diff --git "a/openai_whisper-large-v2_turbo/TextDecoder.mlmodelc/model.mil" "b/openai_whisper-large-v2_turbo/TextDecoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/openai_whisper-large-v2_turbo/TextDecoder.mlmodelc/model.mil" @@ -0,0 +1,5462 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "2.2.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "7.1"}})] +{ + 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(132774528)))]; + 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 var_183 = const()[name = tensor("op_183"), val = tensor(1)]; + tensor var_184 = const()[name = tensor("op_184"), val = tensor(true)]; + tensor var_196 = const()[name = tensor("op_196"), val = tensor([1])]; + tensor channels_mean_1_cast_fp16 = reduce_mean(axes = var_196, keep_dims = var_184, x = inputs_1_cast_fp16)[name = tensor("channels_mean_1_cast_fp16")]; + tensor zero_mean_1_cast_fp16 = sub(x = inputs_1_cast_fp16, y = channels_mean_1_cast_fp16)[name = tensor("zero_mean_1_cast_fp16")]; + tensor zero_mean_sq_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = zero_mean_1_cast_fp16)[name = tensor("zero_mean_sq_1_cast_fp16")]; + tensor var_200 = const()[name = tensor("op_200"), val = tensor([1])]; + tensor var_201_cast_fp16 = reduce_mean(axes = var_200, keep_dims = var_184, x = zero_mean_sq_1_cast_fp16)[name = tensor("op_201_cast_fp16")]; + tensor var_202_to_fp16 = const()[name = tensor("op_202_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_203_cast_fp16 = add(x = var_201_cast_fp16, y = var_202_to_fp16)[name = tensor("op_203_cast_fp16")]; + tensor denom_1_epsilon_0_to_fp16 = const()[name = tensor("denom_1_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_1_cast_fp16 = rsqrt(epsilon = denom_1_epsilon_0_to_fp16, x = var_203_cast_fp16)[name = tensor("denom_1_cast_fp16")]; + tensor out_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = denom_1_cast_fp16)[name = tensor("out_1_cast_fp16")]; + tensor obj_1_mean_0_to_fp16 = const()[name = tensor("obj_1_mean_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133921472)))]; + tensor obj_1_variance_0_to_fp16 = const()[name = tensor("obj_1_variance_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133924096)))]; + tensor obj_1_gamma_0_to_fp16 = const()[name = tensor("obj_1_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133926720)))]; + tensor obj_1_beta_0_to_fp16 = const()[name = tensor("obj_1_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133929344)))]; + tensor obj_1_epsilon_0_to_fp16 = const()[name = tensor("obj_1_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor("obj_1_cast_fp16")]; + tensor var_218 = const()[name = tensor("op_218"), val = tensor([1, 1])]; + tensor var_220 = const()[name = tensor("op_220"), val = tensor([1, 1])]; + tensor query_1_pad_type_0 = const()[name = tensor("query_1_pad_type_0"), val = tensor("custom")]; + tensor query_1_pad_0 = const()[name = tensor("query_1_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(133931968)))]; + 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(137208832)))]; + tensor query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = var_220, groups = var_183, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = var_218, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("query_1_cast_fp16")]; + tensor var_224 = const()[name = tensor("op_224"), val = tensor([1, 1])]; + tensor var_226 = const()[name = tensor("op_226"), val = tensor([1, 1])]; + tensor current_key_1_pad_type_0 = const()[name = tensor("current_key_1_pad_type_0"), val = tensor("custom")]; + tensor current_key_1_pad_0 = const()[name = tensor("current_key_1_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(137211456)))]; + tensor current_key_1_cast_fp16 = conv(dilations = var_226, groups = var_183, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = var_224, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("current_key_1_cast_fp16")]; + tensor var_231 = const()[name = tensor("op_231"), val = tensor([1, 1])]; + tensor var_233 = const()[name = tensor("op_233"), val = tensor([1, 1])]; + tensor current_value_1_pad_type_0 = const()[name = tensor("current_value_1_pad_type_0"), val = tensor("custom")]; + tensor current_value_1_pad_0 = const()[name = tensor("current_value_1_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(140488320)))]; + 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(143765184)))]; + tensor current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = var_233, groups = var_183, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = var_231, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("current_value_1_cast_fp16")]; + tensor var_237_axes_0 = const()[name = tensor("op_237_axes_0"), val = tensor([1])]; + tensor var_237_cast_fp16 = expand_dims(axes = var_237_axes_0, x = kv_cache_update_mask)[name = tensor("op_237_cast_fp16")]; + tensor var_238_axes_0 = const()[name = tensor("op_238_axes_0"), val = tensor([2])]; + tensor var_238_cast_fp16 = expand_dims(axes = var_238_axes_0, x = var_237_cast_fp16)[name = tensor("op_238_cast_fp16")]; + tensor var_240_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_240_cast_fp16")]; + tensor var_177_to_fp16 = const()[name = tensor("op_177_to_fp16"), val = tensor(0x1p+0)]; + tensor var_241_cast_fp16 = sub(x = var_177_to_fp16, y = var_238_cast_fp16)[name = tensor("op_241_cast_fp16")]; + tensor var_242_cast_fp16 = mul(x = var_103_cast_fp16_0, y = var_241_cast_fp16)[name = tensor("op_242_cast_fp16")]; + tensor key_1_cast_fp16 = add(x = var_240_cast_fp16, y = var_242_cast_fp16)[name = tensor("key_1_cast_fp16")]; + tensor var_244_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_244_cast_fp16")]; + tensor var_246_cast_fp16 = mul(x = var_138_cast_fp16_0, y = var_241_cast_fp16)[name = tensor("op_246_cast_fp16")]; + tensor value_1_cast_fp16 = add(x = var_244_cast_fp16, y = var_246_cast_fp16)[name = tensor("value_1_cast_fp16")]; + tensor var_249 = const()[name = tensor("op_249"), val = tensor([1, 20, 64, -1])]; + tensor var_250_cast_fp16 = reshape(shape = var_249, x = query_1_cast_fp16)[name = tensor("op_250_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 = var_250_cast_fp16, y = var_251_to_fp16)[name = tensor("op_252_cast_fp16")]; + tensor var_253 = const()[name = tensor("op_253"), val = tensor([1, 20, 64, -1])]; + tensor var_254_cast_fp16 = reshape(shape = var_253, x = key_1_cast_fp16)[name = tensor("op_254_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_254_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; + tensor var_258_axes_0 = const()[name = tensor("op_258_axes_0"), val = tensor([1])]; + tensor var_258_cast_fp16 = expand_dims(axes = var_258_axes_0, x = decoder_key_padding_mask)[name = tensor("op_258_cast_fp16")]; + tensor var_259_axes_0 = const()[name = tensor("op_259_axes_0"), val = tensor([2])]; + tensor var_259_cast_fp16 = expand_dims(axes = var_259_axes_0, x = var_258_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; + tensor var_262_cast_fp16 = softmax(axis = var_176, x = mh_w_3_cast_fp16)[name = tensor("op_262_cast_fp16")]; + tensor var_263 = const()[name = tensor("op_263"), val = tensor([1, 20, 64, -1])]; + tensor var_264_cast_fp16 = reshape(shape = var_263, x = value_1_cast_fp16)[name = tensor("op_264_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_264_cast_fp16, y = var_262_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_267 = const()[name = tensor("op_267"), val = tensor([1, 1280, 1, -1])]; + tensor input_1_cast_fp16 = reshape(shape = var_267, x = attn_1_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_271 = const()[name = tensor("op_271"), val = tensor([1, 1])]; + tensor var_273 = const()[name = tensor("op_273"), val = tensor([1, 1])]; + tensor obj_7_pad_type_0 = const()[name = tensor("obj_7_pad_type_0"), val = tensor("custom")]; + tensor obj_7_pad_0 = const()[name = tensor("obj_7_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(143767808)))]; + 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(147044672)))]; + tensor obj_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = var_273, groups = var_183, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = var_271, 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 var_283 = const()[name = tensor("op_283"), val = tensor([1])]; + tensor channels_mean_3_cast_fp16 = reduce_mean(axes = var_283, keep_dims = var_184, x = inputs_3_cast_fp16)[name = tensor("channels_mean_3_cast_fp16")]; + tensor zero_mean_3_cast_fp16 = sub(x = inputs_3_cast_fp16, y = channels_mean_3_cast_fp16)[name = tensor("zero_mean_3_cast_fp16")]; + tensor zero_mean_sq_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = zero_mean_3_cast_fp16)[name = tensor("zero_mean_sq_3_cast_fp16")]; + tensor var_287 = const()[name = tensor("op_287"), val = tensor([1])]; + tensor var_288_cast_fp16 = reduce_mean(axes = var_287, keep_dims = var_184, x = zero_mean_sq_3_cast_fp16)[name = tensor("op_288_cast_fp16")]; + tensor var_289_to_fp16 = const()[name = tensor("op_289_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_290_cast_fp16 = add(x = var_288_cast_fp16, y = var_289_to_fp16)[name = tensor("op_290_cast_fp16")]; + tensor denom_3_epsilon_0_to_fp16 = const()[name = tensor("denom_3_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_3_cast_fp16 = rsqrt(epsilon = denom_3_epsilon_0_to_fp16, x = var_290_cast_fp16)[name = tensor("denom_3_cast_fp16")]; + tensor out_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = denom_3_cast_fp16)[name = tensor("out_3_cast_fp16")]; + tensor obj_9_gamma_0_to_fp16 = const()[name = tensor("obj_9_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147047296)))]; + 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(147049920)))]; + tensor obj_9_epsilon_0_to_fp16 = const()[name = tensor("obj_9_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor("obj_9_cast_fp16")]; + tensor var_305 = const()[name = tensor("op_305"), val = tensor([1, 1])]; + tensor var_307 = const()[name = tensor("op_307"), val = tensor([1, 1])]; + tensor query_3_pad_type_0 = const()[name = tensor("query_3_pad_type_0"), val = tensor("custom")]; + tensor query_3_pad_0 = const()[name = tensor("query_3_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(147052544)))]; + 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(150329408)))]; + tensor query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = var_307, groups = var_183, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = var_305, weight = layers_0_encoder_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("query_3_cast_fp16")]; + tensor var_311 = const()[name = tensor("op_311"), val = tensor([1, 1])]; + tensor var_313 = const()[name = tensor("op_313"), val = tensor([1, 1])]; + tensor key_3_pad_type_0 = const()[name = tensor("key_3_pad_type_0"), val = tensor("custom")]; + tensor key_3_pad_0 = const()[name = tensor("key_3_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(150332032)))]; + tensor key_3_cast_fp16 = conv(dilations = var_313, groups = var_183, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = var_311, weight = layers_0_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_3_cast_fp16")]; + tensor var_318 = const()[name = tensor("op_318"), val = tensor([1, 1])]; + tensor var_320 = const()[name = tensor("op_320"), val = tensor([1, 1])]; + tensor value_3_pad_type_0 = const()[name = tensor("value_3_pad_type_0"), val = tensor("custom")]; + tensor value_3_pad_0 = const()[name = tensor("value_3_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(153608896)))]; + 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(156885760)))]; + tensor value_3_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_bias_to_fp16, dilations = var_320, groups = var_183, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = var_318, weight = layers_0_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_3_cast_fp16")]; + tensor var_324 = const()[name = tensor("op_324"), val = tensor([1, 20, 64, -1])]; + tensor var_325_cast_fp16 = reshape(shape = var_324, x = query_3_cast_fp16)[name = tensor("op_325_cast_fp16")]; + tensor var_326_to_fp16 = const()[name = tensor("op_326_to_fp16"), val = tensor(0x1p-3)]; + tensor var_327_cast_fp16 = mul(x = var_325_cast_fp16, y = var_326_to_fp16)[name = tensor("op_327_cast_fp16")]; + tensor var_328 = const()[name = tensor("op_328"), val = tensor([1, 20, 64, -1])]; + tensor var_329_cast_fp16 = reshape(shape = var_328, x = key_3_cast_fp16)[name = tensor("op_329_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_327_cast_fp16, y = var_329_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_333 = const()[name = tensor("op_333"), val = tensor([1, 20, 64, -1])]; + tensor var_334_cast_fp16 = reshape(shape = var_333, x = value_3_cast_fp16)[name = tensor("op_334_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_334_cast_fp16, y = obj_13_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_337 = const()[name = tensor("op_337"), val = tensor([1, 1280, 1, -1])]; + tensor input_3_cast_fp16 = reshape(shape = var_337, x = attn_3_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor var_341 = const()[name = tensor("op_341"), val = tensor([1, 1])]; + tensor var_343 = const()[name = tensor("op_343"), val = tensor([1, 1])]; + tensor obj_11_pad_type_0 = const()[name = tensor("obj_11_pad_type_0"), val = tensor("custom")]; + tensor obj_11_pad_0 = const()[name = tensor("obj_11_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(156888384)))]; + 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(160165248)))]; + tensor obj_11_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = var_343, groups = var_183, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = var_341, 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 var_349 = const()[name = tensor("op_349"), val = tensor([1])]; + tensor channels_mean_5_cast_fp16 = reduce_mean(axes = var_349, keep_dims = var_184, x = inputs_5_cast_fp16)[name = tensor("channels_mean_5_cast_fp16")]; + tensor zero_mean_5_cast_fp16 = sub(x = inputs_5_cast_fp16, y = channels_mean_5_cast_fp16)[name = tensor("zero_mean_5_cast_fp16")]; + tensor zero_mean_sq_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = zero_mean_5_cast_fp16)[name = tensor("zero_mean_sq_5_cast_fp16")]; + tensor var_353 = const()[name = tensor("op_353"), val = tensor([1])]; + tensor var_354_cast_fp16 = reduce_mean(axes = var_353, keep_dims = var_184, x = zero_mean_sq_5_cast_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_355_to_fp16 = const()[name = tensor("op_355_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_356_cast_fp16 = add(x = var_354_cast_fp16, y = var_355_to_fp16)[name = tensor("op_356_cast_fp16")]; + tensor denom_5_epsilon_0_to_fp16 = const()[name = tensor("denom_5_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_5_cast_fp16 = rsqrt(epsilon = denom_5_epsilon_0_to_fp16, x = var_356_cast_fp16)[name = tensor("denom_5_cast_fp16")]; + tensor out_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = denom_5_cast_fp16)[name = tensor("out_5_cast_fp16")]; + tensor input_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(160167872)))]; + 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(160170496)))]; + 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 var_367 = const()[name = tensor("op_367"), val = tensor([1, 1])]; + tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 1])]; + tensor input_7_pad_type_0 = const()[name = tensor("input_7_pad_type_0"), val = tensor("custom")]; + tensor input_7_pad_0 = const()[name = tensor("input_7_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(160173120)))]; + 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(173280384)))]; + tensor input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = var_369, groups = var_183, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = var_367, 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 var_375 = const()[name = tensor("op_375"), val = tensor([1, 1])]; + tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 1])]; + tensor hidden_states_3_pad_type_0 = const()[name = tensor("hidden_states_3_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_3_pad_0 = const()[name = tensor("hidden_states_3_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(173290688)))]; + 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(186397952)))]; + tensor hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = var_377, groups = var_183, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = var_375, 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_390 = const()[name = tensor("op_390"), val = tensor(3)]; + tensor var_397 = const()[name = tensor("op_397"), val = tensor(1)]; + tensor var_398 = const()[name = tensor("op_398"), val = tensor(true)]; + tensor var_410 = const()[name = tensor("op_410"), val = tensor([1])]; + tensor channels_mean_7_cast_fp16 = reduce_mean(axes = var_410, keep_dims = var_398, x = inputs_7_cast_fp16)[name = tensor("channels_mean_7_cast_fp16")]; + tensor zero_mean_7_cast_fp16 = sub(x = inputs_7_cast_fp16, y = channels_mean_7_cast_fp16)[name = tensor("zero_mean_7_cast_fp16")]; + tensor zero_mean_sq_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = zero_mean_7_cast_fp16)[name = tensor("zero_mean_sq_7_cast_fp16")]; + tensor var_414 = const()[name = tensor("op_414"), val = tensor([1])]; + tensor var_415_cast_fp16 = reduce_mean(axes = var_414, keep_dims = var_398, x = zero_mean_sq_7_cast_fp16)[name = tensor("op_415_cast_fp16")]; + tensor var_416_to_fp16 = const()[name = tensor("op_416_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_417_cast_fp16 = add(x = var_415_cast_fp16, y = var_416_to_fp16)[name = tensor("op_417_cast_fp16")]; + tensor denom_7_epsilon_0_to_fp16 = const()[name = tensor("denom_7_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_7_cast_fp16 = rsqrt(epsilon = denom_7_epsilon_0_to_fp16, x = var_417_cast_fp16)[name = tensor("denom_7_cast_fp16")]; + tensor out_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = denom_7_cast_fp16)[name = tensor("out_7_cast_fp16")]; + tensor obj_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(186400576)))]; + 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(186403200)))]; + 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 var_432 = const()[name = tensor("op_432"), val = tensor([1, 1])]; + tensor var_434 = const()[name = tensor("op_434"), val = tensor([1, 1])]; + tensor query_5_pad_type_0 = const()[name = tensor("query_5_pad_type_0"), val = tensor("custom")]; + tensor query_5_pad_0 = const()[name = tensor("query_5_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(186405824)))]; + 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(189682688)))]; + tensor query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = var_434, groups = var_397, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = var_432, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("query_5_cast_fp16")]; + tensor var_438 = const()[name = tensor("op_438"), val = tensor([1, 1])]; + tensor var_440 = const()[name = tensor("op_440"), val = tensor([1, 1])]; + tensor current_key_3_pad_type_0 = const()[name = tensor("current_key_3_pad_type_0"), val = tensor("custom")]; + tensor current_key_3_pad_0 = const()[name = tensor("current_key_3_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(189685312)))]; + tensor current_key_3_cast_fp16 = conv(dilations = var_440, groups = var_397, pad = current_key_3_pad_0, pad_type = current_key_3_pad_type_0, strides = var_438, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("current_key_3_cast_fp16")]; + tensor var_445 = const()[name = tensor("op_445"), val = tensor([1, 1])]; + tensor var_447 = const()[name = tensor("op_447"), val = tensor([1, 1])]; + tensor current_value_3_pad_type_0 = const()[name = tensor("current_value_3_pad_type_0"), val = tensor("custom")]; + tensor current_value_3_pad_0 = const()[name = tensor("current_value_3_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(192962176)))]; + 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(196239040)))]; + tensor current_value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = var_447, groups = var_397, pad = current_value_3_pad_0, pad_type = current_value_3_pad_type_0, strides = var_445, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("current_value_3_cast_fp16")]; + tensor var_454_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_454_cast_fp16")]; + tensor var_456_cast_fp16 = mul(x = var_103_cast_fp16_1, y = var_241_cast_fp16)[name = tensor("op_456_cast_fp16")]; + tensor key_5_cast_fp16 = add(x = var_454_cast_fp16, y = var_456_cast_fp16)[name = tensor("key_5_cast_fp16")]; + tensor var_458_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_458_cast_fp16")]; + tensor var_460_cast_fp16 = mul(x = var_138_cast_fp16_1, y = var_241_cast_fp16)[name = tensor("op_460_cast_fp16")]; + tensor value_5_cast_fp16 = add(x = var_458_cast_fp16, y = var_460_cast_fp16)[name = tensor("value_5_cast_fp16")]; + tensor var_463 = const()[name = tensor("op_463"), val = tensor([1, 20, 64, -1])]; + tensor var_464_cast_fp16 = reshape(shape = var_463, x = query_5_cast_fp16)[name = tensor("op_464_cast_fp16")]; + tensor var_465_to_fp16 = const()[name = tensor("op_465_to_fp16"), val = tensor(0x1p-3)]; + tensor var_466_cast_fp16 = mul(x = var_464_cast_fp16, y = var_465_to_fp16)[name = tensor("op_466_cast_fp16")]; + tensor var_467 = const()[name = tensor("op_467"), val = tensor([1, 20, 64, -1])]; + tensor var_468_cast_fp16 = reshape(shape = var_467, x = key_5_cast_fp16)[name = tensor("op_468_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_466_cast_fp16, y = var_468_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; + tensor mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; + tensor var_476_cast_fp16 = softmax(axis = var_390, x = mh_w_9_cast_fp16)[name = tensor("op_476_cast_fp16")]; + tensor var_477 = const()[name = tensor("op_477"), val = tensor([1, 20, 64, -1])]; + tensor var_478_cast_fp16 = reshape(shape = var_477, x = value_5_cast_fp16)[name = tensor("op_478_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_478_cast_fp16, y = var_476_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_481 = const()[name = tensor("op_481"), val = tensor([1, 1280, 1, -1])]; + tensor input_11_cast_fp16 = reshape(shape = var_481, x = attn_5_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor var_485 = const()[name = tensor("op_485"), val = tensor([1, 1])]; + tensor var_487 = const()[name = tensor("op_487"), val = tensor([1, 1])]; + tensor obj_21_pad_type_0 = const()[name = tensor("obj_21_pad_type_0"), val = tensor("custom")]; + tensor obj_21_pad_0 = const()[name = tensor("obj_21_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(196241664)))]; + 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(199518528)))]; + tensor obj_21_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = var_487, groups = var_397, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = var_485, 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 var_497 = const()[name = tensor("op_497"), val = tensor([1])]; + tensor channels_mean_9_cast_fp16 = reduce_mean(axes = var_497, keep_dims = var_398, x = inputs_9_cast_fp16)[name = tensor("channels_mean_9_cast_fp16")]; + tensor zero_mean_9_cast_fp16 = sub(x = inputs_9_cast_fp16, y = channels_mean_9_cast_fp16)[name = tensor("zero_mean_9_cast_fp16")]; + tensor zero_mean_sq_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = zero_mean_9_cast_fp16)[name = tensor("zero_mean_sq_9_cast_fp16")]; + tensor var_501 = const()[name = tensor("op_501"), val = tensor([1])]; + tensor var_502_cast_fp16 = reduce_mean(axes = var_501, keep_dims = var_398, x = zero_mean_sq_9_cast_fp16)[name = tensor("op_502_cast_fp16")]; + tensor var_503_to_fp16 = const()[name = tensor("op_503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_504_cast_fp16 = add(x = var_502_cast_fp16, y = var_503_to_fp16)[name = tensor("op_504_cast_fp16")]; + tensor denom_9_epsilon_0_to_fp16 = const()[name = tensor("denom_9_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_9_cast_fp16 = rsqrt(epsilon = denom_9_epsilon_0_to_fp16, x = var_504_cast_fp16)[name = tensor("denom_9_cast_fp16")]; + tensor out_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = denom_9_cast_fp16)[name = tensor("out_9_cast_fp16")]; + tensor obj_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(199521152)))]; + 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(199523776)))]; + 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 var_519 = const()[name = tensor("op_519"), val = tensor([1, 1])]; + tensor var_521 = const()[name = tensor("op_521"), val = tensor([1, 1])]; + tensor query_7_pad_type_0 = const()[name = tensor("query_7_pad_type_0"), val = tensor("custom")]; + tensor query_7_pad_0 = const()[name = tensor("query_7_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(199526400)))]; + 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(202803264)))]; + tensor query_7_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = var_521, groups = var_397, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = var_519, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor("query_7_cast_fp16")]; + tensor var_525 = const()[name = tensor("op_525"), val = tensor([1, 1])]; + tensor var_527 = const()[name = tensor("op_527"), val = tensor([1, 1])]; + tensor key_7_pad_type_0 = const()[name = tensor("key_7_pad_type_0"), val = tensor("custom")]; + tensor key_7_pad_0 = const()[name = tensor("key_7_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(202805888)))]; + tensor key_7_cast_fp16 = conv(dilations = var_527, groups = var_397, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = var_525, weight = layers_1_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_7_cast_fp16")]; + tensor var_532 = const()[name = tensor("op_532"), val = tensor([1, 1])]; + tensor var_534 = const()[name = tensor("op_534"), val = tensor([1, 1])]; + tensor value_7_pad_type_0 = const()[name = tensor("value_7_pad_type_0"), val = tensor("custom")]; + tensor value_7_pad_0 = const()[name = tensor("value_7_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(206082752)))]; + 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(209359616)))]; + tensor value_7_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_bias_to_fp16, dilations = var_534, groups = var_397, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = var_532, weight = layers_1_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_7_cast_fp16")]; + tensor var_538 = const()[name = tensor("op_538"), val = tensor([1, 20, 64, -1])]; + tensor var_539_cast_fp16 = reshape(shape = var_538, x = query_7_cast_fp16)[name = tensor("op_539_cast_fp16")]; + tensor var_540_to_fp16 = const()[name = tensor("op_540_to_fp16"), val = tensor(0x1p-3)]; + tensor var_541_cast_fp16 = mul(x = var_539_cast_fp16, y = var_540_to_fp16)[name = tensor("op_541_cast_fp16")]; + tensor var_542 = const()[name = tensor("op_542"), val = tensor([1, 20, 64, -1])]; + tensor var_543_cast_fp16 = reshape(shape = var_542, x = key_7_cast_fp16)[name = tensor("op_543_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_541_cast_fp16, y = var_543_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; + tensor obj_27_cast_fp16 = softmax(axis = var_390, x = mh_w_11_cast_fp16)[name = tensor("obj_27_cast_fp16")]; + tensor var_547 = const()[name = tensor("op_547"), val = tensor([1, 20, 64, -1])]; + tensor var_548_cast_fp16 = reshape(shape = var_547, x = value_7_cast_fp16)[name = tensor("op_548_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_548_cast_fp16, y = obj_27_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_551 = const()[name = tensor("op_551"), val = tensor([1, 1280, 1, -1])]; + tensor input_13_cast_fp16 = reshape(shape = var_551, x = attn_7_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor var_555 = const()[name = tensor("op_555"), val = tensor([1, 1])]; + tensor var_557 = const()[name = tensor("op_557"), val = tensor([1, 1])]; + tensor obj_25_pad_type_0 = const()[name = tensor("obj_25_pad_type_0"), val = tensor("custom")]; + tensor obj_25_pad_0 = const()[name = tensor("obj_25_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(209362240)))]; + 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(212639104)))]; + tensor obj_25_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = var_557, groups = var_397, pad = obj_25_pad_0, pad_type = obj_25_pad_type_0, strides = var_555, 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 var_563 = const()[name = tensor("op_563"), val = tensor([1])]; + tensor channels_mean_11_cast_fp16 = reduce_mean(axes = var_563, keep_dims = var_398, x = inputs_11_cast_fp16)[name = tensor("channels_mean_11_cast_fp16")]; + tensor zero_mean_11_cast_fp16 = sub(x = inputs_11_cast_fp16, y = channels_mean_11_cast_fp16)[name = tensor("zero_mean_11_cast_fp16")]; + tensor zero_mean_sq_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = zero_mean_11_cast_fp16)[name = tensor("zero_mean_sq_11_cast_fp16")]; + tensor var_567 = const()[name = tensor("op_567"), val = tensor([1])]; + tensor var_568_cast_fp16 = reduce_mean(axes = var_567, keep_dims = var_398, x = zero_mean_sq_11_cast_fp16)[name = tensor("op_568_cast_fp16")]; + tensor var_569_to_fp16 = const()[name = tensor("op_569_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_570_cast_fp16 = add(x = var_568_cast_fp16, y = var_569_to_fp16)[name = tensor("op_570_cast_fp16")]; + tensor denom_11_epsilon_0_to_fp16 = const()[name = tensor("denom_11_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_11_cast_fp16 = rsqrt(epsilon = denom_11_epsilon_0_to_fp16, x = var_570_cast_fp16)[name = tensor("denom_11_cast_fp16")]; + tensor out_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = denom_11_cast_fp16)[name = tensor("out_11_cast_fp16")]; + tensor input_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(212641728)))]; + 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(212644352)))]; + 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 var_581 = const()[name = tensor("op_581"), val = tensor([1, 1])]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 1])]; + tensor input_17_pad_type_0 = const()[name = tensor("input_17_pad_type_0"), val = tensor("custom")]; + tensor input_17_pad_0 = const()[name = tensor("input_17_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(212646976)))]; + 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(225754240)))]; + tensor input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = var_583, groups = var_397, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = var_581, 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 var_589 = const()[name = tensor("op_589"), val = tensor([1, 1])]; + tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1])]; + tensor hidden_states_5_pad_type_0 = const()[name = tensor("hidden_states_5_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_5_pad_0 = const()[name = tensor("hidden_states_5_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(225764544)))]; + 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(238871808)))]; + tensor hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = var_591, groups = var_397, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = var_589, 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_604 = const()[name = tensor("op_604"), val = tensor(3)]; + tensor var_611 = const()[name = tensor("op_611"), val = tensor(1)]; + tensor var_612 = const()[name = tensor("op_612"), val = tensor(true)]; + tensor var_624 = const()[name = tensor("op_624"), val = tensor([1])]; + tensor channels_mean_13_cast_fp16 = reduce_mean(axes = var_624, keep_dims = var_612, x = inputs_13_cast_fp16)[name = tensor("channels_mean_13_cast_fp16")]; + tensor zero_mean_13_cast_fp16 = sub(x = inputs_13_cast_fp16, y = channels_mean_13_cast_fp16)[name = tensor("zero_mean_13_cast_fp16")]; + tensor zero_mean_sq_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = zero_mean_13_cast_fp16)[name = tensor("zero_mean_sq_13_cast_fp16")]; + tensor var_628 = const()[name = tensor("op_628"), val = tensor([1])]; + tensor var_629_cast_fp16 = reduce_mean(axes = var_628, keep_dims = var_612, x = zero_mean_sq_13_cast_fp16)[name = tensor("op_629_cast_fp16")]; + tensor var_630_to_fp16 = const()[name = tensor("op_630_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_631_cast_fp16 = add(x = var_629_cast_fp16, y = var_630_to_fp16)[name = tensor("op_631_cast_fp16")]; + tensor denom_13_epsilon_0_to_fp16 = const()[name = tensor("denom_13_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_13_cast_fp16 = rsqrt(epsilon = denom_13_epsilon_0_to_fp16, x = var_631_cast_fp16)[name = tensor("denom_13_cast_fp16")]; + tensor out_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = denom_13_cast_fp16)[name = tensor("out_13_cast_fp16")]; + tensor obj_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(238874432)))]; + 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(238877056)))]; + 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 var_646 = const()[name = tensor("op_646"), val = tensor([1, 1])]; + tensor var_648 = const()[name = tensor("op_648"), val = tensor([1, 1])]; + tensor query_9_pad_type_0 = const()[name = tensor("query_9_pad_type_0"), val = tensor("custom")]; + tensor query_9_pad_0 = const()[name = tensor("query_9_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(238879680)))]; + 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(242156544)))]; + tensor query_9_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = var_648, groups = var_611, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = var_646, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("query_9_cast_fp16")]; + tensor var_652 = const()[name = tensor("op_652"), val = tensor([1, 1])]; + tensor var_654 = const()[name = tensor("op_654"), val = tensor([1, 1])]; + tensor current_key_5_pad_type_0 = const()[name = tensor("current_key_5_pad_type_0"), val = tensor("custom")]; + tensor current_key_5_pad_0 = const()[name = tensor("current_key_5_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(242159168)))]; + tensor current_key_5_cast_fp16 = conv(dilations = var_654, groups = var_611, pad = current_key_5_pad_0, pad_type = current_key_5_pad_type_0, strides = var_652, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("current_key_5_cast_fp16")]; + tensor var_659 = const()[name = tensor("op_659"), val = tensor([1, 1])]; + tensor var_661 = const()[name = tensor("op_661"), val = tensor([1, 1])]; + tensor current_value_5_pad_type_0 = const()[name = tensor("current_value_5_pad_type_0"), val = tensor("custom")]; + tensor current_value_5_pad_0 = const()[name = tensor("current_value_5_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(245436032)))]; + 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(248712896)))]; + tensor current_value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = var_661, groups = var_611, pad = current_value_5_pad_0, pad_type = current_value_5_pad_type_0, strides = var_659, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("current_value_5_cast_fp16")]; + tensor var_668_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_668_cast_fp16")]; + tensor var_670_cast_fp16 = mul(x = var_103_cast_fp16_2, y = var_241_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor key_9_cast_fp16 = add(x = var_668_cast_fp16, y = var_670_cast_fp16)[name = tensor("key_9_cast_fp16")]; + tensor var_672_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_672_cast_fp16")]; + tensor var_674_cast_fp16 = mul(x = var_138_cast_fp16_2, y = var_241_cast_fp16)[name = tensor("op_674_cast_fp16")]; + tensor value_9_cast_fp16 = add(x = var_672_cast_fp16, y = var_674_cast_fp16)[name = tensor("value_9_cast_fp16")]; + tensor var_677 = const()[name = tensor("op_677"), val = tensor([1, 20, 64, -1])]; + tensor var_678_cast_fp16 = reshape(shape = var_677, x = query_9_cast_fp16)[name = tensor("op_678_cast_fp16")]; + tensor var_679_to_fp16 = const()[name = tensor("op_679_to_fp16"), val = tensor(0x1p-3)]; + tensor var_680_cast_fp16 = mul(x = var_678_cast_fp16, y = var_679_to_fp16)[name = tensor("op_680_cast_fp16")]; + tensor var_681 = const()[name = tensor("op_681"), val = tensor([1, 20, 64, -1])]; + tensor var_682_cast_fp16 = reshape(shape = var_681, x = key_9_cast_fp16)[name = tensor("op_682_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_680_cast_fp16, y = var_682_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; + tensor mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; + tensor var_690_cast_fp16 = softmax(axis = var_604, x = mh_w_15_cast_fp16)[name = tensor("op_690_cast_fp16")]; + tensor var_691 = const()[name = tensor("op_691"), val = tensor([1, 20, 64, -1])]; + tensor var_692_cast_fp16 = reshape(shape = var_691, x = value_9_cast_fp16)[name = tensor("op_692_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_692_cast_fp16, y = var_690_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_695 = const()[name = tensor("op_695"), val = tensor([1, 1280, 1, -1])]; + tensor input_21_cast_fp16 = reshape(shape = var_695, x = attn_9_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_699 = const()[name = tensor("op_699"), val = tensor([1, 1])]; + tensor var_701 = const()[name = tensor("op_701"), val = tensor([1, 1])]; + tensor obj_35_pad_type_0 = const()[name = tensor("obj_35_pad_type_0"), val = tensor("custom")]; + tensor obj_35_pad_0 = const()[name = tensor("obj_35_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(248715520)))]; + 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(251992384)))]; + tensor obj_35_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = var_701, groups = var_611, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = var_699, 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 var_711 = const()[name = tensor("op_711"), val = tensor([1])]; + tensor channels_mean_15_cast_fp16 = reduce_mean(axes = var_711, keep_dims = var_612, x = inputs_15_cast_fp16)[name = tensor("channels_mean_15_cast_fp16")]; + tensor zero_mean_15_cast_fp16 = sub(x = inputs_15_cast_fp16, y = channels_mean_15_cast_fp16)[name = tensor("zero_mean_15_cast_fp16")]; + tensor zero_mean_sq_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = zero_mean_15_cast_fp16)[name = tensor("zero_mean_sq_15_cast_fp16")]; + tensor var_715 = const()[name = tensor("op_715"), val = tensor([1])]; + tensor var_716_cast_fp16 = reduce_mean(axes = var_715, keep_dims = var_612, x = zero_mean_sq_15_cast_fp16)[name = tensor("op_716_cast_fp16")]; + tensor var_717_to_fp16 = const()[name = tensor("op_717_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_718_cast_fp16 = add(x = var_716_cast_fp16, y = var_717_to_fp16)[name = tensor("op_718_cast_fp16")]; + tensor denom_15_epsilon_0_to_fp16 = const()[name = tensor("denom_15_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_15_cast_fp16 = rsqrt(epsilon = denom_15_epsilon_0_to_fp16, x = var_718_cast_fp16)[name = tensor("denom_15_cast_fp16")]; + tensor out_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = denom_15_cast_fp16)[name = tensor("out_15_cast_fp16")]; + tensor obj_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(251995008)))]; + 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(251997632)))]; + 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 var_733 = const()[name = tensor("op_733"), val = tensor([1, 1])]; + tensor var_735 = const()[name = tensor("op_735"), val = tensor([1, 1])]; + tensor query_11_pad_type_0 = const()[name = tensor("query_11_pad_type_0"), val = tensor("custom")]; + tensor query_11_pad_0 = const()[name = tensor("query_11_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(252000256)))]; + 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(255277120)))]; + tensor query_11_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_bias_to_fp16, dilations = var_735, groups = var_611, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = var_733, weight = layers_2_encoder_attn_q_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("query_11_cast_fp16")]; + tensor var_739 = const()[name = tensor("op_739"), val = tensor([1, 1])]; + tensor var_741 = const()[name = tensor("op_741"), val = tensor([1, 1])]; + tensor key_11_pad_type_0 = const()[name = tensor("key_11_pad_type_0"), val = tensor("custom")]; + tensor key_11_pad_0 = const()[name = tensor("key_11_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(255279744)))]; + tensor key_11_cast_fp16 = conv(dilations = var_741, groups = var_611, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = var_739, weight = layers_2_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_11_cast_fp16")]; + tensor var_746 = const()[name = tensor("op_746"), val = tensor([1, 1])]; + tensor var_748 = const()[name = tensor("op_748"), val = tensor([1, 1])]; + tensor value_11_pad_type_0 = const()[name = tensor("value_11_pad_type_0"), val = tensor("custom")]; + tensor value_11_pad_0 = const()[name = tensor("value_11_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(258556608)))]; + 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(261833472)))]; + tensor value_11_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_bias_to_fp16, dilations = var_748, groups = var_611, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = var_746, weight = layers_2_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_11_cast_fp16")]; + tensor var_752 = const()[name = tensor("op_752"), val = tensor([1, 20, 64, -1])]; + tensor var_753_cast_fp16 = reshape(shape = var_752, x = query_11_cast_fp16)[name = tensor("op_753_cast_fp16")]; + tensor var_754_to_fp16 = const()[name = tensor("op_754_to_fp16"), val = tensor(0x1p-3)]; + tensor var_755_cast_fp16 = mul(x = var_753_cast_fp16, y = var_754_to_fp16)[name = tensor("op_755_cast_fp16")]; + tensor var_756 = const()[name = tensor("op_756"), val = tensor([1, 20, 64, -1])]; + tensor var_757_cast_fp16 = reshape(shape = var_756, x = key_11_cast_fp16)[name = tensor("op_757_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_755_cast_fp16, y = var_757_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; + tensor obj_41_cast_fp16 = softmax(axis = var_604, x = mh_w_17_cast_fp16)[name = tensor("obj_41_cast_fp16")]; + tensor var_761 = const()[name = tensor("op_761"), val = tensor([1, 20, 64, -1])]; + tensor var_762_cast_fp16 = reshape(shape = var_761, x = value_11_cast_fp16)[name = tensor("op_762_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_762_cast_fp16, y = obj_41_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 1280, 