_self_attn_k_proj_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor("current_key_41_cast_fp16")]; + tensor current_value_41_pad_type_0 = const()[name = tensor("current_value_41_pad_type_0"), val = tensor("valid")]; + tensor current_value_41_strides_0 = const()[name = tensor("current_value_41_strides_0"), val = tensor([1, 1])]; + tensor current_value_41_pad_0 = const()[name = tensor("current_value_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_41_dilations_0 = const()[name = tensor("current_value_41_dilations_0"), val = tensor([1, 1])]; + tensor current_value_41_groups_0 = const()[name = tensor("current_value_41_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1189968000)))]; + tensor layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1193244864)))]; + tensor current_value_41_cast_fp16 = conv(bias = layers_20_self_attn_v_proj_bias_to_fp16, dilations = current_value_41_dilations_0, groups = current_value_41_groups_0, pad = current_value_41_pad_0, pad_type = current_value_41_pad_type_0, strides = current_value_41_strides_0, weight = layers_20_self_attn_v_proj_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor("current_value_41_cast_fp16")]; + tensor var_4648_cast_fp16 = mul(x = var_103_cast_fp16_20, y = var_239_cast_fp16)[name = tensor("op_4648_cast_fp16")]; + tensor var_4649_cast_fp16 = mul(x = current_key_41_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_4649_cast_fp16")]; + tensor key_81_cast_fp16 = add(x = var_4648_cast_fp16, y = var_4649_cast_fp16)[name = tensor("key_81_cast_fp16")]; + tensor var_4652_cast_fp16 = mul(x = var_138_cast_fp16_20, y = var_239_cast_fp16)[name = tensor("op_4652_cast_fp16")]; + tensor var_4653_cast_fp16 = mul(x = current_value_41_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_4653_cast_fp16")]; + tensor value_81_cast_fp16 = add(x = var_4652_cast_fp16, y = var_4653_cast_fp16)[name = tensor("value_81_cast_fp16")]; + tensor var_4657 = const()[name = tensor("op_4657"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_81_cast_fp16 = reshape(shape = var_4657, x = query_81_cast_fp16)[name = tensor("mh_q_81_cast_fp16")]; + tensor var_4659_to_fp16 = const()[name = tensor("op_4659_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4660_cast_fp16 = mul(x = mh_q_81_cast_fp16, y = var_4659_to_fp16)[name = tensor("op_4660_cast_fp16")]; + tensor var_4663 = const()[name = tensor("op_4663"), val = tensor([1, 20, 64, 448])]; + tensor var_4664_cast_fp16 = reshape(shape = var_4663, x = key_81_cast_fp16)[name = tensor("op_4664_cast_fp16")]; + tensor mh_w_121_transpose_x_0 = const()[name = tensor("mh_w_121_transpose_x_0"), val = tensor(true)]; + tensor mh_w_121_transpose_y_0 = const()[name = tensor("mh_w_121_transpose_y_0"), val = tensor(false)]; + tensor mh_w_121_cast_fp16 = matmul(transpose_x = mh_w_121_transpose_x_0, transpose_y = mh_w_121_transpose_y_0, x = var_4660_cast_fp16, y = var_4664_cast_fp16)[name = tensor("mh_w_121_cast_fp16")]; + tensor mh_w_123_cast_fp16 = add(x = mh_w_121_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_123_cast_fp16")]; + tensor var_4672_cast_fp16 = softmax(axis = var_4584, x = mh_w_123_cast_fp16)[name = tensor("op_4672_cast_fp16")]; + tensor var_4673 = const()[name = tensor("op_4673"), val = tensor([1, 20, 64, 448])]; + tensor var_4674_cast_fp16 = reshape(shape = var_4673, x = value_81_cast_fp16)[name = tensor("op_4674_cast_fp16")]; + tensor attn_81_transpose_x_0 = const()[name = tensor("attn_81_transpose_x_0"), val = tensor(false)]; + tensor attn_81_transpose_y_0 = const()[name = tensor("attn_81_transpose_y_0"), val = tensor(true)]; + tensor attn_81_cast_fp16 = matmul(transpose_x = attn_81_transpose_x_0, transpose_y = attn_81_transpose_y_0, x = var_4674_cast_fp16, y = var_4672_cast_fp16)[name = tensor("attn_81_cast_fp16")]; + tensor var_4677 = const()[name = tensor("op_4677"), val = tensor([1, 1280, 1, 1])]; + tensor input_201_cast_fp16 = reshape(shape = var_4677, x = attn_81_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor obj_287_pad_type_0 = const()[name = tensor("obj_287_pad_type_0"), val = tensor("valid")]; + tensor obj_287_strides_0 = const()[name = tensor("obj_287_strides_0"), val = tensor([1, 1])]; + tensor obj_287_pad_0 = const()[name = tensor("obj_287_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_287_dilations_0 = const()[name = tensor("obj_287_dilations_0"), val = tensor([1, 1])]; + tensor obj_287_groups_0 = const()[name = tensor("obj_287_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1193247488)))]; + tensor layers_20_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1196524352)))]; + tensor obj_287_cast_fp16 = conv(bias = layers_20_self_attn_o_proj_bias_to_fp16, dilations = obj_287_dilations_0, groups = obj_287_groups_0, pad = obj_287_pad_0, pad_type = obj_287_pad_type_0, strides = obj_287_strides_0, weight = layers_20_self_attn_o_proj_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("obj_287_cast_fp16")]; + tensor inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = obj_287_cast_fp16)[name = tensor("inputs_123_cast_fp16")]; + tensor out_123_axes_0 = const()[name = tensor("out_123_axes_0"), val = tensor([1])]; + tensor var_4699_to_fp16 = const()[name = tensor("op_4699_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_123_cast_fp16 = layer_norm(axes = out_123_axes_0, epsilon = var_4699_to_fp16, x = inputs_123_cast_fp16)[name = tensor("out_123_cast_fp16")]; + tensor obj_289_gamma_0_to_fp16 = const()[name = tensor("obj_289_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1196526976)))]; + tensor obj_289_beta_0_to_fp16 = const()[name = tensor("obj_289_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1196529600)))]; + tensor obj_289_epsilon_0_to_fp16 = const()[name = tensor("obj_289_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_289_cast_fp16 = batch_norm(beta = obj_289_beta_0_to_fp16, epsilon = obj_289_epsilon_0_to_fp16, gamma = obj_289_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_123_cast_fp16)[name = tensor("obj_289_cast_fp16")]; + tensor query_83_pad_type_0 = const()[name = tensor("query_83_pad_type_0"), val = tensor("valid")]; + tensor query_83_strides_0 = const()[name = tensor("query_83_strides_0"), val = tensor([1, 1])]; + tensor query_83_pad_0 = const()[name = tensor("query_83_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_83_dilations_0 = const()[name = tensor("query_83_dilations_0"), val = tensor([1, 1])]; + tensor query_83_groups_0 = const()[name = tensor("query_83_groups_0"), val = tensor(1)]; + tensor layers_20_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1196532224)))]; + tensor layers_20_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1199809088)))]; + tensor query_83_cast_fp16 = conv(bias = layers_20_encoder_attn_q_proj_bias_to_fp16, dilations = query_83_dilations_0, groups = query_83_groups_0, pad = query_83_pad_0, pad_type = query_83_pad_type_0, strides = query_83_strides_0, weight = layers_20_encoder_attn_q_proj_weight_to_fp16, x = obj_289_cast_fp16)[name = tensor("query_83_cast_fp16")]; + tensor key_83_pad_type_0 = const()[name = tensor("key_83_pad_type_0"), val = tensor("valid")]; + tensor key_83_strides_0 = const()[name = tensor("key_83_strides_0"), val = tensor([1, 1])]; + tensor key_83_pad_0 = const()[name = tensor("key_83_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_83_dilations_0 = const()[name = tensor("key_83_dilations_0"), val = tensor([1, 1])]; + tensor key_83_groups_0 = const()[name = tensor("key_83_groups_0"), val = tensor(1)]; + tensor layers_20_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1199811712)))]; + tensor key_83_cast_fp16 = conv(dilations = key_83_dilations_0, groups = key_83_groups_0, pad = key_83_pad_0, pad_type = key_83_pad_type_0, strides = key_83_strides_0, weight = layers_20_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_83_cast_fp16")]; + tensor value_83_pad_type_0 = const()[name = tensor("value_83_pad_type_0"), val = tensor("valid")]; + tensor value_83_strides_0 = const()[name = tensor("value_83_strides_0"), val = tensor([1, 1])]; + tensor value_83_pad_0 = const()[name = tensor("value_83_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_83_dilations_0 = const()[name = tensor("value_83_dilations_0"), val = tensor([1, 1])]; + tensor value_83_groups_0 = const()[name = tensor("value_83_groups_0"), val = tensor(1)]; + tensor layers_20_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1203088576)))]; + tensor layers_20_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1206365440)))]; + tensor value_83_cast_fp16 = conv(bias = layers_20_encoder_attn_v_proj_bias_to_fp16, dilations = value_83_dilations_0, groups = value_83_groups_0, pad = value_83_pad_0, pad_type = value_83_pad_type_0, strides = value_83_strides_0, weight = layers_20_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_83_cast_fp16")]; + tensor var_4735 = const()[name = tensor("op_4735"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_83_cast_fp16 = reshape(shape = var_4735, x = query_83_cast_fp16)[name = tensor("mh_q_83_cast_fp16")]; + tensor var_4737_to_fp16 = const()[name = tensor("op_4737_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4738_cast_fp16 = mul(x = mh_q_83_cast_fp16, y = var_4737_to_fp16)[name = tensor("op_4738_cast_fp16")]; + tensor var_4741 = const()[name = tensor("op_4741"), val = tensor([1, 20, 64, 1500])]; + tensor var_4742_cast_fp16 = reshape(shape = var_4741, x = key_83_cast_fp16)[name = tensor("op_4742_cast_fp16")]; + tensor mh_w_125_transpose_x_0 = const()[name = tensor("mh_w_125_transpose_x_0"), val = tensor(true)]; + tensor mh_w_125_transpose_y_0 = const()[name = tensor("mh_w_125_transpose_y_0"), val = tensor(false)]; + tensor mh_w_125_cast_fp16 = matmul(transpose_x = mh_w_125_transpose_x_0, transpose_y = mh_w_125_transpose_y_0, x = var_4738_cast_fp16, y = var_4742_cast_fp16)[name = tensor("mh_w_125_cast_fp16")]; + tensor obj_293_cast_fp16 = softmax(axis = var_4584, x = mh_w_125_cast_fp16)[name = tensor("obj_293_cast_fp16")]; + tensor var_4746 = const()[name = tensor("op_4746"), val = tensor([1, 20, 64, 1500])]; + tensor var_4747_cast_fp16 = reshape(shape = var_4746, x = value_83_cast_fp16)[name = tensor("op_4747_cast_fp16")]; + tensor attn_83_transpose_x_0 = const()[name = tensor("attn_83_transpose_x_0"), val = tensor(false)]; + tensor attn_83_transpose_y_0 = const()[name = tensor("attn_83_transpose_y_0"), val = tensor(true)]; + tensor attn_83_cast_fp16 = matmul(transpose_x = attn_83_transpose_x_0, transpose_y = attn_83_transpose_y_0, x = var_4747_cast_fp16, y = obj_293_cast_fp16)[name = tensor("attn_83_cast_fp16")]; + tensor var_4750 = const()[name = tensor("op_4750"), val = tensor([1, 1280, 1, 1])]; + tensor input_203_cast_fp16 = reshape(shape = var_4750, x = attn_83_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor obj_291_pad_type_0 = const()[name = tensor("obj_291_pad_type_0"), val = tensor("valid")]; + tensor obj_291_strides_0 = const()[name = tensor("obj_291_strides_0"), val = tensor([1, 1])]; + tensor obj_291_pad_0 = const()[name = tensor("obj_291_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_291_dilations_0 = const()[name = tensor("obj_291_dilations_0"), val = tensor([1, 1])]; + tensor obj_291_groups_0 = const()[name = tensor("obj_291_groups_0"), val = tensor(1)]; + tensor layers_20_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1206368064)))]; + tensor layers_20_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1209644928)))]; + tensor obj_291_cast_fp16 = conv(bias = layers_20_encoder_attn_o_proj_bias_to_fp16, dilations = obj_291_dilations_0, groups = obj_291_groups_0, pad = obj_291_pad_0, pad_type = obj_291_pad_type_0, strides = obj_291_strides_0, weight = layers_20_encoder_attn_o_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("obj_291_cast_fp16")]; + tensor inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = obj_291_cast_fp16)[name = tensor("inputs_125_cast_fp16")]; + tensor out_125_axes_0 = const()[name = tensor("out_125_axes_0"), val = tensor([1])]; + tensor var_4768_to_fp16 = const()[name = tensor("op_4768_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_125_cast_fp16 = layer_norm(axes = out_125_axes_0, epsilon = var_4768_to_fp16, x = inputs_125_cast_fp16)[name = tensor("out_125_cast_fp16")]; + tensor input_205_gamma_0_to_fp16 = const()[name = tensor("input_205_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1209647552)))]; + tensor input_205_beta_0_to_fp16 = const()[name = tensor("input_205_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1209650176)))]; + tensor input_205_epsilon_0_to_fp16 = const()[name = tensor("input_205_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_205_cast_fp16 = batch_norm(beta = input_205_beta_0_to_fp16, epsilon = input_205_epsilon_0_to_fp16, gamma = input_205_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_125_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor input_207_pad_type_0 = const()[name = tensor("input_207_pad_type_0"), val = tensor("valid")]; + tensor input_207_strides_0 = const()[name = tensor("input_207_strides_0"), val = tensor([1, 1])]; + tensor input_207_pad_0 = const()[name = tensor("input_207_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_207_dilations_0 = const()[name = tensor("input_207_dilations_0"), val = tensor([1, 1])]; + tensor input_207_groups_0 = const()[name = tensor("input_207_groups_0"), val = tensor(1)]; + tensor layers_20_fc1_weight_to_fp16 = const()[name = tensor("layers_20_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1209652800)))]; + tensor layers_20_fc1_bias_to_fp16 = const()[name = tensor("layers_20_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1222760064)))]; + tensor input_207_cast_fp16 = conv(bias = layers_20_fc1_bias_to_fp16, dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = layers_20_fc1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_mode_0 = const()[name = tensor("input_209_mode_0"), val = tensor("EXACT")]; + tensor input_209_cast_fp16 = gelu(mode = input_209_mode_0, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor hidden_states_43_pad_type_0 = const()[name = tensor("hidden_states_43_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_43_strides_0 = const()[name = tensor("hidden_states_43_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_43_pad_0 = const()[name = tensor("hidden_states_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_43_dilations_0 = const()[name = tensor("hidden_states_43_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_43_groups_0 = const()[name = tensor("hidden_states_43_groups_0"), val = tensor(1)]; + tensor layers_20_fc2_weight_to_fp16 = const()[name = tensor("layers_20_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1222770368)))]; + tensor layers_20_fc2_bias_to_fp16 = const()[name = tensor("layers_20_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1235877632)))]; + tensor hidden_states_43_cast_fp16 = conv(bias = layers_20_fc2_bias_to_fp16, dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_20_fc2_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("hidden_states_43_cast_fp16")]; + tensor inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = hidden_states_43_cast_fp16)[name = tensor("inputs_127_cast_fp16")]; + tensor var_4803 = const()[name = tensor("op_4803"), val = tensor(3)]; + tensor out_127_axes_0 = const()[name = tensor("out_127_axes_0"), val = tensor([1])]; + tensor var_4828_to_fp16 = const()[name = tensor("op_4828_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_127_cast_fp16 = layer_norm(axes = out_127_axes_0, epsilon = var_4828_to_fp16, x = inputs_127_cast_fp16)[name = tensor("out_127_cast_fp16")]; + tensor obj_295_gamma_0_to_fp16 = const()[name = tensor("obj_295_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1235880256)))]; + tensor obj_295_beta_0_to_fp16 = const()[name = tensor("obj_295_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1235882880)))]; + tensor obj_295_epsilon_0_to_fp16 = const()[name = tensor("obj_295_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_295_cast_fp16 = batch_norm(beta = obj_295_beta_0_to_fp16, epsilon = obj_295_epsilon_0_to_fp16, gamma = obj_295_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_127_cast_fp16)[name = tensor("obj_295_cast_fp16")]; + tensor query_85_pad_type_0 = const()[name = tensor("query_85_pad_type_0"), val = tensor("valid")]; + tensor query_85_strides_0 = const()[name = tensor("query_85_strides_0"), val = tensor([1, 1])]; + tensor query_85_pad_0 = const()[name = tensor("query_85_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_85_dilations_0 = const()[name = tensor("query_85_dilations_0"), val = tensor([1, 1])]; + tensor query_85_groups_0 = const()[name = tensor("query_85_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1235885504)))]; + tensor layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1239162368)))]; + tensor query_85_cast_fp16 = conv(bias = layers_21_self_attn_q_proj_bias_to_fp16, dilations = query_85_dilations_0, groups = query_85_groups_0, pad = query_85_pad_0, pad_type = query_85_pad_type_0, strides = query_85_strides_0, weight = layers_21_self_attn_q_proj_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor("query_85_cast_fp16")]; + tensor current_key_43_pad_type_0 = const()[name = tensor("current_key_43_pad_type_0"), val = tensor("valid")]; + tensor current_key_43_strides_0 = const()[name = tensor("current_key_43_strides_0"), val = tensor([1, 1])]; + tensor current_key_43_pad_0 = const()[name = tensor("current_key_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_43_dilations_0 = const()[name = tensor("current_key_43_dilations_0"), val = tensor([1, 1])]; + tensor current_key_43_groups_0 = const()[name = tensor("current_key_43_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1239164992)))]; + tensor current_key_43_cast_fp16 = conv(dilations = current_key_43_dilations_0, groups = current_key_43_groups_0, pad = current_key_43_pad_0, pad_type = current_key_43_pad_type_0, strides = current_key_43_strides_0, weight = layers_21_self_attn_k_proj_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor("current_key_43_cast_fp16")]; + tensor current_value_43_pad_type_0 = const()[name = tensor("current_value_43_pad_type_0"), val = tensor("valid")]; + tensor current_value_43_strides_0 = const()[name = tensor("current_value_43_strides_0"), val = tensor([1, 1])]; + tensor current_value_43_pad_0 = const()[name = tensor("current_value_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_43_dilations_0 = const()[name = tensor("current_value_43_dilations_0"), val = tensor([1, 1])]; + tensor current_value_43_groups_0 = const()[name = tensor("current_value_43_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1242441856)))]; + tensor layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1245718720)))]; + tensor current_value_43_cast_fp16 = conv(bias = layers_21_self_attn_v_proj_bias_to_fp16, dilations = current_value_43_dilations_0, groups = current_value_43_groups_0, pad = current_value_43_pad_0, pad_type = current_value_43_pad_type_0, strides = current_value_43_strides_0, weight = layers_21_self_attn_v_proj_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor("current_value_43_cast_fp16")]; + tensor var_4867_cast_fp16 = mul(x = var_103_cast_fp16_21, y = var_239_cast_fp16)[name = tensor("op_4867_cast_fp16")]; + tensor var_4868_cast_fp16 = mul(x = current_key_43_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_4868_cast_fp16")]; + tensor key_85_cast_fp16 = add(x = var_4867_cast_fp16, y = var_4868_cast_fp16)[name = tensor("key_85_cast_fp16")]; + tensor var_4871_cast_fp16 = mul(x = var_138_cast_fp16_21, y = var_239_cast_fp16)[name = tensor("op_4871_cast_fp16")]; + tensor var_4872_cast_fp16 = mul(x = current_value_43_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_4872_cast_fp16")]; + tensor value_85_cast_fp16 = add(x = var_4871_cast_fp16, y = var_4872_cast_fp16)[name = tensor("value_85_cast_fp16")]; + tensor var_4876 = const()[name = tensor("op_4876"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_85_cast_fp16 = reshape(shape = var_4876, x = query_85_cast_fp16)[name = tensor("mh_q_85_cast_fp16")]; + tensor var_4878_to_fp16 = const()[name = tensor("op_4878_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4879_cast_fp16 = mul(x = mh_q_85_cast_fp16, y = var_4878_to_fp16)[name = tensor("op_4879_cast_fp16")]; + tensor var_4882 = const()[name = tensor("op_4882"), val = tensor([1, 20, 64, 448])]; + tensor var_4883_cast_fp16 = reshape(shape = var_4882, x = key_85_cast_fp16)[name = tensor("op_4883_cast_fp16")]; + tensor mh_w_127_transpose_x_0 = const()[name = tensor("mh_w_127_transpose_x_0"), val = tensor(true)]; + tensor mh_w_127_transpose_y_0 = const()[name = tensor("mh_w_127_transpose_y_0"), val = tensor(false)]; + tensor mh_w_127_cast_fp16 = matmul(transpose_x = mh_w_127_transpose_x_0, transpose_y = mh_w_127_transpose_y_0, x = var_4879_cast_fp16, y = var_4883_cast_fp16)[name = tensor("mh_w_127_cast_fp16")]; + tensor mh_w_129_cast_fp16 = add(x = mh_w_127_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_129_cast_fp16")]; + tensor var_4891_cast_fp16 = softmax(axis = var_4803, x = mh_w_129_cast_fp16)[name = tensor("op_4891_cast_fp16")]; + tensor var_4892 = const()[name = tensor("op_4892"), val = tensor([1, 20, 64, 448])]; + tensor var_4893_cast_fp16 = reshape(shape = var_4892, x = value_85_cast_fp16)[name = tensor("op_4893_cast_fp16")]; + tensor attn_85_transpose_x_0 = const()[name = tensor("attn_85_transpose_x_0"), val = tensor(false)]; + tensor attn_85_transpose_y_0 = const()[name = tensor("attn_85_transpose_y_0"), val = tensor(true)]; + tensor attn_85_cast_fp16 = matmul(transpose_x = attn_85_transpose_x_0, transpose_y = attn_85_transpose_y_0, x = var_4893_cast_fp16, y = var_4891_cast_fp16)[name = tensor("attn_85_cast_fp16")]; + tensor var_4896 = const()[name = tensor("op_4896"), val = tensor([1, 1280, 1, 1])]; + tensor input_211_cast_fp16 = reshape(shape = var_4896, x = attn_85_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor obj_301_pad_type_0 = const()[name = tensor("obj_301_pad_type_0"), val = tensor("valid")]; + tensor obj_301_strides_0 = const()[name = tensor("obj_301_strides_0"), val = tensor([1, 1])]; + tensor obj_301_pad_0 = const()[name = tensor("obj_301_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_301_dilations_0 = const()[name = tensor("obj_301_dilations_0"), val = tensor([1, 1])]; + tensor obj_301_groups_0 = const()[name = tensor("obj_301_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1245721344)))]; + tensor layers_21_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1248998208)))]; + tensor obj_301_cast_fp16 = conv(bias = layers_21_self_attn_o_proj_bias_to_fp16, dilations = obj_301_dilations_0, groups = obj_301_groups_0, pad = obj_301_pad_0, pad_type = obj_301_pad_type_0, strides = obj_301_strides_0, weight = layers_21_self_attn_o_proj_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("obj_301_cast_fp16")]; + tensor inputs_129_cast_fp16 = add(x = inputs_127_cast_fp16, y = obj_301_cast_fp16)[name = tensor("inputs_129_cast_fp16")]; + tensor out_129_axes_0 = const()[name = tensor("out_129_axes_0"), val = tensor([1])]; + tensor var_4918_to_fp16 = const()[name = tensor("op_4918_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_129_cast_fp16 = layer_norm(axes = out_129_axes_0, epsilon = var_4918_to_fp16, x = inputs_129_cast_fp16)[name = tensor("out_129_cast_fp16")]; + tensor obj_303_gamma_0_to_fp16 = const()[name = tensor("obj_303_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1249000832)))]; + tensor obj_303_beta_0_to_fp16 = const()[name = tensor("obj_303_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1249003456)))]; + tensor obj_303_epsilon_0_to_fp16 = const()[name = tensor("obj_303_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_303_cast_fp16 = batch_norm(beta = obj_303_beta_0_to_fp16, epsilon = obj_303_epsilon_0_to_fp16, gamma = obj_303_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_129_cast_fp16)[name = tensor("obj_303_cast_fp16")]; + tensor query_87_pad_type_0 = const()[name = tensor("query_87_pad_type_0"), val = tensor("valid")]; + tensor query_87_strides_0 = const()[name = tensor("query_87_strides_0"), val = tensor([1, 1])]; + tensor query_87_pad_0 = const()[name = tensor("query_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_87_dilations_0 = const()[name = tensor("query_87_dilations_0"), val = tensor([1, 1])]; + tensor query_87_groups_0 = const()[name = tensor("query_87_groups_0"), val = tensor(1)]; + tensor layers_21_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1249006080)))]; + tensor layers_21_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1252282944)))]; + tensor query_87_cast_fp16 = conv(bias = layers_21_encoder_attn_q_proj_bias_to_fp16, dilations = query_87_dilations_0, groups = query_87_groups_0, pad = query_87_pad_0, pad_type = query_87_pad_type_0, strides = query_87_strides_0, weight = layers_21_encoder_attn_q_proj_weight_to_fp16, x = obj_303_cast_fp16)[name = tensor("query_87_cast_fp16")]; + tensor key_87_pad_type_0 = const()[name = tensor("key_87_pad_type_0"), val = tensor("valid")]; + tensor key_87_strides_0 = const()[name = tensor("key_87_strides_0"), val = tensor([1, 1])]; + tensor key_87_pad_0 = const()[name = tensor("key_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_87_dilations_0 = const()[name = tensor("key_87_dilations_0"), val = tensor([1, 1])]; + tensor key_87_groups_0 = const()[name = tensor("key_87_groups_0"), val = tensor(1)]; + tensor layers_21_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1252285568)))]; + tensor key_87_cast_fp16 = conv(dilations = key_87_dilations_0, groups = key_87_groups_0, pad = key_87_pad_0, pad_type = key_87_pad_type_0, strides = key_87_strides_0, weight = layers_21_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_87_cast_fp16")]; + tensor value_87_pad_type_0 = const()[name = tensor("value_87_pad_type_0"), val = tensor("valid")]; + tensor value_87_strides_0 = const()[name = tensor("value_87_strides_0"), val = tensor([1, 1])]; + tensor value_87_pad_0 = const()[name = tensor("value_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_87_dilations_0 = const()[name = tensor("value_87_dilations_0"), val = tensor([1, 1])]; + tensor value_87_groups_0 = const()[name = tensor("value_87_groups_0"), val = tensor(1)]; + tensor layers_21_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1255562432)))]; + tensor layers_21_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1258839296)))]; + tensor value_87_cast_fp16 = conv(bias = layers_21_encoder_attn_v_proj_bias_to_fp16, dilations = value_87_dilations_0, groups = value_87_groups_0, pad = value_87_pad_0, pad_type = value_87_pad_type_0, strides = value_87_strides_0, weight = layers_21_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_87_cast_fp16")]; + tensor var_4954 = const()[name = tensor("op_4954"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_87_cast_fp16 = reshape(shape = var_4954, x = query_87_cast_fp16)[name = tensor("mh_q_87_cast_fp16")]; + tensor var_4956_to_fp16 = const()[name = tensor("op_4956_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4957_cast_fp16 = mul(x = mh_q_87_cast_fp16, y = var_4956_to_fp16)[name = tensor("op_4957_cast_fp16")]; + tensor var_4960 = const()[name = tensor("op_4960"), val = tensor([1, 20, 64, 1500])]; + tensor var_4961_cast_fp16 = reshape(shape = var_4960, x = key_87_cast_fp16)[name = tensor("op_4961_cast_fp16")]; + tensor mh_w_131_transpose_x_0 = const()[name = tensor("mh_w_131_transpose_x_0"), val = tensor(true)]; + tensor mh_w_131_transpose_y_0 = const()[name = tensor("mh_w_131_transpose_y_0"), val = tensor(false)]; + tensor mh_w_131_cast_fp16 = matmul(transpose_x = mh_w_131_transpose_x_0, transpose_y = mh_w_131_transpose_y_0, x = var_4957_cast_fp16, y = var_4961_cast_fp16)[name = tensor("mh_w_131_cast_fp16")]; + tensor obj_307_cast_fp16 = softmax(axis = var_4803, x = mh_w_131_cast_fp16)[name = tensor("obj_307_cast_fp16")]; + tensor var_4965 = const()[name = tensor("op_4965"), val = tensor([1, 20, 64, 1500])]; + tensor var_4966_cast_fp16 = reshape(shape = var_4965, x = value_87_cast_fp16)[name = tensor("op_4966_cast_fp16")]; + tensor attn_87_transpose_x_0 = const()[name = tensor("attn_87_transpose_x_0"), val = tensor(false)]; + tensor attn_87_transpose_y_0 = const()[name = tensor("attn_87_transpose_y_0"), val = tensor(true)]; + tensor attn_87_cast_fp16 = matmul(transpose_x = attn_87_transpose_x_0, transpose_y = attn_87_transpose_y_0, x = var_4966_cast_fp16, y = obj_307_cast_fp16)[name = tensor("attn_87_cast_fp16")]; + tensor var_4969 = const()[name = tensor("op_4969"), val = tensor([1, 1280, 1, 1])]; + tensor input_213_cast_fp16 = reshape(shape = var_4969, x = attn_87_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor obj_305_pad_type_0 = const()[name = tensor("obj_305_pad_type_0"), val = tensor("valid")]; + tensor obj_305_strides_0 = const()[name = tensor("obj_305_strides_0"), val = tensor([1, 1])]; + tensor obj_305_pad_0 = const()[name = tensor("obj_305_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_305_dilations_0 = const()[name = tensor("obj_305_dilations_0"), val = tensor([1, 1])]; + tensor obj_305_groups_0 = const()[name = tensor("obj_305_groups_0"), val = tensor(1)]; + tensor layers_21_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1258841920)))]; + tensor layers_21_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1262118784)))]; + tensor obj_305_cast_fp16 = conv(bias = layers_21_encoder_attn_o_proj_bias_to_fp16, dilations = obj_305_dilations_0, groups = obj_305_groups_0, pad = obj_305_pad_0, pad_type = obj_305_pad_type_0, strides = obj_305_strides_0, weight = layers_21_encoder_attn_o_proj_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("obj_305_cast_fp16")]; + tensor inputs_131_cast_fp16 = add(x = inputs_129_cast_fp16, y = obj_305_cast_fp16)[name = tensor("inputs_131_cast_fp16")]; + tensor out_131_axes_0 = const()[name = tensor("out_131_axes_0"), val = tensor([1])]; + tensor var_4990_to_fp16 = const()[name = tensor("op_4990_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_131_cast_fp16 = layer_norm(axes = out_131_axes_0, epsilon = var_4990_to_fp16, x = inputs_131_cast_fp16)[name = tensor("out_131_cast_fp16")]; + tensor input_215_gamma_0_to_fp16 = const()[name = tensor("input_215_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1262121408)))]; + tensor input_215_beta_0_to_fp16 = const()[name = tensor("input_215_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1262124032)))]; + tensor input_215_epsilon_0_to_fp16 = const()[name = tensor("input_215_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_215_cast_fp16 = batch_norm(beta = input_215_beta_0_to_fp16, epsilon = input_215_epsilon_0_to_fp16, gamma = input_215_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_131_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor input_217_pad_type_0 = const()[name = tensor("input_217_pad_type_0"), val = tensor("valid")]; + tensor input_217_strides_0 = const()[name = tensor("input_217_strides_0"), val = tensor([1, 1])]; + tensor input_217_pad_0 = const()[name = tensor("input_217_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_217_dilations_0 = const()[name = tensor("input_217_dilations_0"), val = tensor([1, 1])]; + tensor input_217_groups_0 = const()[name = tensor("input_217_groups_0"), val = tensor(1)]; + tensor layers_21_fc1_weight_to_fp16 = const()[name = tensor("layers_21_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1262126656)))]; + tensor layers_21_fc1_bias_to_fp16 = const()[name = tensor("layers_21_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1275233920)))]; + tensor input_217_cast_fp16 = conv(bias = layers_21_fc1_bias_to_fp16, dilations = input_217_dilations_0, groups = input_217_groups_0, pad = input_217_pad_0, pad_type = input_217_pad_type_0, strides = input_217_strides_0, weight = layers_21_fc1_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_mode_0 = const()[name = tensor("input_219_mode_0"), val = tensor("EXACT")]; + tensor input_219_cast_fp16 = gelu(mode = input_219_mode_0, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor hidden_states_45_pad_type_0 = const()[name = tensor("hidden_states_45_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_45_strides_0 = const()[name = tensor("hidden_states_45_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_45_pad_0 = const()[name = tensor("hidden_states_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_45_dilations_0 = const()[name = tensor("hidden_states_45_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_45_groups_0 = const()[name = tensor("hidden_states_45_groups_0"), val = tensor(1)]; + tensor layers_21_fc2_weight_to_fp16 = const()[name = tensor("layers_21_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1275244224)))]; + tensor layers_21_fc2_bias_to_fp16 = const()[name = tensor("layers_21_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1288351488)))]; + tensor hidden_states_45_cast_fp16 = conv(bias = layers_21_fc2_bias_to_fp16, dilations = hidden_states_45_dilations_0, groups = hidden_states_45_groups_0, pad = hidden_states_45_pad_0, pad_type = hidden_states_45_pad_type_0, strides = hidden_states_45_strides_0, weight = layers_21_fc2_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("hidden_states_45_cast_fp16")]; + tensor inputs_133_cast_fp16 = add(x = inputs_131_cast_fp16, y = hidden_states_45_cast_fp16)[name = tensor("inputs_133_cast_fp16")]; + tensor var_5026 = const()[name = tensor("op_5026"), val = tensor(3)]; + tensor out_133_axes_0 = const()[name = tensor("out_133_axes_0"), val = tensor([1])]; + tensor var_5051_to_fp16 = const()[name = tensor("op_5051_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_133_cast_fp16 = layer_norm(axes = out_133_axes_0, epsilon = var_5051_to_fp16, x = inputs_133_cast_fp16)[name = tensor("out_133_cast_fp16")]; + tensor obj_309_gamma_0_to_fp16 = const()[name = tensor("obj_309_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1288354112)))]; + tensor obj_309_beta_0_to_fp16 = const()[name = tensor("obj_309_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1288356736)))]; + tensor obj_309_epsilon_0_to_fp16 = const()[name = tensor("obj_309_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_309_cast_fp16 = batch_norm(beta = obj_309_beta_0_to_fp16, epsilon = obj_309_epsilon_0_to_fp16, gamma = obj_309_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_133_cast_fp16)[name = tensor("obj_309_cast_fp16")]; + tensor query_89_pad_type_0 = const()[name = tensor("query_89_pad_type_0"), val = tensor("valid")]; + tensor query_89_strides_0 = const()[name = tensor("query_89_strides_0"), val = tensor([1, 1])]; + tensor query_89_pad_0 = const()[name = tensor("query_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_89_dilations_0 = const()[name = tensor("query_89_dilations_0"), val = tensor([1, 1])]; + tensor query_89_groups_0 = const()[name = tensor("query_89_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1288359360)))]; + tensor layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1291636224)))]; + tensor query_89_cast_fp16 = conv(bias = layers_22_self_attn_q_proj_bias_to_fp16, dilations = query_89_dilations_0, groups = query_89_groups_0, pad = query_89_pad_0, pad_type = query_89_pad_type_0, strides = query_89_strides_0, weight = layers_22_self_attn_q_proj_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("query_89_cast_fp16")]; + tensor current_key_45_pad_type_0 = const()[name = tensor("current_key_45_pad_type_0"), val = tensor("valid")]; + tensor current_key_45_strides_0 = const()[name = tensor("current_key_45_strides_0"), val = tensor([1, 1])]; + tensor current_key_45_pad_0 = const()[name = tensor("current_key_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_45_dilations_0 = const()[name = tensor("current_key_45_dilations_0"), val = tensor([1, 1])]; + tensor current_key_45_groups_0 = const()[name = tensor("current_key_45_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1291638848)))]; + tensor current_key_45_cast_fp16 = conv(dilations = current_key_45_dilations_0, groups = current_key_45_groups_0, pad = current_key_45_pad_0, pad_type = current_key_45_pad_type_0, strides = current_key_45_strides_0, weight = layers_22_self_attn_k_proj_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("current_key_45_cast_fp16")]; + tensor current_value_45_pad_type_0 = const()[name = tensor("current_value_45_pad_type_0"), val = tensor("valid")]; + tensor current_value_45_strides_0 = const()[name = tensor("current_value_45_strides_0"), val = tensor([1, 1])]; + tensor current_value_45_pad_0 = const()[name = tensor("current_value_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_45_dilations_0 = const()[name = tensor("current_value_45_dilations_0"), val = tensor([1, 1])]; + tensor current_value_45_groups_0 = const()[name = tensor("current_value_45_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1294915712)))]; + tensor layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1298192576)))]; + tensor current_value_45_cast_fp16 = conv(bias = layers_22_self_attn_v_proj_bias_to_fp16, dilations = current_value_45_dilations_0, groups = current_value_45_groups_0, pad = current_value_45_pad_0, pad_type = current_value_45_pad_type_0, strides = current_value_45_strides_0, weight = layers_22_self_attn_v_proj_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("current_value_45_cast_fp16")]; + tensor var_5090_cast_fp16 = mul(x = var_103_cast_fp16_22, y = var_239_cast_fp16)[name = tensor("op_5090_cast_fp16")]; + tensor var_5091_cast_fp16 = mul(x = current_key_45_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_5091_cast_fp16")]; + tensor key_89_cast_fp16 = add(x = var_5090_cast_fp16, y = var_5091_cast_fp16)[name = tensor("key_89_cast_fp16")]; + tensor var_5094_cast_fp16 = mul(x = var_138_cast_fp16_22, y = var_239_cast_fp16)[name = tensor("op_5094_cast_fp16")]; + tensor var_5095_cast_fp16 = mul(x = current_value_45_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_5095_cast_fp16")]; + tensor value_89_cast_fp16 = add(x = var_5094_cast_fp16, y = var_5095_cast_fp16)[name = tensor("value_89_cast_fp16")]; + tensor var_5099 = const()[name = tensor("op_5099"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_89_cast_fp16 = reshape(shape = var_5099, x = query_89_cast_fp16)[name = tensor("mh_q_89_cast_fp16")]; + tensor var_5101_to_fp16 = const()[name = tensor("op_5101_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5102_cast_fp16 = mul(x = mh_q_89_cast_fp16, y = var_5101_to_fp16)[name = tensor("op_5102_cast_fp16")]; + tensor var_5105 = const()[name = tensor("op_5105"), val = tensor([1, 20, 64, 448])]; + tensor var_5106_cast_fp16 = reshape(shape = var_5105, x = key_89_cast_fp16)[name = tensor("op_5106_cast_fp16")]; + tensor mh_w_133_transpose_x_0 = const()[name = tensor("mh_w_133_transpose_x_0"), val = tensor(true)]; + tensor mh_w_133_transpose_y_0 = const()[name = tensor("mh_w_133_transpose_y_0"), val = tensor(false)]; + tensor mh_w_133_cast_fp16 = matmul(transpose_x = mh_w_133_transpose_x_0, transpose_y = mh_w_133_transpose_y_0, x = var_5102_cast_fp16, y = var_5106_cast_fp16)[name = tensor("mh_w_133_cast_fp16")]; + tensor mh_w_135_cast_fp16 = add(x = mh_w_133_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_135_cast_fp16")]; + tensor var_5114_cast_fp16 = softmax(axis = var_5026, x = mh_w_135_cast_fp16)[name = tensor("op_5114_cast_fp16")]; + tensor var_5115 = const()[name = tensor("op_5115"), val = tensor([1, 20, 64, 448])]; + tensor var_5116_cast_fp16 = reshape(shape = var_5115, x = value_89_cast_fp16)[name = tensor("op_5116_cast_fp16")]; + tensor attn_89_transpose_x_0 = const()[name = tensor("attn_89_transpose_x_0"), val = tensor(false)]; + tensor attn_89_transpose_y_0 = const()[name = tensor("attn_89_transpose_y_0"), val = tensor(true)]; + tensor attn_89_cast_fp16 = matmul(transpose_x = attn_89_transpose_x_0, transpose_y = attn_89_transpose_y_0, x = var_5116_cast_fp16, y = var_5114_cast_fp16)[name = tensor("attn_89_cast_fp16")]; + tensor var_5119 = const()[name = tensor("op_5119"), val = tensor([1, 1280, 1, 1])]; + tensor input_221_cast_fp16 = reshape(shape = var_5119, x = attn_89_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor obj_315_pad_type_0 = const()[name = tensor("obj_315_pad_type_0"), val = tensor("valid")]; + tensor obj_315_strides_0 = const()[name = tensor("obj_315_strides_0"), val = tensor([1, 1])]; + tensor obj_315_pad_0 = const()[name = tensor("obj_315_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_315_dilations_0 = const()[name = tensor("obj_315_dilations_0"), val = tensor([1, 1])]; + tensor obj_315_groups_0 = const()[name = tensor("obj_315_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1298195200)))]; + tensor layers_22_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1301472064)))]; + tensor obj_315_cast_fp16 = conv(bias = layers_22_self_attn_o_proj_bias_to_fp16, dilations = obj_315_dilations_0, groups = obj_315_groups_0, pad = obj_315_pad_0, pad_type = obj_315_pad_type_0, strides = obj_315_strides_0, weight = layers_22_self_attn_o_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("obj_315_cast_fp16")]; + tensor inputs_135_cast_fp16 = add(x = inputs_133_cast_fp16, y = obj_315_cast_fp16)[name = tensor("inputs_135_cast_fp16")]; + tensor out_135_axes_0 = const()[name = tensor("out_135_axes_0"), val = tensor([1])]; + tensor var_5141_to_fp16 = const()[name = tensor("op_5141_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_135_cast_fp16 = layer_norm(axes = out_135_axes_0, epsilon = var_5141_to_fp16, x = inputs_135_cast_fp16)[name = tensor("out_135_cast_fp16")]; + tensor obj_317_gamma_0_to_fp16 = const()[name = tensor("obj_317_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1301474688)))]; + tensor obj_317_beta_0_to_fp16 = const()[name = tensor("obj_317_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1301477312)))]; + tensor obj_317_epsilon_0_to_fp16 = const()[name = tensor("obj_317_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_317_cast_fp16 = batch_norm(beta = obj_317_beta_0_to_fp16, epsilon = obj_317_epsilon_0_to_fp16, gamma = obj_317_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_135_cast_fp16)[name = tensor("obj_317_cast_fp16")]; + tensor query_91_pad_type_0 = const()[name = tensor("query_91_pad_type_0"), val = tensor("valid")]; + tensor query_91_strides_0 = const()[name = tensor("query_91_strides_0"), val = tensor([1, 1])]; + tensor query_91_pad_0 = const()[name = tensor("query_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_91_dilations_0 = const()[name = tensor("query_91_dilations_0"), val = tensor([1, 1])]; + tensor query_91_groups_0 = const()[name = tensor("query_91_groups_0"), val = tensor(1)]; + tensor layers_22_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1301479936)))]; + tensor layers_22_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1304756800)))]; + tensor query_91_cast_fp16 = conv(bias = layers_22_encoder_attn_q_proj_bias_to_fp16, dilations = query_91_dilations_0, groups = query_91_groups_0, pad = query_91_pad_0, pad_type = query_91_pad_type_0, strides = query_91_strides_0, weight = layers_22_encoder_attn_q_proj_weight_to_fp16, x = obj_317_cast_fp16)[name = tensor("query_91_cast_fp16")]; + tensor key_91_pad_type_0 = const()[name = tensor("key_91_pad_type_0"), val = tensor("valid")]; + tensor key_91_strides_0 = const()[name = tensor("key_91_strides_0"), val = tensor([1, 1])]; + tensor key_91_pad_0 = const()[name = tensor("key_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_91_dilations_0 = const()[name = tensor("key_91_dilations_0"), val = tensor([1, 1])]; + tensor key_91_groups_0 = const()[name = tensor("key_91_groups_0"), val = tensor(1)]; + tensor layers_22_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1304759424)))]; + tensor key_91_cast_fp16 = conv(dilations = key_91_dilations_0, groups = key_91_groups_0, pad = key_91_pad_0, pad_type = key_91_pad_type_0, strides = key_91_strides_0, weight = layers_22_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_91_cast_fp16")]; + tensor value_91_pad_type_0 = const()[name = tensor("value_91_pad_type_0"), val = tensor("valid")]; + tensor value_91_strides_0 = const()[name = tensor("value_91_strides_0"), val = tensor([1, 1])]; + tensor value_91_pad_0 = const()[name = tensor("value_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_91_dilations_0 = const()[name = tensor("value_91_dilations_0"), val = tensor([1, 1])]; + tensor value_91_groups_0 = const()[name = tensor("value_91_groups_0"), val = tensor(1)]; + tensor layers_22_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1308036288)))]; + tensor layers_22_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1311313152)))]; + tensor value_91_cast_fp16 = conv(bias = layers_22_encoder_attn_v_proj_bias_to_fp16, dilations = value_91_dilations_0, groups = value_91_groups_0, pad = value_91_pad_0, pad_type = value_91_pad_type_0, strides = value_91_strides_0, weight = layers_22_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_91_cast_fp16")]; + tensor var_5177 = const()[name = tensor("op_5177"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_91_cast_fp16 = reshape(shape = var_5177, x = query_91_cast_fp16)[name = tensor("mh_q_91_cast_fp16")]; + tensor var_5179_to_fp16 = const()[name = tensor("op_5179_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5180_cast_fp16 = mul(x = mh_q_91_cast_fp16, y = var_5179_to_fp16)[name = tensor("op_5180_cast_fp16")]; + tensor var_5183 = const()[name = tensor("op_5183"), val = tensor([1, 20, 64, 1500])]; + tensor var_5184_cast_fp16 = reshape(shape = var_5183, x = key_91_cast_fp16)[name = tensor("op_5184_cast_fp16")]; + tensor mh_w_137_transpose_x_0 = const()[name = tensor("mh_w_137_transpose_x_0"), val = tensor(true)]; + tensor mh_w_137_transpose_y_0 = const()[name = tensor("mh_w_137_transpose_y_0"), val = tensor(false)]; + tensor mh_w_137_cast_fp16 = matmul(transpose_x = mh_w_137_transpose_x_0, transpose_y = mh_w_137_transpose_y_0, x = var_5180_cast_fp16, y = var_5184_cast_fp16)[name = tensor("mh_w_137_cast_fp16")]; + tensor obj_321_cast_fp16 = softmax(axis = var_5026, x = mh_w_137_cast_fp16)[name = tensor("obj_321_cast_fp16")]; + tensor var_5188 = const()[name = tensor("op_5188"), val = tensor([1, 20, 64, 1500])]; + tensor var_5189_cast_fp16 = reshape(shape = var_5188, x = value_91_cast_fp16)[name = tensor("op_5189_cast_fp16")]; + tensor attn_91_transpose_x_0 = const()[name = tensor("attn_91_transpose_x_0"), val = tensor(false)]; + tensor attn_91_transpose_y_0 = const()[name = tensor("attn_91_transpose_y_0"), val = tensor(true)]; + tensor attn_91_cast_fp16 = matmul(transpose_x = attn_91_transpose_x_0, transpose_y = attn_91_transpose_y_0, x = var_5189_cast_fp16, y = obj_321_cast_fp16)[name = tensor("attn_91_cast_fp16")]; + tensor var_5192 = const()[name = tensor("op_5192"), val = tensor([1, 1280, 1, 1])]; + tensor input_223_cast_fp16 = reshape(shape = var_5192, x = attn_91_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor obj_319_pad_type_0 = const()[name = tensor("obj_319_pad_type_0"), val = tensor("valid")]; + tensor obj_319_strides_0 = const()[name = tensor("obj_319_strides_0"), val = tensor([1, 1])]; + tensor obj_319_pad_0 = const()[name = tensor("obj_319_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_319_dilations_0 = const()[name = tensor("obj_319_dilations_0"), val = tensor([1, 1])]; + tensor obj_319_groups_0 = const()[name = tensor("obj_319_groups_0"), val = tensor(1)]; + tensor layers_22_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1311315776)))]; + tensor layers_22_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1314592640)))]; + tensor obj_319_cast_fp16 = conv(bias = layers_22_encoder_attn_o_proj_bias_to_fp16, dilations = obj_319_dilations_0, groups = obj_319_groups_0, pad = obj_319_pad_0, pad_type = obj_319_pad_type_0, strides = obj_319_strides_0, weight = layers_22_encoder_attn_o_proj_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("obj_319_cast_fp16")]; + tensor inputs_137_cast_fp16 = add(x = inputs_135_cast_fp16, y = obj_319_cast_fp16)[name = tensor("inputs_137_cast_fp16")]; + tensor out_137_axes_0 = const()[name = tensor("out_137_axes_0"), val = tensor([1])]; + tensor var_5210_to_fp16 = const()[name = tensor("op_5210_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_137_cast_fp16 = layer_norm(axes = out_137_axes_0, epsilon = var_5210_to_fp16, x = inputs_137_cast_fp16)[name = tensor("out_137_cast_fp16")]; + tensor input_225_gamma_0_to_fp16 = const()[name = tensor("input_225_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1314595264)))]; + tensor input_225_beta_0_to_fp16 = const()[name = tensor("input_225_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1314597888)))]; + tensor input_225_epsilon_0_to_fp16 = const()[name = tensor("input_225_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_225_cast_fp16 = batch_norm(beta = input_225_beta_0_to_fp16, epsilon = input_225_epsilon_0_to_fp16, gamma = input_225_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_137_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor input_227_pad_type_0 = const()[name = tensor("input_227_pad_type_0"), val = tensor("valid")]; + tensor input_227_strides_0 = const()[name = tensor("input_227_strides_0"), val = tensor([1, 1])]; + tensor input_227_pad_0 = const()[name = tensor("input_227_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_227_dilations_0 = const()[name = tensor("input_227_dilations_0"), val = tensor([1, 1])]; + tensor input_227_groups_0 = const()[name = tensor("input_227_groups_0"), val = tensor(1)]; + tensor layers_22_fc1_weight_to_fp16 = const()[name = tensor("layers_22_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1314600512)))]; + tensor layers_22_fc1_bias_to_fp16 = const()[name = tensor("layers_22_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1327707776)))]; + tensor input_227_cast_fp16 = conv(bias = layers_22_fc1_bias_to_fp16, dilations = input_227_dilations_0, groups = input_227_groups_0, pad = input_227_pad_0, pad_type = input_227_pad_type_0, strides = input_227_strides_0, weight = layers_22_fc1_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor input_229_mode_0 = const()[name = tensor("input_229_mode_0"), val = tensor("EXACT")]; + tensor input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor hidden_states_47_pad_type_0 = const()[name = tensor("hidden_states_47_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_47_strides_0 = const()[name = tensor("hidden_states_47_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_47_pad_0 = const()[name = tensor("hidden_states_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_47_dilations_0 = const()[name = tensor("hidden_states_47_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_47_groups_0 = const()[name = tensor("hidden_states_47_groups_0"), val = tensor(1)]; + tensor layers_22_fc2_weight_to_fp16 = const()[name = tensor("layers_22_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1327718080)))]; + tensor layers_22_fc2_bias_to_fp16 = const()[name = tensor("layers_22_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1340825344)))]; + tensor hidden_states_47_cast_fp16 = conv(bias = layers_22_fc2_bias_to_fp16, dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_22_fc2_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("hidden_states_47_cast_fp16")]; + tensor inputs_139_cast_fp16 = add(x = inputs_137_cast_fp16, y = hidden_states_47_cast_fp16)[name = tensor("inputs_139_cast_fp16")]; + tensor var_5245 = const()[name = tensor("op_5245"), val = tensor(3)]; + tensor out_139_axes_0 = const()[name = tensor("out_139_axes_0"), val = tensor([1])]; + tensor var_5270_to_fp16 = const()[name = tensor("op_5270_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_139_cast_fp16 = layer_norm(axes = out_139_axes_0, epsilon = var_5270_to_fp16, x = inputs_139_cast_fp16)[name = tensor("out_139_cast_fp16")]; + tensor obj_323_gamma_0_to_fp16 = const()[name = tensor("obj_323_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1340827968)))]; + tensor obj_323_beta_0_to_fp16 = const()[name = tensor("obj_323_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1340830592)))]; + tensor obj_323_epsilon_0_to_fp16 = const()[name = tensor("obj_323_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_323_cast_fp16 = batch_norm(beta = obj_323_beta_0_to_fp16, epsilon = obj_323_epsilon_0_to_fp16, gamma = obj_323_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_139_cast_fp16)[name = tensor("obj_323_cast_fp16")]; + tensor query_93_pad_type_0 = const()[name = tensor("query_93_pad_type_0"), val = tensor("valid")]; + tensor query_93_strides_0 = const()[name = tensor("query_93_strides_0"), val = tensor([1, 1])]; + tensor query_93_pad_0 = const()[name = tensor("query_93_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_93_dilations_0 = const()[name = tensor("query_93_dilations_0"), val = tensor([1, 1])]; + tensor query_93_groups_0 = const()[name = tensor("query_93_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1340833216)))]; + tensor layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1344110080)))]; + tensor query_93_cast_fp16 = conv(bias = layers_23_self_attn_q_proj_bias_to_fp16, dilations = query_93_dilations_0, groups = query_93_groups_0, pad = query_93_pad_0, pad_type = query_93_pad_type_0, strides = query_93_strides_0, weight = layers_23_self_attn_q_proj_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor("query_93_cast_fp16")]; + tensor current_key_47_pad_type_0 = const()[name = tensor("current_key_47_pad_type_0"), val = tensor("valid")]; + tensor current_key_47_strides_0 = const()[name = tensor("current_key_47_strides_0"), val = tensor([1, 1])]; + tensor current_key_47_pad_0 = const()[name = tensor("current_key_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_47_dilations_0 = const()[name = tensor("current_key_47_dilations_0"), val = tensor([1, 1])]; + tensor current_key_47_groups_0 = const()[name = tensor("current_key_47_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1344112704)))]; + tensor current_key_47_cast_fp16 = conv(dilations = current_key_47_dilations_0, groups = current_key_47_groups_0, pad = current_key_47_pad_0, pad_type = current_key_47_pad_type_0, strides = current_key_47_strides_0, weight = layers_23_self_attn_k_proj_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor("current_key_47_cast_fp16")]; + tensor current_value_47_pad_type_0 = const()[name = tensor("current_value_47_pad_type_0"), val = tensor("valid")]; + tensor current_value_47_strides_0 = const()[name = tensor("current_value_47_strides_0"), val = tensor([1, 1])]; + tensor current_value_47_pad_0 = const()[name = tensor("current_value_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_47_dilations_0 = const()[name = tensor("current_value_47_dilations_0"), val = tensor([1, 1])]; + tensor current_value_47_groups_0 = const()[name = tensor("current_value_47_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1347389568)))]; + tensor layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1350666432)))]; + tensor current_value_47_cast_fp16 = conv(bias = layers_23_self_attn_v_proj_bias_to_fp16, dilations = current_value_47_dilations_0, groups = current_value_47_groups_0, pad = current_value_47_pad_0, pad_type = current_value_47_pad_type_0, strides = current_value_47_strides_0, weight = layers_23_self_attn_v_proj_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor("current_value_47_cast_fp16")]; + tensor var_5309_cast_fp16 = mul(x = var_103_cast_fp16_23, y = var_239_cast_fp16)[name = tensor("op_5309_cast_fp16")]; + tensor var_5310_cast_fp16 = mul(x = current_key_47_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_5310_cast_fp16")]; + tensor key_93_cast_fp16 = add(x = var_5309_cast_fp16, y = var_5310_cast_fp16)[name = tensor("key_93_cast_fp16")]; + tensor var_5313_cast_fp16 = mul(x = var_138_cast_fp16_23, y = var_239_cast_fp16)[name = tensor("op_5313_cast_fp16")]; + tensor var_5314_cast_fp16 = mul(x = current_value_47_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_5314_cast_fp16")]; + tensor value_93_cast_fp16 = add(x = var_5313_cast_fp16, y = var_5314_cast_fp16)[name = tensor("value_93_cast_fp16")]; + tensor var_5318 = const()[name = tensor("op_5318"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_93_cast_fp16 = reshape(shape = var_5318, x = query_93_cast_fp16)[name = tensor("mh_q_93_cast_fp16")]; + tensor var_5320_to_fp16 = const()[name = tensor("op_5320_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5321_cast_fp16 = mul(x = mh_q_93_cast_fp16, y = var_5320_to_fp16)[name = tensor("op_5321_cast_fp16")]; + tensor var_5324 = const()[name = tensor("op_5324"), val = tensor([1, 20, 64, 448])]; + tensor var_5325_cast_fp16 = reshape(shape = var_5324, x = key_93_cast_fp16)[name = tensor("op_5325_cast_fp16")]; + tensor mh_w_139_transpose_x_0 = const()[name = tensor("mh_w_139_transpose_x_0"), val = tensor(true)]; + tensor mh_w_139_transpose_y_0 = const()[name = tensor("mh_w_139_transpose_y_0"), val = tensor(false)]; + tensor mh_w_139_cast_fp16 = matmul(transpose_x = mh_w_139_transpose_x_0, transpose_y = mh_w_139_transpose_y_0, x = var_5321_cast_fp16, y = var_5325_cast_fp16)[name = tensor("mh_w_139_cast_fp16")]; + tensor mh_w_141_cast_fp16 = add(x = mh_w_139_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_141_cast_fp16")]; + tensor var_5333_cast_fp16 = softmax(axis = var_5245, x = mh_w_141_cast_fp16)[name = tensor("op_5333_cast_fp16")]; + tensor var_5334 = const()[name = tensor("op_5334"), val = tensor([1, 20, 64, 448])]; + tensor var_5335_cast_fp16 = reshape(shape = var_5334, x = value_93_cast_fp16)[name = tensor("op_5335_cast_fp16")]; + tensor attn_93_transpose_x_0 = const()[name = tensor("attn_93_transpose_x_0"), val = tensor(false)]; + tensor attn_93_transpose_y_0 = const()[name = tensor("attn_93_transpose_y_0"), val = tensor(true)]; + tensor attn_93_cast_fp16 = matmul(transpose_x = attn_93_transpose_x_0, transpose_y = attn_93_transpose_y_0, x = var_5335_cast_fp16, y = var_5333_cast_fp16)[name = tensor("attn_93_cast_fp16")]; + tensor var_5338 = const()[name = tensor("op_5338"), val = tensor([1, 1280, 1, 1])]; + tensor input_231_cast_fp16 = reshape(shape = var_5338, x = attn_93_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor obj_329_pad_type_0 = const()[name = tensor("obj_329_pad_type_0"), val = tensor("valid")]; + tensor obj_329_strides_0 = const()[name = tensor("obj_329_strides_0"), val = tensor([1, 1])]; + tensor obj_329_pad_0 = const()[name = tensor("obj_329_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_329_dilations_0 = const()[name = tensor("obj_329_dilations_0"), val = tensor([1, 1])]; + tensor obj_329_groups_0 = const()[name = tensor("obj_329_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1350669056)))]; + tensor layers_23_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1353945920)))]; + tensor obj_329_cast_fp16 = conv(bias = layers_23_self_attn_o_proj_bias_to_fp16, dilations = obj_329_dilations_0, groups = obj_329_groups_0, pad = obj_329_pad_0, pad_type = obj_329_pad_type_0, strides = obj_329_strides_0, weight = layers_23_self_attn_o_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("obj_329_cast_fp16")]; + tensor inputs_141_cast_fp16 = add(x = inputs_139_cast_fp16, y = obj_329_cast_fp16)[name = tensor("inputs_141_cast_fp16")]; + tensor out_141_axes_0 = const()[name = tensor("out_141_axes_0"), val = tensor([1])]; + tensor var_5360_to_fp16 = const()[name = tensor("op_5360_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_141_cast_fp16 = layer_norm(axes = out_141_axes_0, epsilon = var_5360_to_fp16, x = inputs_141_cast_fp16)[name = tensor("out_141_cast_fp16")]; + tensor obj_331_gamma_0_to_fp16 = const()[name = tensor("obj_331_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1353948544)))]; + tensor obj_331_beta_0_to_fp16 = const()[name = tensor("obj_331_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1353951168)))]; + tensor obj_331_epsilon_0_to_fp16 = const()[name = tensor("obj_331_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_331_cast_fp16 = batch_norm(beta = obj_331_beta_0_to_fp16, epsilon = obj_331_epsilon_0_to_fp16, gamma = obj_331_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_141_cast_fp16)[name = tensor("obj_331_cast_fp16")]; + tensor query_95_pad_type_0 = const()[name = tensor("query_95_pad_type_0"), val = tensor("valid")]; + tensor query_95_strides_0 = const()[name = tensor("query_95_strides_0"), val = tensor([1, 1])]; + tensor query_95_pad_0 = const()[name = tensor("query_95_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_95_dilations_0 = const()[name = tensor("query_95_dilations_0"), val = tensor([1, 1])]; + tensor query_95_groups_0 = const()[name = tensor("query_95_groups_0"), val = tensor(1)]; + tensor layers_23_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1353953792)))]; + tensor layers_23_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1357230656)))]; + tensor query_95_cast_fp16 = conv(bias = layers_23_encoder_attn_q_proj_bias_to_fp16, dilations = query_95_dilations_0, groups = query_95_groups_0, pad = query_95_pad_0, pad_type = query_95_pad_type_0, strides = query_95_strides_0, weight = layers_23_encoder_attn_q_proj_weight_to_fp16, x = obj_331_cast_fp16)[name = tensor("query_95_cast_fp16")]; + tensor key_95_pad_type_0 = const()[name = tensor("key_95_pad_type_0"), val = tensor("valid")]; + tensor key_95_strides_0 = const()[name = tensor("key_95_strides_0"), val = tensor([1, 1])]; + tensor key_95_pad_0 = const()[name = tensor("key_95_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_95_dilations_0 = const()[name = tensor("key_95_dilations_0"), val = tensor([1, 1])]; + tensor key_95_groups_0 = const()[name = tensor("key_95_groups_0"), val = tensor(1)]; + tensor layers_23_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1357233280)))]; + tensor key_95_cast_fp16 = conv(dilations = key_95_dilations_0, groups = key_95_groups_0, pad = key_95_pad_0, pad_type = key_95_pad_type_0, strides = key_95_strides_0, weight = layers_23_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_95_cast_fp16")]; + tensor value_95_pad_type_0 = const()[name = tensor("value_95_pad_type_0"), val = tensor("valid")]; + tensor value_95_strides_0 = const()[name = tensor("value_95_strides_0"), val = tensor([1, 1])]; + tensor value_95_pad_0 = const()[name = tensor("value_95_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_95_dilations_0 = const()[name = tensor("value_95_dilations_0"), val = tensor([1, 1])]; + tensor value_95_groups_0 = const()[name = tensor("value_95_groups_0"), val = tensor(1)]; + tensor layers_23_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1360510144)))]; + tensor layers_23_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1363787008)))]; + tensor value_95_cast_fp16 = conv(bias = layers_23_encoder_attn_v_proj_bias_to_fp16, dilations = value_95_dilations_0, groups = value_95_groups_0, pad = value_95_pad_0, pad_type = value_95_pad_type_0, strides = value_95_strides_0, weight = layers_23_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_95_cast_fp16")]; + tensor var_5396 = const()[name = tensor("op_5396"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_95_cast_fp16 = reshape(shape = var_5396, x = query_95_cast_fp16)[name = tensor("mh_q_95_cast_fp16")]; + tensor var_5398_to_fp16 = const()[name = tensor("op_5398_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5399_cast_fp16 = mul(x = mh_q_95_cast_fp16, y = var_5398_to_fp16)[name = tensor("op_5399_cast_fp16")]; + tensor var_5402 = const()[name = tensor("op_5402"), val = tensor([1, 20, 64, 1500])]; + tensor var_5403_cast_fp16 = reshape(shape = var_5402, x = key_95_cast_fp16)[name = tensor("op_5403_cast_fp16")]; + tensor mh_w_143_transpose_x_0 = const()[name = tensor("mh_w_143_transpose_x_0"), val = tensor(true)]; + tensor mh_w_143_transpose_y_0 = const()[name = tensor("mh_w_143_transpose_y_0"), val = tensor(false)]; + tensor mh_w_143_cast_fp16 = matmul(transpose_x = mh_w_143_transpose_x_0, transpose_y = mh_w_143_transpose_y_0, x = var_5399_cast_fp16, y = var_5403_cast_fp16)[name = tensor("mh_w_143_cast_fp16")]; + tensor obj_335_cast_fp16 = softmax(axis = var_5245, x = mh_w_143_cast_fp16)[name = tensor("obj_335_cast_fp16")]; + tensor var_5407 = const()[name = tensor("op_5407"), val = tensor([1, 20, 64, 1500])]; + tensor var_5408_cast_fp16 = reshape(shape = var_5407, x = value_95_cast_fp16)[name = tensor("op_5408_cast_fp16")]; + tensor attn_95_transpose_x_0 = const()[name = tensor("attn_95_transpose_x_0"), val = tensor(false)]; + tensor attn_95_transpose_y_0 = const()[name = tensor("attn_95_transpose_y_0"), val = tensor(true)]; + tensor attn_95_cast_fp16 = matmul(transpose_x = attn_95_transpose_x_0, transpose_y = attn_95_transpose_y_0, x = var_5408_cast_fp16, y = obj_335_cast_fp16)[name = tensor("attn_95_cast_fp16")]; + tensor var_5411 = const()[name = tensor("op_5411"), val = tensor([1, 1280, 1, 1])]; + tensor input_233_cast_fp16 = reshape(shape = var_5411, x = attn_95_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor obj_333_pad_type_0 = const()[name = tensor("obj_333_pad_type_0"), val = tensor("valid")]; + tensor obj_333_strides_0 = const()[name = tensor("obj_333_strides_0"), val = tensor([1, 1])]; + tensor obj_333_pad_0 = const()[name = tensor("obj_333_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_333_dilations_0 = const()[name = tensor("obj_333_dilations_0"), val = tensor([1, 1])]; + tensor obj_333_groups_0 = const()[name = tensor("obj_333_groups_0"), val = tensor(1)]; + tensor layers_23_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1363789632)))]; + tensor layers_23_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1367066496)))]; + tensor obj_333_cast_fp16 = conv(bias = layers_23_encoder_attn_o_proj_bias_to_fp16, dilations = obj_333_dilations_0, groups = obj_333_groups_0, pad = obj_333_pad_0, pad_type = obj_333_pad_type_0, strides = obj_333_strides_0, weight = layers_23_encoder_attn_o_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("obj_333_cast_fp16")]; + tensor inputs_143_cast_fp16 = add(x = inputs_141_cast_fp16, y = obj_333_cast_fp16)[name = tensor("inputs_143_cast_fp16")]; + tensor out_143_axes_0 = const()[name = tensor("out_143_axes_0"), val = tensor([1])]; + tensor var_5429_to_fp16 = const()[name = tensor("op_5429_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_143_cast_fp16 = layer_norm(axes = out_143_axes_0, epsilon = var_5429_to_fp16, x = inputs_143_cast_fp16)[name = tensor("out_143_cast_fp16")]; + tensor input_235_gamma_0_to_fp16 = const()[name = tensor("input_235_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1367069120)))]; + tensor input_235_beta_0_to_fp16 = const()[name = tensor("input_235_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1367071744)))]; + tensor input_235_epsilon_0_to_fp16 = const()[name = tensor("input_235_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_235_cast_fp16 = batch_norm(beta = input_235_beta_0_to_fp16, epsilon = input_235_epsilon_0_to_fp16, gamma = input_235_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_143_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor input_237_pad_type_0 = const()[name = tensor("input_237_pad_type_0"), val = tensor("valid")]; + tensor input_237_strides_0 = const()[name = tensor("input_237_strides_0"), val = tensor([1, 1])]; + tensor input_237_pad_0 = const()[name = tensor("input_237_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_237_dilations_0 = const()[name = tensor("input_237_dilations_0"), val = tensor([1, 1])]; + tensor input_237_groups_0 = const()[name = tensor("input_237_groups_0"), val = tensor(1)]; + tensor layers_23_fc1_weight_to_fp16 = const()[name = tensor("layers_23_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1367074368)))]; + tensor layers_23_fc1_bias_to_fp16 = const()[name = tensor("layers_23_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1380181632)))]; + tensor input_237_cast_fp16 = conv(bias = layers_23_fc1_bias_to_fp16, dilations = input_237_dilations_0, groups = input_237_groups_0, pad = input_237_pad_0, pad_type = input_237_pad_type_0, strides = input_237_strides_0, weight = layers_23_fc1_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_mode_0 = const()[name = tensor("input_239_mode_0"), val = tensor("EXACT")]; + tensor input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor hidden_states_49_pad_type_0 = const()[name = tensor("hidden_states_49_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_49_strides_0 = const()[name = tensor("hidden_states_49_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_49_pad_0 = const()[name = tensor("hidden_states_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_49_dilations_0 = const()[name = tensor("hidden_states_49_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_49_groups_0 = const()[name = tensor("hidden_states_49_groups_0"), val = tensor(1)]; + tensor layers_23_fc2_weight_to_fp16 = const()[name = tensor("layers_23_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1380191936)))]; + tensor layers_23_fc2_bias_to_fp16 = const()[name = tensor("layers_23_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1393299200)))]; + tensor hidden_states_49_cast_fp16 = conv(bias = layers_23_fc2_bias_to_fp16, dilations = hidden_states_49_dilations_0, groups = hidden_states_49_groups_0, pad = hidden_states_49_pad_0, pad_type = hidden_states_49_pad_type_0, strides = hidden_states_49_strides_0, weight = layers_23_fc2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("hidden_states_49_cast_fp16")]; + tensor inputs_145_cast_fp16 = add(x = inputs_143_cast_fp16, y = hidden_states_49_cast_fp16)[name = tensor("inputs_145_cast_fp16")]; + tensor var_5464 = const()[name = tensor("op_5464"), val = tensor(3)]; + tensor out_145_axes_0 = const()[name = tensor("out_145_axes_0"), val = tensor([1])]; + tensor var_5489_to_fp16 = const()[name = tensor("op_5489_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_145_cast_fp16 = layer_norm(axes = out_145_axes_0, epsilon = var_5489_to_fp16, x = inputs_145_cast_fp16)[name = tensor("out_145_cast_fp16")]; + tensor obj_337_gamma_0_to_fp16 = const()[name = tensor("obj_337_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1393301824)))]; + tensor obj_337_beta_0_to_fp16 = const()[name = tensor("obj_337_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1393304448)))]; + tensor obj_337_epsilon_0_to_fp16 = const()[name = tensor("obj_337_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_337_cast_fp16 = batch_norm(beta = obj_337_beta_0_to_fp16, epsilon = obj_337_epsilon_0_to_fp16, gamma = obj_337_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_145_cast_fp16)[name = tensor("obj_337_cast_fp16")]; + tensor query_97_pad_type_0 = const()[name = tensor("query_97_pad_type_0"), val = tensor("valid")]; + tensor query_97_strides_0 = const()[name = tensor("query_97_strides_0"), val = tensor([1, 1])]; + tensor query_97_pad_0 = const()[name = tensor("query_97_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_97_dilations_0 = const()[name = tensor("query_97_dilations_0"), val = tensor([1, 1])]; + tensor query_97_groups_0 = const()[name = tensor("query_97_groups_0"), val = tensor(1)]; + tensor layers_24_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1393307072)))]; + tensor layers_24_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1396583936)))]; + tensor query_97_cast_fp16 = conv(bias = layers_24_self_attn_q_proj_bias_to_fp16, dilations = query_97_dilations_0, groups = query_97_groups_0, pad = query_97_pad_0, pad_type = query_97_pad_type_0, strides = query_97_strides_0, weight = layers_24_self_attn_q_proj_weight_to_fp16, x = obj_337_cast_fp16)[name = tensor("query_97_cast_fp16")]; + tensor current_key_49_pad_type_0 = const()[name = tensor("current_key_49_pad_type_0"), val = tensor("valid")]; + tensor current_key_49_strides_0 = const()[name = tensor("current_key_49_strides_0"), val = tensor([1, 1])]; + tensor current_key_49_pad_0 = const()[name = tensor("current_key_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_49_dilations_0 = const()[name = tensor("current_key_49_dilations_0"), val = tensor([1, 1])]; + tensor current_key_49_groups_0 = const()[name = tensor("current_key_49_groups_0"), val = tensor(1)]; + tensor layers_24_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1396586560)))]; + tensor current_key_49_cast_fp16 = conv(dilations = current_key_49_dilations_0, groups = current_key_49_groups_0, pad = current_key_49_pad_0, pad_type = current_key_49_pad_type_0, strides = current_key_49_strides_0, weight = layers_24_self_attn_k_proj_weight_to_fp16, x = obj_337_cast_fp16)[name = tensor("current_key_49_cast_fp16")]; + tensor current_value_49_pad_type_0 = const()[name = tensor("current_value_49_pad_type_0"), val = tensor("valid")]; + tensor current_value_49_strides_0 = const()[name = tensor("current_value_49_strides_0"), val = tensor([1, 1])]; + tensor current_value_49_pad_0 = const()[name = tensor("current_value_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_49_dilations_0 = const()[name = tensor("current_value_49_dilations_0"), val = tensor([1, 1])]; + tensor current_value_49_groups_0 = const()[name = tensor("current_value_49_groups_0"), val = tensor(1)]; + tensor layers_24_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1399863424)))]; + tensor layers_24_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1403140288)))]; + tensor current_value_49_cast_fp16 = conv(bias = layers_24_self_attn_v_proj_bias_to_fp16, dilations = current_value_49_dilations_0, groups = current_value_49_groups_0, pad = current_value_49_pad_0, pad_type = current_value_49_pad_type_0, strides = current_value_49_strides_0, weight = layers_24_self_attn_v_proj_weight_to_fp16, x = obj_337_cast_fp16)[name = tensor("current_value_49_cast_fp16")]; + tensor var_5528_cast_fp16 = mul(x = var_103_cast_fp16_24, y = var_239_cast_fp16)[name = tensor("op_5528_cast_fp16")]; + tensor var_5529_cast_fp16 = mul(x = current_key_49_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_5529_cast_fp16")]; + tensor key_97_cast_fp16 = add(x = var_5528_cast_fp16, y = var_5529_cast_fp16)[name = tensor("key_97_cast_fp16")]; + tensor var_5532_cast_fp16 = mul(x = var_138_cast_fp16_24, y = var_239_cast_fp16)[name = tensor("op_5532_cast_fp16")]; + tensor var_5533_cast_fp16 = mul(x = current_value_49_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_5533_cast_fp16")]; + tensor value_97_cast_fp16 = add(x = var_5532_cast_fp16, y = var_5533_cast_fp16)[name = tensor("value_97_cast_fp16")]; + tensor var_5537 = const()[name = tensor("op_5537"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_97_cast_fp16 = reshape(shape = var_5537, x = query_97_cast_fp16)[name = tensor("mh_q_97_cast_fp16")]; + tensor var_5539_to_fp16 = const()[name = tensor("op_5539_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5540_cast_fp16 = mul(x = mh_q_97_cast_fp16, y = var_5539_to_fp16)[name = tensor("op_5540_cast_fp16")]; + tensor var_5543 = const()[name = tensor("op_5543"), val = tensor([1, 20, 64, 448])]; + tensor var_5544_cast_fp16 = reshape(shape = var_5543, x = key_97_cast_fp16)[name = tensor("op_5544_cast_fp16")]; + tensor mh_w_145_transpose_x_0 = const()[name = tensor("mh_w_145_transpose_x_0"), val = tensor(true)]; + tensor mh_w_145_transpose_y_0 = const()[name = tensor("mh_w_145_transpose_y_0"), val = tensor(false)]; + tensor mh_w_145_cast_fp16 = matmul(transpose_x = mh_w_145_transpose_x_0, transpose_y = mh_w_145_transpose_y_0, x = var_5540_cast_fp16, y = var_5544_cast_fp16)[name = tensor("mh_w_145_cast_fp16")]; + tensor mh_w_147_cast_fp16 = add(x = mh_w_145_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_147_cast_fp16")]; + tensor var_5552_cast_fp16 = softmax(axis = var_5464, x = mh_w_147_cast_fp16)[name = tensor("op_5552_cast_fp16")]; + tensor var_5553 = const()[name = tensor("op_5553"), val = tensor([1, 20, 64, 448])]; + tensor var_5554_cast_fp16 = reshape(shape = var_5553, x = value_97_cast_fp16)[name = tensor("op_5554_cast_fp16")]; + tensor attn_97_transpose_x_0 = const()[name = tensor("attn_97_transpose_x_0"), val = tensor(false)]; + tensor attn_97_transpose_y_0 = const()[name = tensor("attn_97_transpose_y_0"), val = tensor(true)]; + tensor attn_97_cast_fp16 = matmul(transpose_x = attn_97_transpose_x_0, transpose_y = attn_97_transpose_y_0, x = var_5554_cast_fp16, y = var_5552_cast_fp16)[name = tensor("attn_97_cast_fp16")]; + tensor var_5557 = const()[name = tensor("op_5557"), val = tensor([1, 1280, 1, 1])]; + tensor input_241_cast_fp16 = reshape(shape = var_5557, x = attn_97_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor obj_343_pad_type_0 = const()[name = tensor("obj_343_pad_type_0"), val = tensor("valid")]; + tensor obj_343_strides_0 = const()[name = tensor("obj_343_strides_0"), val = tensor([1, 1])]; + tensor obj_343_pad_0 = const()[name = tensor("obj_343_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_343_dilations_0 = const()[name = tensor("obj_343_dilations_0"), val = tensor([1, 1])]; + tensor obj_343_groups_0 = const()[name = tensor("obj_343_groups_0"), val = tensor(1)]; + tensor layers_24_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1403142912)))]; + tensor layers_24_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1406419776)))]; + tensor obj_343_cast_fp16 = conv(bias = layers_24_self_attn_o_proj_bias_to_fp16, dilations = obj_343_dilations_0, groups = obj_343_groups_0, pad = obj_343_pad_0, pad_type = obj_343_pad_type_0, strides = obj_343_strides_0, weight = layers_24_self_attn_o_proj_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("obj_343_cast_fp16")]; + tensor inputs_147_cast_fp16 = add(x = inputs_145_cast_fp16, y = obj_343_cast_fp16)[name = tensor("inputs_147_cast_fp16")]; + tensor out_147_axes_0 = const()[name = tensor("out_147_axes_0"), val = tensor([1])]; + tensor var_5579_to_fp16 = const()[name = tensor("op_5579_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_147_cast_fp16 = layer_norm(axes = out_147_axes_0, epsilon = var_5579_to_fp16, x = inputs_147_cast_fp16)[name = tensor("out_147_cast_fp16")]; + tensor obj_345_gamma_0_to_fp16 = const()[name = tensor("obj_345_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1406422400)))]; + tensor obj_345_beta_0_to_fp16 = const()[name = tensor("obj_345_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1406425024)))]; + tensor obj_345_epsilon_0_to_fp16 = const()[name = tensor("obj_345_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_345_cast_fp16 = batch_norm(beta = obj_345_beta_0_to_fp16, epsilon = obj_345_epsilon_0_to_fp16, gamma = obj_345_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_147_cast_fp16)[name = tensor("obj_345_cast_fp16")]; + tensor query_99_pad_type_0 = const()[name = tensor("query_99_pad_type_0"), val = tensor("valid")]; + tensor query_99_strides_0 = const()[name = tensor("query_99_strides_0"), val = tensor([1, 1])]; + tensor query_99_pad_0 = const()[name = tensor("query_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_99_dilations_0 = const()[name = tensor("query_99_dilations_0"), val = tensor([1, 1])]; + tensor query_99_groups_0 = const()[name = tensor("query_99_groups_0"), val = tensor(1)]; + tensor layers_24_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_24_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1406427648)))]; + tensor layers_24_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_24_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1409704512)))]; + tensor query_99_cast_fp16 = conv(bias = layers_24_encoder_attn_q_proj_bias_to_fp16, dilations = query_99_dilations_0, groups = query_99_groups_0, pad = query_99_pad_0, pad_type = query_99_pad_type_0, strides = query_99_strides_0, weight = layers_24_encoder_attn_q_proj_weight_to_fp16, x = obj_345_cast_fp16)[name = tensor("query_99_cast_fp16")]; + tensor key_99_pad_type_0 = const()[name = tensor("key_99_pad_type_0"), val = tensor("valid")]; + tensor key_99_strides_0 = const()[name = tensor("key_99_strides_0"), val = tensor([1, 1])]; + tensor key_99_pad_0 = const()[name = tensor("key_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_99_dilations_0 = const()[name = tensor("key_99_dilations_0"), val = tensor([1, 1])]; + tensor key_99_groups_0 = const()[name = tensor("key_99_groups_0"), val = tensor(1)]; + tensor layers_24_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_24_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1409707136)))]; + tensor key_99_cast_fp16 = conv(dilations = key_99_dilations_0, groups = key_99_groups_0, pad = key_99_pad_0, pad_type = key_99_pad_type_0, strides = key_99_strides_0, weight = layers_24_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_99_cast_fp16")]; + tensor value_99_pad_type_0 = const()[name = tensor("value_99_pad_type_0"), val = tensor("valid")]; + tensor value_99_strides_0 = const()[name = tensor("value_99_strides_0"), val = tensor([1, 1])]; + tensor value_99_pad_0 = const()[name = tensor("value_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_99_dilations_0 = const()[name = tensor("value_99_dilations_0"), val = tensor([1, 1])]; + tensor value_99_groups_0 = const()[name = tensor("value_99_groups_0"), val = tensor(1)]; + tensor layers_24_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_24_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1412984000)))]; + tensor layers_24_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_24_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1416260864)))]; + tensor value_99_cast_fp16 = conv(bias = layers_24_encoder_attn_v_proj_bias_to_fp16, dilations = value_99_dilations_0, groups = value_99_groups_0, pad = value_99_pad_0, pad_type = value_99_pad_type_0, strides = value_99_strides_0, weight = layers_24_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_99_cast_fp16")]; + tensor var_5615 = const()[name = tensor("op_5615"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_99_cast_fp16 = reshape(shape = var_5615, x = query_99_cast_fp16)[name = tensor("mh_q_99_cast_fp16")]; + tensor var_5617_to_fp16 = const()[name = tensor("op_5617_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5618_cast_fp16 = mul(x = mh_q_99_cast_fp16, y = var_5617_to_fp16)[name = tensor("op_5618_cast_fp16")]; + tensor var_5621 = const()[name = tensor("op_5621"), val = tensor([1, 20, 64, 1500])]; + tensor var_5622_cast_fp16 = reshape(shape = var_5621, x = key_99_cast_fp16)[name = tensor("op_5622_cast_fp16")]; + tensor mh_w_149_transpose_x_0 = const()[name = tensor("mh_w_149_transpose_x_0"), val = tensor(true)]; + tensor mh_w_149_transpose_y_0 = const()[name = tensor("mh_w_149_transpose_y_0"), val = tensor(false)]; + tensor mh_w_149_cast_fp16 = matmul(transpose_x = mh_w_149_transpose_x_0, transpose_y = mh_w_149_transpose_y_0, x = var_5618_cast_fp16, y = var_5622_cast_fp16)[name = tensor("mh_w_149_cast_fp16")]; + tensor obj_349_cast_fp16 = softmax(axis = var_5464, x = mh_w_149_cast_fp16)[name = tensor("obj_349_cast_fp16")]; + tensor var_5626 = const()[name = tensor("op_5626"), val = tensor([1, 20, 64, 1500])]; + tensor var_5627_cast_fp16 = reshape(shape = var_5626, x = value_99_cast_fp16)[name = tensor("op_5627_cast_fp16")]; + tensor attn_99_transpose_x_0 = const()[name = tensor("attn_99_transpose_x_0"), val = tensor(false)]; + tensor attn_99_transpose_y_0 = const()[name = tensor("attn_99_transpose_y_0"), val = tensor(true)]; + tensor attn_99_cast_fp16 = matmul(transpose_x = attn_99_transpose_x_0, transpose_y = attn_99_transpose_y_0, x = var_5627_cast_fp16, y = obj_349_cast_fp16)[name = tensor("attn_99_cast_fp16")]; + tensor var_5630 = const()[name = tensor("op_5630"), val = tensor([1, 1280, 1, 1])]; + tensor input_243_cast_fp16 = reshape(shape = var_5630, x = attn_99_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor obj_347_pad_type_0 = const()[name = tensor("obj_347_pad_type_0"), val = tensor("valid")]; + tensor obj_347_strides_0 = const()[name = tensor("obj_347_strides_0"), val = tensor([1, 1])]; + tensor obj_347_pad_0 = const()[name = tensor("obj_347_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_347_dilations_0 = const()[name = tensor("obj_347_dilations_0"), val = tensor([1, 1])]; + tensor obj_347_groups_0 = const()[name = tensor("obj_347_groups_0"), val = tensor(1)]; + tensor layers_24_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_24_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1416263488)))]; + tensor layers_24_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_24_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1419540352)))]; + tensor obj_347_cast_fp16 = conv(bias = layers_24_encoder_attn_o_proj_bias_to_fp16, dilations = obj_347_dilations_0, groups = obj_347_groups_0, pad = obj_347_pad_0, pad_type = obj_347_pad_type_0, strides = obj_347_strides_0, weight = layers_24_encoder_attn_o_proj_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("obj_347_cast_fp16")]; + tensor inputs_149_cast_fp16 = add(x = inputs_147_cast_fp16, y = obj_347_cast_fp16)[name = tensor("inputs_149_cast_fp16")]; + tensor out_149_axes_0 = const()[name = tensor("out_149_axes_0"), val = tensor([1])]; + tensor var_5651_to_fp16 = const()[name = tensor("op_5651_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_149_cast_fp16 = layer_norm(axes = out_149_axes_0, epsilon = var_5651_to_fp16, x = inputs_149_cast_fp16)[name = tensor("out_149_cast_fp16")]; + tensor input_245_gamma_0_to_fp16 = const()[name = tensor("input_245_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1419542976)))]; + tensor input_245_beta_0_to_fp16 = const()[name = tensor("input_245_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1419545600)))]; + tensor input_245_epsilon_0_to_fp16 = const()[name = tensor("input_245_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_245_cast_fp16 = batch_norm(beta = input_245_beta_0_to_fp16, epsilon = input_245_epsilon_0_to_fp16, gamma = input_245_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_149_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor input_247_pad_type_0 = const()[name = tensor("input_247_pad_type_0"), val = tensor("valid")]; + tensor input_247_strides_0 = const()[name = tensor("input_247_strides_0"), val = tensor([1, 1])]; + tensor input_247_pad_0 = const()[name = tensor("input_247_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_247_dilations_0 = const()[name = tensor("input_247_dilations_0"), val = tensor([1, 1])]; + tensor input_247_groups_0 = const()[name = tensor("input_247_groups_0"), val = tensor(1)]; + tensor layers_24_fc1_weight_to_fp16 = const()[name = tensor("layers_24_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1419548224)))]; + tensor layers_24_fc1_bias_to_fp16 = const()[name = tensor("layers_24_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1432655488)))]; + tensor input_247_cast_fp16 = conv(bias = layers_24_fc1_bias_to_fp16, dilations = input_247_dilations_0, groups = input_247_groups_0, pad = input_247_pad_0, pad_type = input_247_pad_type_0, strides = input_247_strides_0, weight = layers_24_fc1_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor input_249_mode_0 = const()[name = tensor("input_249_mode_0"), val = tensor("EXACT")]; + tensor input_249_cast_fp16 = gelu(mode = input_249_mode_0, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor hidden_states_51_pad_type_0 = const()[name = tensor("hidden_states_51_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_51_strides_0 = const()[name = tensor("hidden_states_51_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_51_pad_0 = const()[name = tensor("hidden_states_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_51_dilations_0 = const()[name = tensor("hidden_states_51_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_51_groups_0 = const()[name = tensor("hidden_states_51_groups_0"), val = tensor(1)]; + tensor layers_24_fc2_weight_to_fp16 = const()[name = tensor("layers_24_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1432665792)))]; + tensor layers_24_fc2_bias_to_fp16 = const()[name = tensor("layers_24_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1445773056)))]; + tensor hidden_states_51_cast_fp16 = conv(bias = layers_24_fc2_bias_to_fp16, dilations = hidden_states_51_dilations_0, groups = hidden_states_51_groups_0, pad = hidden_states_51_pad_0, pad_type = hidden_states_51_pad_type_0, strides = hidden_states_51_strides_0, weight = layers_24_fc2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("hidden_states_51_cast_fp16")]; + tensor inputs_151_cast_fp16 = add(x = inputs_149_cast_fp16, y = hidden_states_51_cast_fp16)[name = tensor("inputs_151_cast_fp16")]; + tensor var_5687 = const()[name = tensor("op_5687"), val = tensor(3)]; + tensor out_151_axes_0 = const()[name = tensor("out_151_axes_0"), val = tensor([1])]; + tensor var_5712_to_fp16 = const()[name = tensor("op_5712_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_151_cast_fp16 = layer_norm(axes = out_151_axes_0, epsilon = var_5712_to_fp16, x = inputs_151_cast_fp16)[name = tensor("out_151_cast_fp16")]; + tensor obj_351_gamma_0_to_fp16 = const()[name = tensor("obj_351_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1445775680)))]; + tensor obj_351_beta_0_to_fp16 = const()[name = tensor("obj_351_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1445778304)))]; + tensor obj_351_epsilon_0_to_fp16 = const()[name = tensor("obj_351_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_351_cast_fp16 = batch_norm(beta = obj_351_beta_0_to_fp16, epsilon = obj_351_epsilon_0_to_fp16, gamma = obj_351_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_151_cast_fp16)[name = tensor("obj_351_cast_fp16")]; + tensor query_101_pad_type_0 = const()[name = tensor("query_101_pad_type_0"), val = tensor("valid")]; + tensor query_101_strides_0 = const()[name = tensor("query_101_strides_0"), val = tensor([1, 1])]; + tensor query_101_pad_0 = const()[name = tensor("query_101_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_101_dilations_0 = const()[name = tensor("query_101_dilations_0"), val = tensor([1, 1])]; + tensor query_101_groups_0 = const()[name = tensor("query_101_groups_0"), val = tensor(1)]; + tensor layers_25_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1445780928)))]; + tensor layers_25_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1449057792)))]; + tensor query_101_cast_fp16 = conv(bias = layers_25_self_attn_q_proj_bias_to_fp16, dilations = query_101_dilations_0, groups = query_101_groups_0, pad = query_101_pad_0, pad_type = query_101_pad_type_0, strides = query_101_strides_0, weight = layers_25_self_attn_q_proj_weight_to_fp16, x = obj_351_cast_fp16)[name = tensor("query_101_cast_fp16")]; + tensor current_key_51_pad_type_0 = const()[name = tensor("current_key_51_pad_type_0"), val = tensor("valid")]; + tensor current_key_51_strides_0 = const()[name = tensor("current_key_51_strides_0"), val = tensor([1, 1])]; + tensor current_key_51_pad_0 = const()[name = tensor("current_key_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_51_dilations_0 = const()[name = tensor("current_key_51_dilations_0"), val = tensor([1, 1])]; + tensor current_key_51_groups_0 = const()[name = tensor("current_key_51_groups_0"), val = tensor(1)]; + tensor layers_25_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1449060416)))]; + tensor current_key_51_cast_fp16 = conv(dilations = current_key_51_dilations_0, groups = current_key_51_groups_0, pad = current_key_51_pad_0, pad_type = current_key_51_pad_type_0, strides = current_key_51_strides_0, weight = layers_25_self_attn_k_proj_weight_to_fp16, x = obj_351_cast_fp16)[name = tensor("current_key_51_cast_fp16")]; + tensor current_value_51_pad_type_0 = const()[name = tensor("current_value_51_pad_type_0"), val = tensor("valid")]; + tensor current_value_51_strides_0 = const()[name = tensor("current_value_51_strides_0"), val = tensor([1, 1])]; + tensor current_value_51_pad_0 = const()[name = tensor("current_value_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_51_dilations_0 = const()[name = tensor("current_value_51_dilations_0"), val = tensor([1, 1])]; + tensor current_value_51_groups_0 = const()[name = tensor("current_value_51_groups_0"), val = tensor(1)]; + tensor layers_25_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1452337280)))]; + tensor layers_25_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1455614144)))]; + tensor current_value_51_cast_fp16 = conv(bias = layers_25_self_attn_v_proj_bias_to_fp16, dilations = current_value_51_dilations_0, groups = current_value_51_groups_0, pad = current_value_51_pad_0, pad_type = current_value_51_pad_type_0, strides = current_value_51_strides_0, weight = layers_25_self_attn_v_proj_weight_to_fp16, x = obj_351_cast_fp16)[name = tensor("current_value_51_cast_fp16")]; + tensor var_5751_cast_fp16 = mul(x = var_103_cast_fp16_25, y = var_239_cast_fp16)[name = tensor("op_5751_cast_fp16")]; + tensor var_5752_cast_fp16 = mul(x = current_key_51_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_5752_cast_fp16")]; + tensor key_101_cast_fp16 = add(x = var_5751_cast_fp16, y = var_5752_cast_fp16)[name = tensor("key_101_cast_fp16")]; + tensor var_5755_cast_fp16 = mul(x = var_138_cast_fp16_25, y = var_239_cast_fp16)[name = tensor("op_5755_cast_fp16")]; + tensor var_5756_cast_fp16 = mul(x = current_value_51_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_5756_cast_fp16")]; + tensor value_101_cast_fp16 = add(x = var_5755_cast_fp16, y = var_5756_cast_fp16)[name = tensor("value_101_cast_fp16")]; + tensor var_5760 = const()[name = tensor("op_5760"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_101_cast_fp16 = reshape(shape = var_5760, x = query_101_cast_fp16)[name = tensor("mh_q_101_cast_fp16")]; + tensor var_5762_to_fp16 = const()[name = tensor("op_5762_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5763_cast_fp16 = mul(x = mh_q_101_cast_fp16, y = var_5762_to_fp16)[name = tensor("op_5763_cast_fp16")]; + tensor var_5766 = const()[name = tensor("op_5766"), val = tensor([1, 20, 64, 448])]; + tensor var_5767_cast_fp16 = reshape(shape = var_5766, x = key_101_cast_fp16)[name = tensor("op_5767_cast_fp16")]; + tensor mh_w_151_transpose_x_0 = const()[name = tensor("mh_w_151_transpose_x_0"), val = tensor(true)]; + tensor mh_w_151_transpose_y_0 = const()[name = tensor("mh_w_151_transpose_y_0"), val = tensor(false)]; + tensor mh_w_151_cast_fp16 = matmul(transpose_x = mh_w_151_transpose_x_0, transpose_y = mh_w_151_transpose_y_0, x = var_5763_cast_fp16, y = var_5767_cast_fp16)[name = tensor("mh_w_151_cast_fp16")]; + tensor mh_w_153_cast_fp16 = add(x = mh_w_151_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_153_cast_fp16")]; + tensor var_5775_cast_fp16 = softmax(axis = var_5687, x = mh_w_153_cast_fp16)[name = tensor("op_5775_cast_fp16")]; + tensor var_5776 = const()[name = tensor("op_5776"), val = tensor([1, 20, 64, 448])]; + tensor var_5777_cast_fp16 = reshape(shape = var_5776, x = value_101_cast_fp16)[name = tensor("op_5777_cast_fp16")]; + tensor attn_101_transpose_x_0 = const()[name = tensor("attn_101_transpose_x_0"), val = tensor(false)]; + tensor attn_101_transpose_y_0 = const()[name = tensor("attn_101_transpose_y_0"), val = tensor(true)]; + tensor attn_101_cast_fp16 = matmul(transpose_x = attn_101_transpose_x_0, transpose_y = attn_101_transpose_y_0, x = var_5777_cast_fp16, y = var_5775_cast_fp16)[name = tensor("attn_101_cast_fp16")]; + tensor var_5780 = const()[name = tensor("op_5780"), val = tensor([1, 1280, 1, 1])]; + tensor input_251_cast_fp16 = reshape(shape = var_5780, x = attn_101_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor obj_357_pad_type_0 = const()[name = tensor("obj_357_pad_type_0"), val = tensor("valid")]; + tensor obj_357_strides_0 = const()[name = tensor("obj_357_strides_0"), val = tensor([1, 1])]; + tensor obj_357_pad_0 = const()[name = tensor("obj_357_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_357_dilations_0 = const()[name = tensor("obj_357_dilations_0"), val = tensor([1, 1])]; + tensor obj_357_groups_0 = const()[name = tensor("obj_357_groups_0"), val = tensor(1)]; + tensor layers_25_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1455616768)))]; + tensor layers_25_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1458893632)))]; + tensor obj_357_cast_fp16 = conv(bias = layers_25_self_attn_o_proj_bias_to_fp16, dilations = obj_357_dilations_0, groups = obj_357_groups_0, pad = obj_357_pad_0, pad_type = obj_357_pad_type_0, strides = obj_357_strides_0, weight = layers_25_self_attn_o_proj_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("obj_357_cast_fp16")]; + tensor inputs_153_cast_fp16 = add(x = inputs_151_cast_fp16, y = obj_357_cast_fp16)[name = tensor("inputs_153_cast_fp16")]; + tensor out_153_axes_0 = const()[name = tensor("out_153_axes_0"), val = tensor([1])]; + tensor var_5802_to_fp16 = const()[name = tensor("op_5802_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_153_cast_fp16 = layer_norm(axes = out_153_axes_0, epsilon = var_5802_to_fp16, x = inputs_153_cast_fp16)[name = tensor("out_153_cast_fp16")]; + tensor obj_359_gamma_0_to_fp16 = const()[name = tensor("obj_359_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1458896256)))]; + tensor obj_359_beta_0_to_fp16 = const()[name = tensor("obj_359_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1458898880)))]; + tensor obj_359_epsilon_0_to_fp16 = const()[name = tensor("obj_359_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_359_cast_fp16 = batch_norm(beta = obj_359_beta_0_to_fp16, epsilon = obj_359_epsilon_0_to_fp16, gamma = obj_359_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_153_cast_fp16)[name = tensor("obj_359_cast_fp16")]; + tensor query_103_pad_type_0 = const()[name = tensor("query_103_pad_type_0"), val = tensor("valid")]; + tensor query_103_strides_0 = const()[name = tensor("query_103_strides_0"), val = tensor([1, 1])]; + tensor query_103_pad_0 = const()[name = tensor("query_103_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_103_dilations_0 = const()[name = tensor("query_103_dilations_0"), val = tensor([1, 1])]; + tensor query_103_groups_0 = const()[name = tensor("query_103_groups_0"), val = tensor(1)]; + tensor layers_25_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_25_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1458901504)))]; + tensor layers_25_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_25_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1462178368)))]; + tensor query_103_cast_fp16 = conv(bias = layers_25_encoder_attn_q_proj_bias_to_fp16, dilations = query_103_dilations_0, groups = query_103_groups_0, pad = query_103_pad_0, pad_type = query_103_pad_type_0, strides = query_103_strides_0, weight = layers_25_encoder_attn_q_proj_weight_to_fp16, x = obj_359_cast_fp16)[name = tensor("query_103_cast_fp16")]; + tensor key_103_pad_type_0 = const()[name = tensor("key_103_pad_type_0"), val = tensor("valid")]; + tensor key_103_strides_0 = const()[name = tensor("key_103_strides_0"), val = tensor([1, 1])]; + tensor key_103_pad_0 = const()[name = tensor("key_103_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_103_dilations_0 = const()[name = tensor("key_103_dilations_0"), val = tensor([1, 1])]; + tensor key_103_groups_0 = const()[name = tensor("key_103_groups_0"), val = tensor(1)]; + tensor layers_25_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_25_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1462180992)))]; + tensor key_103_cast_fp16 = conv(dilations = key_103_dilations_0, groups = key_103_groups_0, pad = key_103_pad_0, pad_type = key_103_pad_type_0, strides = key_103_strides_0, weight = layers_25_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_103_cast_fp16")]; + tensor value_103_pad_type_0 = const()[name = tensor("value_103_pad_type_0"), val = tensor("valid")]; + tensor value_103_strides_0 = const()[name = tensor("value_103_strides_0"), val = tensor([1, 1])]; + tensor value_103_pad_0 = const()[name = tensor("value_103_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_103_dilations_0 = const()[name = tensor("value_103_dilations_0"), val = tensor([1, 1])]; + tensor value_103_groups_0 = const()[name = tensor("value_103_groups_0"), val = tensor(1)]; + tensor layers_25_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_25_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1465457856)))]; + tensor layers_25_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_25_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1468734720)))]; + tensor value_103_cast_fp16 = conv(bias = layers_25_encoder_attn_v_proj_bias_to_fp16, dilations = value_103_dilations_0, groups = value_103_groups_0, pad = value_103_pad_0, pad_type = value_103_pad_type_0, strides = value_103_strides_0, weight = layers_25_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_103_cast_fp16")]; + tensor var_5838 = const()[name = tensor("op_5838"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_103_cast_fp16 = reshape(shape = var_5838, x = query_103_cast_fp16)[name = tensor("mh_q_103_cast_fp16")]; + tensor var_5840_to_fp16 = const()[name = tensor("op_5840_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5841_cast_fp16 = mul(x = mh_q_103_cast_fp16, y = var_5840_to_fp16)[name = tensor("op_5841_cast_fp16")]; + tensor var_5844 = const()[name = tensor("op_5844"), val = tensor([1, 20, 64, 1500])]; + tensor var_5845_cast_fp16 = reshape(shape = var_5844, x = key_103_cast_fp16)[name = tensor("op_5845_cast_fp16")]; + tensor mh_w_155_transpose_x_0 = const()[name = tensor("mh_w_155_transpose_x_0"), val = tensor(true)]; + tensor mh_w_155_transpose_y_0 = const()[name = tensor("mh_w_155_transpose_y_0"), val = tensor(false)]; + tensor mh_w_155_cast_fp16 = matmul(transpose_x = mh_w_155_transpose_x_0, transpose_y = mh_w_155_transpose_y_0, x = var_5841_cast_fp16, y = var_5845_cast_fp16)[name = tensor("mh_w_155_cast_fp16")]; + tensor obj_363_cast_fp16 = softmax(axis = var_5687, x = mh_w_155_cast_fp16)[name = tensor("obj_363_cast_fp16")]; + tensor var_5849 = const()[name = tensor("op_5849"), val = tensor([1, 20, 64, 1500])]; + tensor var_5850_cast_fp16 = reshape(shape = var_5849, x = value_103_cast_fp16)[name = tensor("op_5850_cast_fp16")]; + tensor attn_103_transpose_x_0 = const()[name = tensor("attn_103_transpose_x_0"), val = tensor(false)]; + tensor attn_103_transpose_y_0 = const()[name = tensor("attn_103_transpose_y_0"), val = tensor(true)]; + tensor attn_103_cast_fp16 = matmul(transpose_x = attn_103_transpose_x_0, transpose_y = attn_103_transpose_y_0, x = var_5850_cast_fp16, y = obj_363_cast_fp16)[name = tensor("attn_103_cast_fp16")]; + tensor var_5853 = const()[name = tensor("op_5853"), val = tensor([1, 1280, 1, 1])]; + tensor input_253_cast_fp16 = reshape(shape = var_5853, x = attn_103_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor obj_361_pad_type_0 = const()[name = tensor("obj_361_pad_type_0"), val = tensor("valid")]; + tensor obj_361_strides_0 = const()[name = tensor("obj_361_strides_0"), val = tensor([1, 1])]; + tensor obj_361_pad_0 = const()[name = tensor("obj_361_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_361_dilations_0 = const()[name = tensor("obj_361_dilations_0"), val = tensor([1, 1])]; + tensor obj_361_groups_0 = const()[name = tensor("obj_361_groups_0"), val = tensor(1)]; + tensor layers_25_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_25_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1468737344)))]; + tensor layers_25_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_25_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1472014208)))]; + tensor obj_361_cast_fp16 = conv(bias = layers_25_encoder_attn_o_proj_bias_to_fp16, dilations = obj_361_dilations_0, groups = obj_361_groups_0, pad = obj_361_pad_0, pad_type = obj_361_pad_type_0, strides = obj_361_strides_0, weight = layers_25_encoder_attn_o_proj_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("obj_361_cast_fp16")]; + tensor inputs_155_cast_fp16 = add(x = inputs_153_cast_fp16, y = obj_361_cast_fp16)[name = tensor("inputs_155_cast_fp16")]; + tensor out_155_axes_0 = const()[name = tensor("out_155_axes_0"), val = tensor([1])]; + tensor var_5874_to_fp16 = const()[name = tensor("op_5874_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_155_cast_fp16 = layer_norm(axes = out_155_axes_0, epsilon = var_5874_to_fp16, x = inputs_155_cast_fp16)[name = tensor("out_155_cast_fp16")]; + tensor input_255_gamma_0_to_fp16 = const()[name = tensor("input_255_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1472016832)))]; + tensor input_255_beta_0_to_fp16 = const()[name = tensor("input_255_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1472019456)))]; + tensor input_255_epsilon_0_to_fp16 = const()[name = tensor("input_255_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_255_cast_fp16 = batch_norm(beta = input_255_beta_0_to_fp16, epsilon = input_255_epsilon_0_to_fp16, gamma = input_255_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_155_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor input_257_pad_type_0 = const()[name = tensor("input_257_pad_type_0"), val = tensor("valid")]; + tensor input_257_strides_0 = const()[name = tensor("input_257_strides_0"), val = tensor([1, 1])]; + tensor input_257_pad_0 = const()[name = tensor("input_257_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_257_dilations_0 = const()[name = tensor("input_257_dilations_0"), val = tensor([1, 1])]; + tensor input_257_groups_0 = const()[name = tensor("input_257_groups_0"), val = tensor(1)]; + tensor layers_25_fc1_weight_to_fp16 = const()[name = tensor("layers_25_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1472022080)))]; + tensor layers_25_fc1_bias_to_fp16 = const()[name = tensor("layers_25_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1485129344)))]; + tensor input_257_cast_fp16 = conv(bias = layers_25_fc1_bias_to_fp16, dilations = input_257_dilations_0, groups = input_257_groups_0, pad = input_257_pad_0, pad_type = input_257_pad_type_0, strides = input_257_strides_0, weight = layers_25_fc1_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor input_259_mode_0 = const()[name = tensor("input_259_mode_0"), val = tensor("EXACT")]; + tensor input_259_cast_fp16 = gelu(mode = input_259_mode_0, x = input_257_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor hidden_states_53_pad_type_0 = const()[name = tensor("hidden_states_53_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_53_strides_0 = const()[name = tensor("hidden_states_53_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_53_pad_0 = const()[name = tensor("hidden_states_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_53_dilations_0 = const()[name = tensor("hidden_states_53_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_53_groups_0 = const()[name = tensor("hidden_states_53_groups_0"), val = tensor(1)]; + tensor layers_25_fc2_weight_to_fp16 = const()[name = tensor("layers_25_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1485139648)))]; + tensor layers_25_fc2_bias_to_fp16 = const()[name = tensor("layers_25_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1498246912)))]; + tensor hidden_states_53_cast_fp16 = conv(bias = layers_25_fc2_bias_to_fp16, dilations = hidden_states_53_dilations_0, groups = hidden_states_53_groups_0, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = hidden_states_53_strides_0, weight = layers_25_fc2_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("hidden_states_53_cast_fp16")]; + tensor inputs_157_cast_fp16 = add(x = inputs_155_cast_fp16, y = hidden_states_53_cast_fp16)[name = tensor("inputs_157_cast_fp16")]; + tensor var_5910 = const()[name = tensor("op_5910"), val = tensor(3)]; + tensor out_157_axes_0 = const()[name = tensor("out_157_axes_0"), val = tensor([1])]; + tensor var_5935_to_fp16 = const()[name = tensor("op_5935_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_157_cast_fp16 = layer_norm(axes = out_157_axes_0, epsilon = var_5935_to_fp16, x = inputs_157_cast_fp16)[name = tensor("out_157_cast_fp16")]; + tensor obj_365_gamma_0_to_fp16 = const()[name = tensor("obj_365_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1498249536)))]; + tensor obj_365_beta_0_to_fp16 = const()[name = tensor("obj_365_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1498252160)))]; + tensor obj_365_epsilon_0_to_fp16 = const()[name = tensor("obj_365_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_365_cast_fp16 = batch_norm(beta = obj_365_beta_0_to_fp16, epsilon = obj_365_epsilon_0_to_fp16, gamma = obj_365_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_157_cast_fp16)[name = tensor("obj_365_cast_fp16")]; + tensor query_105_pad_type_0 = const()[name = tensor("query_105_pad_type_0"), val = tensor("valid")]; + tensor query_105_strides_0 = const()[name = tensor("query_105_strides_0"), val = tensor([1, 1])]; + tensor query_105_pad_0 = const()[name = tensor("query_105_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_105_dilations_0 = const()[name = tensor("query_105_dilations_0"), val = tensor([1, 1])]; + tensor query_105_groups_0 = const()[name = tensor("query_105_groups_0"), val = tensor(1)]; + tensor layers_26_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1498254784)))]; + tensor layers_26_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1501531648)))]; + tensor query_105_cast_fp16 = conv(bias = layers_26_self_attn_q_proj_bias_to_fp16, dilations = query_105_dilations_0, groups = query_105_groups_0, pad = query_105_pad_0, pad_type = query_105_pad_type_0, strides = query_105_strides_0, weight = layers_26_self_attn_q_proj_weight_to_fp16, x = obj_365_cast_fp16)[name = tensor("query_105_cast_fp16")]; + tensor current_key_53_pad_type_0 = const()[name = tensor("current_key_53_pad_type_0"), val = tensor("valid")]; + tensor current_key_53_strides_0 = const()[name = tensor("current_key_53_strides_0"), val = tensor([1, 1])]; + tensor current_key_53_pad_0 = const()[name = tensor("current_key_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_53_dilations_0 = const()[name = tensor("current_key_53_dilations_0"), val = tensor([1, 1])]; + tensor current_key_53_groups_0 = const()[name = tensor("current_key_53_groups_0"), val = tensor(1)]; + tensor layers_26_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1501534272)))]; + tensor current_key_53_cast_fp16 = conv(dilations = current_key_53_dilations_0, groups = current_key_53_groups_0, pad = current_key_53_pad_0, pad_type = current_key_53_pad_type_0, strides = current_key_53_strides_0, weight = layers_26_self_attn_k_proj_weight_to_fp16, x = obj_365_cast_fp16)[name = tensor("current_key_53_cast_fp16")]; + tensor current_value_53_pad_type_0 = const()[name = tensor("current_value_53_pad_type_0"), val = tensor("valid")]; + tensor current_value_53_strides_0 = const()[name = tensor("current_value_53_strides_0"), val = tensor([1, 1])]; + tensor current_value_53_pad_0 = const()[name = tensor("current_value_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_53_dilations_0 = const()[name = tensor("current_value_53_dilations_0"), val = tensor([1, 1])]; + tensor current_value_53_groups_0 = const()[name = tensor("current_value_53_groups_0"), val = tensor(1)]; + tensor layers_26_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1504811136)))]; + tensor layers_26_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1508088000)))]; + tensor current_value_53_cast_fp16 = conv(bias = layers_26_self_attn_v_proj_bias_to_fp16, dilations = current_value_53_dilations_0, groups = current_value_53_groups_0, pad = current_value_53_pad_0, pad_type = current_value_53_pad_type_0, strides = current_value_53_strides_0, weight = layers_26_self_attn_v_proj_weight_to_fp16, x = obj_365_cast_fp16)[name = tensor("current_value_53_cast_fp16")]; + tensor var_5974_cast_fp16 = mul(x = var_103_cast_fp16_26, y = var_239_cast_fp16)[name = tensor("op_5974_cast_fp16")]; + tensor var_5975_cast_fp16 = mul(x = current_key_53_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_5975_cast_fp16")]; + tensor key_105_cast_fp16 = add(x = var_5974_cast_fp16, y = var_5975_cast_fp16)[name = tensor("key_105_cast_fp16")]; + tensor var_5978_cast_fp16 = mul(x = var_138_cast_fp16_26, y = var_239_cast_fp16)[name = tensor("op_5978_cast_fp16")]; + tensor var_5979_cast_fp16 = mul(x = current_value_53_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_5979_cast_fp16")]; + tensor value_105_cast_fp16 = add(x = var_5978_cast_fp16, y = var_5979_cast_fp16)[name = tensor("value_105_cast_fp16")]; + tensor var_5983 = const()[name = tensor("op_5983"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_105_cast_fp16 = reshape(shape = var_5983, x = query_105_cast_fp16)[name = tensor("mh_q_105_cast_fp16")]; + tensor var_5985_to_fp16 = const()[name = tensor("op_5985_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5986_cast_fp16 = mul(x = mh_q_105_cast_fp16, y = var_5985_to_fp16)[name = tensor("op_5986_cast_fp16")]; + tensor var_5989 = const()[name = tensor("op_5989"), val = tensor([1, 20, 64, 448])]; + tensor var_5990_cast_fp16 = reshape(shape = var_5989, x = key_105_cast_fp16)[name = tensor("op_5990_cast_fp16")]; + tensor mh_w_157_transpose_x_0 = const()[name = tensor("mh_w_157_transpose_x_0"), val = tensor(true)]; + tensor mh_w_157_transpose_y_0 = const()[name = tensor("mh_w_157_transpose_y_0"), val = tensor(false)]; + tensor mh_w_157_cast_fp16 = matmul(transpose_x = mh_w_157_transpose_x_0, transpose_y = mh_w_157_transpose_y_0, x = var_5986_cast_fp16, y = var_5990_cast_fp16)[name = tensor("mh_w_157_cast_fp16")]; + tensor mh_w_159_cast_fp16 = add(x = mh_w_157_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_159_cast_fp16")]; + tensor var_5998_cast_fp16 = softmax(axis = var_5910, x = mh_w_159_cast_fp16)[name = tensor("op_5998_cast_fp16")]; + tensor var_5999 = const()[name = tensor("op_5999"), val = tensor([1, 20, 64, 448])]; + tensor var_6000_cast_fp16 = reshape(shape = var_5999, x = value_105_cast_fp16)[name = tensor("op_6000_cast_fp16")]; + tensor attn_105_transpose_x_0 = const()[name = tensor("attn_105_transpose_x_0"), val = tensor(false)]; + tensor attn_105_transpose_y_0 = const()[name = tensor("attn_105_transpose_y_0"), val = tensor(true)]; + tensor attn_105_cast_fp16 = matmul(transpose_x = attn_105_transpose_x_0, transpose_y = attn_105_transpose_y_0, x = var_6000_cast_fp16, y = var_5998_cast_fp16)[name = tensor("attn_105_cast_fp16")]; + tensor var_6003 = const()[name = tensor("op_6003"), val = tensor([1, 1280, 1, 1])]; + tensor input_261_cast_fp16 = reshape(shape = var_6003, x = attn_105_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor obj_371_pad_type_0 = const()[name = tensor("obj_371_pad_type_0"), val = tensor("valid")]; + tensor obj_371_strides_0 = const()[name = tensor("obj_371_strides_0"), val = tensor([1, 1])]; + tensor obj_371_pad_0 = const()[name = tensor("obj_371_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_371_dilations_0 = const()[name = tensor("obj_371_dilations_0"), val = tensor([1, 1])]; + tensor obj_371_groups_0 = const()[name = tensor("obj_371_groups_0"), val = tensor(1)]; + tensor layers_26_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1508090624)))]; + tensor layers_26_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1511367488)))]; + tensor obj_371_cast_fp16 = conv(bias = layers_26_self_attn_o_proj_bias_to_fp16, dilations = obj_371_dilations_0, groups = obj_371_groups_0, pad = obj_371_pad_0, pad_type = obj_371_pad_type_0, strides = obj_371_strides_0, weight = layers_26_self_attn_o_proj_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("obj_371_cast_fp16")]; + tensor inputs_159_cast_fp16 = add(x = inputs_157_cast_fp16, y = obj_371_cast_fp16)[name = tensor("inputs_159_cast_fp16")]; + tensor out_159_axes_0 = const()[name = tensor("out_159_axes_0"), val = tensor([1])]; + tensor var_6025_to_fp16 = const()[name = tensor("op_6025_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_159_cast_fp16 = layer_norm(axes = out_159_axes_0, epsilon = var_6025_to_fp16, x = inputs_159_cast_fp16)[name = tensor("out_159_cast_fp16")]; + tensor obj_373_gamma_0_to_fp16 = const()[name = tensor("obj_373_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1511370112)))]; + tensor obj_373_beta_0_to_fp16 = const()[name = tensor("obj_373_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1511372736)))]; + tensor obj_373_epsilon_0_to_fp16 = const()[name = tensor("obj_373_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_373_cast_fp16 = batch_norm(beta = obj_373_beta_0_to_fp16, epsilon = obj_373_epsilon_0_to_fp16, gamma = obj_373_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_159_cast_fp16)[name = tensor("obj_373_cast_fp16")]; + tensor query_107_pad_type_0 = const()[name = tensor("query_107_pad_type_0"), val = tensor("valid")]; + tensor query_107_strides_0 = const()[name = tensor("query_107_strides_0"), val = tensor([1, 1])]; + tensor query_107_pad_0 = const()[name = tensor("query_107_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_107_dilations_0 = const()[name = tensor("query_107_dilations_0"), val = tensor([1, 1])]; + tensor query_107_groups_0 = const()[name = tensor("query_107_groups_0"), val = tensor(1)]; + tensor layers_26_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_26_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1511375360)))]; + tensor layers_26_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_26_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1514652224)))]; + tensor query_107_cast_fp16 = conv(bias = layers_26_encoder_attn_q_proj_bias_to_fp16, dilations = query_107_dilations_0, groups = query_107_groups_0, pad = query_107_pad_0, pad_type = query_107_pad_type_0, strides = query_107_strides_0, weight = layers_26_encoder_attn_q_proj_weight_to_fp16, x = obj_373_cast_fp16)[name = tensor("query_107_cast_fp16")]; + tensor key_107_pad_type_0 = const()[name = tensor("key_107_pad_type_0"), val = tensor("valid")]; + tensor key_107_strides_0 = const()[name = tensor("key_107_strides_0"), val = tensor([1, 1])]; + tensor key_107_pad_0 = const()[name = tensor("key_107_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_107_dilations_0 = const()[name = tensor("key_107_dilations_0"), val = tensor([1, 1])]; + tensor key_107_groups_0 = const()[name = tensor("key_107_groups_0"), val = tensor(1)]; + tensor layers_26_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_26_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1514654848)))]; + tensor key_107_cast_fp16 = conv(dilations = key_107_dilations_0, groups = key_107_groups_0, pad = key_107_pad_0, pad_type = key_107_pad_type_0, strides = key_107_strides_0, weight = layers_26_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_107_cast_fp16")]; + tensor value_107_pad_type_0 = const()[name = tensor("value_107_pad_type_0"), val = tensor("valid")]; + tensor value_107_strides_0 = const()[name = tensor("value_107_strides_0"), val = tensor([1, 1])]; + tensor value_107_pad_0 = const()[name = tensor("value_107_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_107_dilations_0 = const()[name = tensor("value_107_dilations_0"), val = tensor([1, 1])]; + tensor value_107_groups_0 = const()[name = tensor("value_107_groups_0"), val = tensor(1)]; + tensor layers_26_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_26_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1517931712)))]; + tensor layers_26_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_26_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1521208576)))]; + tensor value_107_cast_fp16 = conv(bias = layers_26_encoder_attn_v_proj_bias_to_fp16, dilations = value_107_dilations_0, groups = value_107_groups_0, pad = value_107_pad_0, pad_type = value_107_pad_type_0, strides = value_107_strides_0, weight = layers_26_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_107_cast_fp16")]; + tensor var_6061 = const()[name = tensor("op_6061"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_107_cast_fp16 = reshape(shape = var_6061, x = query_107_cast_fp16)[name = tensor("mh_q_107_cast_fp16")]; + tensor var_6063_to_fp16 = const()[name = tensor("op_6063_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6064_cast_fp16 = mul(x = mh_q_107_cast_fp16, y = var_6063_to_fp16)[name = tensor("op_6064_cast_fp16")]; + tensor var_6067 = const()[name = tensor("op_6067"), val = tensor([1, 20, 64, 1500])]; + tensor var_6068_cast_fp16 = reshape(shape = var_6067, x = key_107_cast_fp16)[name = tensor("op_6068_cast_fp16")]; + tensor mh_w_161_transpose_x_0 = const()[name = tensor("mh_w_161_transpose_x_0"), val = tensor(true)]; + tensor mh_w_161_transpose_y_0 = const()[name = tensor("mh_w_161_transpose_y_0"), val = tensor(false)]; + tensor mh_w_161_cast_fp16 = matmul(transpose_x = mh_w_161_transpose_x_0, transpose_y = mh_w_161_transpose_y_0, x = var_6064_cast_fp16, y = var_6068_cast_fp16)[name = tensor("mh_w_161_cast_fp16")]; + tensor obj_377_cast_fp16 = softmax(axis = var_5910, x = mh_w_161_cast_fp16)[name = tensor("obj_377_cast_fp16")]; + tensor var_6072 = const()[name = tensor("op_6072"), val = tensor([1, 20, 64, 1500])]; + tensor var_6073_cast_fp16 = reshape(shape = var_6072, x = value_107_cast_fp16)[name = tensor("op_6073_cast_fp16")]; + tensor attn_107_transpose_x_0 = const()[name = tensor("attn_107_transpose_x_0"), val = tensor(false)]; + tensor attn_107_transpose_y_0 = const()[name = tensor("attn_107_transpose_y_0"), val = tensor(true)]; + tensor attn_107_cast_fp16 = matmul(transpose_x = attn_107_transpose_x_0, transpose_y = attn_107_transpose_y_0, x = var_6073_cast_fp16, y = obj_377_cast_fp16)[name = tensor("attn_107_cast_fp16")]; + tensor var_6076 = const()[name = tensor("op_6076"), val = tensor([1, 1280, 1, 1])]; + tensor input_263_cast_fp16 = reshape(shape = var_6076, x = attn_107_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor obj_375_pad_type_0 = const()[name = tensor("obj_375_pad_type_0"), val = tensor("valid")]; + tensor obj_375_strides_0 = const()[name = tensor("obj_375_strides_0"), val = tensor([1, 1])]; + tensor obj_375_pad_0 = const()[name = tensor("obj_375_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_375_dilations_0 = const()[name = tensor("obj_375_dilations_0"), val = tensor([1, 1])]; + tensor obj_375_groups_0 = const()[name = tensor("obj_375_groups_0"), val = tensor(1)]; + tensor layers_26_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_26_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1521211200)))]; + tensor layers_26_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_26_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1524488064)))]; + tensor obj_375_cast_fp16 = conv(bias = layers_26_encoder_attn_o_proj_bias_to_fp16, dilations = obj_375_dilations_0, groups = obj_375_groups_0, pad = obj_375_pad_0, pad_type = obj_375_pad_type_0, strides = obj_375_strides_0, weight = layers_26_encoder_attn_o_proj_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("obj_375_cast_fp16")]; + tensor inputs_161_cast_fp16 = add(x = inputs_159_cast_fp16, y = obj_375_cast_fp16)[name = tensor("inputs_161_cast_fp16")]; + tensor out_161_axes_0 = const()[name = tensor("out_161_axes_0"), val = tensor([1])]; + tensor var_6094_to_fp16 = const()[name = tensor("op_6094_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_161_cast_fp16 = layer_norm(axes = out_161_axes_0, epsilon = var_6094_to_fp16, x = inputs_161_cast_fp16)[name = tensor("out_161_cast_fp16")]; + tensor input_265_gamma_0_to_fp16 = const()[name = tensor("input_265_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1524490688)))]; + tensor input_265_beta_0_to_fp16 = const()[name = tensor("input_265_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1524493312)))]; + tensor input_265_epsilon_0_to_fp16 = const()[name = tensor("input_265_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_265_cast_fp16 = batch_norm(beta = input_265_beta_0_to_fp16, epsilon = input_265_epsilon_0_to_fp16, gamma = input_265_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_161_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor input_267_pad_type_0 = const()[name = tensor("input_267_pad_type_0"), val = tensor("valid")]; + tensor input_267_strides_0 = const()[name = tensor("input_267_strides_0"), val = tensor([1, 1])]; + tensor input_267_pad_0 = const()[name = tensor("input_267_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_267_dilations_0 = const()[name = tensor("input_267_dilations_0"), val = tensor([1, 1])]; + tensor input_267_groups_0 = const()[name = tensor("input_267_groups_0"), val = tensor(1)]; + tensor layers_26_fc1_weight_to_fp16 = const()[name = tensor("layers_26_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1524495936)))]; + tensor layers_26_fc1_bias_to_fp16 = const()[name = tensor("layers_26_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1537603200)))]; + tensor input_267_cast_fp16 = conv(bias = layers_26_fc1_bias_to_fp16, dilations = input_267_dilations_0, groups = input_267_groups_0, pad = input_267_pad_0, pad_type = input_267_pad_type_0, strides = input_267_strides_0, weight = layers_26_fc1_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor input_269_mode_0 = const()[name = tensor("input_269_mode_0"), val = tensor("EXACT")]; + tensor input_269_cast_fp16 = gelu(mode = input_269_mode_0, x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor hidden_states_55_pad_type_0 = const()[name = tensor("hidden_states_55_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_55_strides_0 = const()[name = tensor("hidden_states_55_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_55_pad_0 = const()[name = tensor("hidden_states_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_55_dilations_0 = const()[name = tensor("hidden_states_55_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_55_groups_0 = const()[name = tensor("hidden_states_55_groups_0"), val = tensor(1)]; + tensor layers_26_fc2_weight_to_fp16 = const()[name = tensor("layers_26_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1537613504)))]; + tensor layers_26_fc2_bias_to_fp16 = const()[name = tensor("layers_26_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1550720768)))]; + tensor hidden_states_55_cast_fp16 = conv(bias = layers_26_fc2_bias_to_fp16, dilations = hidden_states_55_dilations_0, groups = hidden_states_55_groups_0, pad = hidden_states_55_pad_0, pad_type = hidden_states_55_pad_type_0, strides = hidden_states_55_strides_0, weight = layers_26_fc2_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("hidden_states_55_cast_fp16")]; + tensor inputs_163_cast_fp16 = add(x = inputs_161_cast_fp16, y = hidden_states_55_cast_fp16)[name = tensor("inputs_163_cast_fp16")]; + tensor var_6129 = const()[name = tensor("op_6129"), val = tensor(3)]; + tensor out_163_axes_0 = const()[name = tensor("out_163_axes_0"), val = tensor([1])]; + tensor var_6154_to_fp16 = const()[name = tensor("op_6154_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_163_cast_fp16 = layer_norm(axes = out_163_axes_0, epsilon = var_6154_to_fp16, x = inputs_163_cast_fp16)[name = tensor("out_163_cast_fp16")]; + tensor obj_379_gamma_0_to_fp16 = const()[name = tensor("obj_379_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1550723392)))]; + tensor obj_379_beta_0_to_fp16 = const()[name = tensor("obj_379_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1550726016)))]; + tensor obj_379_epsilon_0_to_fp16 = const()[name = tensor("obj_379_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_379_cast_fp16 = batch_norm(beta = obj_379_beta_0_to_fp16, epsilon = obj_379_epsilon_0_to_fp16, gamma = obj_379_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_163_cast_fp16)[name = tensor("obj_379_cast_fp16")]; + tensor query_109_pad_type_0 = const()[name = tensor("query_109_pad_type_0"), val = tensor("valid")]; + tensor query_109_strides_0 = const()[name = tensor("query_109_strides_0"), val = tensor([1, 1])]; + tensor query_109_pad_0 = const()[name = tensor("query_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_109_dilations_0 = const()[name = tensor("query_109_dilations_0"), val = tensor([1, 1])]; + tensor query_109_groups_0 = const()[name = tensor("query_109_groups_0"), val = tensor(1)]; + tensor layers_27_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1550728640)))]; + tensor layers_27_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1554005504)))]; + tensor query_109_cast_fp16 = conv(bias = layers_27_self_attn_q_proj_bias_to_fp16, dilations = query_109_dilations_0, groups = query_109_groups_0, pad = query_109_pad_0, pad_type = query_109_pad_type_0, strides = query_109_strides_0, weight = layers_27_self_attn_q_proj_weight_to_fp16, x = obj_379_cast_fp16)[name = tensor("query_109_cast_fp16")]; + tensor current_key_55_pad_type_0 = const()[name = tensor("current_key_55_pad_type_0"), val = tensor("valid")]; + tensor current_key_55_strides_0 = const()[name = tensor("current_key_55_strides_0"), val = tensor([1, 1])]; + tensor current_key_55_pad_0 = const()[name = tensor("current_key_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_55_dilations_0 = const()[name = tensor("current_key_55_dilations_0"), val = tensor([1, 1])]; + tensor current_key_55_groups_0 = const()[name = tensor("current_key_55_groups_0"), val = tensor(1)]; + tensor layers_27_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1554008128)))]; + tensor current_key_55_cast_fp16 = conv(dilations = current_key_55_dilations_0, groups = current_key_55_groups_0, pad = current_key_55_pad_0, pad_type = current_key_55_pad_type_0, strides = current_key_55_strides_0, weight = layers_27_self_attn_k_proj_weight_to_fp16, x = obj_379_cast_fp16)[name = tensor("current_key_55_cast_fp16")]; + tensor current_value_55_pad_type_0 = const()[name = tensor("current_value_55_pad_type_0"), val = tensor("valid")]; + tensor current_value_55_strides_0 = const()[name = tensor("current_value_55_strides_0"), val = tensor([1, 1])]; + tensor current_value_55_pad_0 = const()[name = tensor("current_value_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_55_dilations_0 = const()[name = tensor("current_value_55_dilations_0"), val = tensor([1, 1])]; + tensor current_value_55_groups_0 = const()[name = tensor("current_value_55_groups_0"), val = tensor(1)]; + tensor layers_27_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1557284992)))]; + tensor layers_27_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1560561856)))]; + tensor current_value_55_cast_fp16 = conv(bias = layers_27_self_attn_v_proj_bias_to_fp16, dilations = current_value_55_dilations_0, groups = current_value_55_groups_0, pad = current_value_55_pad_0, pad_type = current_value_55_pad_type_0, strides = current_value_55_strides_0, weight = layers_27_self_attn_v_proj_weight_to_fp16, x = obj_379_cast_fp16)[name = tensor("current_value_55_cast_fp16")]; + tensor var_6193_cast_fp16 = mul(x = var_103_cast_fp16_27, y = var_239_cast_fp16)[name = tensor("op_6193_cast_fp16")]; + tensor var_6194_cast_fp16 = mul(x = current_key_55_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_6194_cast_fp16")]; + tensor key_109_cast_fp16 = add(x = var_6193_cast_fp16, y = var_6194_cast_fp16)[name = tensor("key_109_cast_fp16")]; + tensor var_6197_cast_fp16 = mul(x = var_138_cast_fp16_27, y = var_239_cast_fp16)[name = tensor("op_6197_cast_fp16")]; + tensor var_6198_cast_fp16 = mul(x = current_value_55_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_6198_cast_fp16")]; + tensor value_109_cast_fp16 = add(x = var_6197_cast_fp16, y = var_6198_cast_fp16)[name = tensor("value_109_cast_fp16")]; + tensor var_6202 = const()[name = tensor("op_6202"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_109_cast_fp16 = reshape(shape = var_6202, x = query_109_cast_fp16)[name = tensor("mh_q_109_cast_fp16")]; + tensor var_6204_to_fp16 = const()[name = tensor("op_6204_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6205_cast_fp16 = mul(x = mh_q_109_cast_fp16, y = var_6204_to_fp16)[name = tensor("op_6205_cast_fp16")]; + tensor var_6208 = const()[name = tensor("op_6208"), val = tensor([1, 20, 64, 448])]; + tensor var_6209_cast_fp16 = reshape(shape = var_6208, x = key_109_cast_fp16)[name = tensor("op_6209_cast_fp16")]; + tensor mh_w_163_transpose_x_0 = const()[name = tensor("mh_w_163_transpose_x_0"), val = tensor(true)]; + tensor mh_w_163_transpose_y_0 = const()[name = tensor("mh_w_163_transpose_y_0"), val = tensor(false)]; + tensor mh_w_163_cast_fp16 = matmul(transpose_x = mh_w_163_transpose_x_0, transpose_y = mh_w_163_transpose_y_0, x = var_6205_cast_fp16, y = var_6209_cast_fp16)[name = tensor("mh_w_163_cast_fp16")]; + tensor mh_w_165_cast_fp16 = add(x = mh_w_163_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_165_cast_fp16")]; + tensor var_6217_cast_fp16 = softmax(axis = var_6129, x = mh_w_165_cast_fp16)[name = tensor("op_6217_cast_fp16")]; + tensor var_6218 = const()[name = tensor("op_6218"), val = tensor([1, 20, 64, 448])]; + tensor var_6219_cast_fp16 = reshape(shape = var_6218, x = value_109_cast_fp16)[name = tensor("op_6219_cast_fp16")]; + tensor attn_109_transpose_x_0 = const()[name = tensor("attn_109_transpose_x_0"), val = tensor(false)]; + tensor attn_109_transpose_y_0 = const()[name = tensor("attn_109_transpose_y_0"), val = tensor(true)]; + tensor attn_109_cast_fp16 = matmul(transpose_x = attn_109_transpose_x_0, transpose_y = attn_109_transpose_y_0, x = var_6219_cast_fp16, y = var_6217_cast_fp16)[name = tensor("attn_109_cast_fp16")]; + tensor var_6222 = const()[name = tensor("op_6222"), val = tensor([1, 1280, 1, 1])]; + tensor input_271_cast_fp16 = reshape(shape = var_6222, x = attn_109_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor obj_385_pad_type_0 = const()[name = tensor("obj_385_pad_type_0"), val = tensor("valid")]; + tensor obj_385_strides_0 = const()[name = tensor("obj_385_strides_0"), val = tensor([1, 1])]; + tensor obj_385_pad_0 = const()[name = tensor("obj_385_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_385_dilations_0 = const()[name = tensor("obj_385_dilations_0"), val = tensor([1, 1])]; + tensor obj_385_groups_0 = const()[name = tensor("obj_385_groups_0"), val = tensor(1)]; + tensor layers_27_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1560564480)))]; + tensor layers_27_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1563841344)))]; + tensor obj_385_cast_fp16 = conv(bias = layers_27_self_attn_o_proj_bias_to_fp16, dilations = obj_385_dilations_0, groups = obj_385_groups_0, pad = obj_385_pad_0, pad_type = obj_385_pad_type_0, strides = obj_385_strides_0, weight = layers_27_self_attn_o_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("obj_385_cast_fp16")]; + tensor inputs_165_cast_fp16 = add(x = inputs_163_cast_fp16, y = obj_385_cast_fp16)[name = tensor("inputs_165_cast_fp16")]; + tensor out_165_axes_0 = const()[name = tensor("out_165_axes_0"), val = tensor([1])]; + tensor var_6244_to_fp16 = const()[name = tensor("op_6244_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_165_cast_fp16 = layer_norm(axes = out_165_axes_0, epsilon = var_6244_to_fp16, x = inputs_165_cast_fp16)[name = tensor("out_165_cast_fp16")]; + tensor obj_387_gamma_0_to_fp16 = const()[name = tensor("obj_387_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1563843968)))]; + tensor obj_387_beta_0_to_fp16 = const()[name = tensor("obj_387_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1563846592)))]; + tensor obj_387_epsilon_0_to_fp16 = const()[name = tensor("obj_387_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_387_cast_fp16 = batch_norm(beta = obj_387_beta_0_to_fp16, epsilon = obj_387_epsilon_0_to_fp16, gamma = obj_387_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_165_cast_fp16)[name = tensor("obj_387_cast_fp16")]; + tensor query_111_pad_type_0 = const()[name = tensor("query_111_pad_type_0"), val = tensor("valid")]; + tensor query_111_strides_0 = const()[name = tensor("query_111_strides_0"), val = tensor([1, 1])]; + tensor query_111_pad_0 = const()[name = tensor("query_111_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_111_dilations_0 = const()[name = tensor("query_111_dilations_0"), val = tensor([1, 1])]; + tensor query_111_groups_0 = const()[name = tensor("query_111_groups_0"), val = tensor(1)]; + tensor layers_27_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_27_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1563849216)))]; + tensor layers_27_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_27_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1567126080)))]; + tensor query_111_cast_fp16 = conv(bias = layers_27_encoder_attn_q_proj_bias_to_fp16, dilations = query_111_dilations_0, groups = query_111_groups_0, pad = query_111_pad_0, pad_type = query_111_pad_type_0, strides = query_111_strides_0, weight = layers_27_encoder_attn_q_proj_weight_to_fp16, x = obj_387_cast_fp16)[name = tensor("query_111_cast_fp16")]; + tensor key_111_pad_type_0 = const()[name = tensor("key_111_pad_type_0"), val = tensor("valid")]; + tensor key_111_strides_0 = const()[name = tensor("key_111_strides_0"), val = tensor([1, 1])]; + tensor key_111_pad_0 = const()[name = tensor("key_111_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_111_dilations_0 = const()[name = tensor("key_111_dilations_0"), val = tensor([1, 1])]; + tensor key_111_groups_0 = const()[name = tensor("key_111_groups_0"), val = tensor(1)]; + tensor layers_27_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_27_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1567128704)))]; + tensor key_111_cast_fp16 = conv(dilations = key_111_dilations_0, groups = key_111_groups_0, pad = key_111_pad_0, pad_type = key_111_pad_type_0, strides = key_111_strides_0, weight = layers_27_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_111_cast_fp16")]; + tensor value_111_pad_type_0 = const()[name = tensor("value_111_pad_type_0"), val = tensor("valid")]; + tensor value_111_strides_0 = const()[name = tensor("value_111_strides_0"), val = tensor([1, 1])]; + tensor value_111_pad_0 = const()[name = tensor("value_111_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_111_dilations_0 = const()[name = tensor("value_111_dilations_0"), val = tensor([1, 1])]; + tensor value_111_groups_0 = const()[name = tensor("value_111_groups_0"), val = tensor(1)]; + tensor layers_27_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_27_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1570405568)))]; + tensor layers_27_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_27_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1573682432)))]; + tensor value_111_cast_fp16 = conv(bias = layers_27_encoder_attn_v_proj_bias_to_fp16, dilations = value_111_dilations_0, groups = value_111_groups_0, pad = value_111_pad_0, pad_type = value_111_pad_type_0, strides = value_111_strides_0, weight = layers_27_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_111_cast_fp16")]; + tensor var_6280 = const()[name = tensor("op_6280"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_111_cast_fp16 = reshape(shape = var_6280, x = query_111_cast_fp16)[name = tensor("mh_q_111_cast_fp16")]; + tensor var_6282_to_fp16 = const()[name = tensor("op_6282_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6283_cast_fp16 = mul(x = mh_q_111_cast_fp16, y = var_6282_to_fp16)[name = tensor("op_6283_cast_fp16")]; + tensor var_6286 = const()[name = tensor("op_6286"), val = tensor([1, 20, 64, 1500])]; + tensor var_6287_cast_fp16 = reshape(shape = var_6286, x = key_111_cast_fp16)[name = tensor("op_6287_cast_fp16")]; + tensor mh_w_167_transpose_x_0 = const()[name = tensor("mh_w_167_transpose_x_0"), val = tensor(true)]; + tensor mh_w_167_transpose_y_0 = const()[name = tensor("mh_w_167_transpose_y_0"), val = tensor(false)]; + tensor mh_w_167_cast_fp16 = matmul(transpose_x = mh_w_167_transpose_x_0, transpose_y = mh_w_167_transpose_y_0, x = var_6283_cast_fp16, y = var_6287_cast_fp16)[name = tensor("mh_w_167_cast_fp16")]; + tensor obj_391_cast_fp16 = softmax(axis = var_6129, x = mh_w_167_cast_fp16)[name = tensor("obj_391_cast_fp16")]; + tensor var_6291 = const()[name = tensor("op_6291"), val = tensor([1, 20, 64, 1500])]; + tensor var_6292_cast_fp16 = reshape(shape = var_6291, x = value_111_cast_fp16)[name = tensor("op_6292_cast_fp16")]; + tensor attn_111_transpose_x_0 = const()[name = tensor("attn_111_transpose_x_0"), val = tensor(false)]; + tensor attn_111_transpose_y_0 = const()[name = tensor("attn_111_transpose_y_0"), val = tensor(true)]; + tensor attn_111_cast_fp16 = matmul(transpose_x = attn_111_transpose_x_0, transpose_y = attn_111_transpose_y_0, x = var_6292_cast_fp16, y = obj_391_cast_fp16)[name = tensor("attn_111_cast_fp16")]; + tensor var_6295 = const()[name = tensor("op_6295"), val = tensor([1, 1280, 1, 1])]; + tensor input_273_cast_fp16 = reshape(shape = var_6295, x = attn_111_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor obj_389_pad_type_0 = const()[name = tensor("obj_389_pad_type_0"), val = tensor("valid")]; + tensor obj_389_strides_0 = const()[name = tensor("obj_389_strides_0"), val = tensor([1, 1])]; + tensor obj_389_pad_0 = const()[name = tensor("obj_389_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_389_dilations_0 = const()[name = tensor("obj_389_dilations_0"), val = tensor([1, 1])]; + tensor obj_389_groups_0 = const()[name = tensor("obj_389_groups_0"), val = tensor(1)]; + tensor layers_27_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_27_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1573685056)))]; + tensor layers_27_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_27_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1576961920)))]; + tensor obj_389_cast_fp16 = conv(bias = layers_27_encoder_attn_o_proj_bias_to_fp16, dilations = obj_389_dilations_0, groups = obj_389_groups_0, pad = obj_389_pad_0, pad_type = obj_389_pad_type_0, strides = obj_389_strides_0, weight = layers_27_encoder_attn_o_proj_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("obj_389_cast_fp16")]; + tensor inputs_167_cast_fp16 = add(x = inputs_165_cast_fp16, y = obj_389_cast_fp16)[name = tensor("inputs_167_cast_fp16")]; + tensor out_167_axes_0 = const()[name = tensor("out_167_axes_0"), val = tensor([1])]; + tensor var_6313_to_fp16 = const()[name = tensor("op_6313_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_167_cast_fp16 = layer_norm(axes = out_167_axes_0, epsilon = var_6313_to_fp16, x = inputs_167_cast_fp16)[name = tensor("out_167_cast_fp16")]; + tensor input_275_gamma_0_to_fp16 = const()[name = tensor("input_275_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1576964544)))]; + tensor input_275_beta_0_to_fp16 = const()[name = tensor("input_275_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1576967168)))]; + tensor input_275_epsilon_0_to_fp16 = const()[name = tensor("input_275_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_275_cast_fp16 = batch_norm(beta = input_275_beta_0_to_fp16, epsilon = input_275_epsilon_0_to_fp16, gamma = input_275_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_167_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor input_277_pad_type_0 = const()[name = tensor("input_277_pad_type_0"), val = tensor("valid")]; + tensor input_277_strides_0 = const()[name = tensor("input_277_strides_0"), val = tensor([1, 1])]; + tensor input_277_pad_0 = const()[name = tensor("input_277_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_277_dilations_0 = const()[name = tensor("input_277_dilations_0"), val = tensor([1, 1])]; + tensor input_277_groups_0 = const()[name = tensor("input_277_groups_0"), val = tensor(1)]; + tensor layers_27_fc1_weight_to_fp16 = const()[name = tensor("layers_27_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1576969792)))]; + tensor layers_27_fc1_bias_to_fp16 = const()[name = tensor("layers_27_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1590077056)))]; + tensor input_277_cast_fp16 = conv(bias = layers_27_fc1_bias_to_fp16, dilations = input_277_dilations_0, groups = input_277_groups_0, pad = input_277_pad_0, pad_type = input_277_pad_type_0, strides = input_277_strides_0, weight = layers_27_fc1_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor input_279_mode_0 = const()[name = tensor("input_279_mode_0"), val = tensor("EXACT")]; + tensor input_279_cast_fp16 = gelu(mode = input_279_mode_0, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor hidden_states_57_pad_type_0 = const()[name = tensor("hidden_states_57_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_57_strides_0 = const()[name = tensor("hidden_states_57_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_57_pad_0 = const()[name = tensor("hidden_states_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_57_dilations_0 = const()[name = tensor("hidden_states_57_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_57_groups_0 = const()[name = tensor("hidden_states_57_groups_0"), val = tensor(1)]; + tensor layers_27_fc2_weight_to_fp16 = const()[name = tensor("layers_27_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1590087360)))]; + tensor layers_27_fc2_bias_to_fp16 = const()[name = tensor("layers_27_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1603194624)))]; + tensor hidden_states_57_cast_fp16 = conv(bias = layers_27_fc2_bias_to_fp16, dilations = hidden_states_57_dilations_0, groups = hidden_states_57_groups_0, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = hidden_states_57_strides_0, weight = layers_27_fc2_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("hidden_states_57_cast_fp16")]; + tensor inputs_169_cast_fp16 = add(x = inputs_167_cast_fp16, y = hidden_states_57_cast_fp16)[name = tensor("inputs_169_cast_fp16")]; + tensor var_6348 = const()[name = tensor("op_6348"), val = tensor(3)]; + tensor out_169_axes_0 = const()[name = tensor("out_169_axes_0"), val = tensor([1])]; + tensor var_6373_to_fp16 = const()[name = tensor("op_6373_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_169_cast_fp16 = layer_norm(axes = out_169_axes_0, epsilon = var_6373_to_fp16, x = inputs_169_cast_fp16)[name = tensor("out_169_cast_fp16")]; + tensor obj_393_gamma_0_to_fp16 = const()[name = tensor("obj_393_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1603197248)))]; + tensor obj_393_beta_0_to_fp16 = const()[name = tensor("obj_393_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1603199872)))]; + tensor obj_393_epsilon_0_to_fp16 = const()[name = tensor("obj_393_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_393_cast_fp16 = batch_norm(beta = obj_393_beta_0_to_fp16, epsilon = obj_393_epsilon_0_to_fp16, gamma = obj_393_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_169_cast_fp16)[name = tensor("obj_393_cast_fp16")]; + tensor query_113_pad_type_0 = const()[name = tensor("query_113_pad_type_0"), val = tensor("valid")]; + tensor query_113_strides_0 = const()[name = tensor("query_113_strides_0"), val = tensor([1, 1])]; + tensor query_113_pad_0 = const()[name = tensor("query_113_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_113_dilations_0 = const()[name = tensor("query_113_dilations_0"), val = tensor([1, 1])]; + tensor query_113_groups_0 = const()[name = tensor("query_113_groups_0"), val = tensor(1)]; + tensor layers_28_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1603202496)))]; + tensor layers_28_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1606479360)))]; + tensor query_113_cast_fp16 = conv(bias = layers_28_self_attn_q_proj_bias_to_fp16, dilations = query_113_dilations_0, groups = query_113_groups_0, pad = query_113_pad_0, pad_type = query_113_pad_type_0, strides = query_113_strides_0, weight = layers_28_self_attn_q_proj_weight_to_fp16, x = obj_393_cast_fp16)[name = tensor("query_113_cast_fp16")]; + tensor current_key_57_pad_type_0 = const()[name = tensor("current_key_57_pad_type_0"), val = tensor("valid")]; + tensor current_key_57_strides_0 = const()[name = tensor("current_key_57_strides_0"), val = tensor([1, 1])]; + tensor current_key_57_pad_0 = const()[name = tensor("current_key_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_57_dilations_0 = const()[name = tensor("current_key_57_dilations_0"), val = tensor([1, 1])]; + tensor current_key_57_groups_0 = const()[name = tensor("current_key_57_groups_0"), val = tensor(1)]; + tensor layers_28_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1606481984)))]; + tensor current_key_57_cast_fp16 = conv(dilations = current_key_57_dilations_0, groups = current_key_57_groups_0, pad = current_key_57_pad_0, pad_type = current_key_57_pad_type_0, strides = current_key_57_strides_0, weight = layers_28_self_attn_k_proj_weight_to_fp16, x = obj_393_cast_fp16)[name = tensor("current_key_57_cast_fp16")]; + tensor current_value_57_pad_type_0 = const()[name = tensor("current_value_57_pad_type_0"), val = tensor("valid")]; + tensor current_value_57_strides_0 = const()[name = tensor("current_value_57_strides_0"), val = tensor([1, 1])]; + tensor current_value_57_pad_0 = const()[name = tensor("current_value_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_57_dilations_0 = const()[name = tensor("current_value_57_dilations_0"), val = tensor([1, 1])]; + tensor current_value_57_groups_0 = const()[name = tensor("current_value_57_groups_0"), val = tensor(1)]; + tensor layers_28_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1609758848)))]; + tensor layers_28_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1613035712)))]; + tensor current_value_57_cast_fp16 = conv(bias = layers_28_self_attn_v_proj_bias_to_fp16, dilations = current_value_57_dilations_0, groups = current_value_57_groups_0, pad = current_value_57_pad_0, pad_type = current_value_57_pad_type_0, strides = current_value_57_strides_0, weight = layers_28_self_attn_v_proj_weight_to_fp16, x = obj_393_cast_fp16)[name = tensor("current_value_57_cast_fp16")]; + tensor var_6412_cast_fp16 = mul(x = var_103_cast_fp16_28, y = var_239_cast_fp16)[name = tensor("op_6412_cast_fp16")]; + tensor var_6413_cast_fp16 = mul(x = current_key_57_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_6413_cast_fp16")]; + tensor key_113_cast_fp16 = add(x = var_6412_cast_fp16, y = var_6413_cast_fp16)[name = tensor("key_113_cast_fp16")]; + tensor var_6416_cast_fp16 = mul(x = var_138_cast_fp16_28, y = var_239_cast_fp16)[name = tensor("op_6416_cast_fp16")]; + tensor var_6417_cast_fp16 = mul(x = current_value_57_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_6417_cast_fp16")]; + tensor value_113_cast_fp16 = add(x = var_6416_cast_fp16, y = var_6417_cast_fp16)[name = tensor("value_113_cast_fp16")]; + tensor var_6421 = const()[name = tensor("op_6421"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_113_cast_fp16 = reshape(shape = var_6421, x = query_113_cast_fp16)[name = tensor("mh_q_113_cast_fp16")]; + tensor var_6423_to_fp16 = const()[name = tensor("op_6423_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6424_cast_fp16 = mul(x = mh_q_113_cast_fp16, y = var_6423_to_fp16)[name = tensor("op_6424_cast_fp16")]; + tensor var_6427 = const()[name = tensor("op_6427"), val = tensor([1, 20, 64, 448])]; + tensor var_6428_cast_fp16 = reshape(shape = var_6427, x = key_113_cast_fp16)[name = tensor("op_6428_cast_fp16")]; + tensor mh_w_169_transpose_x_0 = const()[name = tensor("mh_w_169_transpose_x_0"), val = tensor(true)]; + tensor mh_w_169_transpose_y_0 = const()[name = tensor("mh_w_169_transpose_y_0"), val = tensor(false)]; + tensor mh_w_169_cast_fp16 = matmul(transpose_x = mh_w_169_transpose_x_0, transpose_y = mh_w_169_transpose_y_0, x = var_6424_cast_fp16, y = var_6428_cast_fp16)[name = tensor("mh_w_169_cast_fp16")]; + tensor mh_w_171_cast_fp16 = add(x = mh_w_169_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_171_cast_fp16")]; + tensor var_6436_cast_fp16 = softmax(axis = var_6348, x = mh_w_171_cast_fp16)[name = tensor("op_6436_cast_fp16")]; + tensor var_6437 = const()[name = tensor("op_6437"), val = tensor([1, 20, 64, 448])]; + tensor var_6438_cast_fp16 = reshape(shape = var_6437, x = value_113_cast_fp16)[name = tensor("op_6438_cast_fp16")]; + tensor attn_113_transpose_x_0 = const()[name = tensor("attn_113_transpose_x_0"), val = tensor(false)]; + tensor attn_113_transpose_y_0 = const()[name = tensor("attn_113_transpose_y_0"), val = tensor(true)]; + tensor attn_113_cast_fp16 = matmul(transpose_x = attn_113_transpose_x_0, transpose_y = attn_113_transpose_y_0, x = var_6438_cast_fp16, y = var_6436_cast_fp16)[name = tensor("attn_113_cast_fp16")]; + tensor var_6441 = const()[name = tensor("op_6441"), val = tensor([1, 1280, 1, 1])]; + tensor input_281_cast_fp16 = reshape(shape = var_6441, x = attn_113_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor obj_399_pad_type_0 = const()[name = tensor("obj_399_pad_type_0"), val = tensor("valid")]; + tensor obj_399_strides_0 = const()[name = tensor("obj_399_strides_0"), val = tensor([1, 1])]; + tensor obj_399_pad_0 = const()[name = tensor("obj_399_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_399_dilations_0 = const()[name = tensor("obj_399_dilations_0"), val = tensor([1, 1])]; + tensor obj_399_groups_0 = const()[name = tensor("obj_399_groups_0"), val = tensor(1)]; + tensor layers_28_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1613038336)))]; + tensor layers_28_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1616315200)))]; + tensor obj_399_cast_fp16 = conv(bias = layers_28_self_attn_o_proj_bias_to_fp16, dilations = obj_399_dilations_0, groups = obj_399_groups_0, pad = obj_399_pad_0, pad_type = obj_399_pad_type_0, strides = obj_399_strides_0, weight = layers_28_self_attn_o_proj_weight_to_fp16, x = input_281_cast_fp16)[name = tensor("obj_399_cast_fp16")]; + tensor inputs_171_cast_fp16 = add(x = inputs_169_cast_fp16, y = obj_399_cast_fp16)[name = tensor("inputs_171_cast_fp16")]; + tensor out_171_axes_0 = const()[name = tensor("out_171_axes_0"), val = tensor([1])]; + tensor var_6463_to_fp16 = const()[name = tensor("op_6463_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_171_cast_fp16 = layer_norm(axes = out_171_axes_0, epsilon = var_6463_to_fp16, x = inputs_171_cast_fp16)[name = tensor("out_171_cast_fp16")]; + tensor obj_401_gamma_0_to_fp16 = const()[name = tensor("obj_401_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1616317824)))]; + tensor obj_401_beta_0_to_fp16 = const()[name = tensor("obj_401_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1616320448)))]; + tensor obj_401_epsilon_0_to_fp16 = const()[name = tensor("obj_401_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_401_cast_fp16 = batch_norm(beta = obj_401_beta_0_to_fp16, epsilon = obj_401_epsilon_0_to_fp16, gamma = obj_401_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_171_cast_fp16)[name = tensor("obj_401_cast_fp16")]; + tensor query_115_pad_type_0 = const()[name = tensor("query_115_pad_type_0"), val = tensor("valid")]; + tensor query_115_strides_0 = const()[name = tensor("query_115_strides_0"), val = tensor([1, 1])]; + tensor query_115_pad_0 = const()[name = tensor("query_115_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_115_dilations_0 = const()[name = tensor("query_115_dilations_0"), val = tensor([1, 1])]; + tensor query_115_groups_0 = const()[name = tensor("query_115_groups_0"), val = tensor(1)]; + tensor layers_28_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_28_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1616323072)))]; + tensor layers_28_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_28_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1619599936)))]; + tensor query_115_cast_fp16 = conv(bias = layers_28_encoder_attn_q_proj_bias_to_fp16, dilations = query_115_dilations_0, groups = query_115_groups_0, pad = query_115_pad_0, pad_type = query_115_pad_type_0, strides = query_115_strides_0, weight = layers_28_encoder_attn_q_proj_weight_to_fp16, x = obj_401_cast_fp16)[name = tensor("query_115_cast_fp16")]; + tensor key_115_pad_type_0 = const()[name = tensor("key_115_pad_type_0"), val = tensor("valid")]; + tensor key_115_strides_0 = const()[name = tensor("key_115_strides_0"), val = tensor([1, 1])]; + tensor key_115_pad_0 = const()[name = tensor("key_115_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_115_dilations_0 = const()[name = tensor("key_115_dilations_0"), val = tensor([1, 1])]; + tensor key_115_groups_0 = const()[name = tensor("key_115_groups_0"), val = tensor(1)]; + tensor layers_28_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_28_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1619602560)))]; + tensor key_115_cast_fp16 = conv(dilations = key_115_dilations_0, groups = key_115_groups_0, pad = key_115_pad_0, pad_type = key_115_pad_type_0, strides = key_115_strides_0, weight = layers_28_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_115_cast_fp16")]; + tensor value_115_pad_type_0 = const()[name = tensor("value_115_pad_type_0"), val = tensor("valid")]; + tensor value_115_strides_0 = const()[name = tensor("value_115_strides_0"), val = tensor([1, 1])]; + tensor value_115_pad_0 = const()[name = tensor("value_115_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_115_dilations_0 = const()[name = tensor("value_115_dilations_0"), val = tensor([1, 1])]; + tensor value_115_groups_0 = const()[name = tensor("value_115_groups_0"), val = tensor(1)]; + tensor layers_28_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_28_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1622879424)))]; + tensor layers_28_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_28_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1626156288)))]; + tensor value_115_cast_fp16 = conv(bias = layers_28_encoder_attn_v_proj_bias_to_fp16, dilations = value_115_dilations_0, groups = value_115_groups_0, pad = value_115_pad_0, pad_type = value_115_pad_type_0, strides = value_115_strides_0, weight = layers_28_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_115_cast_fp16")]; + tensor var_6499 = const()[name = tensor("op_6499"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_115_cast_fp16 = reshape(shape = var_6499, x = query_115_cast_fp16)[name = tensor("mh_q_115_cast_fp16")]; + tensor var_6501_to_fp16 = const()[name = tensor("op_6501_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6502_cast_fp16 = mul(x = mh_q_115_cast_fp16, y = var_6501_to_fp16)[name = tensor("op_6502_cast_fp16")]; + tensor var_6505 = const()[name = tensor("op_6505"), val = tensor([1, 20, 64, 1500])]; + tensor var_6506_cast_fp16 = reshape(shape = var_6505, x = key_115_cast_fp16)[name = tensor("op_6506_cast_fp16")]; + tensor mh_w_173_transpose_x_0 = const()[name = tensor("mh_w_173_transpose_x_0"), val = tensor(true)]; + tensor mh_w_173_transpose_y_0 = const()[name = tensor("mh_w_173_transpose_y_0"), val = tensor(false)]; + tensor mh_w_173_cast_fp16 = matmul(transpose_x = mh_w_173_transpose_x_0, transpose_y = mh_w_173_transpose_y_0, x = var_6502_cast_fp16, y = var_6506_cast_fp16)[name = tensor("mh_w_173_cast_fp16")]; + tensor obj_405_cast_fp16 = softmax(axis = var_6348, x = mh_w_173_cast_fp16)[name = tensor("obj_405_cast_fp16")]; + tensor var_6510 = const()[name = tensor("op_6510"), val = tensor([1, 20, 64, 1500])]; + tensor var_6511_cast_fp16 = reshape(shape = var_6510, x = value_115_cast_fp16)[name = tensor("op_6511_cast_fp16")]; + tensor attn_115_transpose_x_0 = const()[name = tensor("attn_115_transpose_x_0"), val = tensor(false)]; + tensor attn_115_transpose_y_0 = const()[name = tensor("attn_115_transpose_y_0"), val = tensor(true)]; + tensor attn_115_cast_fp16 = matmul(transpose_x = attn_115_transpose_x_0, transpose_y = attn_115_transpose_y_0, x = var_6511_cast_fp16, y = obj_405_cast_fp16)[name = tensor("attn_115_cast_fp16")]; + tensor var_6514 = const()[name = tensor("op_6514"), val = tensor([1, 1280, 1, 1])]; + tensor input_283_cast_fp16 = reshape(shape = var_6514, x = attn_115_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor obj_403_pad_type_0 = const()[name = tensor("obj_403_pad_type_0"), val = tensor("valid")]; + tensor obj_403_strides_0 = const()[name = tensor("obj_403_strides_0"), val = tensor([1, 1])]; + tensor obj_403_pad_0 = const()[name = tensor("obj_403_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_403_dilations_0 = const()[name = tensor("obj_403_dilations_0"), val = tensor([1, 1])]; + tensor obj_403_groups_0 = const()[name = tensor("obj_403_groups_0"), val = tensor(1)]; + tensor layers_28_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_28_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1626158912)))]; + tensor layers_28_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_28_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1629435776)))]; + tensor obj_403_cast_fp16 = conv(bias = layers_28_encoder_attn_o_proj_bias_to_fp16, dilations = obj_403_dilations_0, groups = obj_403_groups_0, pad = obj_403_pad_0, pad_type = obj_403_pad_type_0, strides = obj_403_strides_0, weight = layers_28_encoder_attn_o_proj_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("obj_403_cast_fp16")]; + tensor inputs_173_cast_fp16 = add(x = inputs_171_cast_fp16, y = obj_403_cast_fp16)[name = tensor("inputs_173_cast_fp16")]; + tensor out_173_axes_0 = const()[name = tensor("out_173_axes_0"), val = tensor([1])]; + tensor var_6532_to_fp16 = const()[name = tensor("op_6532_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_173_cast_fp16 = layer_norm(axes = out_173_axes_0, epsilon = var_6532_to_fp16, x = inputs_173_cast_fp16)[name = tensor("out_173_cast_fp16")]; + tensor input_285_gamma_0_to_fp16 = const()[name = tensor("input_285_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1629438400)))]; + tensor input_285_beta_0_to_fp16 = const()[name = tensor("input_285_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1629441024)))]; + tensor input_285_epsilon_0_to_fp16 = const()[name = tensor("input_285_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_285_cast_fp16 = batch_norm(beta = input_285_beta_0_to_fp16, epsilon = input_285_epsilon_0_to_fp16, gamma = input_285_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_173_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_pad_type_0 = const()[name = tensor("input_287_pad_type_0"), val = tensor("valid")]; + tensor input_287_strides_0 = const()[name = tensor("input_287_strides_0"), val = tensor([1, 1])]; + tensor input_287_pad_0 = const()[name = tensor("input_287_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_287_dilations_0 = const()[name = tensor("input_287_dilations_0"), val = tensor([1, 1])]; + tensor input_287_groups_0 = const()[name = tensor("input_287_groups_0"), val = tensor(1)]; + tensor layers_28_fc1_weight_to_fp16 = const()[name = tensor("layers_28_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1629443648)))]; + tensor layers_28_fc1_bias_to_fp16 = const()[name = tensor("layers_28_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1642550912)))]; + tensor input_287_cast_fp16 = conv(bias = layers_28_fc1_bias_to_fp16, dilations = input_287_dilations_0, groups = input_287_groups_0, pad = input_287_pad_0, pad_type = input_287_pad_type_0, strides = input_287_strides_0, weight = layers_28_fc1_weight_to_fp16, x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor input_289_mode_0 = const()[name = tensor("input_289_mode_0"), val = tensor("EXACT")]; + tensor input_289_cast_fp16 = gelu(mode = input_289_mode_0, x = input_287_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor hidden_states_59_pad_type_0 = const()[name = tensor("hidden_states_59_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_59_strides_0 = const()[name = tensor("hidden_states_59_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_59_pad_0 = const()[name = tensor("hidden_states_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_59_dilations_0 = const()[name = tensor("hidden_states_59_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_59_groups_0 = const()[name = tensor("hidden_states_59_groups_0"), val = tensor(1)]; + tensor layers_28_fc2_weight_to_fp16 = const()[name = tensor("layers_28_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1642561216)))]; + tensor layers_28_fc2_bias_to_fp16 = const()[name = tensor("layers_28_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1655668480)))]; + tensor hidden_states_59_cast_fp16 = conv(bias = layers_28_fc2_bias_to_fp16, dilations = hidden_states_59_dilations_0, groups = hidden_states_59_groups_0, pad = hidden_states_59_pad_0, pad_type = hidden_states_59_pad_type_0, strides = hidden_states_59_strides_0, weight = layers_28_fc2_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("hidden_states_59_cast_fp16")]; + tensor inputs_175_cast_fp16 = add(x = inputs_173_cast_fp16, y = hidden_states_59_cast_fp16)[name = tensor("inputs_175_cast_fp16")]; + tensor var_6567 = const()[name = tensor("op_6567"), val = tensor(3)]; + tensor out_175_axes_0 = const()[name = tensor("out_175_axes_0"), val = tensor([1])]; + tensor var_6592_to_fp16 = const()[name = tensor("op_6592_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_175_cast_fp16 = layer_norm(axes = out_175_axes_0, epsilon = var_6592_to_fp16, x = inputs_175_cast_fp16)[name = tensor("out_175_cast_fp16")]; + tensor obj_407_gamma_0_to_fp16 = const()[name = tensor("obj_407_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1655671104)))]; + tensor obj_407_beta_0_to_fp16 = const()[name = tensor("obj_407_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1655673728)))]; + tensor obj_407_epsilon_0_to_fp16 = const()[name = tensor("obj_407_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_407_cast_fp16 = batch_norm(beta = obj_407_beta_0_to_fp16, epsilon = obj_407_epsilon_0_to_fp16, gamma = obj_407_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_175_cast_fp16)[name = tensor("obj_407_cast_fp16")]; + tensor query_117_pad_type_0 = const()[name = tensor("query_117_pad_type_0"), val = tensor("valid")]; + tensor query_117_strides_0 = const()[name = tensor("query_117_strides_0"), val = tensor([1, 1])]; + tensor query_117_pad_0 = const()[name = tensor("query_117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_117_dilations_0 = const()[name = tensor("query_117_dilations_0"), val = tensor([1, 1])]; + tensor query_117_groups_0 = const()[name = tensor("query_117_groups_0"), val = tensor(1)]; + tensor layers_29_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1655676352)))]; + tensor layers_29_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1658953216)))]; + tensor query_117_cast_fp16 = conv(bias = layers_29_self_attn_q_proj_bias_to_fp16, dilations = query_117_dilations_0, groups = query_117_groups_0, pad = query_117_pad_0, pad_type = query_117_pad_type_0, strides = query_117_strides_0, weight = layers_29_self_attn_q_proj_weight_to_fp16, x = obj_407_cast_fp16)[name = tensor("query_117_cast_fp16")]; + tensor current_key_59_pad_type_0 = const()[name = tensor("current_key_59_pad_type_0"), val = tensor("valid")]; + tensor current_key_59_strides_0 = const()[name = tensor("current_key_59_strides_0"), val = tensor([1, 1])]; + tensor current_key_59_pad_0 = const()[name = tensor("current_key_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_59_dilations_0 = const()[name = tensor("current_key_59_dilations_0"), val = tensor([1, 1])]; + tensor current_key_59_groups_0 = const()[name = tensor("current_key_59_groups_0"), val = tensor(1)]; + tensor layers_29_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1658955840)))]; + tensor current_key_59_cast_fp16 = conv(dilations = current_key_59_dilations_0, groups = current_key_59_groups_0, pad = current_key_59_pad_0, pad_type = current_key_59_pad_type_0, strides = current_key_59_strides_0, weight = layers_29_self_attn_k_proj_weight_to_fp16, x = obj_407_cast_fp16)[name = tensor("current_key_59_cast_fp16")]; + tensor current_value_59_pad_type_0 = const()[name = tensor("current_value_59_pad_type_0"), val = tensor("valid")]; + tensor current_value_59_strides_0 = const()[name = tensor("current_value_59_strides_0"), val = tensor([1, 1])]; + tensor current_value_59_pad_0 = const()[name = tensor("current_value_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_59_dilations_0 = const()[name = tensor("current_value_59_dilations_0"), val = tensor([1, 1])]; + tensor current_value_59_groups_0 = const()[name = tensor("current_value_59_groups_0"), val = tensor(1)]; + tensor layers_29_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1662232704)))]; + tensor layers_29_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1665509568)))]; + tensor current_value_59_cast_fp16 = conv(bias = layers_29_self_attn_v_proj_bias_to_fp16, dilations = current_value_59_dilations_0, groups = current_value_59_groups_0, pad = current_value_59_pad_0, pad_type = current_value_59_pad_type_0, strides = current_value_59_strides_0, weight = layers_29_self_attn_v_proj_weight_to_fp16, x = obj_407_cast_fp16)[name = tensor("current_value_59_cast_fp16")]; + tensor var_6631_cast_fp16 = mul(x = var_103_cast_fp16_29, y = var_239_cast_fp16)[name = tensor("op_6631_cast_fp16")]; + tensor var_6632_cast_fp16 = mul(x = current_key_59_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_6632_cast_fp16")]; + tensor key_117_cast_fp16 = add(x = var_6631_cast_fp16, y = var_6632_cast_fp16)[name = tensor("key_117_cast_fp16")]; + tensor var_6635_cast_fp16 = mul(x = var_138_cast_fp16_29, y = var_239_cast_fp16)[name = tensor("op_6635_cast_fp16")]; + tensor var_6636_cast_fp16 = mul(x = current_value_59_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_6636_cast_fp16")]; + tensor value_117_cast_fp16 = add(x = var_6635_cast_fp16, y = var_6636_cast_fp16)[name = tensor("value_117_cast_fp16")]; + tensor var_6640 = const()[name = tensor("op_6640"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_117_cast_fp16 = reshape(shape = var_6640, x = query_117_cast_fp16)[name = tensor("mh_q_117_cast_fp16")]; + tensor var_6642_to_fp16 = const()[name = tensor("op_6642_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6643_cast_fp16 = mul(x = mh_q_117_cast_fp16, y = var_6642_to_fp16)[name = tensor("op_6643_cast_fp16")]; + tensor var_6646 = const()[name = tensor("op_6646"), val = tensor([1, 20, 64, 448])]; + tensor var_6647_cast_fp16 = reshape(shape = var_6646, x = key_117_cast_fp16)[name = tensor("op_6647_cast_fp16")]; + tensor mh_w_175_transpose_x_0 = const()[name = tensor("mh_w_175_transpose_x_0"), val = tensor(true)]; + tensor mh_w_175_transpose_y_0 = const()[name = tensor("mh_w_175_transpose_y_0"), val = tensor(false)]; + tensor mh_w_175_cast_fp16 = matmul(transpose_x = mh_w_175_transpose_x_0, transpose_y = mh_w_175_transpose_y_0, x = var_6643_cast_fp16, y = var_6647_cast_fp16)[name = tensor("mh_w_175_cast_fp16")]; + tensor mh_w_177_cast_fp16 = add(x = mh_w_175_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_177_cast_fp16")]; + tensor var_6655_cast_fp16 = softmax(axis = var_6567, x = mh_w_177_cast_fp16)[name = tensor("op_6655_cast_fp16")]; + tensor var_6656 = const()[name = tensor("op_6656"), val = tensor([1, 20, 64, 448])]; + tensor var_6657_cast_fp16 = reshape(shape = var_6656, x = value_117_cast_fp16)[name = tensor("op_6657_cast_fp16")]; + tensor attn_117_transpose_x_0 = const()[name = tensor("attn_117_transpose_x_0"), val = tensor(false)]; + tensor attn_117_transpose_y_0 = const()[name = tensor("attn_117_transpose_y_0"), val = tensor(true)]; + tensor attn_117_cast_fp16 = matmul(transpose_x = attn_117_transpose_x_0, transpose_y = attn_117_transpose_y_0, x = var_6657_cast_fp16, y = var_6655_cast_fp16)[name = tensor("attn_117_cast_fp16")]; + tensor var_6660 = const()[name = tensor("op_6660"), val = tensor([1, 1280, 1, 1])]; + tensor input_291_cast_fp16 = reshape(shape = var_6660, x = attn_117_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor obj_413_pad_type_0 = const()[name = tensor("obj_413_pad_type_0"), val = tensor("valid")]; + tensor obj_413_strides_0 = const()[name = tensor("obj_413_strides_0"), val = tensor([1, 1])]; + tensor obj_413_pad_0 = const()[name = tensor("obj_413_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_413_dilations_0 = const()[name = tensor("obj_413_dilations_0"), val = tensor([1, 1])]; + tensor obj_413_groups_0 = const()[name = tensor("obj_413_groups_0"), val = tensor(1)]; + tensor layers_29_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1665512192)))]; + tensor layers_29_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1668789056)))]; + tensor obj_413_cast_fp16 = conv(bias = layers_29_self_attn_o_proj_bias_to_fp16, dilations = obj_413_dilations_0, groups = obj_413_groups_0, pad = obj_413_pad_0, pad_type = obj_413_pad_type_0, strides = obj_413_strides_0, weight = layers_29_self_attn_o_proj_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("obj_413_cast_fp16")]; + tensor inputs_177_cast_fp16 = add(x = inputs_175_cast_fp16, y = obj_413_cast_fp16)[name = tensor("inputs_177_cast_fp16")]; + tensor out_177_axes_0 = const()[name = tensor("out_177_axes_0"), val = tensor([1])]; + tensor var_6682_to_fp16 = const()[name = tensor("op_6682_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_177_cast_fp16 = layer_norm(axes = out_177_axes_0, epsilon = var_6682_to_fp16, x = inputs_177_cast_fp16)[name = tensor("out_177_cast_fp16")]; + tensor obj_415_gamma_0_to_fp16 = const()[name = tensor("obj_415_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1668791680)))]; + tensor obj_415_beta_0_to_fp16 = const()[name = tensor("obj_415_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1668794304)))]; + tensor obj_415_epsilon_0_to_fp16 = const()[name = tensor("obj_415_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_415_cast_fp16 = batch_norm(beta = obj_415_beta_0_to_fp16, epsilon = obj_415_epsilon_0_to_fp16, gamma = obj_415_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_177_cast_fp16)[name = tensor("obj_415_cast_fp16")]; + tensor query_119_pad_type_0 = const()[name = tensor("query_119_pad_type_0"), val = tensor("valid")]; + tensor query_119_strides_0 = const()[name = tensor("query_119_strides_0"), val = tensor([1, 1])]; + tensor query_119_pad_0 = const()[name = tensor("query_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_119_dilations_0 = const()[name = tensor("query_119_dilations_0"), val = tensor([1, 1])]; + tensor query_119_groups_0 = const()[name = tensor("query_119_groups_0"), val = tensor(1)]; + tensor layers_29_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_29_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1668796928)))]; + tensor layers_29_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_29_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1672073792)))]; + tensor query_119_cast_fp16 = conv(bias = layers_29_encoder_attn_q_proj_bias_to_fp16, dilations = query_119_dilations_0, groups = query_119_groups_0, pad = query_119_pad_0, pad_type = query_119_pad_type_0, strides = query_119_strides_0, weight = layers_29_encoder_attn_q_proj_weight_to_fp16, x = obj_415_cast_fp16)[name = tensor("query_119_cast_fp16")]; + tensor key_119_pad_type_0 = const()[name = tensor("key_119_pad_type_0"), val = tensor("valid")]; + tensor key_119_strides_0 = const()[name = tensor("key_119_strides_0"), val = tensor([1, 1])]; + tensor key_119_pad_0 = const()[name = tensor("key_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_119_dilations_0 = const()[name = tensor("key_119_dilations_0"), val = tensor([1, 1])]; + tensor key_119_groups_0 = const()[name = tensor("key_119_groups_0"), val = tensor(1)]; + tensor layers_29_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_29_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1672076416)))]; + tensor key_119_cast_fp16 = conv(dilations = key_119_dilations_0, groups = key_119_groups_0, pad = key_119_pad_0, pad_type = key_119_pad_type_0, strides = key_119_strides_0, weight = layers_29_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_119_cast_fp16")]; + tensor value_119_pad_type_0 = const()[name = tensor("value_119_pad_type_0"), val = tensor("valid")]; + tensor value_119_strides_0 = const()[name = tensor("value_119_strides_0"), val = tensor([1, 1])]; + tensor value_119_pad_0 = const()[name = tensor("value_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_119_dilations_0 = const()[name = tensor("value_119_dilations_0"), val = tensor([1, 1])]; + tensor value_119_groups_0 = const()[name = tensor("value_119_groups_0"), val = tensor(1)]; + tensor layers_29_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_29_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1675353280)))]; + tensor layers_29_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_29_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1678630144)))]; + tensor value_119_cast_fp16 = conv(bias = layers_29_encoder_attn_v_proj_bias_to_fp16, dilations = value_119_dilations_0, groups = value_119_groups_0, pad = value_119_pad_0, pad_type = value_119_pad_type_0, strides = value_119_strides_0, weight = layers_29_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_119_cast_fp16")]; + tensor var_6718 = const()[name = tensor("op_6718"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_119_cast_fp16 = reshape(shape = var_6718, x = query_119_cast_fp16)[name = tensor("mh_q_119_cast_fp16")]; + tensor var_6720_to_fp16 = const()[name = tensor("op_6720_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6721_cast_fp16 = mul(x = mh_q_119_cast_fp16, y = var_6720_to_fp16)[name = tensor("op_6721_cast_fp16")]; + tensor var_6724 = const()[name = tensor("op_6724"), val = tensor([1, 20, 64, 1500])]; + tensor var_6725_cast_fp16 = reshape(shape = var_6724, x = key_119_cast_fp16)[name = tensor("op_6725_cast_fp16")]; + tensor mh_w_179_transpose_x_0 = const()[name = tensor("mh_w_179_transpose_x_0"), val = tensor(true)]; + tensor mh_w_179_transpose_y_0 = const()[name = tensor("mh_w_179_transpose_y_0"), val = tensor(false)]; + tensor mh_w_179_cast_fp16 = matmul(transpose_x = mh_w_179_transpose_x_0, transpose_y = mh_w_179_transpose_y_0, x = var_6721_cast_fp16, y = var_6725_cast_fp16)[name = tensor("mh_w_179_cast_fp16")]; + tensor obj_419_cast_fp16 = softmax(axis = var_6567, x = mh_w_179_cast_fp16)[name = tensor("obj_419_cast_fp16")]; + tensor var_6729 = const()[name = tensor("op_6729"), val = tensor([1, 20, 64, 1500])]; + tensor var_6730_cast_fp16 = reshape(shape = var_6729, x = value_119_cast_fp16)[name = tensor("op_6730_cast_fp16")]; + tensor attn_119_transpose_x_0 = const()[name = tensor("attn_119_transpose_x_0"), val = tensor(false)]; + tensor attn_119_transpose_y_0 = const()[name = tensor("attn_119_transpose_y_0"), val = tensor(true)]; + tensor attn_119_cast_fp16 = matmul(transpose_x = attn_119_transpose_x_0, transpose_y = attn_119_transpose_y_0, x = var_6730_cast_fp16, y = obj_419_cast_fp16)[name = tensor("attn_119_cast_fp16")]; + tensor var_6733 = const()[name = tensor("op_6733"), val = tensor([1, 1280, 1, 1])]; + tensor input_293_cast_fp16 = reshape(shape = var_6733, x = attn_119_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor obj_417_pad_type_0 = const()[name = tensor("obj_417_pad_type_0"), val = tensor("valid")]; + tensor obj_417_strides_0 = const()[name = tensor("obj_417_strides_0"), val = tensor([1, 1])]; + tensor obj_417_pad_0 = const()[name = tensor("obj_417_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_417_dilations_0 = const()[name = tensor("obj_417_dilations_0"), val = tensor([1, 1])]; + tensor obj_417_groups_0 = const()[name = tensor("obj_417_groups_0"), val = tensor(1)]; + tensor layers_29_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_29_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1678632768)))]; + tensor layers_29_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_29_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1681909632)))]; + tensor obj_417_cast_fp16 = conv(bias = layers_29_encoder_attn_o_proj_bias_to_fp16, dilations = obj_417_dilations_0, groups = obj_417_groups_0, pad = obj_417_pad_0, pad_type = obj_417_pad_type_0, strides = obj_417_strides_0, weight = layers_29_encoder_attn_o_proj_weight_to_fp16, x = input_293_cast_fp16)[name = tensor("obj_417_cast_fp16")]; + tensor inputs_179_cast_fp16 = add(x = inputs_177_cast_fp16, y = obj_417_cast_fp16)[name = tensor("inputs_179_cast_fp16")]; + tensor out_179_axes_0 = const()[name = tensor("out_179_axes_0"), val = tensor([1])]; + tensor var_6751_to_fp16 = const()[name = tensor("op_6751_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_179_cast_fp16 = layer_norm(axes = out_179_axes_0, epsilon = var_6751_to_fp16, x = inputs_179_cast_fp16)[name = tensor("out_179_cast_fp16")]; + tensor input_295_gamma_0_to_fp16 = const()[name = tensor("input_295_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1681912256)))]; + tensor input_295_beta_0_to_fp16 = const()[name = tensor("input_295_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1681914880)))]; + tensor input_295_epsilon_0_to_fp16 = const()[name = tensor("input_295_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_295_cast_fp16 = batch_norm(beta = input_295_beta_0_to_fp16, epsilon = input_295_epsilon_0_to_fp16, gamma = input_295_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_179_cast_fp16)[name = tensor("input_295_cast_fp16")]; + tensor input_297_pad_type_0 = const()[name = tensor("input_297_pad_type_0"), val = tensor("valid")]; + tensor input_297_strides_0 = const()[name = tensor("input_297_strides_0"), val = tensor([1, 1])]; + tensor input_297_pad_0 = const()[name = tensor("input_297_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_297_dilations_0 = const()[name = tensor("input_297_dilations_0"), val = tensor([1, 1])]; + tensor input_297_groups_0 = const()[name = tensor("input_297_groups_0"), val = tensor(1)]; + tensor layers_29_fc1_weight_to_fp16 = const()[name = tensor("layers_29_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1681917504)))]; + tensor layers_29_fc1_bias_to_fp16 = const()[name = tensor("layers_29_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1695024768)))]; + tensor input_297_cast_fp16 = conv(bias = layers_29_fc1_bias_to_fp16, dilations = input_297_dilations_0, groups = input_297_groups_0, pad = input_297_pad_0, pad_type = input_297_pad_type_0, strides = input_297_strides_0, weight = layers_29_fc1_weight_to_fp16, x = input_295_cast_fp16)[name = tensor("input_297_cast_fp16")]; + tensor input_299_mode_0 = const()[name = tensor("input_299_mode_0"), val = tensor("EXACT")]; + tensor input_299_cast_fp16 = gelu(mode = input_299_mode_0, x = input_297_cast_fp16)[name = tensor("input_299_cast_fp16")]; + tensor hidden_states_61_pad_type_0 = const()[name = tensor("hidden_states_61_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_61_strides_0 = const()[name = tensor("hidden_states_61_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_61_pad_0 = const()[name = tensor("hidden_states_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_61_dilations_0 = const()[name = tensor("hidden_states_61_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_61_groups_0 = const()[name = tensor("hidden_states_61_groups_0"), val = tensor(1)]; + tensor layers_29_fc2_weight_to_fp16 = const()[name = tensor("layers_29_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1695035072)))]; + tensor layers_29_fc2_bias_to_fp16 = const()[name = tensor("layers_29_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1708142336)))]; + tensor hidden_states_61_cast_fp16 = conv(bias = layers_29_fc2_bias_to_fp16, dilations = hidden_states_61_dilations_0, groups = hidden_states_61_groups_0, pad = hidden_states_61_pad_0, pad_type = hidden_states_61_pad_type_0, strides = hidden_states_61_strides_0, weight = layers_29_fc2_weight_to_fp16, x = input_299_cast_fp16)[name = tensor("hidden_states_61_cast_fp16")]; + tensor inputs_181_cast_fp16 = add(x = inputs_179_cast_fp16, y = hidden_states_61_cast_fp16)[name = tensor("inputs_181_cast_fp16")]; + tensor var_6786 = const()[name = tensor("op_6786"), val = tensor(3)]; + tensor out_181_axes_0 = const()[name = tensor("out_181_axes_0"), val = tensor([1])]; + tensor var_6811_to_fp16 = const()[name = tensor("op_6811_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_181_cast_fp16 = layer_norm(axes = out_181_axes_0, epsilon = var_6811_to_fp16, x = inputs_181_cast_fp16)[name = tensor("out_181_cast_fp16")]; + tensor obj_421_gamma_0_to_fp16 = const()[name = tensor("obj_421_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1708144960)))]; + tensor obj_421_beta_0_to_fp16 = const()[name = tensor("obj_421_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1708147584)))]; + tensor obj_421_epsilon_0_to_fp16 = const()[name = tensor("obj_421_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_421_cast_fp16 = batch_norm(beta = obj_421_beta_0_to_fp16, epsilon = obj_421_epsilon_0_to_fp16, gamma = obj_421_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_181_cast_fp16)[name = tensor("obj_421_cast_fp16")]; + tensor query_121_pad_type_0 = const()[name = tensor("query_121_pad_type_0"), val = tensor("valid")]; + tensor query_121_strides_0 = const()[name = tensor("query_121_strides_0"), val = tensor([1, 1])]; + tensor query_121_pad_0 = const()[name = tensor("query_121_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_121_dilations_0 = const()[name = tensor("query_121_dilations_0"), val = tensor([1, 1])]; + tensor query_121_groups_0 = const()[name = tensor("query_121_groups_0"), val = tensor(1)]; + tensor layers_30_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1708150208)))]; + tensor layers_30_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1711427072)))]; + tensor query_121_cast_fp16 = conv(bias = layers_30_self_attn_q_proj_bias_to_fp16, dilations = query_121_dilations_0, groups = query_121_groups_0, pad = query_121_pad_0, pad_type = query_121_pad_type_0, strides = query_121_strides_0, weight = layers_30_self_attn_q_proj_weight_to_fp16, x = obj_421_cast_fp16)[name = tensor("query_121_cast_fp16")]; + tensor current_key_61_pad_type_0 = const()[name = tensor("current_key_61_pad_type_0"), val = tensor("valid")]; + tensor current_key_61_strides_0 = const()[name = tensor("current_key_61_strides_0"), val = tensor([1, 1])]; + tensor current_key_61_pad_0 = const()[name = tensor("current_key_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_61_dilations_0 = const()[name = tensor("current_key_61_dilations_0"), val = tensor([1, 1])]; + tensor current_key_61_groups_0 = const()[name = tensor("current_key_61_groups_0"), val = tensor(1)]; + tensor layers_30_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1711429696)))]; + tensor current_key_61_cast_fp16 = conv(dilations = current_key_61_dilations_0, groups = current_key_61_groups_0, pad = current_key_61_pad_0, pad_type = current_key_61_pad_type_0, strides = current_key_61_strides_0, weight = layers_30_self_attn_k_proj_weight_to_fp16, x = obj_421_cast_fp16)[name = tensor("current_key_61_cast_fp16")]; + tensor current_value_61_pad_type_0 = const()[name = tensor("current_value_61_pad_type_0"), val = tensor("valid")]; + tensor current_value_61_strides_0 = const()[name = tensor("current_value_61_strides_0"), val = tensor([1, 1])]; + tensor current_value_61_pad_0 = const()[name = tensor("current_value_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_61_dilations_0 = const()[name = tensor("current_value_61_dilations_0"), val = tensor([1, 1])]; + tensor current_value_61_groups_0 = const()[name = tensor("current_value_61_groups_0"), val = tensor(1)]; + tensor layers_30_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1714706560)))]; + tensor layers_30_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1717983424)))]; + tensor current_value_61_cast_fp16 = conv(bias = layers_30_self_attn_v_proj_bias_to_fp16, dilations = current_value_61_dilations_0, groups = current_value_61_groups_0, pad = current_value_61_pad_0, pad_type = current_value_61_pad_type_0, strides = current_value_61_strides_0, weight = layers_30_self_attn_v_proj_weight_to_fp16, x = obj_421_cast_fp16)[name = tensor("current_value_61_cast_fp16")]; + tensor var_6850_cast_fp16 = mul(x = var_103_cast_fp16_30, y = var_239_cast_fp16)[name = tensor("op_6850_cast_fp16")]; + tensor var_6851_cast_fp16 = mul(x = current_key_61_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_6851_cast_fp16")]; + tensor key_121_cast_fp16 = add(x = var_6850_cast_fp16, y = var_6851_cast_fp16)[name = tensor("key_121_cast_fp16")]; + tensor var_6854_cast_fp16 = mul(x = var_138_cast_fp16_30, y = var_239_cast_fp16)[name = tensor("op_6854_cast_fp16")]; + tensor var_6855_cast_fp16 = mul(x = current_value_61_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_6855_cast_fp16")]; + tensor value_121_cast_fp16 = add(x = var_6854_cast_fp16, y = var_6855_cast_fp16)[name = tensor("value_121_cast_fp16")]; + tensor var_6859 = const()[name = tensor("op_6859"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_121_cast_fp16 = reshape(shape = var_6859, x = query_121_cast_fp16)[name = tensor("mh_q_121_cast_fp16")]; + tensor var_6861_to_fp16 = const()[name = tensor("op_6861_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6862_cast_fp16 = mul(x = mh_q_121_cast_fp16, y = var_6861_to_fp16)[name = tensor("op_6862_cast_fp16")]; + tensor var_6865 = const()[name = tensor("op_6865"), val = tensor([1, 20, 64, 448])]; + tensor var_6866_cast_fp16 = reshape(shape = var_6865, x = key_121_cast_fp16)[name = tensor("op_6866_cast_fp16")]; + tensor mh_w_181_transpose_x_0 = const()[name = tensor("mh_w_181_transpose_x_0"), val = tensor(true)]; + tensor mh_w_181_transpose_y_0 = const()[name = tensor("mh_w_181_transpose_y_0"), val = tensor(false)]; + tensor mh_w_181_cast_fp16 = matmul(transpose_x = mh_w_181_transpose_x_0, transpose_y = mh_w_181_transpose_y_0, x = var_6862_cast_fp16, y = var_6866_cast_fp16)[name = tensor("mh_w_181_cast_fp16")]; + tensor mh_w_183_cast_fp16 = add(x = mh_w_181_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_183_cast_fp16")]; + tensor var_6874_cast_fp16 = softmax(axis = var_6786, x = mh_w_183_cast_fp16)[name = tensor("op_6874_cast_fp16")]; + tensor var_6875 = const()[name = tensor("op_6875"), val = tensor([1, 20, 64, 448])]; + tensor var_6876_cast_fp16 = reshape(shape = var_6875, x = value_121_cast_fp16)[name = tensor("op_6876_cast_fp16")]; + tensor attn_121_transpose_x_0 = const()[name = tensor("attn_121_transpose_x_0"), val = tensor(false)]; + tensor attn_121_transpose_y_0 = const()[name = tensor("attn_121_transpose_y_0"), val = tensor(true)]; + tensor attn_121_cast_fp16 = matmul(transpose_x = attn_121_transpose_x_0, transpose_y = attn_121_transpose_y_0, x = var_6876_cast_fp16, y = var_6874_cast_fp16)[name = tensor("attn_121_cast_fp16")]; + tensor var_6879 = const()[name = tensor("op_6879"), val = tensor([1, 1280, 1, 1])]; + tensor input_301_cast_fp16 = reshape(shape = var_6879, x = attn_121_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor obj_427_pad_type_0 = const()[name = tensor("obj_427_pad_type_0"), val = tensor("valid")]; + tensor obj_427_strides_0 = const()[name = tensor("obj_427_strides_0"), val = tensor([1, 1])]; + tensor obj_427_pad_0 = const()[name = tensor("obj_427_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_427_dilations_0 = const()[name = tensor("obj_427_dilations_0"), val = tensor([1, 1])]; + tensor obj_427_groups_0 = const()[name = tensor("obj_427_groups_0"), val = tensor(1)]; + tensor layers_30_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1717986048)))]; + tensor layers_30_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1721262912)))]; + tensor obj_427_cast_fp16 = conv(bias = layers_30_self_attn_o_proj_bias_to_fp16, dilations = obj_427_dilations_0, groups = obj_427_groups_0, pad = obj_427_pad_0, pad_type = obj_427_pad_type_0, strides = obj_427_strides_0, weight = layers_30_self_attn_o_proj_weight_to_fp16, x = input_301_cast_fp16)[name = tensor("obj_427_cast_fp16")]; + tensor inputs_183_cast_fp16 = add(x = inputs_181_cast_fp16, y = obj_427_cast_fp16)[name = tensor("inputs_183_cast_fp16")]; + tensor out_183_axes_0 = const()[name = tensor("out_183_axes_0"), val = tensor([1])]; + tensor var_6901_to_fp16 = const()[name = tensor("op_6901_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_183_cast_fp16 = layer_norm(axes = out_183_axes_0, epsilon = var_6901_to_fp16, x = inputs_183_cast_fp16)[name = tensor("out_183_cast_fp16")]; + tensor obj_429_gamma_0_to_fp16 = const()[name = tensor("obj_429_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1721265536)))]; + tensor obj_429_beta_0_to_fp16 = const()[name = tensor("obj_429_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1721268160)))]; + tensor obj_429_epsilon_0_to_fp16 = const()[name = tensor("obj_429_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_429_cast_fp16 = batch_norm(beta = obj_429_beta_0_to_fp16, epsilon = obj_429_epsilon_0_to_fp16, gamma = obj_429_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_183_cast_fp16)[name = tensor("obj_429_cast_fp16")]; + tensor query_123_pad_type_0 = const()[name = tensor("query_123_pad_type_0"), val = tensor("valid")]; + tensor query_123_strides_0 = const()[name = tensor("query_123_strides_0"), val = tensor([1, 1])]; + tensor query_123_pad_0 = const()[name = tensor("query_123_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_123_dilations_0 = const()[name = tensor("query_123_dilations_0"), val = tensor([1, 1])]; + tensor query_123_groups_0 = const()[name = tensor("query_123_groups_0"), val = tensor(1)]; + tensor layers_30_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_30_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1721270784)))]; + tensor layers_30_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_30_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1724547648)))]; + tensor query_123_cast_fp16 = conv(bias = layers_30_encoder_attn_q_proj_bias_to_fp16, dilations = query_123_dilations_0, groups = query_123_groups_0, pad = query_123_pad_0, pad_type = query_123_pad_type_0, strides = query_123_strides_0, weight = layers_30_encoder_attn_q_proj_weight_to_fp16, x = obj_429_cast_fp16)[name = tensor("query_123_cast_fp16")]; + tensor key_123_pad_type_0 = const()[name = tensor("key_123_pad_type_0"), val = tensor("valid")]; + tensor key_123_strides_0 = const()[name = tensor("key_123_strides_0"), val = tensor([1, 1])]; + tensor key_123_pad_0 = const()[name = tensor("key_123_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_123_dilations_0 = const()[name = tensor("key_123_dilations_0"), val = tensor([1, 1])]; + tensor key_123_groups_0 = const()[name = tensor("key_123_groups_0"), val = tensor(1)]; + tensor layers_30_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_30_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1724550272)))]; + tensor key_123_cast_fp16 = conv(dilations = key_123_dilations_0, groups = key_123_groups_0, pad = key_123_pad_0, pad_type = key_123_pad_type_0, strides = key_123_strides_0, weight = layers_30_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_123_cast_fp16")]; + tensor value_123_pad_type_0 = const()[name = tensor("value_123_pad_type_0"), val = tensor("valid")]; + tensor value_123_strides_0 = const()[name = tensor("value_123_strides_0"), val = tensor([1, 1])]; + tensor value_123_pad_0 = const()[name = tensor("value_123_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_123_dilations_0 = const()[name = tensor("value_123_dilations_0"), val = tensor([1, 1])]; + tensor value_123_groups_0 = const()[name = tensor("value_123_groups_0"), val = tensor(1)]; + tensor layers_30_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_30_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1727827136)))]; + tensor layers_30_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_30_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1731104000)))]; + tensor value_123_cast_fp16 = conv(bias = layers_30_encoder_attn_v_proj_bias_to_fp16, dilations = value_123_dilations_0, groups = value_123_groups_0, pad = value_123_pad_0, pad_type = value_123_pad_type_0, strides = value_123_strides_0, weight = layers_30_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_123_cast_fp16")]; + tensor var_6937 = const()[name = tensor("op_6937"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_123_cast_fp16 = reshape(shape = var_6937, x = query_123_cast_fp16)[name = tensor("mh_q_123_cast_fp16")]; + tensor var_6939_to_fp16 = const()[name = tensor("op_6939_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6940_cast_fp16 = mul(x = mh_q_123_cast_fp16, y = var_6939_to_fp16)[name = tensor("op_6940_cast_fp16")]; + tensor var_6943 = const()[name = tensor("op_6943"), val = tensor([1, 20, 64, 1500])]; + tensor var_6944_cast_fp16 = reshape(shape = var_6943, x = key_123_cast_fp16)[name = tensor("op_6944_cast_fp16")]; + tensor mh_w_185_transpose_x_0 = const()[name = tensor("mh_w_185_transpose_x_0"), val = tensor(true)]; + tensor mh_w_185_transpose_y_0 = const()[name = tensor("mh_w_185_transpose_y_0"), val = tensor(false)]; + tensor mh_w_185_cast_fp16 = matmul(transpose_x = mh_w_185_transpose_x_0, transpose_y = mh_w_185_transpose_y_0, x = var_6940_cast_fp16, y = var_6944_cast_fp16)[name = tensor("mh_w_185_cast_fp16")]; + tensor obj_433_cast_fp16 = softmax(axis = var_6786, x = mh_w_185_cast_fp16)[name = tensor("obj_433_cast_fp16")]; + tensor var_6948 = const()[name = tensor("op_6948"), val = tensor([1, 20, 64, 1500])]; + tensor var_6949_cast_fp16 = reshape(shape = var_6948, x = value_123_cast_fp16)[name = tensor("op_6949_cast_fp16")]; + tensor attn_123_transpose_x_0 = const()[name = tensor("attn_123_transpose_x_0"), val = tensor(false)]; + tensor attn_123_transpose_y_0 = const()[name = tensor("attn_123_transpose_y_0"), val = tensor(true)]; + tensor attn_123_cast_fp16 = matmul(transpose_x = attn_123_transpose_x_0, transpose_y = attn_123_transpose_y_0, x = var_6949_cast_fp16, y = obj_433_cast_fp16)[name = tensor("attn_123_cast_fp16")]; + tensor var_6952 = const()[name = tensor("op_6952"), val = tensor([1, 1280, 1, 1])]; + tensor input_303_cast_fp16 = reshape(shape = var_6952, x = attn_123_cast_fp16)[name = tensor("input_303_cast_fp16")]; + tensor obj_431_pad_type_0 = const()[name = tensor("obj_431_pad_type_0"), val = tensor("valid")]; + tensor obj_431_strides_0 = const()[name = tensor("obj_431_strides_0"), val = tensor([1, 1])]; + tensor obj_431_pad_0 = const()[name = tensor("obj_431_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_431_dilations_0 = const()[name = tensor("obj_431_dilations_0"), val = tensor([1, 1])]; + tensor obj_431_groups_0 = const()[name = tensor("obj_431_groups_0"), val = tensor(1)]; + tensor layers_30_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_30_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1731106624)))]; + tensor layers_30_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_30_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1734383488)))]; + tensor obj_431_cast_fp16 = conv(bias = layers_30_encoder_attn_o_proj_bias_to_fp16, dilations = obj_431_dilations_0, groups = obj_431_groups_0, pad = obj_431_pad_0, pad_type = obj_431_pad_type_0, strides = obj_431_strides_0, weight = layers_30_encoder_attn_o_proj_weight_to_fp16, x = input_303_cast_fp16)[name = tensor("obj_431_cast_fp16")]; + tensor inputs_185_cast_fp16 = add(x = inputs_183_cast_fp16, y = obj_431_cast_fp16)[name = tensor("inputs_185_cast_fp16")]; + tensor out_185_axes_0 = const()[name = tensor("out_185_axes_0"), val = tensor([1])]; + tensor var_6970_to_fp16 = const()[name = tensor("op_6970_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_185_cast_fp16 = layer_norm(axes = out_185_axes_0, epsilon = var_6970_to_fp16, x = inputs_185_cast_fp16)[name = tensor("out_185_cast_fp16")]; + tensor input_305_gamma_0_to_fp16 = const()[name = tensor("input_305_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1734386112)))]; + tensor input_305_beta_0_to_fp16 = const()[name = tensor("input_305_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1734388736)))]; + tensor input_305_epsilon_0_to_fp16 = const()[name = tensor("input_305_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_305_cast_fp16 = batch_norm(beta = input_305_beta_0_to_fp16, epsilon = input_305_epsilon_0_to_fp16, gamma = input_305_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_185_cast_fp16)[name = tensor("input_305_cast_fp16")]; + tensor input_307_pad_type_0 = const()[name = tensor("input_307_pad_type_0"), val = tensor("valid")]; + tensor input_307_strides_0 = const()[name = tensor("input_307_strides_0"), val = tensor([1, 1])]; + tensor input_307_pad_0 = const()[name = tensor("input_307_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_307_dilations_0 = const()[name = tensor("input_307_dilations_0"), val = tensor([1, 1])]; + tensor input_307_groups_0 = const()[name = tensor("input_307_groups_0"), val = tensor(1)]; + tensor layers_30_fc1_weight_to_fp16 = const()[name = tensor("layers_30_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1734391360)))]; + tensor layers_30_fc1_bias_to_fp16 = const()[name = tensor("layers_30_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1747498624)))]; + tensor input_307_cast_fp16 = conv(bias = layers_30_fc1_bias_to_fp16, dilations = input_307_dilations_0, groups = input_307_groups_0, pad = input_307_pad_0, pad_type = input_307_pad_type_0, strides = input_307_strides_0, weight = layers_30_fc1_weight_to_fp16, x = input_305_cast_fp16)[name = tensor("input_307_cast_fp16")]; + tensor input_309_mode_0 = const()[name = tensor("input_309_mode_0"), val = tensor("EXACT")]; + tensor input_309_cast_fp16 = gelu(mode = input_309_mode_0, x = input_307_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor hidden_states_63_pad_type_0 = const()[name = tensor("hidden_states_63_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_63_strides_0 = const()[name = tensor("hidden_states_63_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_63_pad_0 = const()[name = tensor("hidden_states_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_63_dilations_0 = const()[name = tensor("hidden_states_63_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_63_groups_0 = const()[name = tensor("hidden_states_63_groups_0"), val = tensor(1)]; + tensor layers_30_fc2_weight_to_fp16 = const()[name = tensor("layers_30_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1747508928)))]; + tensor layers_30_fc2_bias_to_fp16 = const()[name = tensor("layers_30_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1760616192)))]; + tensor hidden_states_63_cast_fp16 = conv(bias = layers_30_fc2_bias_to_fp16, dilations = hidden_states_63_dilations_0, groups = hidden_states_63_groups_0, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = hidden_states_63_strides_0, weight = layers_30_fc2_weight_to_fp16, x = input_309_cast_fp16)[name = tensor("hidden_states_63_cast_fp16")]; + tensor inputs_187_cast_fp16 = add(x = inputs_185_cast_fp16, y = hidden_states_63_cast_fp16)[name = tensor("inputs_187_cast_fp16")]; + tensor var_7005 = const()[name = tensor("op_7005"), val = tensor(3)]; + tensor out_187_axes_0 = const()[name = tensor("out_187_axes_0"), val = tensor([1])]; + tensor var_7030_to_fp16 = const()[name = tensor("op_7030_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_187_cast_fp16 = layer_norm(axes = out_187_axes_0, epsilon = var_7030_to_fp16, x = inputs_187_cast_fp16)[name = tensor("out_187_cast_fp16")]; + tensor obj_435_gamma_0_to_fp16 = const()[name = tensor("obj_435_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1760618816)))]; + tensor obj_435_beta_0_to_fp16 = const()[name = tensor("obj_435_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1760621440)))]; + tensor obj_435_epsilon_0_to_fp16 = const()[name = tensor("obj_435_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_435_cast_fp16 = batch_norm(beta = obj_435_beta_0_to_fp16, epsilon = obj_435_epsilon_0_to_fp16, gamma = obj_435_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_187_cast_fp16)[name = tensor("obj_435_cast_fp16")]; + tensor query_125_pad_type_0 = const()[name = tensor("query_125_pad_type_0"), val = tensor("valid")]; + tensor query_125_strides_0 = const()[name = tensor("query_125_strides_0"), val = tensor([1, 1])]; + tensor query_125_pad_0 = const()[name = tensor("query_125_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_125_dilations_0 = const()[name = tensor("query_125_dilations_0"), val = tensor([1, 1])]; + tensor