1, -1])]; + tensor input_23_cast_fp16 = reshape(shape = var_765, x = attn_11_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor var_769 = const()[name = tensor("op_769"), val = tensor([1, 1])]; + tensor var_771 = const()[name = tensor("op_771"), val = tensor([1, 1])]; + tensor obj_39_pad_type_0 = const()[name = tensor("obj_39_pad_type_0"), val = tensor("custom")]; + tensor obj_39_pad_0 = const()[name = tensor("obj_39_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(261836096)))]; + 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(265112960)))]; + tensor obj_39_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = var_771, groups = var_611, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = var_769, 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 var_777 = const()[name = tensor("op_777"), val = tensor([1])]; + tensor channels_mean_17_cast_fp16 = reduce_mean(axes = var_777, keep_dims = var_612, x = inputs_17_cast_fp16)[name = tensor("channels_mean_17_cast_fp16")]; + tensor zero_mean_17_cast_fp16 = sub(x = inputs_17_cast_fp16, y = channels_mean_17_cast_fp16)[name = tensor("zero_mean_17_cast_fp16")]; + tensor zero_mean_sq_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = zero_mean_17_cast_fp16)[name = tensor("zero_mean_sq_17_cast_fp16")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor([1])]; + tensor var_782_cast_fp16 = reduce_mean(axes = var_781, keep_dims = var_612, x = zero_mean_sq_17_cast_fp16)[name = tensor("op_782_cast_fp16")]; + tensor var_783_to_fp16 = const()[name = tensor("op_783_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_784_cast_fp16 = add(x = var_782_cast_fp16, y = var_783_to_fp16)[name = tensor("op_784_cast_fp16")]; + tensor denom_17_epsilon_0_to_fp16 = const()[name = tensor("denom_17_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_17_cast_fp16 = rsqrt(epsilon = denom_17_epsilon_0_to_fp16, x = var_784_cast_fp16)[name = tensor("denom_17_cast_fp16")]; + tensor out_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = denom_17_cast_fp16)[name = tensor("out_17_cast_fp16")]; + tensor input_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(265115584)))]; + 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(265118208)))]; + 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 var_795 = const()[name = tensor("op_795"), val = tensor([1, 1])]; + tensor var_797 = const()[name = tensor("op_797"), val = tensor([1, 1])]; + tensor input_27_pad_type_0 = const()[name = tensor("input_27_pad_type_0"), val = tensor("custom")]; + tensor input_27_pad_0 = const()[name = tensor("input_27_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(265120832)))]; + 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(278228096)))]; + tensor input_27_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = var_797, groups = var_611, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = var_795, 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 var_803 = const()[name = tensor("op_803"), val = tensor([1, 1])]; + tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 1])]; + tensor hidden_states_7_pad_type_0 = const()[name = tensor("hidden_states_7_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_7_pad_0 = const()[name = tensor("hidden_states_7_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(278238400)))]; + 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(291345664)))]; + tensor hidden_states_7_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = var_805, groups = var_611, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = var_803, 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_818 = const()[name = tensor("op_818"), val = tensor(3)]; + tensor var_825 = const()[name = tensor("op_825"), val = tensor(1)]; + tensor var_826 = const()[name = tensor("op_826"), val = tensor(true)]; + tensor var_838 = const()[name = tensor("op_838"), val = tensor([1])]; + tensor channels_mean_19_cast_fp16 = reduce_mean(axes = var_838, keep_dims = var_826, x = inputs_19_cast_fp16)[name = tensor("channels_mean_19_cast_fp16")]; + tensor zero_mean_19_cast_fp16 = sub(x = inputs_19_cast_fp16, y = channels_mean_19_cast_fp16)[name = tensor("zero_mean_19_cast_fp16")]; + tensor zero_mean_sq_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = zero_mean_19_cast_fp16)[name = tensor("zero_mean_sq_19_cast_fp16")]; + tensor var_842 = const()[name = tensor("op_842"), val = tensor([1])]; + tensor var_843_cast_fp16 = reduce_mean(axes = var_842, keep_dims = var_826, x = zero_mean_sq_19_cast_fp16)[name = tensor("op_843_cast_fp16")]; + tensor var_844_to_fp16 = const()[name = tensor("op_844_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_845_cast_fp16 = add(x = var_843_cast_fp16, y = var_844_to_fp16)[name = tensor("op_845_cast_fp16")]; + tensor denom_19_epsilon_0_to_fp16 = const()[name = tensor("denom_19_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_19_cast_fp16 = rsqrt(epsilon = denom_19_epsilon_0_to_fp16, x = var_845_cast_fp16)[name = tensor("denom_19_cast_fp16")]; + tensor out_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = denom_19_cast_fp16)[name = tensor("out_19_cast_fp16")]; + tensor obj_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(291348288)))]; + 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(291350912)))]; + 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 var_860 = const()[name = tensor("op_860"), val = tensor([1, 1])]; + tensor var_862 = const()[name = tensor("op_862"), val = tensor([1, 1])]; + tensor query_13_pad_type_0 = const()[name = tensor("query_13_pad_type_0"), val = tensor("custom")]; + tensor query_13_pad_0 = const()[name = tensor("query_13_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(291353536)))]; + 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(294630400)))]; + tensor query_13_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = var_862, groups = var_825, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = var_860, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("query_13_cast_fp16")]; + tensor var_866 = const()[name = tensor("op_866"), val = tensor([1, 1])]; + tensor var_868 = const()[name = tensor("op_868"), val = tensor([1, 1])]; + tensor current_key_7_pad_type_0 = const()[name = tensor("current_key_7_pad_type_0"), val = tensor("custom")]; + tensor current_key_7_pad_0 = const()[name = tensor("current_key_7_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(294633024)))]; + tensor current_key_7_cast_fp16 = conv(dilations = var_868, groups = var_825, pad = current_key_7_pad_0, pad_type = current_key_7_pad_type_0, strides = var_866, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("current_key_7_cast_fp16")]; + tensor var_873 = const()[name = tensor("op_873"), val = tensor([1, 1])]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([1, 1])]; + tensor current_value_7_pad_type_0 = const()[name = tensor("current_value_7_pad_type_0"), val = tensor("custom")]; + tensor current_value_7_pad_0 = const()[name = tensor("current_value_7_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(297909888)))]; + 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(301186752)))]; + tensor current_value_7_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = var_875, groups = var_825, pad = current_value_7_pad_0, pad_type = current_value_7_pad_type_0, strides = var_873, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("current_value_7_cast_fp16")]; + tensor var_882_cast_fp16 = mul(x = current_key_7_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_882_cast_fp16")]; + tensor var_884_cast_fp16 = mul(x = var_103_cast_fp16_3, y = var_241_cast_fp16)[name = tensor("op_884_cast_fp16")]; + tensor key_13_cast_fp16 = add(x = var_882_cast_fp16, y = var_884_cast_fp16)[name = tensor("key_13_cast_fp16")]; + tensor var_886_cast_fp16 = mul(x = current_value_7_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_886_cast_fp16")]; + tensor var_888_cast_fp16 = mul(x = var_138_cast_fp16_3, y = var_241_cast_fp16)[name = tensor("op_888_cast_fp16")]; + tensor value_13_cast_fp16 = add(x = var_886_cast_fp16, y = var_888_cast_fp16)[name = tensor("value_13_cast_fp16")]; + tensor var_891 = const()[name = tensor("op_891"), val = tensor([1, 20, 64, -1])]; + tensor var_892_cast_fp16 = reshape(shape = var_891, x = query_13_cast_fp16)[name = tensor("op_892_cast_fp16")]; + tensor var_893_to_fp16 = const()[name = tensor("op_893_to_fp16"), val = tensor(0x1p-3)]; + tensor var_894_cast_fp16 = mul(x = var_892_cast_fp16, y = var_893_to_fp16)[name = tensor("op_894_cast_fp16")]; + tensor var_895 = const()[name = tensor("op_895"), val = tensor([1, 20, 64, -1])]; + tensor var_896_cast_fp16 = reshape(shape = var_895, x = key_13_cast_fp16)[name = tensor("op_896_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_894_cast_fp16, y = var_896_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; + tensor mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; + tensor var_904_cast_fp16 = softmax(axis = var_818, x = mh_w_21_cast_fp16)[name = tensor("op_904_cast_fp16")]; + tensor var_905 = const()[name = tensor("op_905"), val = tensor([1, 20, 64, -1])]; + tensor var_906_cast_fp16 = reshape(shape = var_905, x = value_13_cast_fp16)[name = tensor("op_906_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_906_cast_fp16, y = var_904_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_909 = const()[name = tensor("op_909"), val = tensor([1, 1280, 1, -1])]; + tensor input_31_cast_fp16 = reshape(shape = var_909, x = attn_13_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor var_913 = const()[name = tensor("op_913"), val = tensor([1, 1])]; + tensor var_915 = const()[name = tensor("op_915"), val = tensor([1, 1])]; + tensor obj_49_pad_type_0 = const()[name = tensor("obj_49_pad_type_0"), val = tensor("custom")]; + tensor obj_49_pad_0 = const()[name = tensor("obj_49_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(301189376)))]; + 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(304466240)))]; + tensor obj_49_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = var_915, groups = var_825, pad = obj_49_pad_0, pad_type = obj_49_pad_type_0, strides = var_913, 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 var_925 = const()[name = tensor("op_925"), val = tensor([1])]; + tensor channels_mean_21_cast_fp16 = reduce_mean(axes = var_925, keep_dims = var_826, x = inputs_21_cast_fp16)[name = tensor("channels_mean_21_cast_fp16")]; + tensor zero_mean_21_cast_fp16 = sub(x = inputs_21_cast_fp16, y = channels_mean_21_cast_fp16)[name = tensor("zero_mean_21_cast_fp16")]; + tensor zero_mean_sq_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = zero_mean_21_cast_fp16)[name = tensor("zero_mean_sq_21_cast_fp16")]; + tensor var_929 = const()[name = tensor("op_929"), val = tensor([1])]; + tensor var_930_cast_fp16 = reduce_mean(axes = var_929, keep_dims = var_826, x = zero_mean_sq_21_cast_fp16)[name = tensor("op_930_cast_fp16")]; + tensor var_931_to_fp16 = const()[name = tensor("op_931_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_932_cast_fp16 = add(x = var_930_cast_fp16, y = var_931_to_fp16)[name = tensor("op_932_cast_fp16")]; + tensor denom_21_epsilon_0_to_fp16 = const()[name = tensor("denom_21_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_21_cast_fp16 = rsqrt(epsilon = denom_21_epsilon_0_to_fp16, x = var_932_cast_fp16)[name = tensor("denom_21_cast_fp16")]; + tensor out_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = denom_21_cast_fp16)[name = tensor("out_21_cast_fp16")]; + tensor obj_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(304468864)))]; + 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(304471488)))]; + 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 var_947 = const()[name = tensor("op_947"), val = tensor([1, 1])]; + tensor var_949 = const()[name = tensor("op_949"), val = tensor([1, 1])]; + tensor query_15_pad_type_0 = const()[name = tensor("query_15_pad_type_0"), val = tensor("custom")]; + tensor query_15_pad_0 = const()[name = tensor("query_15_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(304474112)))]; + 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(307750976)))]; + tensor query_15_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_bias_to_fp16, dilations = var_949, groups = var_825, pad = query_15_pad_0, pad_type = query_15_pad_type_0, strides = var_947, weight = layers_3_encoder_attn_q_proj_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor("query_15_cast_fp16")]; + tensor var_953 = const()[name = tensor("op_953"), val = tensor([1, 1])]; + tensor var_955 = const()[name = tensor("op_955"), val = tensor([1, 1])]; + tensor key_15_pad_type_0 = const()[name = tensor("key_15_pad_type_0"), val = tensor("custom")]; + tensor key_15_pad_0 = const()[name = tensor("key_15_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(307753600)))]; + tensor key_15_cast_fp16 = conv(dilations = var_955, groups = var_825, pad = key_15_pad_0, pad_type = key_15_pad_type_0, strides = var_953, weight = layers_3_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_15_cast_fp16")]; + tensor var_960 = const()[name = tensor("op_960"), val = tensor([1, 1])]; + tensor var_962 = const()[name = tensor("op_962"), val = tensor([1, 1])]; + tensor value_15_pad_type_0 = const()[name = tensor("value_15_pad_type_0"), val = tensor("custom")]; + tensor value_15_pad_0 = const()[name = tensor("value_15_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(311030464)))]; + 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(314307328)))]; + tensor value_15_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_bias_to_fp16, dilations = var_962, groups = var_825, pad = value_15_pad_0, pad_type = value_15_pad_type_0, strides = var_960, weight = layers_3_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_15_cast_fp16")]; + tensor var_966 = const()[name = tensor("op_966"), val = tensor([1, 20, 64, -1])]; + tensor var_967_cast_fp16 = reshape(shape = var_966, x = query_15_cast_fp16)[name = tensor("op_967_cast_fp16")]; + tensor var_968_to_fp16 = const()[name = tensor("op_968_to_fp16"), val = tensor(0x1p-3)]; + tensor var_969_cast_fp16 = mul(x = var_967_cast_fp16, y = var_968_to_fp16)[name = tensor("op_969_cast_fp16")]; + tensor var_970 = const()[name = tensor("op_970"), val = tensor([1, 20, 64, -1])]; + tensor var_971_cast_fp16 = reshape(shape = var_970, x = key_15_cast_fp16)[name = tensor("op_971_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_969_cast_fp16, y = var_971_cast_fp16)[name = tensor("mh_w_23_cast_fp16")]; + tensor obj_55_cast_fp16 = softmax(axis = var_818, x = mh_w_23_cast_fp16)[name = tensor("obj_55_cast_fp16")]; + tensor var_975 = const()[name = tensor("op_975"), val = tensor([1, 20, 64, -1])]; + tensor var_976_cast_fp16 = reshape(shape = var_975, x = value_15_cast_fp16)[name = tensor("op_976_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_976_cast_fp16, y = obj_55_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_979 = const()[name = tensor("op_979"), val = tensor([1, 1280, 1, -1])]; + tensor input_33_cast_fp16 = reshape(shape = var_979, x = attn_15_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor var_983 = const()[name = tensor("op_983"), val = tensor([1, 1])]; + tensor var_985 = const()[name = tensor("op_985"), val = tensor([1, 1])]; + tensor obj_53_pad_type_0 = const()[name = tensor("obj_53_pad_type_0"), val = tensor("custom")]; + tensor obj_53_pad_0 = const()[name = tensor("obj_53_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(314309952)))]; + 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(317586816)))]; + tensor obj_53_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = var_985, groups = var_825, pad = obj_53_pad_0, pad_type = obj_53_pad_type_0, strides = var_983, 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 var_991 = const()[name = tensor("op_991"), val = tensor([1])]; + tensor channels_mean_23_cast_fp16 = reduce_mean(axes = var_991, keep_dims = var_826, x = inputs_23_cast_fp16)[name = tensor("channels_mean_23_cast_fp16")]; + tensor zero_mean_23_cast_fp16 = sub(x = inputs_23_cast_fp16, y = channels_mean_23_cast_fp16)[name = tensor("zero_mean_23_cast_fp16")]; + tensor zero_mean_sq_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = zero_mean_23_cast_fp16)[name = tensor("zero_mean_sq_23_cast_fp16")]; + tensor var_995 = const()[name = tensor("op_995"), val = tensor([1])]; + tensor var_996_cast_fp16 = reduce_mean(axes = var_995, keep_dims = var_826, x = zero_mean_sq_23_cast_fp16)[name = tensor("op_996_cast_fp16")]; + tensor var_997_to_fp16 = const()[name = tensor("op_997_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_998_cast_fp16 = add(x = var_996_cast_fp16, y = var_997_to_fp16)[name = tensor("op_998_cast_fp16")]; + tensor denom_23_epsilon_0_to_fp16 = const()[name = tensor("denom_23_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_23_cast_fp16 = rsqrt(epsilon = denom_23_epsilon_0_to_fp16, x = var_998_cast_fp16)[name = tensor("denom_23_cast_fp16")]; + tensor out_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = denom_23_cast_fp16)[name = tensor("out_23_cast_fp16")]; + tensor input_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(317589440)))]; + 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(317592064)))]; + 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 var_1009 = const()[name = tensor("op_1009"), val = tensor([1, 1])]; + tensor var_1011 = const()[name = tensor("op_1011"), val = tensor([1, 1])]; + tensor input_37_pad_type_0 = const()[name = tensor("input_37_pad_type_0"), val = tensor("custom")]; + tensor input_37_pad_0 = const()[name = tensor("input_37_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(317594688)))]; + 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(330701952)))]; + tensor input_37_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = var_1011, groups = var_825, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = var_1009, 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 var_1017 = const()[name = tensor("op_1017"), val = tensor([1, 1])]; + tensor var_1019 = const()[name = tensor("op_1019"), val = tensor([1, 1])]; + tensor hidden_states_9_pad_type_0 = const()[name = tensor("hidden_states_9_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_9_pad_0 = const()[name = tensor("hidden_states_9_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(330712256)))]; + 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(343819520)))]; + tensor hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = var_1019, groups = var_825, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = var_1017, 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_1032 = const()[name = tensor("op_1032"), val = tensor(3)]; + tensor var_1039 = const()[name = tensor("op_1039"), val = tensor(1)]; + tensor var_1040 = const()[name = tensor("op_1040"), val = tensor(true)]; + tensor var_1052 = const()[name = tensor("op_1052"), val = tensor([1])]; + tensor channels_mean_25_cast_fp16 = reduce_mean(axes = var_1052, keep_dims = var_1040, x = inputs_25_cast_fp16)[name = tensor("channels_mean_25_cast_fp16")]; + tensor zero_mean_25_cast_fp16 = sub(x = inputs_25_cast_fp16, y = channels_mean_25_cast_fp16)[name = tensor("zero_mean_25_cast_fp16")]; + tensor zero_mean_sq_25_cast_fp16 = mul(x = zero_mean_25_cast_fp16, y = zero_mean_25_cast_fp16)[name = tensor("zero_mean_sq_25_cast_fp16")]; + tensor var_1056 = const()[name = tensor("op_1056"), val = tensor([1])]; + tensor var_1057_cast_fp16 = reduce_mean(axes = var_1056, keep_dims = var_1040, x = zero_mean_sq_25_cast_fp16)[name = tensor("op_1057_cast_fp16")]; + tensor var_1058_to_fp16 = const()[name = tensor("op_1058_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1059_cast_fp16 = add(x = var_1057_cast_fp16, y = var_1058_to_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor denom_25_epsilon_0_to_fp16 = const()[name = tensor("denom_25_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_25_cast_fp16 = rsqrt(epsilon = denom_25_epsilon_0_to_fp16, x = var_1059_cast_fp16)[name = tensor("denom_25_cast_fp16")]; + tensor out_25_cast_fp16 = mul(x = zero_mean_25_cast_fp16, y = denom_25_cast_fp16)[name = tensor("out_25_cast_fp16")]; + tensor obj_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(343822144)))]; + 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(343824768)))]; + 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 var_1074 = const()[name = tensor("op_1074"), val = tensor([1, 1])]; + tensor var_1076 = const()[name = tensor("op_1076"), val = tensor([1, 1])]; + tensor query_17_pad_type_0 = const()[name = tensor("query_17_pad_type_0"), val = tensor("custom")]; + tensor query_17_pad_0 = const()[name = tensor("query_17_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(343827392)))]; + 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(347104256)))]; + tensor query_17_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_bias_to_fp16, dilations = var_1076, groups = var_1039, pad = query_17_pad_0, pad_type = query_17_pad_type_0, strides = var_1074, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("query_17_cast_fp16")]; + tensor var_1080 = const()[name = tensor("op_1080"), val = tensor([1, 1])]; + tensor var_1082 = const()[name = tensor("op_1082"), val = tensor([1, 1])]; + tensor current_key_9_pad_type_0 = const()[name = tensor("current_key_9_pad_type_0"), val = tensor("custom")]; + tensor current_key_9_pad_0 = const()[name = tensor("current_key_9_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(347106880)))]; + tensor current_key_9_cast_fp16 = conv(dilations = var_1082, groups = var_1039, pad = current_key_9_pad_0, pad_type = current_key_9_pad_type_0, strides = var_1080, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("current_key_9_cast_fp16")]; + tensor var_1087 = const()[name = tensor("op_1087"), val = tensor([1, 1])]; + tensor var_1089 = const()[name = tensor("op_1089"), val = tensor([1, 1])]; + tensor current_value_9_pad_type_0 = const()[name = tensor("current_value_9_pad_type_0"), val = tensor("custom")]; + tensor current_value_9_pad_0 = const()[name = tensor("current_value_9_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(350383744)))]; + 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(353660608)))]; + tensor current_value_9_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_bias_to_fp16, dilations = var_1089, groups = var_1039, pad = current_value_9_pad_0, pad_type = current_value_9_pad_type_0, strides = var_1087, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("current_value_9_cast_fp16")]; + tensor var_1096_cast_fp16 = mul(x = current_key_9_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor var_1098_cast_fp16 = mul(x = var_103_cast_fp16_4, y = var_241_cast_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor key_17_cast_fp16 = add(x = var_1096_cast_fp16, y = var_1098_cast_fp16)[name = tensor("key_17_cast_fp16")]; + tensor var_1100_cast_fp16 = mul(x = current_value_9_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1100_cast_fp16")]; + tensor var_1102_cast_fp16 = mul(x = var_138_cast_fp16_4, y = var_241_cast_fp16)[name = tensor("op_1102_cast_fp16")]; + tensor value_17_cast_fp16 = add(x = var_1100_cast_fp16, y = var_1102_cast_fp16)[name = tensor("value_17_cast_fp16")]; + tensor var_1105 = const()[name = tensor("op_1105"), val = tensor([1, 20, 64, -1])]; + tensor var_1106_cast_fp16 = reshape(shape = var_1105, x = query_17_cast_fp16)[name = tensor("op_1106_cast_fp16")]; + tensor var_1107_to_fp16 = const()[name = tensor("op_1107_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1108_cast_fp16 = mul(x = var_1106_cast_fp16, y = var_1107_to_fp16)[name = tensor("op_1108_cast_fp16")]; + tensor var_1109 = const()[name = tensor("op_1109"), val = tensor([1, 20, 64, -1])]; + tensor var_1110_cast_fp16 = reshape(shape = var_1109, x = key_17_cast_fp16)[name = tensor("op_1110_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_1108_cast_fp16, y = var_1110_cast_fp16)[name = tensor("mh_w_25_cast_fp16")]; + tensor mh_w_27_cast_fp16 = add(x = mh_w_25_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_27_cast_fp16")]; + tensor var_1118_cast_fp16 = softmax(axis = var_1032, x = mh_w_27_cast_fp16)[name = tensor("op_1118_cast_fp16")]; + tensor var_1119 = const()[name = tensor("op_1119"), val = tensor([1, 20, 64, -1])]; + tensor var_1120_cast_fp16 = reshape(shape = var_1119, x = value_17_cast_fp16)[name = tensor("op_1120_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_1120_cast_fp16, y = var_1118_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_1123 = const()[name = tensor("op_1123"), val = tensor([1, 1280, 1, -1])]; + tensor input_41_cast_fp16 = reshape(shape = var_1123, x = attn_17_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor var_1127 = const()[name = tensor("op_1127"), val = tensor([1, 1])]; + tensor var_1129 = const()[name = tensor("op_1129"), val = tensor([1, 1])]; + tensor obj_63_pad_type_0 = const()[name = tensor("obj_63_pad_type_0"), val = tensor("custom")]; + tensor obj_63_pad_0 = const()[name = tensor("obj_63_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(353663232)))]; + 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(356940096)))]; + tensor obj_63_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_bias_to_fp16, dilations = var_1129, groups = var_1039, pad = obj_63_pad_0, pad_type = obj_63_pad_type_0, strides = var_1127, 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 var_1139 = const()[name = tensor("op_1139"), val = tensor([1])]; + tensor channels_mean_27_cast_fp16 = reduce_mean(axes = var_1139, keep_dims = var_1040, x = inputs_27_cast_fp16)[name = tensor("channels_mean_27_cast_fp16")]; + tensor zero_mean_27_cast_fp16 = sub(x = inputs_27_cast_fp16, y = channels_mean_27_cast_fp16)[name = tensor("zero_mean_27_cast_fp16")]; + tensor zero_mean_sq_27_cast_fp16 = mul(x = zero_mean_27_cast_fp16, y = zero_mean_27_cast_fp16)[name = tensor("zero_mean_sq_27_cast_fp16")]; + tensor var_1143 = const()[name = tensor("op_1143"), val = tensor([1])]; + tensor var_1144_cast_fp16 = reduce_mean(axes = var_1143, keep_dims = var_1040, x = zero_mean_sq_27_cast_fp16)[name = tensor("op_1144_cast_fp16")]; + tensor var_1145_to_fp16 = const()[name = tensor("op_1145_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1146_cast_fp16 = add(x = var_1144_cast_fp16, y = var_1145_to_fp16)[name = tensor("op_1146_cast_fp16")]; + tensor denom_27_epsilon_0_to_fp16 = const()[name = tensor("denom_27_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_27_cast_fp16 = rsqrt(epsilon = denom_27_epsilon_0_to_fp16, x = var_1146_cast_fp16)[name = tensor("denom_27_cast_fp16")]; + tensor out_27_cast_fp16 = mul(x = zero_mean_27_cast_fp16, y = denom_27_cast_fp16)[name = tensor("out_27_cast_fp16")]; + tensor obj_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(356942720)))]; + 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(356945344)))]; + 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 var_1161 = const()[name = tensor("op_1161"), val = tensor([1, 1])]; + tensor var_1163 = const()[name = tensor("op_1163"), val = tensor([1, 1])]; + tensor query_19_pad_type_0 = const()[name = tensor("query_19_pad_type_0"), val = tensor("custom")]; + tensor query_19_pad_0 = const()[name = tensor("query_19_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(356947968)))]; + 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(360224832)))]; + tensor query_19_cast_fp16 = conv(bias = layers_4_encoder_attn_q_proj_bias_to_fp16, dilations = var_1163, groups = var_1039, pad = query_19_pad_0, pad_type = query_19_pad_type_0, strides = var_1161, weight = layers_4_encoder_attn_q_proj_weight_to_fp16, x = obj_65_cast_fp16)[name = tensor("query_19_cast_fp16")]; + tensor var_1167 = const()[name = tensor("op_1167"), val = tensor([1, 1])]; + tensor var_1169 = const()[name = tensor("op_1169"), val = tensor([1, 1])]; + tensor key_19_pad_type_0 = const()[name = tensor("key_19_pad_type_0"), val = tensor("custom")]; + tensor key_19_pad_0 = const()[name = tensor("key_19_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(360227456)))]; + tensor key_19_cast_fp16 = conv(dilations = var_1169, groups = var_1039, pad = key_19_pad_0, pad_type = key_19_pad_type_0, strides = var_1167, weight = layers_4_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_19_cast_fp16")]; + tensor var_1174 = const()[name = tensor("op_1174"), val = tensor([1, 1])]; + tensor var_1176 = const()[name = tensor("op_1176"), val = tensor([1, 1])]; + tensor value_19_pad_type_0 = const()[name = tensor("value_19_pad_type_0"), val = tensor("custom")]; + tensor value_19_pad_0 = const()[name = tensor("value_19_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(363504320)))]; + 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(366781184)))]; + tensor value_19_cast_fp16 = conv(bias = layers_4_encoder_attn_v_proj_bias_to_fp16, dilations = var_1176, groups = var_1039, pad = value_19_pad_0, pad_type = value_19_pad_type_0, strides = var_1174, weight = layers_4_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_19_cast_fp16")]; + tensor var_1180 = const()[name = tensor("op_1180"), val = tensor([1, 20, 64, -1])]; + tensor var_1181_cast_fp16 = reshape(shape = var_1180, x = query_19_cast_fp16)[name = tensor("op_1181_cast_fp16")]; + tensor var_1182_to_fp16 = const()[name = tensor("op_1182_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1183_cast_fp16 = mul(x = var_1181_cast_fp16, y = var_1182_to_fp16)[name = tensor("op_1183_cast_fp16")]; + tensor var_1184 = const()[name = tensor("op_1184"), val = tensor([1, 20, 64, -1])]; + tensor var_1185_cast_fp16 = reshape(shape = var_1184, x = key_19_cast_fp16)[name = tensor("op_1185_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_1183_cast_fp16, y = var_1185_cast_fp16)[name = tensor("mh_w_29_cast_fp16")]; + tensor obj_69_cast_fp16 = softmax(axis = var_1032, x = mh_w_29_cast_fp16)[name = tensor("obj_69_cast_fp16")]; + tensor var_1189 = const()[name = tensor("op_1189"), val = tensor([1, 20, 64, -1])]; + tensor var_1190_cast_fp16 = reshape(shape = var_1189, x = value_19_cast_fp16)[name = tensor("op_1190_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_1190_cast_fp16, y = obj_69_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_1193 = const()[name = tensor("op_1193"), val = tensor([1, 1280, 1, -1])]; + tensor input_43_cast_fp16 = reshape(shape = var_1193, x = attn_19_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([1, 1])]; + tensor var_1199 = const()[name = tensor("op_1199"), val = tensor([1, 1])]; + tensor obj_67_pad_type_0 = const()[name = tensor("obj_67_pad_type_0"), val = tensor("custom")]; + tensor obj_67_pad_0 = const()[name = tensor("obj_67_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(366783808)))]; + 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(370060672)))]; + tensor obj_67_cast_fp16 = conv(bias = layers_4_encoder_attn_o_proj_bias_to_fp16, dilations = var_1199, groups = var_1039, pad = obj_67_pad_0, pad_type = obj_67_pad_type_0, strides = var_1197, 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 var_1205 = const()[name = tensor("op_1205"), val = tensor([1])]; + tensor channels_mean_29_cast_fp16 = reduce_mean(axes = var_1205, keep_dims = var_1040, x = inputs_29_cast_fp16)[name = tensor("channels_mean_29_cast_fp16")]; + tensor zero_mean_29_cast_fp16 = sub(x = inputs_29_cast_fp16, y = channels_mean_29_cast_fp16)[name = tensor("zero_mean_29_cast_fp16")]; + tensor zero_mean_sq_29_cast_fp16 = mul(x = zero_mean_29_cast_fp16, y = zero_mean_29_cast_fp16)[name = tensor("zero_mean_sq_29_cast_fp16")]; + tensor var_1209 = const()[name = tensor("op_1209"), val = tensor([1])]; + tensor var_1210_cast_fp16 = reduce_mean(axes = var_1209, keep_dims = var_1040, x = zero_mean_sq_29_cast_fp16)[name = tensor("op_1210_cast_fp16")]; + tensor var_1211_to_fp16 = const()[name = tensor("op_1211_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1212_cast_fp16 = add(x = var_1210_cast_fp16, y = var_1211_to_fp16)[name = tensor("op_1212_cast_fp16")]; + tensor denom_29_epsilon_0_to_fp16 = const()[name = tensor("denom_29_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_29_cast_fp16 = rsqrt(epsilon = denom_29_epsilon_0_to_fp16, x = var_1212_cast_fp16)[name = tensor("denom_29_cast_fp16")]; + tensor out_29_cast_fp16 = mul(x = zero_mean_29_cast_fp16, y = denom_29_cast_fp16)[name = tensor("out_29_cast_fp16")]; + tensor input_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(370063296)))]; + 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(370065920)))]; + 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 var_1223 = const()[name = tensor("op_1223"), val = tensor([1, 1])]; + tensor var_1225 = const()[name = tensor("op_1225"), val = tensor([1, 1])]; + tensor input_47_pad_type_0 = const()[name = tensor("input_47_pad_type_0"), val = tensor("custom")]; + tensor input_47_pad_0 = const()[name = tensor("input_47_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(370068544)))]; + 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(383175808)))]; + tensor input_47_cast_fp16 = conv(bias = layers_4_fc1_bias_to_fp16, dilations = var_1225, groups = var_1039, pad = input_47_pad_0, pad_type = input_47_pad_type_0, strides = var_1223, 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 var_1231 = const()[name = tensor("op_1231"), val = tensor([1, 1])]; + tensor var_1233 = const()[name = tensor("op_1233"), val = tensor([1, 1])]; + tensor hidden_states_11_pad_type_0 = const()[name = tensor("hidden_states_11_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_11_pad_0 = const()[name = tensor("hidden_states_11_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(383186112)))]; + 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(396293376)))]; + tensor hidden_states_11_cast_fp16 = conv(bias = layers_4_fc2_bias_to_fp16, dilations = var_1233, groups = var_1039, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = var_1231, 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_1246 = const()[name = tensor("op_1246"), val = tensor(3)]; + tensor var_1253 = const()[name = tensor("op_1253"), val = tensor(1)]; + tensor var_1254 = const()[name = tensor("op_1254"), val = tensor(true)]; + tensor var_1266 = const()[name = tensor("op_1266"), val = tensor([1])]; + tensor channels_mean_31_cast_fp16 = reduce_mean(axes = var_1266, keep_dims = var_1254, x = inputs_31_cast_fp16)[name = tensor("channels_mean_31_cast_fp16")]; + tensor zero_mean_31_cast_fp16 = sub(x = inputs_31_cast_fp16, y = channels_mean_31_cast_fp16)[name = tensor("zero_mean_31_cast_fp16")]; + tensor zero_mean_sq_31_cast_fp16 = mul(x = zero_mean_31_cast_fp16, y = zero_mean_31_cast_fp16)[name = tensor("zero_mean_sq_31_cast_fp16")]; + tensor var_1270 = const()[name = tensor("op_1270"), val = tensor([1])]; + tensor var_1271_cast_fp16 = reduce_mean(axes = var_1270, keep_dims = var_1254, x = zero_mean_sq_31_cast_fp16)[name = tensor("op_1271_cast_fp16")]; + tensor var_1272_to_fp16 = const()[name = tensor("op_1272_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1273_cast_fp16 = add(x = var_1271_cast_fp16, y = var_1272_to_fp16)[name = tensor("op_1273_cast_fp16")]; + tensor denom_31_epsilon_0_to_fp16 = const()[name = tensor("denom_31_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_31_cast_fp16 = rsqrt(epsilon = denom_31_epsilon_0_to_fp16, x = var_1273_cast_fp16)[name = tensor("denom_31_cast_fp16")]; + tensor out_31_cast_fp16 = mul(x = zero_mean_31_cast_fp16, y = denom_31_cast_fp16)[name = tensor("out_31_cast_fp16")]; + tensor obj_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(396296000)))]; + 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(396298624)))]; + 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 var_1288 = const()[name = tensor("op_1288"), val = tensor([1, 1])]; + tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([1, 1])]; + tensor query_21_pad_type_0 = const()[name = tensor("query_21_pad_type_0"), val = tensor("custom")]; + tensor query_21_pad_0 = const()[name = tensor("query_21_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(396301248)))]; + 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(399578112)))]; + tensor query_21_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_bias_to_fp16, dilations = var_1290, groups = var_1253, pad = query_21_pad_0, pad_type = query_21_pad_type_0, strides = var_1288, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("query_21_cast_fp16")]; + tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([1, 1])]; + tensor var_1296 = const()[name = tensor("op_1296"), val = tensor([1, 1])]; + tensor current_key_11_pad_type_0 = const()[name = tensor("current_key_11_pad_type_0"), val = tensor("custom")]; + tensor current_key_11_pad_0 = const()[name = tensor("current_key_11_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(399580736)))]; + tensor current_key_11_cast_fp16 = conv(dilations = var_1296, groups = var_1253, pad = current_key_11_pad_0, pad_type = current_key_11_pad_type_0, strides = var_1294, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("current_key_11_cast_fp16")]; + tensor var_1301 = const()[name = tensor("op_1301"), val = tensor([1, 1])]; + tensor var_1303 = const()[name = tensor("op_1303"), val = tensor([1, 1])]; + tensor current_value_11_pad_type_0 = const()[name = tensor("current_value_11_pad_type_0"), val = tensor("custom")]; + tensor current_value_11_pad_0 = const()[name = tensor("current_value_11_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(402857600)))]; + 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(406134464)))]; + tensor current_value_11_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_bias_to_fp16, dilations = var_1303, groups = var_1253, pad = current_value_11_pad_0, pad_type = current_value_11_pad_type_0, strides = var_1301, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("current_value_11_cast_fp16")]; + tensor var_1310_cast_fp16 = mul(x = current_key_11_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1310_cast_fp16")]; + tensor var_1312_cast_fp16 = mul(x = var_103_cast_fp16_5, y = var_241_cast_fp16)[name = tensor("op_1312_cast_fp16")]; + tensor key_21_cast_fp16 = add(x = var_1310_cast_fp16, y = var_1312_cast_fp16)[name = tensor("key_21_cast_fp16")]; + tensor var_1314_cast_fp16 = mul(x = current_value_11_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1314_cast_fp16")]; + tensor var_1316_cast_fp16 = mul(x = var_138_cast_fp16_5, y = var_241_cast_fp16)[name = tensor("op_1316_cast_fp16")]; + tensor value_21_cast_fp16 = add(x = var_1314_cast_fp16, y = var_1316_cast_fp16)[name = tensor("value_21_cast_fp16")]; + tensor var_1319 = const()[name = tensor("op_1319"), val = tensor([1, 20, 64, -1])]; + tensor var_1320_cast_fp16 = reshape(shape = var_1319, x = query_21_cast_fp16)[name = tensor("op_1320_cast_fp16")]; + tensor var_1321_to_fp16 = const()[name = tensor("op_1321_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1322_cast_fp16 = mul(x = var_1320_cast_fp16, y = var_1321_to_fp16)[name = tensor("op_1322_cast_fp16")]; + tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([1, 20, 64, -1])]; + tensor var_1324_cast_fp16 = reshape(shape = var_1323, x = key_21_cast_fp16)[name = tensor("op_1324_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_1322_cast_fp16, y = var_1324_cast_fp16)[name = tensor("mh_w_31_cast_fp16")]; + tensor mh_w_33_cast_fp16 = add(x = mh_w_31_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_33_cast_fp16")]; + tensor var_1332_cast_fp16 = softmax(axis = var_1246, x = mh_w_33_cast_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([1, 20, 64, -1])]; + tensor var_1334_cast_fp16 = reshape(shape = var_1333, x = value_21_cast_fp16)[name = tensor("op_1334_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_1334_cast_fp16, y = var_1332_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([1, 1280, 1, -1])]; + tensor input_51_cast_fp16 = reshape(shape = var_1337, x = attn_21_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor var_1341 = const()[name = tensor("op_1341"), val = tensor([1, 1])]; + tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([1, 1])]; + tensor obj_77_pad_type_0 = const()[name = tensor("obj_77_pad_type_0"), val = tensor("custom")]; + tensor obj_77_pad_0 = const()[name = tensor("obj_77_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(406137088)))]; + 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(409413952)))]; + tensor obj_77_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_bias_to_fp16, dilations = var_1343, groups = var_1253, pad = obj_77_pad_0, pad_type = obj_77_pad_type_0, strides = var_1341, 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 var_1353 = const()[name = tensor("op_1353"), val = tensor([1])]; + tensor channels_mean_33_cast_fp16 = reduce_mean(axes = var_1353, keep_dims = var_1254, x = inputs_33_cast_fp16)[name = tensor("channels_mean_33_cast_fp16")]; + tensor zero_mean_33_cast_fp16 = sub(x = inputs_33_cast_fp16, y = channels_mean_33_cast_fp16)[name = tensor("zero_mean_33_cast_fp16")]; + tensor zero_mean_sq_33_cast_fp16 = mul(x = zero_mean_33_cast_fp16, y = zero_mean_33_cast_fp16)[name = tensor("zero_mean_sq_33_cast_fp16")]; + tensor var_1357 = const()[name = tensor("op_1357"), val = tensor([1])]; + tensor var_1358_cast_fp16 = reduce_mean(axes = var_1357, keep_dims = var_1254, x = zero_mean_sq_33_cast_fp16)[name = tensor("op_1358_cast_fp16")]; + tensor var_1359_to_fp16 = const()[name = tensor("op_1359_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1360_cast_fp16 = add(x = var_1358_cast_fp16, y = var_1359_to_fp16)[name = tensor("op_1360_cast_fp16")]; + tensor denom_33_epsilon_0_to_fp16 = const()[name = tensor("denom_33_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_33_cast_fp16 = rsqrt(epsilon = denom_33_epsilon_0_to_fp16, x = var_1360_cast_fp16)[name = tensor("denom_33_cast_fp16")]; + tensor out_33_cast_fp16 = mul(x = zero_mean_33_cast_fp16, y = denom_33_cast_fp16)[name = tensor("out_33_cast_fp16")]; + tensor obj_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(409416576)))]; + 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(409419200)))]; + 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 var_1375 = const()[name = tensor("op_1375"), val = tensor([1, 1])]; + tensor var_1377 = const()[name = tensor("op_1377"), val = tensor([1, 1])]; + tensor query_23_pad_type_0 = const()[name = tensor("query_23_pad_type_0"), val = tensor("custom")]; + tensor query_23_pad_0 = const()[name = tensor("query_23_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(409421824)))]; + 