query_125_groups_0 = const()[name = tensor("query_125_groups_0"), val = tensor(1)]; + tensor layers_31_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1760624064)))]; + tensor layers_31_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1763900928)))]; + tensor query_125_cast_fp16 = conv(bias = layers_31_self_attn_q_proj_bias_to_fp16, dilations = query_125_dilations_0, groups = query_125_groups_0, pad = query_125_pad_0, pad_type = query_125_pad_type_0, strides = query_125_strides_0, weight = layers_31_self_attn_q_proj_weight_to_fp16, x = obj_435_cast_fp16)[name = tensor("query_125_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_31_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1763903552)))]; + tensor current_key_cast_fp16 = conv(dilations = 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_31_self_attn_k_proj_weight_to_fp16, x = obj_435_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_31_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1767180416)))]; + tensor layers_31_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1770457280)))]; + tensor current_value_cast_fp16 = conv(bias = layers_31_self_attn_v_proj_bias_to_fp16, dilations = 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_31_self_attn_v_proj_weight_to_fp16, x = obj_435_cast_fp16)[name = tensor("current_value_cast_fp16")]; + tensor var_7069_cast_fp16 = mul(x = var_103_cast_fp16_31, y = var_239_cast_fp16)[name = tensor("op_7069_cast_fp16")]; + tensor var_7070_cast_fp16 = mul(x = current_key_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_7070_cast_fp16")]; + tensor key_125_cast_fp16 = add(x = var_7069_cast_fp16, y = var_7070_cast_fp16)[name = tensor("key_125_cast_fp16")]; + tensor var_7073_cast_fp16 = mul(x = var_138_cast_fp16_31, y = var_239_cast_fp16)[name = tensor("op_7073_cast_fp16")]; + tensor var_7074_cast_fp16 = mul(x = current_value_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_7074_cast_fp16")]; + tensor value_125_cast_fp16 = add(x = var_7073_cast_fp16, y = var_7074_cast_fp16)[name = tensor("value_125_cast_fp16")]; + tensor var_7078 = const()[name = tensor("op_7078"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_125_cast_fp16 = reshape(shape = var_7078, x = query_125_cast_fp16)[name = tensor("mh_q_125_cast_fp16")]; + tensor var_7080_to_fp16 = const()[name = tensor("op_7080_to_fp16"), val = tensor(0x1p-3)]; + tensor var_7081_cast_fp16 = mul(x = mh_q_125_cast_fp16, y = var_7080_to_fp16)[name = tensor("op_7081_cast_fp16")]; + tensor var_7084 = const()[name = tensor("op_7084"), val = tensor([1, 20, 64, 448])]; + tensor var_7085_cast_fp16 = reshape(shape = var_7084, x = key_125_cast_fp16)[name = tensor("op_7085_cast_fp16")]; + tensor mh_w_187_transpose_x_0 = const()[name = tensor("mh_w_187_transpose_x_0"), val = tensor(true)]; + tensor mh_w_187_transpose_y_0 = const()[name = tensor("mh_w_187_transpose_y_0"), val = tensor(false)]; + tensor mh_w_187_cast_fp16 = matmul(transpose_x = mh_w_187_transpose_x_0, transpose_y = mh_w_187_transpose_y_0, x = var_7081_cast_fp16, y = var_7085_cast_fp16)[name = tensor("mh_w_187_cast_fp16")]; + tensor mh_w_189_cast_fp16 = add(x = mh_w_187_cast_fp16, y = var_261_cast_fp16)[name = tensor("mh_w_189_cast_fp16")]; + tensor var_7093_cast_fp16 = softmax(axis = var_7005, x = mh_w_189_cast_fp16)[name = tensor("op_7093_cast_fp16")]; + tensor var_7094 = const()[name = tensor("op_7094"), val = tensor([1, 20, 64, 448])]; + tensor var_7095_cast_fp16 = reshape(shape = var_7094, x = value_125_cast_fp16)[name = tensor("op_7095_cast_fp16")]; + tensor attn_125_transpose_x_0 = const()[name = tensor("attn_125_transpose_x_0"), val = tensor(false)]; + tensor attn_125_transpose_y_0 = const()[name = tensor("attn_125_transpose_y_0"), val = tensor(true)]; + tensor attn_125_cast_fp16 = matmul(transpose_x = attn_125_transpose_x_0, transpose_y = attn_125_transpose_y_0, x = var_7095_cast_fp16, y = var_7093_cast_fp16)[name = tensor("attn_125_cast_fp16")]; + tensor var_7098 = const()[name = tensor("op_7098"), val = tensor([1, 1280, 1, 1])]; + tensor input_311_cast_fp16 = reshape(shape = var_7098, x = attn_125_cast_fp16)[name = tensor("input_311_cast_fp16")]; + tensor obj_441_pad_type_0 = const()[name = tensor("obj_441_pad_type_0"), val = tensor("valid")]; + tensor obj_441_strides_0 = const()[name = tensor("obj_441_strides_0"), val = tensor([1, 1])]; + tensor obj_441_pad_0 = const()[name = tensor("obj_441_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_441_dilations_0 = const()[name = tensor("obj_441_dilations_0"), val = tensor([1, 1])]; + tensor obj_441_groups_0 = const()[name = tensor("obj_441_groups_0"), val = tensor(1)]; + tensor layers_31_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1770459904)))]; + tensor layers_31_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1773736768)))]; + tensor obj_441_cast_fp16 = conv(bias = layers_31_self_attn_o_proj_bias_to_fp16, dilations = obj_441_dilations_0, groups = obj_441_groups_0, pad = obj_441_pad_0, pad_type = obj_441_pad_type_0, strides = obj_441_strides_0, weight = layers_31_self_attn_o_proj_weight_to_fp16, x = input_311_cast_fp16)[name = tensor("obj_441_cast_fp16")]; + tensor inputs_189_cast_fp16 = add(x = inputs_187_cast_fp16, y = obj_441_cast_fp16)[name = tensor("inputs_189_cast_fp16")]; + tensor out_189_axes_0 = const()[name = tensor("out_189_axes_0"), val = tensor([1])]; + tensor var_7120_to_fp16 = const()[name = tensor("op_7120_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_189_cast_fp16 = layer_norm(axes = out_189_axes_0, epsilon = var_7120_to_fp16, x = inputs_189_cast_fp16)[name = tensor("out_189_cast_fp16")]; + tensor obj_443_gamma_0_to_fp16 = const()[name = tensor("obj_443_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1773739392)))]; + tensor obj_443_beta_0_to_fp16 = const()[name = tensor("obj_443_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1773742016)))]; + tensor obj_443_epsilon_0_to_fp16 = const()[name = tensor("obj_443_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_443_cast_fp16 = batch_norm(beta = obj_443_beta_0_to_fp16, epsilon = obj_443_epsilon_0_to_fp16, gamma = obj_443_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_189_cast_fp16)[name = tensor("obj_443_cast_fp16")]; + tensor 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_31_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_31_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1773744640)))]; + tensor layers_31_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_31_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1777021504)))]; + tensor query_cast_fp16 = conv(bias = layers_31_encoder_attn_q_proj_bias_to_fp16, dilations = query_dilations_0, groups = query_groups_0, pad = query_pad_0, pad_type = query_pad_type_0, strides = query_strides_0, weight = layers_31_encoder_attn_q_proj_weight_to_fp16, x = obj_443_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_31_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_31_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1777024128)))]; + tensor key_cast_fp16 = conv(dilations = key_dilations_0, groups = key_groups_0, pad = key_pad_0, pad_type = key_pad_type_0, strides = key_strides_0, weight = layers_31_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_31_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_31_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1780300992)))]; + tensor layers_31_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_31_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1783577856)))]; + tensor value_cast_fp16 = conv(bias = layers_31_encoder_attn_v_proj_bias_to_fp16, dilations = value_dilations_0, groups = value_groups_0, pad = value_pad_0, pad_type = value_pad_type_0, strides = value_strides_0, weight = layers_31_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_cast_fp16")]; + tensor var_7156 = const()[name = tensor("op_7156"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_cast_fp16 = reshape(shape = var_7156, x = query_cast_fp16)[name = tensor("mh_q_cast_fp16")]; + tensor var_7158_to_fp16 = const()[name = tensor("op_7158_to_fp16"), val = tensor(0x1p-3)]; + tensor var_7159_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_7158_to_fp16)[name = tensor("op_7159_cast_fp16")]; + tensor var_7162 = const()[name = tensor("op_7162"), val = tensor([1, 20, 64, 1500])]; + tensor var_7163_cast_fp16 = reshape(shape = var_7162, x = key_cast_fp16)[name = tensor("op_7163_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_7159_cast_fp16, y = var_7163_cast_fp16)[name = tensor("mh_w_cast_fp16")]; + tensor obj_447_cast_fp16 = softmax(axis = var_7005, x = mh_w_cast_fp16)[name = tensor("obj_447_cast_fp16")]; + tensor var_7167 = const()[name = tensor("op_7167"), val = tensor([1, 20, 64, 1500])]; + tensor var_7168_cast_fp16 = reshape(shape = var_7167, x = value_cast_fp16)[name = tensor("op_7168_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_7168_cast_fp16, y = obj_447_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_7171 = const()[name = tensor("op_7171"), val = tensor([1, 1280, 1, 1])]; + tensor input_313_cast_fp16 = reshape(shape = var_7171, x = attn_cast_fp16)[name = tensor("input_313_cast_fp16")]; + tensor obj_445_pad_type_0 = const()[name = tensor("obj_445_pad_type_0"), val = tensor("valid")]; + tensor obj_445_strides_0 = const()[name = tensor("obj_445_strides_0"), val = tensor([1, 1])]; + tensor obj_445_pad_0 = const()[name = tensor("obj_445_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_445_dilations_0 = const()[name = tensor("obj_445_dilations_0"), val = tensor([1, 1])]; + tensor obj_445_groups_0 = const()[name = tensor("obj_445_groups_0"), val = tensor(1)]; + tensor layers_31_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_31_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1783580480)))]; + tensor layers_31_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_31_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1786857344)))]; + tensor obj_445_cast_fp16 = conv(bias = layers_31_encoder_attn_o_proj_bias_to_fp16, dilations = obj_445_dilations_0, groups = obj_445_groups_0, pad = obj_445_pad_0, pad_type = obj_445_pad_type_0, strides = obj_445_strides_0, weight = layers_31_encoder_attn_o_proj_weight_to_fp16, x = input_313_cast_fp16)[name = tensor("obj_445_cast_fp16")]; + tensor inputs_191_cast_fp16 = add(x = inputs_189_cast_fp16, y = obj_445_cast_fp16)[name = tensor("inputs_191_cast_fp16")]; + tensor out_191_axes_0 = const()[name = tensor("out_191_axes_0"), val = tensor([1])]; + tensor var_7189_to_fp16 = const()[name = tensor("op_7189_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_191_cast_fp16 = layer_norm(axes = out_191_axes_0, epsilon = var_7189_to_fp16, x = inputs_191_cast_fp16)[name = tensor("out_191_cast_fp16")]; + tensor input_315_gamma_0_to_fp16 = const()[name = tensor("input_315_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1786859968)))]; + tensor input_315_beta_0_to_fp16 = const()[name = tensor("input_315_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1786862592)))]; + tensor input_315_epsilon_0_to_fp16 = const()[name = tensor("input_315_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_315_cast_fp16 = batch_norm(beta = input_315_beta_0_to_fp16, epsilon = input_315_epsilon_0_to_fp16, gamma = input_315_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_191_cast_fp16)[name = tensor("input_315_cast_fp16")]; + tensor input_317_pad_type_0 = const()[name = tensor("input_317_pad_type_0"), val = tensor("valid")]; + tensor input_317_strides_0 = const()[name = tensor("input_317_strides_0"), val = tensor([1, 1])]; + tensor input_317_pad_0 = const()[name = tensor("input_317_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_317_dilations_0 = const()[name = tensor("input_317_dilations_0"), val = tensor([1, 1])]; + tensor input_317_groups_0 = const()[name = tensor("input_317_groups_0"), val = tensor(1)]; + tensor layers_31_fc1_weight_to_fp16 = const()[name = tensor("layers_31_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1786865216)))]; + tensor layers_31_fc1_bias_to_fp16 = const()[name = tensor("layers_31_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1799972480)))]; + tensor input_317_cast_fp16 = conv(bias = layers_31_fc1_bias_to_fp16, dilations = input_317_dilations_0, groups = input_317_groups_0, pad = input_317_pad_0, pad_type = input_317_pad_type_0, strides = input_317_strides_0, weight = layers_31_fc1_weight_to_fp16, x = input_315_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_317_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor hidden_states_65_pad_type_0 = const()[name = tensor("hidden_states_65_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_65_strides_0 = const()[name = tensor("hidden_states_65_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_65_pad_0 = const()[name = tensor("hidden_states_65_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_65_dilations_0 = const()[name = tensor("hidden_states_65_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_65_groups_0 = const()[name = tensor("hidden_states_65_groups_0"), val = tensor(1)]; + tensor layers_31_fc2_weight_to_fp16 = const()[name = tensor("layers_31_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1799982784)))]; + tensor layers_31_fc2_bias_to_fp16 = const()[name = tensor("layers_31_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1813090048)))]; + tensor hidden_states_65_cast_fp16 = conv(bias = layers_31_fc2_bias_to_fp16, dilations = hidden_states_65_dilations_0, groups = hidden_states_65_groups_0, pad = hidden_states_65_pad_0, pad_type = hidden_states_65_pad_type_0, strides = hidden_states_65_strides_0, weight = layers_31_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor("hidden_states_65_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_191_cast_fp16, y = hidden_states_65_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor out_axes_0 = const()[name = tensor("out_axes_0"), val = tensor([1])]; + tensor var_7231_to_fp16 = const()[name = tensor("op_7231_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_7231_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(1813092672)))]; + tensor hidden_states_beta_0_to_fp16 = const()[name = tensor("hidden_states_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1813095296)))]; + tensor hidden_states_epsilon_0_to_fp16 = const()[name = tensor("hidden_states_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor("hidden_states_cast_fp16")]; + tensor var_7242_axes_0 = const()[name = tensor("op_7242_axes_0"), val = tensor([2])]; + tensor var_7242_cast_fp16 = squeeze(axes = var_7242_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_7242_cast_fp16")]; + tensor var_7245_perm_0 = const()[name = tensor("op_7245_perm_0"), val = tensor([0, 2, 1])]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1813097920)))]; + tensor var_7245_cast_fp16 = transpose(perm = var_7245_perm_0, x = var_7242_cast_fp16)[name = tensor("transpose_0")]; + tensor logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_7245_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_7249 = const()[name = tensor("op_7249"), val = tensor(1)]; + tensor obj_451_interleave_0 = const()[name = tensor("obj_451_interleave_0"), val = tensor(false)]; + tensor key_cache_updates = concat(axis = var_7249, interleave = obj_451_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_7_cast_fp16, current_key_9_cast_fp16, current_key_11_cast_fp16, current_key_13_cast_fp16, current_key_15_cast_fp16, current_key_17_cast_fp16, current_key_19_cast_fp16, current_key_21_cast_fp16, current_key_23_cast_fp16, current_key_25_cast_fp16, current_key_27_cast_fp16, current_key_29_cast_fp16, current_key_31_cast_fp16, current_key_33_cast_fp16, current_key_35_cast_fp16, current_key_37_cast_fp16, current_key_39_cast_fp16, current_key_41_cast_fp16, current_key_43_cast_fp16, current_key_45_cast_fp16, current_key_47_cast_fp16, current_key_49_cast_fp16, current_key_51_cast_fp16, current_key_53_cast_fp16, current_key_55_cast_fp16, current_key_57_cast_fp16, current_key_59_cast_fp16, current_key_61_cast_fp16, current_key_cast_fp16))[name = tensor("obj_451_cast_fp16")]; + tensor var_7252 = const()[name = tensor("op_7252"), val = tensor(1)]; + tensor obj_453_interleave_0 = const()[name = tensor("obj_453_interleave_0"), val = tensor(false)]; + tensor value_cache_updates = concat(axis = var_7252, interleave = obj_453_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_7_cast_fp16, current_value_9_cast_fp16, current_value_11_cast_fp16, current_value_13_cast_fp16, current_value_15_cast_fp16, current_value_17_cast_fp16, current_value_19_cast_fp16, current_value_21_cast_fp16, current_value_23_cast_fp16, current_value_25_cast_fp16, current_value_27_cast_fp16, current_value_29_cast_fp16, current_value_31_cast_fp16, current_value_33_cast_fp16, current_value_35_cast_fp16, current_value_37_cast_fp16, current_value_39_cast_fp16, current_value_41_cast_fp16, current_value_43_cast_fp16, current_value_45_cast_fp16, current_value_47_cast_fp16, current_value_49_cast_fp16, current_value_51_cast_fp16, current_value_53_cast_fp16, current_value_55_cast_fp16, current_value_57_cast_fp16, current_value_59_cast_fp16, current_value_61_cast_fp16, current_value_cast_fp16))[name = tensor("obj_453_cast_fp16")]; + tensor var_7263_begin_0 = const()[name = tensor("op_7263_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7263_end_0 = const()[name = tensor("op_7263_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7263_end_mask_0 = const()[name = tensor("op_7263_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7263_cast_fp16 = slice_by_index(begin = var_7263_begin_0, end = var_7263_end_0, end_mask = var_7263_end_mask_0, x = obj_111_cast_fp16)[name = tensor("op_7263_cast_fp16")]; + tensor var_7266_begin_0 = const()[name = tensor("op_7266_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7266_end_0 = const()[name = tensor("op_7266_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7266_end_mask_0 = const()[name = tensor("op_7266_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7266_squeeze_mask_0 = const()[name = tensor("op_7266_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7266_cast_fp16 = slice_by_index(begin = var_7266_begin_0, end = var_7266_end_0, end_mask = var_7266_end_mask_0, squeeze_mask = var_7266_squeeze_mask_0, x = var_7263_cast_fp16)[name = tensor("op_7266_cast_fp16")]; + tensor var_7281_begin_0 = const()[name = tensor("op_7281_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_7281_end_0 = const()[name = tensor("op_7281_end_0"), val = tensor([1, 18, 1, 1500])]; + tensor var_7281_end_mask_0 = const()[name = tensor("op_7281_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7281_cast_fp16 = slice_by_index(begin = var_7281_begin_0, end = var_7281_end_0, end_mask = var_7281_end_mask_0, x = obj_153_cast_fp16)[name = tensor("op_7281_cast_fp16")]; + tensor var_7284_begin_0 = const()[name = tensor("op_7284_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7284_end_0 = const()[name = tensor("op_7284_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7284_end_mask_0 = const()[name = tensor("op_7284_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7284_squeeze_mask_0 = const()[name = tensor("op_7284_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7284_cast_fp16 = slice_by_index(begin = var_7284_begin_0, end = var_7284_end_0, end_mask = var_7284_end_mask_0, squeeze_mask = var_7284_squeeze_mask_0, x = var_7281_cast_fp16)[name = tensor("op_7284_cast_fp16")]; + tensor var_7299_begin_0 = const()[name = tensor("op_7299_begin_0"), val = tensor([0, 18, 0, 0])]; + tensor var_7299_end_0 = const()[name = tensor("op_7299_end_0"), val = tensor([1, 19, 1, 1500])]; + tensor var_7299_end_mask_0 = const()[name = tensor("op_7299_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7299_cast_fp16 = slice_by_index(begin = var_7299_begin_0, end = var_7299_end_0, end_mask = var_7299_end_mask_0, x = obj_181_cast_fp16)[name = tensor("op_7299_cast_fp16")]; + tensor var_7302_begin_0 = const()[name = tensor("op_7302_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7302_end_0 = const()[name = tensor("op_7302_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7302_end_mask_0 = const()[name = tensor("op_7302_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7302_squeeze_mask_0 = const()[name = tensor("op_7302_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7302_cast_fp16 = slice_by_index(begin = var_7302_begin_0, end = var_7302_end_0, end_mask = var_7302_end_mask_0, squeeze_mask = var_7302_squeeze_mask_0, x = var_7299_cast_fp16)[name = tensor("op_7302_cast_fp16")]; + tensor var_7317_begin_0 = const()[name = tensor("op_7317_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_7317_end_0 = const()[name = tensor("op_7317_end_0"), val = tensor([1, 13, 1, 1500])]; + tensor var_7317_end_mask_0 = const()[name = tensor("op_7317_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7317_cast_fp16 = slice_by_index(begin = var_7317_begin_0, end = var_7317_end_0, end_mask = var_7317_end_mask_0, x = obj_195_cast_fp16)[name = tensor("op_7317_cast_fp16")]; + tensor var_7320_begin_0 = const()[name = tensor("op_7320_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7320_end_0 = const()[name = tensor("op_7320_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7320_end_mask_0 = const()[name = tensor("op_7320_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7320_squeeze_mask_0 = const()[name = tensor("op_7320_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7320_cast_fp16 = slice_by_index(begin = var_7320_begin_0, end = var_7320_end_0, end_mask = var_7320_end_mask_0, squeeze_mask = var_7320_squeeze_mask_0, x = var_7317_cast_fp16)[name = tensor("op_7320_cast_fp16")]; + tensor var_7335_begin_0 = const()[name = tensor("op_7335_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_7335_end_0 = const()[name = tensor("op_7335_end_0"), val = tensor([1, 2, 1, 1500])]; + tensor var_7335_end_mask_0 = const()[name = tensor("op_7335_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7335_cast_fp16 = slice_by_index(begin = var_7335_begin_0, end = var_7335_end_0, end_mask = var_7335_end_mask_0, x = obj_237_cast_fp16)[name = tensor("op_7335_cast_fp16")]; + tensor var_7338_begin_0 = const()[name = tensor("op_7338_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7338_end_0 = const()[name = tensor("op_7338_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7338_end_mask_0 = const()[name = tensor("op_7338_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7338_squeeze_mask_0 = const()[name = tensor("op_7338_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7338_cast_fp16 = slice_by_index(begin = var_7338_begin_0, end = var_7338_end_0, end_mask = var_7338_end_mask_0, squeeze_mask = var_7338_squeeze_mask_0, x = var_7335_cast_fp16)[name = tensor("op_7338_cast_fp16")]; + tensor var_7353_begin_0 = const()[name = tensor("op_7353_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_7353_end_0 = const()[name = tensor("op_7353_end_0"), val = tensor([1, 15, 1, 1500])]; + tensor var_7353_end_mask_0 = const()[name = tensor("op_7353_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7353_cast_fp16 = slice_by_index(begin = var_7353_begin_0, end = var_7353_end_0, end_mask = var_7353_end_mask_0, x = obj_251_cast_fp16)[name = tensor("op_7353_cast_fp16")]; + tensor var_7356_begin_0 = const()[name = tensor("op_7356_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7356_end_0 = const()[name = tensor("op_7356_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7356_end_mask_0 = const()[name = tensor("op_7356_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7356_squeeze_mask_0 = const()[name = tensor("op_7356_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7356_cast_fp16 = slice_by_index(begin = var_7356_begin_0, end = var_7356_end_0, end_mask = var_7356_end_mask_0, squeeze_mask = var_7356_squeeze_mask_0, x = var_7353_cast_fp16)[name = tensor("op_7356_cast_fp16")]; + tensor var_7371_begin_0 = const()[name = tensor("op_7371_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_7371_end_0 = const()[name = tensor("op_7371_end_0"), val = tensor([1, 12, 1, 1500])]; + tensor var_7371_end_mask_0 = const()[name = tensor("op_7371_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7371_cast_fp16 = slice_by_index(begin = var_7371_begin_0, end = var_7371_end_0, end_mask = var_7371_end_mask_0, x = obj_279_cast_fp16)[name = tensor("op_7371_cast_fp16")]; + tensor var_7374_begin_0 = const()[name = tensor("op_7374_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7374_end_0 = const()[name = tensor("op_7374_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7374_end_mask_0 = const()[name = tensor("op_7374_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7374_squeeze_mask_0 = const()[name = tensor("op_7374_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7374_cast_fp16 = slice_by_index(begin = var_7374_begin_0, end = var_7374_end_0, end_mask = var_7374_end_mask_0, squeeze_mask = var_7374_squeeze_mask_0, x = var_7371_cast_fp16)[name = tensor("op_7374_cast_fp16")]; + tensor var_7389_begin_0 = const()[name = tensor("op_7389_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_7389_end_0 = const()[name = tensor("op_7389_end_0"), val = tensor([1, 5, 1, 1500])]; + tensor var_7389_end_mask_0 = const()[name = tensor("op_7389_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7389_cast_fp16 = slice_by_index(begin = var_7389_begin_0, end = var_7389_end_0, end_mask = var_7389_end_mask_0, x = obj_307_cast_fp16)[name = tensor("op_7389_cast_fp16")]; + tensor var_7392_begin_0 = const()[name = tensor("op_7392_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7392_end_0 = const()[name = tensor("op_7392_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7392_end_mask_0 = const()[name = tensor("op_7392_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7392_squeeze_mask_0 = const()[name = tensor("op_7392_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7392_cast_fp16 = slice_by_index(begin = var_7392_begin_0, end = var_7392_end_0, end_mask = var_7392_end_mask_0, squeeze_mask = var_7392_squeeze_mask_0, x = var_7389_cast_fp16)[name = tensor("op_7392_cast_fp16")]; + tensor var_7407_begin_0 = const()[name = tensor("op_7407_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_7407_end_0 = const()[name = tensor("op_7407_end_0"), val = tensor([1, 2, 1, 1500])]; + tensor var_7407_end_mask_0 = const()[name = tensor("op_7407_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7407_cast_fp16 = slice_by_index(begin = var_7407_begin_0, end = var_7407_end_0, end_mask = var_7407_end_mask_0, x = obj_349_cast_fp16)[name = tensor("op_7407_cast_fp16")]; + tensor var_7410_begin_0 = const()[name = tensor("op_7410_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7410_end_0 = const()[name = tensor("op_7410_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7410_end_mask_0 = const()[name = tensor("op_7410_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7410_squeeze_mask_0 = const()[name = tensor("op_7410_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7410_cast_fp16 = slice_by_index(begin = var_7410_begin_0, end = var_7410_end_0, end_mask = var_7410_end_mask_0, squeeze_mask = var_7410_squeeze_mask_0, x = var_7407_cast_fp16)[name = tensor("op_7410_cast_fp16")]; + tensor var_7425_begin_0 = const()[name = tensor("op_7425_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_7425_end_0 = const()[name = tensor("op_7425_end_0"), val = tensor([1, 7, 1, 1500])]; + tensor var_7425_end_mask_0 = const()[name = tensor("op_7425_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7425_cast_fp16 = slice_by_index(begin = var_7425_begin_0, end = var_7425_end_0, end_mask = var_7425_end_mask_0, x = obj_363_cast_fp16)[name = tensor("op_7425_cast_fp16")]; + tensor var_7428_begin_0 = const()[name = tensor("op_7428_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7428_end_0 = const()[name = tensor("op_7428_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7428_end_mask_0 = const()[name = tensor("op_7428_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7428_squeeze_mask_0 = const()[name = tensor("op_7428_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7428_cast_fp16 = slice_by_index(begin = var_7428_begin_0, end = var_7428_end_0, end_mask = var_7428_end_mask_0, squeeze_mask = var_7428_squeeze_mask_0, x = var_7425_cast_fp16)[name = tensor("op_7428_cast_fp16")]; + tensor var_7435 = const()[name = tensor("op_7435"), val = tensor(1)]; + tensor var_7436_interleave_0 = const()[name = tensor("op_7436_interleave_0"), val = tensor(false)]; + tensor var_7436_cast_fp16 = concat(axis = var_7435, interleave = var_7436_interleave_0, values = (var_7266_cast_fp16, var_7284_cast_fp16, var_7302_cast_fp16, var_7320_cast_fp16, var_7338_cast_fp16, var_7356_cast_fp16, var_7374_cast_fp16, var_7392_cast_fp16, var_7410_cast_fp16, var_7428_cast_fp16))[name = tensor("op_7436_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_7436_cast_fp16)[name = tensor("obj_cast_fp16")]; + } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); +} \ No newline at end of file