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(412698688)))]; + tensor query_23_cast_fp16 = conv(bias = layers_5_encoder_attn_q_proj_bias_to_fp16, dilations = var_1377, groups = var_1253, pad = query_23_pad_0, pad_type = query_23_pad_type_0, strides = var_1375, weight = layers_5_encoder_attn_q_proj_weight_to_fp16, x = obj_79_cast_fp16)[name = tensor("query_23_cast_fp16")]; + tensor var_1381 = const()[name = tensor("op_1381"), val = tensor([1, 1])]; + tensor var_1383 = const()[name = tensor("op_1383"), val = tensor([1, 1])]; + tensor key_23_pad_type_0 = const()[name = tensor("key_23_pad_type_0"), val = tensor("custom")]; + tensor key_23_pad_0 = const()[name = tensor("key_23_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(412701312)))]; + tensor key_23_cast_fp16 = conv(dilations = var_1383, groups = var_1253, pad = key_23_pad_0, pad_type = key_23_pad_type_0, strides = var_1381, weight = layers_5_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_23_cast_fp16")]; + tensor var_1388 = const()[name = tensor("op_1388"), val = tensor([1, 1])]; + tensor var_1390 = const()[name = tensor("op_1390"), val = tensor([1, 1])]; + tensor value_23_pad_type_0 = const()[name = tensor("value_23_pad_type_0"), val = tensor("custom")]; + tensor value_23_pad_0 = const()[name = tensor("value_23_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(415978176)))]; + 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(419255040)))]; + tensor value_23_cast_fp16 = conv(bias = layers_5_encoder_attn_v_proj_bias_to_fp16, dilations = var_1390, groups = var_1253, pad = value_23_pad_0, pad_type = value_23_pad_type_0, strides = var_1388, weight = layers_5_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_23_cast_fp16")]; + tensor var_1394 = const()[name = tensor("op_1394"), val = tensor([1, 20, 64, -1])]; + tensor var_1395_cast_fp16 = reshape(shape = var_1394, x = query_23_cast_fp16)[name = tensor("op_1395_cast_fp16")]; + tensor var_1396_to_fp16 = const()[name = tensor("op_1396_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1397_cast_fp16 = mul(x = var_1395_cast_fp16, y = var_1396_to_fp16)[name = tensor("op_1397_cast_fp16")]; + tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([1, 20, 64, -1])]; + tensor var_1399_cast_fp16 = reshape(shape = var_1398, x = key_23_cast_fp16)[name = tensor("op_1399_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_1397_cast_fp16, y = var_1399_cast_fp16)[name = tensor("mh_w_35_cast_fp16")]; + tensor obj_83_cast_fp16 = softmax(axis = var_1246, x = mh_w_35_cast_fp16)[name = tensor("obj_83_cast_fp16")]; + tensor var_1403 = const()[name = tensor("op_1403"), val = tensor([1, 20, 64, -1])]; + tensor var_1404_cast_fp16 = reshape(shape = var_1403, x = value_23_cast_fp16)[name = tensor("op_1404_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_1404_cast_fp16, y = obj_83_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_1407 = const()[name = tensor("op_1407"), val = tensor([1, 1280, 1, -1])]; + tensor input_53_cast_fp16 = reshape(shape = var_1407, x = attn_23_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor var_1411 = const()[name = tensor("op_1411"), val = tensor([1, 1])]; + tensor var_1413 = const()[name = tensor("op_1413"), val = tensor([1, 1])]; + tensor obj_81_pad_type_0 = const()[name = tensor("obj_81_pad_type_0"), val = tensor("custom")]; + tensor obj_81_pad_0 = const()[name = tensor("obj_81_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(419257664)))]; + 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(422534528)))]; + tensor obj_81_cast_fp16 = conv(bias = layers_5_encoder_attn_o_proj_bias_to_fp16, dilations = var_1413, groups = var_1253, pad = obj_81_pad_0, pad_type = obj_81_pad_type_0, strides = var_1411, 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 var_1419 = const()[name = tensor("op_1419"), val = tensor([1])]; + tensor channels_mean_35_cast_fp16 = reduce_mean(axes = var_1419, keep_dims = var_1254, x = inputs_35_cast_fp16)[name = tensor("channels_mean_35_cast_fp16")]; + tensor zero_mean_35_cast_fp16 = sub(x = inputs_35_cast_fp16, y = channels_mean_35_cast_fp16)[name = tensor("zero_mean_35_cast_fp16")]; + tensor zero_mean_sq_35_cast_fp16 = mul(x = zero_mean_35_cast_fp16, y = zero_mean_35_cast_fp16)[name = tensor("zero_mean_sq_35_cast_fp16")]; + tensor var_1423 = const()[name = tensor("op_1423"), val = tensor([1])]; + tensor var_1424_cast_fp16 = reduce_mean(axes = var_1423, keep_dims = var_1254, x = zero_mean_sq_35_cast_fp16)[name = tensor("op_1424_cast_fp16")]; + tensor var_1425_to_fp16 = const()[name = tensor("op_1425_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1426_cast_fp16 = add(x = var_1424_cast_fp16, y = var_1425_to_fp16)[name = tensor("op_1426_cast_fp16")]; + tensor denom_35_epsilon_0_to_fp16 = const()[name = tensor("denom_35_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_35_cast_fp16 = rsqrt(epsilon = denom_35_epsilon_0_to_fp16, x = var_1426_cast_fp16)[name = tensor("denom_35_cast_fp16")]; + tensor out_35_cast_fp16 = mul(x = zero_mean_35_cast_fp16, y = denom_35_cast_fp16)[name = tensor("out_35_cast_fp16")]; + tensor input_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(422537152)))]; + 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(422539776)))]; + 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 var_1437 = const()[name = tensor("op_1437"), val = tensor([1, 1])]; + tensor var_1439 = const()[name = tensor("op_1439"), val = tensor([1, 1])]; + tensor input_57_pad_type_0 = const()[name = tensor("input_57_pad_type_0"), val = tensor("custom")]; + tensor input_57_pad_0 = const()[name = tensor("input_57_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(422542400)))]; + 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(435649664)))]; + tensor input_57_cast_fp16 = conv(bias = layers_5_fc1_bias_to_fp16, dilations = var_1439, groups = var_1253, pad = input_57_pad_0, pad_type = input_57_pad_type_0, strides = var_1437, 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 var_1445 = const()[name = tensor("op_1445"), val = tensor([1, 1])]; + tensor var_1447 = const()[name = tensor("op_1447"), val = tensor([1, 1])]; + tensor hidden_states_13_pad_type_0 = const()[name = tensor("hidden_states_13_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_13_pad_0 = const()[name = tensor("hidden_states_13_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(435659968)))]; + 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(448767232)))]; + tensor hidden_states_13_cast_fp16 = conv(bias = layers_5_fc2_bias_to_fp16, dilations = var_1447, groups = var_1253, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = var_1445, 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_1460 = const()[name = tensor("op_1460"), val = tensor(3)]; + tensor var_1467 = const()[name = tensor("op_1467"), val = tensor(1)]; + tensor var_1468 = const()[name = tensor("op_1468"), val = tensor(true)]; + tensor var_1480 = const()[name = tensor("op_1480"), val = tensor([1])]; + tensor channels_mean_37_cast_fp16 = reduce_mean(axes = var_1480, keep_dims = var_1468, x = inputs_37_cast_fp16)[name = tensor("channels_mean_37_cast_fp16")]; + tensor zero_mean_37_cast_fp16 = sub(x = inputs_37_cast_fp16, y = channels_mean_37_cast_fp16)[name = tensor("zero_mean_37_cast_fp16")]; + tensor zero_mean_sq_37_cast_fp16 = mul(x = zero_mean_37_cast_fp16, y = zero_mean_37_cast_fp16)[name = tensor("zero_mean_sq_37_cast_fp16")]; + tensor var_1484 = const()[name = tensor("op_1484"), val = tensor([1])]; + tensor var_1485_cast_fp16 = reduce_mean(axes = var_1484, keep_dims = var_1468, x = zero_mean_sq_37_cast_fp16)[name = tensor("op_1485_cast_fp16")]; + tensor var_1486_to_fp16 = const()[name = tensor("op_1486_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1487_cast_fp16 = add(x = var_1485_cast_fp16, y = var_1486_to_fp16)[name = tensor("op_1487_cast_fp16")]; + tensor denom_37_epsilon_0_to_fp16 = const()[name = tensor("denom_37_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_37_cast_fp16 = rsqrt(epsilon = denom_37_epsilon_0_to_fp16, x = var_1487_cast_fp16)[name = tensor("denom_37_cast_fp16")]; + tensor out_37_cast_fp16 = mul(x = zero_mean_37_cast_fp16, y = denom_37_cast_fp16)[name = tensor("out_37_cast_fp16")]; + tensor obj_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(448769856)))]; + 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(448772480)))]; + 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 var_1502 = const()[name = tensor("op_1502"), val = tensor([1, 1])]; + tensor var_1504 = const()[name = tensor("op_1504"), val = tensor([1, 1])]; + tensor query_25_pad_type_0 = const()[name = tensor("query_25_pad_type_0"), val = tensor("custom")]; + tensor query_25_pad_0 = const()[name = tensor("query_25_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(448775104)))]; + 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(452051968)))]; + tensor query_25_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_bias_to_fp16, dilations = var_1504, groups = var_1467, pad = query_25_pad_0, pad_type = query_25_pad_type_0, strides = var_1502, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("query_25_cast_fp16")]; + tensor var_1508 = const()[name = tensor("op_1508"), val = tensor([1, 1])]; + tensor var_1510 = const()[name = tensor("op_1510"), val = tensor([1, 1])]; + tensor current_key_13_pad_type_0 = const()[name = tensor("current_key_13_pad_type_0"), val = tensor("custom")]; + tensor current_key_13_pad_0 = const()[name = tensor("current_key_13_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(452054592)))]; + tensor current_key_13_cast_fp16 = conv(dilations = var_1510, groups = var_1467, pad = current_key_13_pad_0, pad_type = current_key_13_pad_type_0, strides = var_1508, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("current_key_13_cast_fp16")]; + tensor var_1515 = const()[name = tensor("op_1515"), val = tensor([1, 1])]; + tensor var_1517 = const()[name = tensor("op_1517"), val = tensor([1, 1])]; + tensor current_value_13_pad_type_0 = const()[name = tensor("current_value_13_pad_type_0"), val = tensor("custom")]; + tensor current_value_13_pad_0 = const()[name = tensor("current_value_13_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(455331456)))]; + 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(458608320)))]; + tensor current_value_13_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_bias_to_fp16, dilations = var_1517, groups = var_1467, pad = current_value_13_pad_0, pad_type = current_value_13_pad_type_0, strides = var_1515, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("current_value_13_cast_fp16")]; + tensor var_1524_cast_fp16 = mul(x = current_key_13_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1524_cast_fp16")]; + tensor var_1526_cast_fp16 = mul(x = var_103_cast_fp16_6, y = var_241_cast_fp16)[name = tensor("op_1526_cast_fp16")]; + tensor key_25_cast_fp16 = add(x = var_1524_cast_fp16, y = var_1526_cast_fp16)[name = tensor("key_25_cast_fp16")]; + tensor var_1528_cast_fp16 = mul(x = current_value_13_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1528_cast_fp16")]; + tensor var_1530_cast_fp16 = mul(x = var_138_cast_fp16_6, y = var_241_cast_fp16)[name = tensor("op_1530_cast_fp16")]; + tensor value_25_cast_fp16 = add(x = var_1528_cast_fp16, y = var_1530_cast_fp16)[name = tensor("value_25_cast_fp16")]; + tensor var_1533 = const()[name = tensor("op_1533"), val = tensor([1, 20, 64, -1])]; + tensor var_1534_cast_fp16 = reshape(shape = var_1533, x = query_25_cast_fp16)[name = tensor("op_1534_cast_fp16")]; + tensor var_1535_to_fp16 = const()[name = tensor("op_1535_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1536_cast_fp16 = mul(x = var_1534_cast_fp16, y = var_1535_to_fp16)[name = tensor("op_1536_cast_fp16")]; + tensor var_1537 = const()[name = tensor("op_1537"), val = tensor([1, 20, 64, -1])]; + tensor var_1538_cast_fp16 = reshape(shape = var_1537, x = key_25_cast_fp16)[name = tensor("op_1538_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_1536_cast_fp16, y = var_1538_cast_fp16)[name = tensor("mh_w_37_cast_fp16")]; + tensor mh_w_39_cast_fp16 = add(x = mh_w_37_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_39_cast_fp16")]; + tensor var_1546_cast_fp16 = softmax(axis = var_1460, x = mh_w_39_cast_fp16)[name = tensor("op_1546_cast_fp16")]; + tensor var_1547 = const()[name = tensor("op_1547"), val = tensor([1, 20, 64, -1])]; + tensor var_1548_cast_fp16 = reshape(shape = var_1547, x = value_25_cast_fp16)[name = tensor("op_1548_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_1548_cast_fp16, y = var_1546_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_1551 = const()[name = tensor("op_1551"), val = tensor([1, 1280, 1, -1])]; + tensor input_61_cast_fp16 = reshape(shape = var_1551, x = attn_25_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor var_1555 = const()[name = tensor("op_1555"), val = tensor([1, 1])]; + tensor var_1557 = const()[name = tensor("op_1557"), val = tensor([1, 1])]; + tensor obj_91_pad_type_0 = const()[name = tensor("obj_91_pad_type_0"), val = tensor("custom")]; + tensor obj_91_pad_0 = const()[name = tensor("obj_91_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(458610944)))]; + 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(461887808)))]; + tensor obj_91_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_bias_to_fp16, dilations = var_1557, groups = var_1467, pad = obj_91_pad_0, pad_type = obj_91_pad_type_0, strides = var_1555, 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 var_1567 = const()[name = tensor("op_1567"), val = tensor([1])]; + tensor channels_mean_39_cast_fp16 = reduce_mean(axes = var_1567, keep_dims = var_1468, x = inputs_39_cast_fp16)[name = tensor("channels_mean_39_cast_fp16")]; + tensor zero_mean_39_cast_fp16 = sub(x = inputs_39_cast_fp16, y = channels_mean_39_cast_fp16)[name = tensor("zero_mean_39_cast_fp16")]; + tensor zero_mean_sq_39_cast_fp16 = mul(x = zero_mean_39_cast_fp16, y = zero_mean_39_cast_fp16)[name = tensor("zero_mean_sq_39_cast_fp16")]; + tensor var_1571 = const()[name = tensor("op_1571"), val = tensor([1])]; + tensor var_1572_cast_fp16 = reduce_mean(axes = var_1571, keep_dims = var_1468, x = zero_mean_sq_39_cast_fp16)[name = tensor("op_1572_cast_fp16")]; + tensor var_1573_to_fp16 = const()[name = tensor("op_1573_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1574_cast_fp16 = add(x = var_1572_cast_fp16, y = var_1573_to_fp16)[name = tensor("op_1574_cast_fp16")]; + tensor denom_39_epsilon_0_to_fp16 = const()[name = tensor("denom_39_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_39_cast_fp16 = rsqrt(epsilon = denom_39_epsilon_0_to_fp16, x = var_1574_cast_fp16)[name = tensor("denom_39_cast_fp16")]; + tensor out_39_cast_fp16 = mul(x = zero_mean_39_cast_fp16, y = denom_39_cast_fp16)[name = tensor("out_39_cast_fp16")]; + tensor obj_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(461890432)))]; + 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(461893056)))]; + 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 var_1589 = const()[name = tensor("op_1589"), val = tensor([1, 1])]; + tensor var_1591 = const()[name = tensor("op_1591"), val = tensor([1, 1])]; + tensor query_27_pad_type_0 = const()[name = tensor("query_27_pad_type_0"), val = tensor("custom")]; + tensor query_27_pad_0 = const()[name = tensor("query_27_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(461895680)))]; + 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(465172544)))]; + tensor query_27_cast_fp16 = conv(bias = layers_6_encoder_attn_q_proj_bias_to_fp16, dilations = var_1591, groups = var_1467, pad = query_27_pad_0, pad_type = query_27_pad_type_0, strides = var_1589, weight = layers_6_encoder_attn_q_proj_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor("query_27_cast_fp16")]; + tensor var_1595 = const()[name = tensor("op_1595"), val = tensor([1, 1])]; + tensor var_1597 = const()[name = tensor("op_1597"), val = tensor([1, 1])]; + tensor key_27_pad_type_0 = const()[name = tensor("key_27_pad_type_0"), val = tensor("custom")]; + tensor key_27_pad_0 = const()[name = tensor("key_27_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(465175168)))]; + tensor key_27_cast_fp16 = conv(dilations = var_1597, groups = var_1467, pad = key_27_pad_0, pad_type = key_27_pad_type_0, strides = var_1595, weight = layers_6_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_27_cast_fp16")]; + tensor var_1602 = const()[name = tensor("op_1602"), val = tensor([1, 1])]; + tensor var_1604 = const()[name = tensor("op_1604"), val = tensor([1, 1])]; + tensor value_27_pad_type_0 = const()[name = tensor("value_27_pad_type_0"), val = tensor("custom")]; + tensor value_27_pad_0 = const()[name = tensor("value_27_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(468452032)))]; + 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(471728896)))]; + tensor value_27_cast_fp16 = conv(bias = layers_6_encoder_attn_v_proj_bias_to_fp16, dilations = var_1604, groups = var_1467, pad = value_27_pad_0, pad_type = value_27_pad_type_0, strides = var_1602, weight = layers_6_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_27_cast_fp16")]; + tensor var_1608 = const()[name = tensor("op_1608"), val = tensor([1, 20, 64, -1])]; + tensor var_1609_cast_fp16 = reshape(shape = var_1608, x = query_27_cast_fp16)[name = tensor("op_1609_cast_fp16")]; + tensor var_1610_to_fp16 = const()[name = tensor("op_1610_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1611_cast_fp16 = mul(x = var_1609_cast_fp16, y = var_1610_to_fp16)[name = tensor("op_1611_cast_fp16")]; + tensor var_1612 = const()[name = tensor("op_1612"), val = tensor([1, 20, 64, -1])]; + tensor var_1613_cast_fp16 = reshape(shape = var_1612, x = key_27_cast_fp16)[name = tensor("op_1613_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_1611_cast_fp16, y = var_1613_cast_fp16)[name = tensor("mh_w_41_cast_fp16")]; + tensor obj_97_cast_fp16 = softmax(axis = var_1460, x = mh_w_41_cast_fp16)[name = tensor("obj_97_cast_fp16")]; + tensor var_1617 = const()[name = tensor("op_1617"), val = tensor([1, 20, 64, -1])]; + tensor var_1618_cast_fp16 = reshape(shape = var_1617, x = value_27_cast_fp16)[name = tensor("op_1618_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_1618_cast_fp16, y = obj_97_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_1621 = const()[name = tensor("op_1621"), val = tensor([1, 1280, 1, -1])]; + tensor input_63_cast_fp16 = reshape(shape = var_1621, x = attn_27_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor var_1625 = const()[name = tensor("op_1625"), val = tensor([1, 1])]; + tensor var_1627 = const()[name = tensor("op_1627"), val = tensor([1, 1])]; + tensor obj_95_pad_type_0 = const()[name = tensor("obj_95_pad_type_0"), val = tensor("custom")]; + tensor obj_95_pad_0 = const()[name = tensor("obj_95_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(471731520)))]; + 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(475008384)))]; + tensor obj_95_cast_fp16 = conv(bias = layers_6_encoder_attn_o_proj_bias_to_fp16, dilations = var_1627, groups = var_1467, pad = obj_95_pad_0, pad_type = obj_95_pad_type_0, strides = var_1625, 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 var_1633 = const()[name = tensor("op_1633"), val = tensor([1])]; + tensor channels_mean_41_cast_fp16 = reduce_mean(axes = var_1633, keep_dims = var_1468, x = inputs_41_cast_fp16)[name = tensor("channels_mean_41_cast_fp16")]; + tensor zero_mean_41_cast_fp16 = sub(x = inputs_41_cast_fp16, y = channels_mean_41_cast_fp16)[name = tensor("zero_mean_41_cast_fp16")]; + tensor zero_mean_sq_41_cast_fp16 = mul(x = zero_mean_41_cast_fp16, y = zero_mean_41_cast_fp16)[name = tensor("zero_mean_sq_41_cast_fp16")]; + tensor var_1637 = const()[name = tensor("op_1637"), val = tensor([1])]; + tensor var_1638_cast_fp16 = reduce_mean(axes = var_1637, keep_dims = var_1468, x = zero_mean_sq_41_cast_fp16)[name = tensor("op_1638_cast_fp16")]; + tensor var_1639_to_fp16 = const()[name = tensor("op_1639_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1640_cast_fp16 = add(x = var_1638_cast_fp16, y = var_1639_to_fp16)[name = tensor("op_1640_cast_fp16")]; + tensor denom_41_epsilon_0_to_fp16 = const()[name = tensor("denom_41_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_41_cast_fp16 = rsqrt(epsilon = denom_41_epsilon_0_to_fp16, x = var_1640_cast_fp16)[name = tensor("denom_41_cast_fp16")]; + tensor out_41_cast_fp16 = mul(x = zero_mean_41_cast_fp16, y = denom_41_cast_fp16)[name = tensor("out_41_cast_fp16")]; + tensor input_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(475011008)))]; + 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(475013632)))]; + 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 var_1651 = const()[name = tensor("op_1651"), val = tensor([1, 1])]; + tensor var_1653 = const()[name = tensor("op_1653"), val = tensor([1, 1])]; + tensor input_67_pad_type_0 = const()[name = tensor("input_67_pad_type_0"), val = tensor("custom")]; + tensor input_67_pad_0 = const()[name = tensor("input_67_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(475016256)))]; + 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(488123520)))]; + tensor input_67_cast_fp16 = conv(bias = layers_6_fc1_bias_to_fp16, dilations = var_1653, groups = var_1467, pad = input_67_pad_0, pad_type = input_67_pad_type_0, strides = var_1651, 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 var_1659 = const()[name = tensor("op_1659"), val = tensor([1, 1])]; + tensor var_1661 = const()[name = tensor("op_1661"), val = tensor([1, 1])]; + tensor hidden_states_15_pad_type_0 = const()[name = tensor("hidden_states_15_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_15_pad_0 = const()[name = tensor("hidden_states_15_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(488133824)))]; + 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(501241088)))]; + tensor hidden_states_15_cast_fp16 = conv(bias = layers_6_fc2_bias_to_fp16, dilations = var_1661, groups = var_1467, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = var_1659, 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_1674 = const()[name = tensor("op_1674"), val = tensor(3)]; + tensor var_1681 = const()[name = tensor("op_1681"), val = tensor(1)]; + tensor var_1682 = const()[name = tensor("op_1682"), val = tensor(true)]; + tensor var_1694 = const()[name = tensor("op_1694"), val = tensor([1])]; + tensor channels_mean_43_cast_fp16 = reduce_mean(axes = var_1694, keep_dims = var_1682, x = inputs_43_cast_fp16)[name = tensor("channels_mean_43_cast_fp16")]; + tensor zero_mean_43_cast_fp16 = sub(x = inputs_43_cast_fp16, y = channels_mean_43_cast_fp16)[name = tensor("zero_mean_43_cast_fp16")]; + tensor zero_mean_sq_43_cast_fp16 = mul(x = zero_mean_43_cast_fp16, y = zero_mean_43_cast_fp16)[name = tensor("zero_mean_sq_43_cast_fp16")]; + tensor var_1698 = const()[name = tensor("op_1698"), val = tensor([1])]; + tensor var_1699_cast_fp16 = reduce_mean(axes = var_1698, keep_dims = var_1682, x = zero_mean_sq_43_cast_fp16)[name = tensor("op_1699_cast_fp16")]; + tensor var_1700_to_fp16 = const()[name = tensor("op_1700_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1701_cast_fp16 = add(x = var_1699_cast_fp16, y = var_1700_to_fp16)[name = tensor("op_1701_cast_fp16")]; + tensor denom_43_epsilon_0_to_fp16 = const()[name = tensor("denom_43_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_43_cast_fp16 = rsqrt(epsilon = denom_43_epsilon_0_to_fp16, x = var_1701_cast_fp16)[name = tensor("denom_43_cast_fp16")]; + tensor out_43_cast_fp16 = mul(x = zero_mean_43_cast_fp16, y = denom_43_cast_fp16)[name = tensor("out_43_cast_fp16")]; + tensor obj_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(501243712)))]; + 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(501246336)))]; + 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 var_1716 = const()[name = tensor("op_1716"), val = tensor([1, 1])]; + tensor var_1718 = const()[name = tensor("op_1718"), val = tensor([1, 1])]; + tensor query_29_pad_type_0 = const()[name = tensor("query_29_pad_type_0"), val = tensor("custom")]; + tensor query_29_pad_0 = const()[name = tensor("query_29_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(501248960)))]; + 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(504525824)))]; + tensor query_29_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_bias_to_fp16, dilations = var_1718, groups = var_1681, pad = query_29_pad_0, pad_type = query_29_pad_type_0, strides = var_1716, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("query_29_cast_fp16")]; + tensor var_1722 = const()[name = tensor("op_1722"), val = tensor([1, 1])]; + tensor var_1724 = const()[name = tensor("op_1724"), val = tensor([1, 1])]; + tensor current_key_15_pad_type_0 = const()[name = tensor("current_key_15_pad_type_0"), val = tensor("custom")]; + tensor current_key_15_pad_0 = const()[name = tensor("current_key_15_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(504528448)))]; + tensor current_key_15_cast_fp16 = conv(dilations = var_1724, groups = var_1681, pad = current_key_15_pad_0, pad_type = current_key_15_pad_type_0, strides = var_1722, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("current_key_15_cast_fp16")]; + tensor var_1729 = const()[name = tensor("op_1729"), val = tensor([1, 1])]; + tensor var_1731 = const()[name = tensor("op_1731"), val = tensor([1, 1])]; + tensor current_value_15_pad_type_0 = const()[name = tensor("current_value_15_pad_type_0"), val = tensor("custom")]; + tensor current_value_15_pad_0 = const()[name = tensor("current_value_15_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(507805312)))]; + 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(511082176)))]; + tensor current_value_15_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_bias_to_fp16, dilations = var_1731, groups = var_1681, pad = current_value_15_pad_0, pad_type = current_value_15_pad_type_0, strides = var_1729, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("current_value_15_cast_fp16")]; + tensor var_1738_cast_fp16 = mul(x = current_key_15_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1738_cast_fp16")]; + tensor var_1740_cast_fp16 = mul(x = var_103_cast_fp16_7, y = var_241_cast_fp16)[name = tensor("op_1740_cast_fp16")]; + tensor key_29_cast_fp16 = add(x = var_1738_cast_fp16, y = var_1740_cast_fp16)[name = tensor("key_29_cast_fp16")]; + tensor var_1742_cast_fp16 = mul(x = current_value_15_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1742_cast_fp16")]; + tensor var_1744_cast_fp16 = mul(x = var_138_cast_fp16_7, y = var_241_cast_fp16)[name = tensor("op_1744_cast_fp16")]; + tensor value_29_cast_fp16 = add(x = var_1742_cast_fp16, y = var_1744_cast_fp16)[name = tensor("value_29_cast_fp16")]; + tensor var_1747 = const()[name = tensor("op_1747"), val = tensor([1, 20, 64, -1])]; + tensor var_1748_cast_fp16 = reshape(shape = var_1747, x = query_29_cast_fp16)[name = tensor("op_1748_cast_fp16")]; + tensor var_1749_to_fp16 = const()[name = tensor("op_1749_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1750_cast_fp16 = mul(x = var_1748_cast_fp16, y = var_1749_to_fp16)[name = tensor("op_1750_cast_fp16")]; + tensor var_1751 = const()[name = tensor("op_1751"), val = tensor([1, 20, 64, -1])]; + tensor var_1752_cast_fp16 = reshape(shape = var_1751, x = key_29_cast_fp16)[name = tensor("op_1752_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_1750_cast_fp16, y = var_1752_cast_fp16)[name = tensor("mh_w_43_cast_fp16")]; + tensor mh_w_45_cast_fp16 = add(x = mh_w_43_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_45_cast_fp16")]; + tensor var_1760_cast_fp16 = softmax(axis = var_1674, x = mh_w_45_cast_fp16)[name = tensor("op_1760_cast_fp16")]; + tensor var_1761 = const()[name = tensor("op_1761"), val = tensor([1, 20, 64, -1])]; + tensor var_1762_cast_fp16 = reshape(shape = var_1761, x = value_29_cast_fp16)[name = tensor("op_1762_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_1762_cast_fp16, y = var_1760_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_1765 = const()[name = tensor("op_1765"), val = tensor([1, 1280, 1, -1])]; + tensor input_71_cast_fp16 = reshape(shape = var_1765, x = attn_29_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor var_1769 = const()[name = tensor("op_1769"), val = tensor([1, 1])]; + tensor var_1771 = const()[name = tensor("op_1771"), val = tensor([1, 1])]; + tensor obj_105_pad_type_0 = const()[name = tensor("obj_105_pad_type_0"), val = tensor("custom")]; + tensor obj_105_pad_0 = const()[name = tensor("obj_105_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(511084800)))]; + 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(514361664)))]; + tensor obj_105_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_bias_to_fp16, dilations = var_1771, groups = var_1681, pad = obj_105_pad_0, pad_type = obj_105_pad_type_0, strides = var_1769, 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 var_1781 = const()[name = tensor("op_1781"), val = tensor([1])]; + tensor channels_mean_45_cast_fp16 = reduce_mean(axes = var_1781, keep_dims = var_1682, x = inputs_45_cast_fp16)[name = tensor("channels_mean_45_cast_fp16")]; + tensor zero_mean_45_cast_fp16 = sub(x = inputs_45_cast_fp16, y = channels_mean_45_cast_fp16)[name = tensor("zero_mean_45_cast_fp16")]; + tensor zero_mean_sq_45_cast_fp16 = mul(x = zero_mean_45_cast_fp16, y = zero_mean_45_cast_fp16)[name = tensor("zero_mean_sq_45_cast_fp16")]; + tensor var_1785 = const()[name = tensor("op_1785"), val = tensor([1])]; + tensor var_1786_cast_fp16 = reduce_mean(axes = var_1785, keep_dims = var_1682, x = zero_mean_sq_45_cast_fp16)[name = tensor("op_1786_cast_fp16")]; + tensor var_1787_to_fp16 = const()[name = tensor("op_1787_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1788_cast_fp16 = add(x = var_1786_cast_fp16, y = var_1787_to_fp16)[name = tensor("op_1788_cast_fp16")]; + tensor denom_45_epsilon_0_to_fp16 = const()[name = tensor("denom_45_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_45_cast_fp16 = rsqrt(epsilon = denom_45_epsilon_0_to_fp16, x = var_1788_cast_fp16)[name = tensor("denom_45_cast_fp16")]; + tensor out_45_cast_fp16 = mul(x = zero_mean_45_cast_fp16, y = denom_45_cast_fp16)[name = tensor("out_45_cast_fp16")]; + tensor obj_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(514364288)))]; + 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(514366912)))]; + 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 var_1803 = const()[name = tensor("op_1803"), val = tensor([1, 1])]; + tensor var_1805 = const()[name = tensor("op_1805"), val = tensor([1, 1])]; + tensor query_31_pad_type_0 = const()[name = tensor("query_31_pad_type_0"), val = tensor("custom")]; + tensor query_31_pad_0 = const()[name = tensor("query_31_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(514369536)))]; + 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(517646400)))]; + tensor query_31_cast_fp16 = conv(bias = layers_7_encoder_attn_q_proj_bias_to_fp16, dilations = var_1805, groups = var_1681, pad = query_31_pad_0, pad_type = query_31_pad_type_0, strides = var_1803, weight = layers_7_encoder_attn_q_proj_weight_to_fp16, x = obj_107_cast_fp16)[name = tensor("query_31_cast_fp16")]; + tensor var_1809 = const()[name = tensor("op_1809"), val = tensor([1, 1])]; + tensor var_1811 = const()[name = tensor("op_1811"), val = tensor([1, 1])]; + tensor key_31_pad_type_0 = const()[name = tensor("key_31_pad_type_0"), val = tensor("custom")]; + tensor key_31_pad_0 = const()[name = tensor("key_31_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(517649024)))]; + tensor key_31_cast_fp16 = conv(dilations = var_1811, groups = var_1681, pad = key_31_pad_0, pad_type = key_31_pad_type_0, strides = var_1809, weight = layers_7_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_31_cast_fp16")]; + tensor var_1816 = const()[name = tensor("op_1816"), val = tensor([1, 1])]; + tensor var_1818 = const()[name = tensor("op_1818"), val = tensor([1, 1])]; + tensor value_31_pad_type_0 = const()[name = tensor("value_31_pad_type_0"), val = tensor("custom")]; + tensor value_31_pad_0 = const()[name = tensor("value_31_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(520925888)))]; + 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(524202752)))]; + tensor value_31_cast_fp16 = conv(bias = layers_7_encoder_attn_v_proj_bias_to_fp16, dilations = var_1818, groups = var_1681, pad = value_31_pad_0, pad_type = value_31_pad_type_0, strides = var_1816, weight = layers_7_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_31_cast_fp16")]; + tensor var_1822 = const()[name = tensor("op_1822"), val = tensor([1, 20, 64, -1])]; + tensor var_1823_cast_fp16 = reshape(shape = var_1822, x = query_31_cast_fp16)[name = tensor("op_1823_cast_fp16")]; + tensor var_1824_to_fp16 = const()[name = tensor("op_1824_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1825_cast_fp16 = mul(x = var_1823_cast_fp16, y = var_1824_to_fp16)[name = tensor("op_1825_cast_fp16")]; + tensor var_1826 = const()[name = tensor("op_1826"), val = tensor([1, 20, 64, -1])]; + tensor var_1827_cast_fp16 = reshape(shape = var_1826, x = key_31_cast_fp16)[name = tensor("op_1827_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_1825_cast_fp16, y = var_1827_cast_fp16)[name = tensor("mh_w_47_cast_fp16")]; + tensor obj_111_cast_fp16 = softmax(axis = var_1674, x = mh_w_47_cast_fp16)[name = tensor("obj_111_cast_fp16")]; + tensor var_1831 = const()[name = tensor("op_1831"), val = tensor([1, 20, 64, -1])]; + tensor var_1832_cast_fp16 = reshape(shape = var_1831, x = value_31_cast_fp16)[name = tensor("op_1832_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_1832_cast_fp16, y = obj_111_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_1835 = const()[name = tensor("op_1835"), val = tensor([1, 1280, 1, -1])]; + tensor input_73_cast_fp16 = reshape(shape = var_1835, x = attn_31_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor var_1839 = const()[name = tensor("op_1839"), val = tensor([1, 1])]; + tensor var_1841 = const()[name = tensor("op_1841"), val = tensor([1, 1])]; + tensor obj_109_pad_type_0 = const()[name = tensor("obj_109_pad_type_0"), val = tensor("custom")]; + tensor obj_109_pad_0 = const()[name = tensor("obj_109_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(524205376)))]; + 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(527482240)))]; + tensor obj_109_cast_fp16 = conv(bias = layers_7_encoder_attn_o_proj_bias_to_fp16, dilations = var_1841, groups = var_1681, pad = obj_109_pad_0, pad_type = obj_109_pad_type_0, strides = var_1839, 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 var_1847 = const()[name = tensor("op_1847"), val = tensor([1])]; + tensor channels_mean_47_cast_fp16 = reduce_mean(axes = var_1847, keep_dims = var_1682, x = inputs_47_cast_fp16)[name = tensor("channels_mean_47_cast_fp16")]; + tensor zero_mean_47_cast_fp16 = sub(x = inputs_47_cast_fp16, y = channels_mean_47_cast_fp16)[name = tensor("zero_mean_47_cast_fp16")]; + tensor zero_mean_sq_47_cast_fp16 = mul(x = zero_mean_47_cast_fp16, y = zero_mean_47_cast_fp16)[name = tensor("zero_mean_sq_47_cast_fp16")]; + tensor var_1851 = const()[name = tensor("op_1851"), val = tensor([1])]; + tensor var_1852_cast_fp16 = reduce_mean(axes = var_1851, keep_dims = var_1682, x = zero_mean_sq_47_cast_fp16)[name = tensor("op_1852_cast_fp16")]; + tensor var_1853_to_fp16 = const()[name = tensor("op_1853_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1854_cast_fp16 = add(x = var_1852_cast_fp16, y = var_1853_to_fp16)[name = tensor("op_1854_cast_fp16")]; + tensor denom_47_epsilon_0_to_fp16 = const()[name = tensor("denom_47_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_47_cast_fp16 = rsqrt(epsilon = denom_47_epsilon_0_to_fp16, x = var_1854_cast_fp16)[name = tensor("denom_47_cast_fp16")]; + tensor out_47_cast_fp16 = mul(x = zero_mean_47_cast_fp16, y = denom_47_cast_fp16)[name = tensor("out_47_cast_fp16")]; + tensor input_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(527484864)))]; + 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(527487488)))]; + 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 var_1865 = const()[name = tensor("op_1865"), val = tensor([1, 1])]; + tensor var_1867 = const()[name = tensor("op_1867"), val = tensor([1, 1])]; + tensor input_77_pad_type_0 = const()[name = tensor("input_77_pad_type_0"), val = tensor("custom")]; + tensor input_77_pad_0 = const()[name = tensor("input_77_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(527490112)))]; + 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(540597376)))]; + tensor input_77_cast_fp16 = conv(bias = layers_7_fc1_bias_to_fp16, dilations = var_1867, groups = var_1681, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = var_1865, 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 var_1873 = const()[name = tensor("op_1873"), val = tensor([1, 1])]; + tensor var_1875 = const()[name = tensor("op_1875"), val = tensor([1, 1])]; + tensor hidden_states_17_pad_type_0 = const()[name = tensor("hidden_states_17_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_17_pad_0 = const()[name = tensor("hidden_states_17_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(540607680)))]; + 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(553714944)))]; + tensor hidden_states_17_cast_fp16 = conv(bias = layers_7_fc2_bias_to_fp16, dilations = var_1875, groups = var_1681, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = var_1873, 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_1888 = const()[name = tensor("op_1888"), val = tensor(3)]; + tensor var_1895 = const()[name = tensor("op_1895"), val = tensor(1)]; + tensor var_1896 = const()[name = tensor("op_1896"), val = tensor(true)]; + tensor var_1908 = const()[name = tensor("op_1908"), val = tensor([1])]; + tensor channels_mean_49_cast_fp16 = reduce_mean(axes = var_1908, keep_dims = var_1896, x = inputs_49_cast_fp16)[name = tensor("channels_mean_49_cast_fp16")]; + tensor zero_mean_49_cast_fp16 = sub(x = inputs_49_cast_fp16, y = channels_mean_49_cast_fp16)[name = tensor("zero_mean_49_cast_fp16")]; + tensor zero_mean_sq_49_cast_fp16 = mul(x = zero_mean_49_cast_fp16, y = zero_mean_49_cast_fp16)[name = tensor("zero_mean_sq_49_cast_fp16")]; + tensor var_1912 = const()[name = tensor("op_1912"), val = tensor([1])]; + tensor var_1913_cast_fp16 = reduce_mean(axes = var_1912, keep_dims = var_1896, x = zero_mean_sq_49_cast_fp16)[name = tensor("op_1913_cast_fp16")]; + tensor var_1914_to_fp16 = const()[name = tensor("op_1914_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1915_cast_fp16 = add(x = var_1913_cast_fp16, y = var_1914_to_fp16)[name = tensor("op_1915_cast_fp16")]; + tensor denom_49_epsilon_0_to_fp16 = const()[name = tensor("denom_49_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_49_cast_fp16 = rsqrt(epsilon = denom_49_epsilon_0_to_fp16, x = var_1915_cast_fp16)[name = tensor("denom_49_cast_fp16")]; + tensor out_49_cast_fp16 = mul(x = zero_mean_49_cast_fp16, y = denom_49_cast_fp16)[name = tensor("out_49_cast_fp16")]; + tensor obj_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(553717568)))]; + 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(553720192)))]; + 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 var_1930 = const()[name = tensor("op_1930"), val = tensor([1, 1])]; + tensor var_1932 = const()[name = tensor("op_1932"), val = tensor([1, 1])]; + tensor query_33_pad_type_0 = const()[name = tensor("query_33_pad_type_0"), val = tensor("custom")]; + tensor query_33_pad_0 = const()[name = tensor("query_33_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(553722816)))]; + 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(556999680)))]; + tensor query_33_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_bias_to_fp16, dilations = var_1932, groups = var_1895, pad = query_33_pad_0, pad_type = query_33_pad_type_0, strides = var_1930, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("query_33_cast_fp16")]; + tensor var_1936 = const()[name = tensor("op_1936"), val = tensor([1, 1])]; + tensor var_1938 = const()[name = tensor("op_1938"), val = tensor([1, 1])]; + tensor current_key_17_pad_type_0 = const()[name = tensor("current_key_17_pad_type_0"), val = tensor("custom")]; + tensor current_key_17_pad_0 = const()[name = tensor("current_key_17_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(557002304)))]; + tensor current_key_17_cast_fp16 = conv(dilations = var_1938, groups = var_1895, pad = current_key_17_pad_0, pad_type = current_key_17_pad_type_0, strides = var_1936, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("current_key_17_cast_fp16")]; + tensor var_1943 = const()[name = tensor("op_1943"), val = tensor([1, 1])]; + tensor var_1945 = const()[name = tensor("op_1945"), val = tensor([1, 1])]; + tensor current_value_17_pad_type_0 = const()[name = tensor("current_value_17_pad_type_0"), val = tensor("custom")]; + tensor current_value_17_pad_0 = const()[name = tensor("current_value_17_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(560279168)))]; + 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(563556032)))]; + tensor current_value_17_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_bias_to_fp16, dilations = var_1945, groups = var_1895, pad = current_value_17_pad_0, pad_type = current_value_17_pad_type_0, strides = var_1943, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("current_value_17_cast_fp16")]; + tensor var_1952_cast_fp16 = mul(x = current_key_17_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1952_cast_fp16")]; + tensor var_1954_cast_fp16 = mul(x = var_103_cast_fp16_8, y = var_241_cast_fp16)[name = tensor("op_1954_cast_fp16")]; + tensor key_33_cast_fp16 = add(x = var_1952_cast_fp16, y = var_1954_cast_fp16)[name = tensor("key_33_cast_fp16")]; + tensor var_1956_cast_fp16 = mul(x = current_value_17_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_1956_cast_fp16")]; + tensor var_1958_cast_fp16 = mul(x = var_138_cast_fp16_8, y = var_241_cast_fp16)[name = tensor("op_1958_cast_fp16")]; + tensor value_33_cast_fp16 = add(x = var_1956_cast_fp16, y = var_1958_cast_fp16)[name = tensor("value_33_cast_fp16")]; + tensor var_1961 = const()[name = tensor("op_1961"), val = tensor([1, 20, 64, -1])]; + tensor var_1962_cast_fp16 = reshape(shape = var_1961, x = query_33_cast_fp16)[name = tensor("op_1962_cast_fp16")]; + tensor var_1963_to_fp16 = const()[name = tensor("op_1963_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1964_cast_fp16 = mul(x = var_1962_cast_fp16, y = var_1963_to_fp16)[name = tensor("op_1964_cast_fp16")]; + tensor var_1965 = const()[name = tensor("op_1965"), val = tensor([1, 20, 64, -1])]; + tensor var_1966_cast_fp16 = reshape(shape = var_1965, x = key_33_cast_fp16)[name = tensor("op_1966_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_1964_cast_fp16, y = var_1966_cast_fp16)[name = tensor("mh_w_49_cast_fp16")]; + tensor mh_w_51_cast_fp16 = add(x = mh_w_49_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_51_cast_fp16")]; + tensor var_1974_cast_fp16 = softmax(axis = var_1888, x = mh_w_51_cast_fp16)[name = tensor("op_1974_cast_fp16")]; + tensor var_1975 = const()[name = tensor("op_1975"), val = tensor([1, 20, 64, -1])]; + tensor var_1976_cast_fp16 = reshape(shape = var_1975, x = value_33_cast_fp16)[name = tensor("op_1976_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_1976_cast_fp16, y = var_1974_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_1979 = const()[name = tensor("op_1979"), val = tensor([1, 1280, 1, -1])]; + tensor input_81_cast_fp16 = reshape(shape = var_1979, x = attn_33_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor var_1983 = const()[name = tensor("op_1983"), val = tensor([1, 1])]; + tensor var_1985 = const()[name = tensor("op_1985"), val = tensor([1, 1])]; + tensor obj_119_pad_type_0 = const()[name = tensor("obj_119_pad_type_0"), val = tensor("custom")]; + tensor obj_119_pad_0 = const()[name = tensor("obj_119_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(563558656)))]; + 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(566835520)))]; + tensor obj_119_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_bias_to_fp16, dilations = var_1985, groups = var_1895, pad = obj_119_pad_0, pad_type = obj_119_pad_type_0, strides = var_1983, 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 var_1995 = const()[name = tensor("op_1995"), val = tensor([1])]; + tensor channels_mean_51_cast_fp16 = reduce_mean(axes = var_1995, keep_dims = var_1896, x = inputs_51_cast_fp16)[name = tensor("channels_mean_51_cast_fp16")]; + tensor zero_mean_51_cast_fp16 = sub(x = inputs_51_cast_fp16, y = channels_mean_51_cast_fp16)[name = tensor("zero_mean_51_cast_fp16")]; + tensor zero_mean_sq_51_cast_fp16 = mul(x = zero_mean_51_cast_fp16, y = zero_mean_51_cast_fp16)[name = tensor("zero_mean_sq_51_cast_fp16")]; + tensor var_1999 = const()[name = tensor("op_1999"), val = tensor([1])]; + tensor var_2000_cast_fp16 = reduce_mean(axes = var_1999, keep_dims = var_1896, x = zero_mean_sq_51_cast_fp16)[name = tensor("op_2000_cast_fp16")]; + tensor var_2001_to_fp16 = const()[name = tensor("op_2001_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2002_cast_fp16 = add(x = var_2000_cast_fp16, y = var_2001_to_fp16)[name = tensor("op_2002_cast_fp16")]; + tensor denom_51_epsilon_0_to_fp16 = const()[name = tensor("denom_51_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_51_cast_fp16 = rsqrt(epsilon = denom_51_epsilon_0_to_fp16, x = var_2002_cast_fp16)[name = tensor("denom_51_cast_fp16")]; + tensor out_51_cast_fp16 = mul(x = zero_mean_51_cast_fp16, y = denom_51_cast_fp16)[name = tensor("out_51_cast_fp16")]; + tensor obj_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(566838144)))]; + 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(566840768)))]; + 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 var_2017 = const()[name = tensor("op_2017"), val = tensor([1, 1])]; + tensor var_2019 = const()[name = tensor("op_2019"), val = tensor([1, 1])]; + tensor query_35_pad_type_0 = const()[name = tensor("query_35_pad_type_0"), val = tensor("custom")]; + tensor query_35_pad_0 = const()[name = tensor("query_35_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(566843392)))]; + 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(570120256)))]; + tensor query_35_cast_fp16 = conv(bias = layers_8_encoder_attn_q_proj_bias_to_fp16, dilations = var_2019, groups = var_1895, pad = query_35_pad_0, pad_type = query_35_pad_type_0, strides = var_2017, weight = layers_8_encoder_attn_q_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("query_35_cast_fp16")]; + tensor var_2023 = const()[name = tensor("op_2023"), val = tensor([1, 1])]; + tensor var_2025 = const()[name = tensor("op_2025"), val = tensor([1, 1])]; + tensor key_35_pad_type_0 = const()[name = tensor("key_35_pad_type_0"), val = tensor("custom")]; + tensor key_35_pad_0 = const()[name = tensor("key_35_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(570122880)))]; + tensor key_35_cast_fp16 = conv(dilations = var_2025, groups = var_1895, pad = key_35_pad_0, pad_type = key_35_pad_type_0, strides = var_2023, weight = layers_8_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_35_cast_fp16")]; + tensor var_2030 = const()[name = tensor("op_2030"), val = tensor([1, 1])]; + tensor var_2032 = const()[name = tensor("op_2032"), val = tensor([1, 1])]; + tensor value_35_pad_type_0 = const()[name = tensor("value_35_pad_type_0"), val = tensor("custom")]; + tensor value_35_pad_0 = const()[name = tensor("value_35_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(573399744)))]; + 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(576676608)))]; + tensor value_35_cast_fp16 = conv(bias = layers_8_encoder_attn_v_proj_bias_to_fp16, dilations = var_2032, groups = var_1895, pad = value_35_pad_0, pad_type = value_35_pad_type_0, strides = var_2030, weight = layers_8_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_35_cast_fp16")]; + tensor var_2036 = const()[name = tensor("op_2036"), val = tensor([1, 20, 64, -1])]; + tensor var_2037_cast_fp16 = reshape(shape = var_2036, x = query_35_cast_fp16)[name = tensor("op_2037_cast_fp16")]; + tensor var_2038_to_fp16 = const()[name = tensor("op_2038_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2039_cast_fp16 = mul(x = var_2037_cast_fp16, y = var_2038_to_fp16)[name = tensor("op_2039_cast_fp16")]; + tensor var_2040 = const()[name = tensor("op_2040"), val = tensor([1, 20, 64, -1])]; + tensor var_2041_cast_fp16 = reshape(shape = var_2040, x = key_35_cast_fp16)[name = tensor("op_2041_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_2039_cast_fp16, y = var_2041_cast_fp16)[name = tensor("mh_w_53_cast_fp16")]; + tensor obj_125_cast_fp16 = softmax(axis = var_1888, x = mh_w_53_cast_fp16)[name = tensor("obj_125_cast_fp16")]; + tensor var_2045 = const()[name = tensor("op_2045"), val = tensor([1, 20, 64, -1])]; + tensor var_2046_cast_fp16 = reshape(shape = var_2045, x = value_35_cast_fp16)[name = tensor("op_2046_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_2046_cast_fp16, y = obj_125_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_2049 = const()[name = tensor("op_2049"), val = tensor([1, 1280, 1, -1])]; + tensor input_83_cast_fp16 = reshape(shape = var_2049, x = attn_35_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor var_2053 = const()[name = tensor("op_2053"), val = tensor([1, 1])]; + tensor var_2055 = const()[name = tensor("op_2055"), val = tensor([1, 1])]; + tensor obj_123_pad_type_0 = const()[name = tensor("obj_123_pad_type_0"), val = tensor("custom")]; + tensor obj_123_pad_0 = const()[name = tensor("obj_123_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(576679232)))]; + 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(579956096)))]; + tensor obj_123_cast_fp16 = conv(bias = layers_8_encoder_attn_o_proj_bias_to_fp16, dilations = var_2055, groups = var_1895, pad = obj_123_pad_0, pad_type = obj_123_pad_type_0, strides = var_2053, 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 var_2061 = const()[name = tensor("op_2061"), val = tensor([1])]; + tensor channels_mean_53_cast_fp16 = reduce_mean(axes = var_2061, keep_dims = var_1896, x = inputs_53_cast_fp16)[name = tensor("channels_mean_53_cast_fp16")]; + tensor zero_mean_53_cast_fp16 = sub(x = inputs_53_cast_fp16, y = channels_mean_53_cast_fp16)[name = tensor("zero_mean_53_cast_fp16")]; + tensor zero_mean_sq_53_cast_fp16 = mul(x = zero_mean_53_cast_fp16, y = zero_mean_53_cast_fp16)[name = tensor("zero_mean_sq_53_cast_fp16")]; + tensor var_2065 = const()[name = tensor("op_2065"), val = tensor([1])]; + tensor var_2066_cast_fp16 = reduce_mean(axes = var_2065, keep_dims = var_1896, x = zero_mean_sq_53_cast_fp16)[name = tensor("op_2066_cast_fp16")]; + tensor var_2067_to_fp16 = const()[name = tensor("op_2067_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2068_cast_fp16 = add(x = var_2066_cast_fp16, y = var_2067_to_fp16)[name = tensor("op_2068_cast_fp16")]; + tensor denom_53_epsilon_0_to_fp16 = const()[name = tensor("denom_53_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_53_cast_fp16 = rsqrt(epsilon = denom_53_epsilon_0_to_fp16, x = var_2068_cast_fp16)[name = tensor("denom_53_cast_fp16")]; + tensor out_53_cast_fp16 = mul(x = zero_mean_53_cast_fp16, y = denom_53_cast_fp16)[name = tensor("out_53_cast_fp16")]; + tensor input_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(579958720)))]; + 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(579961344)))]; + 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 var_2079 = const()[name = tensor("op_2079"), val = tensor([1, 1])]; + tensor var_2081 = const()[name = tensor("op_2081"), val = tensor([1, 1])]; + tensor input_87_pad_type_0 = const()[name = tensor("input_87_pad_type_0"), val = tensor("custom")]; + tensor input_87_pad_0 = const()[name = tensor("input_87_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(579963968)))]; + 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(593071232)))]; + tensor input_87_cast_fp16 = conv(bias = layers_8_fc1_bias_to_fp16, dilations = var_2081, groups = var_1895, pad = input_87_pad_0, pad_type = input_87_pad_type_0, strides = var_2079, 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 var_2087 = const()[name = tensor("op_2087"), val = tensor([1, 1])]; + tensor var_2089 = const()[name = tensor("op_2089"), val = tensor([1, 1])]; + tensor hidden_states_19_pad_type_0 = const()[name = tensor("hidden_states_19_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_19_pad_0 = const()[name = tensor("hidden_states_19_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(593081536)))]; + 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(606188800)))]; + tensor hidden_states_19_cast_fp16 = conv(bias = layers_8_fc2_bias_to_fp16, dilations = var_2089, groups = var_1895, pad = hidden_states_19_pad_0, pad_type = hidden_states_19_pad_type_0, strides = var_2087, 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_2102 = const()[name = tensor("op_2102"), val = tensor(3)]; + tensor var_2109 = const()[name = tensor("op_2109"), val = tensor(1)]; + tensor var_2110 = const()[name = tensor("op_2110"), val = tensor(true)]; + tensor var_2122 = const()[name = tensor("op_2122"), val = tensor([1])]; + tensor channels_mean_55_cast_fp16 = reduce_mean(axes = var_2122, keep_dims = var_2110, x = inputs_55_cast_fp16)[name = tensor("channels_mean_55_cast_fp16")]; + tensor zero_mean_55_cast_fp16 = sub(x = inputs_55_cast_fp16, y = channels_mean_55_cast_fp16)[name = tensor("zero_mean_55_cast_fp16")]; + tensor zero_mean_sq_55_cast_fp16 = mul(x = zero_mean_55_cast_fp16, y = zero_mean_55_cast_fp16)[name = tensor("zero_mean_sq_55_cast_fp16")]; + tensor var_2126 = const()[name = tensor("op_2126"), val = tensor([1])]; + tensor var_2127_cast_fp16 = reduce_mean(axes = var_2126, keep_dims = var_2110, x = zero_mean_sq_55_cast_fp16)[name = tensor("op_2127_cast_fp16")]; + tensor var_2128_to_fp16 = const()[name = tensor("op_2128_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2129_cast_fp16 = add(x = var_2127_cast_fp16, y = var_2128_to_fp16)[name = tensor("op_2129_cast_fp16")]; + tensor denom_55_epsilon_0_to_fp16 = const()[name = tensor("denom_55_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_55_cast_fp16 = rsqrt(epsilon = denom_55_epsilon_0_to_fp16, x = var_2129_cast_fp16)[name = tensor("denom_55_cast_fp16")]; + tensor out_55_cast_fp16 = mul(x = zero_mean_55_cast_fp16, y = denom_55_cast_fp16)[name = tensor("out_55_cast_fp16")]; + tensor obj_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(606191424)))]; + 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(606194048)))]; + 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 var_2144 = const()[name = tensor("op_2144"), val = tensor([1, 1])]; + tensor var_2146 = const()[name = tensor("op_2146"), val = tensor([1, 1])]; + tensor query_37_pad_type_0 = const()[name = tensor("query_37_pad_type_0"), val = tensor("custom")]; + tensor query_37_pad_0 = const()[name = tensor("query_37_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(606196672)))]; + 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(609473536)))]; + tensor query_37_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_bias_to_fp16, dilations = var_2146, groups = var_2109, pad = query_37_pad_0, pad_type = query_37_pad_type_0, strides = var_2144, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("query_37_cast_fp16")]; + tensor var_2150 = const()[name = tensor("op_2150"), val = tensor([1, 1])]; + tensor var_2152 = const()[name = tensor("op_2152"), val = tensor([1, 1])]; + tensor current_key_19_pad_type_0 = const()[name = tensor("current_key_19_pad_type_0"), val = tensor("custom")]; + tensor current_key_19_pad_0 = const()[name = tensor("current_key_19_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(609476160)))]; + tensor current_key_19_cast_fp16 = conv(dilations = var_2152, groups = var_2109, pad = current_key_19_pad_0, pad_type = current_key_19_pad_type_0, strides = var_2150, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("current_key_19_cast_fp16")]; + tensor var_2157 = const()[name = tensor("op_2157"), val = tensor([1, 1])]; + tensor var_2159 = const()[name = tensor("op_2159"), val = tensor([1, 1])]; + tensor current_value_19_pad_type_0 = const()[name = tensor("current_value_19_pad_type_0"), val = tensor("custom")]; + tensor current_value_19_pad_0 = const()[name = tensor("current_value_19_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(612753024)))]; + 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(616029888)))]; + tensor current_value_19_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_bias_to_fp16, dilations = var_2159, groups = var_2109, pad = current_value_19_pad_0, pad_type = current_value_19_pad_type_0, strides = var_2157, weight = layers_9_self_attn_v_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("current_value_19_cast_fp16")]; + tensor var_2166_cast_fp16 = mul(x = current_key_19_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_2166_cast_fp16")]; + tensor var_2168_cast_fp16 = mul(x = var_103_cast_fp16_9, y = var_241_cast_fp16)[name = tensor("op_2168_cast_fp16")]; + tensor key_37_cast_fp16 = add(x = var_2166_cast_fp16, y = var_2168_cast_fp16)[name = tensor("key_37_cast_fp16")]; + tensor var_2170_cast_fp16 = mul(x = current_value_19_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_2170_cast_fp16")]; + tensor var_2172_cast_fp16 = mul(x = var_138_cast_fp16_9, y = var_241_cast_fp16)[name = tensor("op_2172_cast_fp16")]; + tensor value_37_cast_fp16 = add(x = var_2170_cast_fp16, y = var_2172_cast_fp16)[name = tensor("value_37_cast_fp16")]; + tensor var_2175 = const()[name = tensor("op_2175"), val = tensor([1, 20, 64, -1])]; + tensor var_2176_cast_fp16 = reshape(shape = var_2175, x = query_37_cast_fp16)[name = tensor("op_2176_cast_fp16")]; + tensor var_2177_to_fp16 = const()[name = tensor("op_2177_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2178_cast_fp16 = mul(x = var_2176_cast_fp16, y = var_2177_to_fp16)[name = tensor("op_2178_cast_fp16")]; + tensor var_2179 = const()[name = tensor("op_2179"), val = tensor([1, 20, 64, -1])]; + tensor var_2180_cast_fp16 = reshape(shape = var_2179, x = key_37_cast_fp16)[name = tensor("op_2180_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_2178_cast_fp16, y = var_2180_cast_fp16)[name = tensor("mh_w_55_cast_fp16")]; + tensor mh_w_57_cast_fp16 = add(x = mh_w_55_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_57_cast_fp16")]; + tensor var_2188_cast_fp16 = softmax(axis = var_2102, x = mh_w_57_cast_fp16)[name = tensor("op_2188_cast_fp16")]; + tensor var_2189 = const()[name = tensor("op_2189"), val = tensor([1, 20, 64, -1])]; + tensor var_2190_cast_fp16 = reshape(shape = var_2189, x = value_37_cast_fp16)[name = tensor("op_2190_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_2190_cast_fp16, y = var_2188_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_2193 = const()[name = tensor("op_2193"), val = tensor([1, 1280, 1, -1])]; + tensor input_91_cast_fp16 = reshape(shape = var_2193, x = attn_37_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor var_2197 = const()[name = tensor("op_2197"), val = tensor([1, 1])]; + tensor var_2199 = const()[name = tensor("op_2199"), val = tensor([1, 1])]; + tensor obj_133_pad_type_0 = const()[name = tensor("obj_133_pad_type_0"), val = tensor("custom")]; + tensor obj_133_pad_0 = const()[name = tensor("obj_133_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(616032512)))]; + 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(619309376)))]; + tensor obj_133_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_bias_to_fp16, dilations = var_2199, groups = var_2109, pad = obj_133_pad_0, pad_type = obj_133_pad_type_0, strides = var_2197, 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 var_2209 = const()[name = tensor("op_2209"), val = tensor([1])]; + tensor channels_mean_57_cast_fp16 = reduce_mean(axes = var_2209, keep_dims = var_2110, x = inputs_57_cast_fp16)[name = tensor("channels_mean_57_cast_fp16")]; + tensor zero_mean_57_cast_fp16 = sub(x = inputs_57_cast_fp16, y = channels_mean_57_cast_fp16)[name = tensor("zero_mean_57_cast_fp16")]; + tensor zero_mean_sq_57_cast_fp16 = mul(x = zero_mean_57_cast_fp16, y = zero_mean_57_cast_fp16)[name = tensor("zero_mean_sq_57_cast_fp16")]; + tensor var_2213 = const()[name = tensor("op_2213"), val = tensor([1])]; + tensor var_2214_cast_fp16 = reduce_mean(axes = var_2213, keep_dims = var_2110, x = zero_mean_sq_57_cast_fp16)[name = tensor("op_2214_cast_fp16")]; + tensor var_2215_to_fp16 = const()[name = tensor("op_2215_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2216_cast_fp16 = add(x = var_2214_cast_fp16, y = var_2215_to_fp16)[name = tensor("op_2216_cast_fp16")]; + tensor denom_57_epsilon_0_to_fp16 = const()[name = tensor("denom_57_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_57_cast_fp16 = rsqrt(epsilon = denom_57_epsilon_0_to_fp16, x = var_2216_cast_fp16)[name = tensor("denom_57_cast_fp16")]; + tensor out_57_cast_fp16 = mul(x = zero_mean_57_cast_fp16, y = denom_57_cast_fp16)[name = tensor("out_57_cast_fp16")]; + tensor obj_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(619312000)))]; + 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(619314624)))]; + 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 var_2231 = const()[name = tensor("op_2231"), val = tensor([1, 1])]; + tensor var_2233 = const()[name = tensor("op_2233"), val = tensor([1, 1])]; + tensor query_39_pad_type_0 = const()[name = tensor("query_39_pad_type_0"), val = tensor("custom")]; + tensor query_39_pad_0 = const()[name = tensor("query_39_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(619317248)))]; + 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(622594112)))]; + tensor query_39_cast_fp16 = conv(bias = layers_9_encoder_attn_q_proj_bias_to_fp16, dilations = var_2233, groups = var_2109, pad = query_39_pad_0, pad_type = query_39_pad_type_0, strides = var_2231, weight = layers_9_encoder_attn_q_proj_weight_to_fp16, x = obj_135_cast_fp16)[name = tensor("query_39_cast_fp16")]; + tensor var_2237 = const()[name = tensor("op_2237"), val = tensor([1, 1])]; + tensor var_2239 = const()[name = tensor("op_2239"), val = tensor([1, 1])]; + tensor key_39_pad_type_0 = const()[name = tensor("key_39_pad_type_0"), val = tensor("custom")]; + tensor key_39_pad_0 = const()[name = tensor("key_39_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(622596736)))]; + tensor key_39_cast_fp16 = conv(dilations = var_2239, groups = var_2109, pad = key_39_pad_0, pad_type = key_39_pad_type_0, strides = var_2237, weight = layers_9_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_39_cast_fp16")]; + tensor var_2244 = const()[name = tensor("op_2244"), val = tensor([1, 1])]; + tensor var_2246 = const()[name = tensor("op_2246"), val = tensor([1, 1])]; + tensor value_39_pad_type_0 = const()[name = tensor("value_39_pad_type_0"), val = tensor("custom")]; + tensor value_39_pad_0 = const()[name = tensor("value_39_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(625873600)))]; + 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(629150464)))]; + tensor value_39_cast_fp16 = conv(bias = layers_9_encoder_attn_v_proj_bias_to_fp16, dilations = var_2246, groups = var_2109, pad = value_39_pad_0, pad_type = value_39_pad_type_0, strides = var_2244, weight = layers_9_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_39_cast_fp16")]; + tensor var_2250 = const()[name = tensor("op_2250"), val = tensor([1, 20, 64, -1])]; + tensor var_2251_cast_fp16 = reshape(shape = var_2250, x = query_39_cast_fp16)[name = tensor("op_2251_cast_fp16")]; + tensor var_2252_to_fp16 = const()[name = tensor("op_2252_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2253_cast_fp16 = mul(x = var_2251_cast_fp16, y = var_2252_to_fp16)[name = tensor("op_2253_cast_fp16")]; + tensor var_2254 = const()[name = tensor("op_2254"), val = tensor([1, 20, 64, -1])]; + tensor var_2255_cast_fp16 = reshape(shape = var_2254, x = key_39_cast_fp16)[name = tensor("op_2255_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_2253_cast_fp16, y = var_2255_cast_fp16)[name = tensor("mh_w_59_cast_fp16")]; + tensor obj_139_cast_fp16 = softmax(axis = var_2102, x = mh_w_59_cast_fp16)[name = tensor("obj_139_cast_fp16")]; + tensor var_2259 = const()[name = tensor("op_2259"), val = tensor([1, 20, 64, -1])]; + tensor var_2260_cast_fp16 = reshape(shape = var_2259, x = value_39_cast_fp16)[name = tensor("op_2260_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_2260_cast_fp16, y = obj_139_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_2263 = const()[name = tensor("op_2263"), val = tensor([1, 1280, 1, -1])]; + tensor input_93_cast_fp16 = reshape(shape = var_2263, x = attn_39_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor var_2267 = const()[name = tensor("op_2267"), val = tensor([1, 1])]; + tensor var_2269 = const()[name = tensor("op_2269"), val = tensor([1, 1])]; + tensor obj_137_pad_type_0 = const()[name = tensor("obj_137_pad_type_0"), val = tensor("custom")]; + tensor obj_137_pad_0 = const()[name = tensor("obj_137_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(629153088)))]; + 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(632429952)))]; + tensor obj_137_cast_fp16 = conv(bias = layers_9_encoder_attn_o_proj_bias_to_fp16, dilations = var_2269, groups = var_2109, pad = obj_137_pad_0, pad_type = obj_137_pad_type_0, strides = var_2267, 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 var_2275 = const()[name = tensor("op_2275"), val = tensor([1])]; + tensor channels_mean_59_cast_fp16 = reduce_mean(axes = var_2275, keep_dims = var_2110, x = inputs_59_cast_fp16)[name = tensor("channels_mean_59_cast_fp16")]; + tensor zero_mean_59_cast_fp16 = sub(x = inputs_59_cast_fp16, y = channels_mean_59_cast_fp16)[name = tensor("zero_mean_59_cast_fp16")]; + tensor zero_mean_sq_59_cast_fp16 = mul(x = zero_mean_59_cast_fp16, y = zero_mean_59_cast_fp16)[name = tensor("zero_mean_sq_59_cast_fp16")]; + tensor var_2279 = const()[name = tensor("op_2279"), val = tensor([1])]; + tensor var_2280_cast_fp16 = reduce_mean(axes = var_2279, keep_dims = var_2110, x = zero_mean_sq_59_cast_fp16)[name = tensor("op_2280_cast_fp16")]; + tensor var_2281_to_fp16 = const()[name = tensor("op_2281_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2282_cast_fp16 = add(x = var_2280_cast_fp16, y = var_2281_to_fp16)[name = tensor("op_2282_cast_fp16")]; + tensor denom_59_epsilon_0_to_fp16 = const()[name = tensor("denom_59_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_59_cast_fp16 = rsqrt(epsilon = denom_59_epsilon_0_to_fp16, x = var_2282_cast_fp16)[name = tensor("denom_59_cast_fp16")]; + tensor out_59_cast_fp16 = mul(x = zero_mean_59_cast_fp16, y = denom_59_cast_fp16)[name = tensor("out_59_cast_fp16")]; + tensor input_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(632432576)))]; + 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(632435200)))]; + 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 var_2293 = const()[name = tensor("op_2293"), val = tensor([1, 1])]; + tensor var_2295 = const()[name = tensor("op_2295"), val = tensor([1, 1])]; + tensor input_97_pad_type_0 = const()[name = tensor("input_97_pad_type_0"), val = tensor("custom")]; + tensor input_97_pad_0 = const()[name = tensor("input_97_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(632437824)))]; + 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(645545088)))]; + tensor input_97_cast_fp16 = conv(bias = layers_9_fc1_bias_to_fp16, dilations = var_2295, groups = var_2109, pad = input_97_pad_0, pad_type = input_97_pad_type_0, strides = var_2293, 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 var_2301 = const()[name = tensor("op_2301"), val = tensor([1, 1])]; + tensor var_2303 = const()[name = tensor("op_2303"), val = tensor([1, 1])]; + tensor hidden_states_21_pad_type_0 = const()[name = tensor("hidden_states_21_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_21_pad_0 = const()[name = tensor("hidden_states_21_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(645555392)))]; + 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(658662656)))]; + tensor hidden_states_21_cast_fp16 = conv(bias = layers_9_fc2_bias_to_fp16, dilations = var_2303, groups = var_2109, pad = hidden_states_21_pad_0, pad_type = hidden_states_21_pad_type_0, strides = var_2301, 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_2316 = const()[name = tensor("op_2316"), val = tensor(3)]; + tensor var_2323 = const()[name = tensor("op_2323"), val = tensor(1)]; + tensor var_2324 = const()[name = tensor("op_2324"), val = tensor(true)]; + tensor var_2336 = const()[name = tensor("op_2336"), val = tensor([1])]; + tensor channels_mean_61_cast_fp16 = reduce_mean(axes = var_2336, keep_dims = var_2324, x = inputs_61_cast_fp16)[name = tensor("channels_mean_61_cast_fp16")]; + tensor zero_mean_61_cast_fp16 = sub(x = inputs_61_cast_fp16, y = channels_mean_61_cast_fp16)[name = tensor("zero_mean_61_cast_fp16")]; + tensor zero_mean_sq_61_cast_fp16 = mul(x = zero_mean_61_cast_fp16, y = zero_mean_61_cast_fp16)[name = tensor("zero_mean_sq_61_cast_fp16")]; + tensor var_2340 = const()[name = tensor("op_2340"), val = tensor([1])]; + tensor var_2341_cast_fp16 = reduce_mean(axes = var_2340, keep_dims = var_2324, x = zero_mean_sq_61_cast_fp16)[name = tensor("op_2341_cast_fp16")]; + tensor var_2342_to_fp16 = const()[name = tensor("op_2342_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2343_cast_fp16 = add(x = var_2341_cast_fp16, y = var_2342_to_fp16)[name = tensor("op_2343_cast_fp16")]; + tensor denom_61_epsilon_0_to_fp16 = const()[name = tensor("denom_61_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_61_cast_fp16 = rsqrt(epsilon = denom_61_epsilon_0_to_fp16, x = var_2343_cast_fp16)[name = tensor("denom_61_cast_fp16")]; + tensor out_61_cast_fp16 = mul(x = zero_mean_61_cast_fp16, y = denom_61_cast_fp16)[name = tensor("out_61_cast_fp16")]; + tensor obj_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(658665280)))]; + 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(658667904)))]; + 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 var_2358 = const()[name = tensor("op_2358"), val = tensor([1, 1])]; + tensor var_2360 = const()[name = tensor("op_2360"), val = tensor([1, 1])]; + tensor query_41_pad_type_0 = const()[name = tensor("query_41_pad_type_0"), val = tensor("custom")]; + tensor query_41_pad_0 = const()[name = tensor("query_41_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(658670528)))]; + 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(661947392)))]; + tensor query_41_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_bias_to_fp16, dilations = var_2360, groups = var_2323, pad = query_41_pad_0, pad_type = query_41_pad_type_0, strides = var_2358, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("query_41_cast_fp16")]; + tensor var_2364 = const()[name = tensor("op_2364"), val = tensor([1, 1])]; + tensor var_2366 = const()[name = tensor("op_2366"), val = tensor([1, 1])]; + tensor current_key_21_pad_type_0 = const()[name = tensor("current_key_21_pad_type_0"), val = tensor("custom")]; + tensor current_key_21_pad_0 = const()[name = tensor("current_key_21_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(661950016)))]; + tensor current_key_21_cast_fp16 = conv(dilations = var_2366, groups = var_2323, pad = current_key_21_pad_0, pad_type = current_key_21_pad_type_0, strides = var_2364, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("current_key_21_cast_fp16")]; + tensor var_2371 = const()[name = tensor("op_2371"), val = tensor([1, 1])]; + tensor var_2373 = const()[name = tensor("op_2373"), val = tensor([1, 1])]; + tensor current_value_21_pad_type_0 = const()[name = tensor("current_value_21_pad_type_0"), val = tensor("custom")]; + tensor current_value_21_pad_0 = const()[name = tensor("current_value_21_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(665226880)))]; + 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(668503744)))]; + tensor current_value_21_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_bias_to_fp16, dilations = var_2373, groups = var_2323, pad = current_value_21_pad_0, pad_type = current_value_21_pad_type_0, strides = var_2371, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("current_value_21_cast_fp16")]; + tensor var_2380_cast_fp16 = mul(x = current_key_21_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_2380_cast_fp16")]; + tensor var_2382_cast_fp16 = mul(x = var_103_cast_fp16_10, y = var_241_cast_fp16)[name = tensor("op_2382_cast_fp16")]; + tensor key_41_cast_fp16 = add(x = var_2380_cast_fp16, y = var_2382_cast_fp16)[name = tensor("key_41_cast_fp16")]; + tensor var_2384_cast_fp16 = mul(x = current_value_21_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_2384_cast_fp16")]; + tensor var_2386_cast_fp16 = mul(x = var_138_cast_fp16_10, y = var_241_cast_fp16)[name = tensor("op_2386_cast_fp16")]; + tensor value_41_cast_fp16 = add(x = var_2384_cast_fp16, y = var_2386_cast_fp16)[name = tensor("value_41_cast_fp16")]; + tensor var_2389 = const()[name = tensor("op_2389"), val = tensor([1, 20, 64, -1])]; + tensor var_2390_cast_fp16 = reshape(shape = var_2389, x = query_41_cast_fp16)[name = tensor("op_2390_cast_fp16")]; + tensor var_2391_to_fp16 = const()[name = tensor("op_2391_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2392_cast_fp16 = mul(x = var_2390_cast_fp16, y = var_2391_to_fp16)[name = tensor("op_2392_cast_fp16")]; + tensor var_2393 = const()[name = tensor("op_2393"), val = tensor([1, 20, 64, -1])]; + tensor var_2394_cast_fp16 = reshape(shape = var_2393, x = key_41_cast_fp16)[name = tensor("op_2394_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_2392_cast_fp16, y = var_2394_cast_fp16)[name = tensor("mh_w_61_cast_fp16")]; + tensor mh_w_63_cast_fp16 = add(x = mh_w_61_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_63_cast_fp16")]; + tensor var_2402_cast_fp16 = softmax(axis = var_2316, x = mh_w_63_cast_fp16)[name = tensor("op_2402_cast_fp16")]; + tensor var_2403 = const()[name = tensor("op_2403"), val = tensor([1, 20, 64, -1])]; + tensor var_2404_cast_fp16 = reshape(shape = var_2403, x = value_41_cast_fp16)[name = tensor("op_2404_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_2404_cast_fp16, y = var_2402_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_2407 = const()[name = tensor("op_2407"), val = tensor([1, 1280, 1, -1])]; + tensor input_101_cast_fp16 = reshape(shape = var_2407, x = attn_41_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor var_2411 = const()[name = tensor("op_2411"), val = tensor([1, 1])]; + tensor var_2413 = const()[name = tensor("op_2413"), val = tensor([1, 1])]; + tensor obj_147_pad_type_0 = const()[name = tensor("obj_147_pad_type_0"), val = tensor("custom")]; + tensor obj_147_pad_0 = const()[name = tensor("obj_147_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(668506368)))]; + 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(671783232)))]; + tensor obj_147_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_bias_to_fp16, dilations = var_2413, groups = var_2323, pad = obj_147_pad_0, pad_type = obj_147_pad_type_0, strides = var_2411, 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 var_2423 = const()[name = tensor("op_2423"), val = tensor([1])]; + tensor channels_mean_63_cast_fp16 = reduce_mean(axes = var_2423, keep_dims = var_2324, x = inputs_63_cast_fp16)[name = tensor("channels_mean_63_cast_fp16")]; + tensor zero_mean_63_cast_fp16 = sub(x = inputs_63_cast_fp16, y = channels_mean_63_cast_fp16)[name = tensor("zero_mean_63_cast_fp16")]; + tensor zero_mean_sq_63_cast_fp16 = mul(x = zero_mean_63_cast_fp16, y = zero_mean_63_cast_fp16)[name = tensor("zero_mean_sq_63_cast_fp16")]; + tensor var_2427 = const()[name = tensor("op_2427"), val = tensor([1])]; + tensor var_2428_cast_fp16 = reduce_mean(axes = var_2427, keep_dims = var_2324, x = zero_mean_sq_63_cast_fp16)[name = tensor("op_2428_cast_fp16")]; + tensor var_2429_to_fp16 = const()[name = tensor("op_2429_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2430_cast_fp16 = add(x = var_2428_cast_fp16, y = var_2429_to_fp16)[name = tensor("op_2430_cast_fp16")]; + tensor denom_63_epsilon_0_to_fp16 = const()[name = tensor("denom_63_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_63_cast_fp16 = rsqrt(epsilon = denom_63_epsilon_0_to_fp16, x = var_2430_cast_fp16)[name = tensor("denom_63_cast_fp16")]; + tensor out_63_cast_fp16 = mul(x = zero_mean_63_cast_fp16, y = denom_63_cast_fp16)[name = tensor("out_63_cast_fp16")]; + tensor obj_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(671785856)))]; + 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(671788480)))]; + 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 var_2445 = const()[name = tensor("op_2445"), val = tensor([1, 1])]; + tensor var_2447 = const()[name = tensor("op_2447"), val = tensor([1, 1])]; + tensor query_43_pad_type_0 = const()[name = tensor("query_43_pad_type_0"), val = tensor("custom")]; + tensor query_43_pad_0 = const()[name = tensor("query_43_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(671791104)))]; + 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(675067968)))]; + tensor query_43_cast_fp16 = conv(bias = layers_10_encoder_attn_q_proj_bias_to_fp16, dilations = var_2447, groups = var_2323, pad = query_43_pad_0, pad_type = query_43_pad_type_0, strides = var_2445, weight = layers_10_encoder_attn_q_proj_weight_to_fp16, x = obj_149_cast_fp16)[name = tensor("query_43_cast_fp16")]; + tensor var_2451 = const()[name = tensor("op_2451"), val = tensor([1, 1])]; + tensor var_2453 = const()[name = tensor("op_2453"), val = tensor([1, 1])]; + tensor key_43_pad_type_0 = const()[name = tensor("key_43_pad_type_0"), val = tensor("custom")]; + tensor key_43_pad_0 = const()[name = tensor("key_43_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(675070592)))]; + tensor key_43_cast_fp16 = conv(dilations = var_2453, groups = var_2323, pad = key_43_pad_0, pad_type = key_43_pad_type_0, strides = var_2451, weight = layers_10_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_43_cast_fp16")]; + tensor var_2458 = const()[name = tensor("op_2458"), val = tensor([1, 1])]; + tensor var_2460 = const()[name = tensor("op_2460"), val = tensor([1, 1])]; + tensor value_43_pad_type_0 = const()[name = tensor("value_43_pad_type_0"), val = tensor("custom")]; + tensor value_43_pad_0 = const()[name = tensor("value_43_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(678347456)))]; + 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(681624320)))]; + tensor value_43_cast_fp16 = conv(bias = layers_10_encoder_attn_v_proj_bias_to_fp16, dilations = var_2460, groups = var_2323, pad = value_43_pad_0, pad_type = value_43_pad_type_0, strides = var_2458, weight = layers_10_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_43_cast_fp16")]; + tensor var_2464 = const()[name = tensor("op_2464"), val = tensor([1, 20, 64, -1])]; + tensor var_2465_cast_fp16 = reshape(shape = var_2464, x = query_43_cast_fp16)[name = tensor("op_2465_cast_fp16")]; + tensor var_2466_to_fp16 = const()[name = tensor("op_2466_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2467_cast_fp16 = mul(x = var_2465_cast_fp16, y = var_2466_to_fp16)[name = tensor("op_2467_cast_fp16")]; + tensor var_2468 = const()[name = tensor("op_2468"), val = tensor([1, 20, 64, -1])]; + tensor var_2469_cast_fp16 = reshape(shape = var_2468, x = key_43_cast_fp16)[name = tensor("op_2469_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_2467_cast_fp16, y = var_2469_cast_fp16)[name = tensor("mh_w_65_cast_fp16")]; + tensor obj_153_cast_fp16 = softmax(axis = var_2316, x = mh_w_65_cast_fp16)[name = tensor("obj_153_cast_fp16")]; + tensor var_2473 = const()[name = tensor("op_2473"), val = tensor([1, 20, 64, -1])]; + tensor var_2474_cast_fp16 = reshape(shape = var_2473, x = value_43_cast_fp16)[name = tensor("op_2474_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_2474_cast_fp16, y = obj_153_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_2477 = const()[name = tensor("op_2477"), val = tensor([1, 1280, 1, -1])]; + tensor input_103_cast_fp16 = reshape(shape = var_2477, x = attn_43_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor var_2481 = const()[name = tensor("op_2481"), val = tensor([1, 1])]; + tensor var_2483 = const()[name = tensor("op_2483"), val = tensor([1, 1])]; + tensor obj_151_pad_type_0 = const()[name = tensor("obj_151_pad_type_0"), val = tensor("custom")]; + tensor obj_151_pad_0 = const()[name = tensor("obj_151_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(681626944)))]; + 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(684903808)))]; + tensor obj_151_cast_fp16 = conv(bias = layers_10_encoder_attn_o_proj_bias_to_fp16, dilations = var_2483, groups = var_2323, pad = obj_151_pad_0, pad_type = obj_151_pad_type_0, strides = var_2481, 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 var_2492 = const()[name = tensor("op_2492"), val = tensor([1])]; + tensor channels_mean_65_cast_fp16 = reduce_mean(axes = var_2492, keep_dims = var_2324, x = inputs_65_cast_fp16)[name = tensor("channels_mean_65_cast_fp16")]; + tensor zero_mean_65_cast_fp16 = sub(x = inputs_65_cast_fp16, y = channels_mean_65_cast_fp16)[name = tensor("zero_mean_65_cast_fp16")]; + tensor zero_mean_sq_65_cast_fp16 = mul(x = zero_mean_65_cast_fp16, y = zero_mean_65_cast_fp16)[name = tensor("zero_mean_sq_65_cast_fp16")]; + tensor var_2496 = const()[name = tensor("op_2496"), val = tensor([1])]; + tensor var_2497_cast_fp16 = reduce_mean(axes = var_2496, keep_dims = var_2324, x = zero_mean_sq_65_cast_fp16)[name = tensor("op_2497_cast_fp16")]; + tensor var_2498_to_fp16 = const()[name = tensor("op_2498_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2499_cast_fp16 = add(x = var_2497_cast_fp16, y = var_2498_to_fp16)[name = tensor("op_2499_cast_fp16")]; + tensor denom_65_epsilon_0_to_fp16 = const()[name = tensor("denom_65_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_65_cast_fp16 = rsqrt(epsilon = denom_65_epsilon_0_to_fp16, x = var_2499_cast_fp16)[name = tensor("denom_65_cast_fp16")]; + tensor out_65_cast_fp16 = mul(x = zero_mean_65_cast_fp16, y = denom_65_cast_fp16)[name = tensor("out_65_cast_fp16")]; + tensor input_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(684906432)))]; + 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(684909056)))]; + 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 var_2510 = const()[name = tensor("op_2510"), val = tensor([1, 1])]; + tensor var_2512 = const()[name = tensor("op_2512"), val = tensor([1, 1])]; + tensor input_107_pad_type_0 = const()[name = tensor("input_107_pad_type_0"), val = tensor("custom")]; + tensor input_107_pad_0 = const()[name = tensor("input_107_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(684911680)))]; + 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(698018944)))]; + tensor input_107_cast_fp16 = conv(bias = layers_10_fc1_bias_to_fp16, dilations = var_2512, groups = var_2323, pad = input_107_pad_0, pad_type = input_107_pad_type_0, strides = var_2510, 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 var_2518 = const()[name = tensor("op_2518"), val = tensor([1, 1])]; + tensor var_2520 = const()[name = tensor("op_2520"), val = tensor([1, 1])]; + tensor hidden_states_23_pad_type_0 = const()[name = tensor("hidden_states_23_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_23_pad_0 = const()[name = tensor("hidden_states_23_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(698029248)))]; + 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(711136512)))]; + tensor hidden_states_23_cast_fp16 = conv(bias = layers_10_fc2_bias_to_fp16, dilations = var_2520, groups = var_2323, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = var_2518, 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_2534 = const()[name = tensor("op_2534"), val = tensor(3)]; + tensor var_2541 = const()[name = tensor("op_2541"), val = tensor(1)]; + tensor var_2542 = const()[name = tensor("op_2542"), val = tensor(true)]; + tensor var_2554 = const()[name = tensor("op_2554"), val = tensor([1])]; + tensor channels_mean_67_cast_fp16 = reduce_mean(axes = var_2554, keep_dims = var_2542, x = inputs_67_cast_fp16)[name = tensor("channels_mean_67_cast_fp16")]; + tensor zero_mean_67_cast_fp16 = sub(x = inputs_67_cast_fp16, y = channels_mean_67_cast_fp16)[name = tensor("zero_mean_67_cast_fp16")]; + tensor zero_mean_sq_67_cast_fp16 = mul(x = zero_mean_67_cast_fp16, y = zero_mean_67_cast_fp16)[name = tensor("zero_mean_sq_67_cast_fp16")]; + tensor var_2558 = const()[name = tensor("op_2558"), val = tensor([1])]; + tensor var_2559_cast_fp16 = reduce_mean(axes = var_2558, keep_dims = var_2542, x = zero_mean_sq_67_cast_fp16)[name = tensor("op_2559_cast_fp16")]; + tensor var_2560_to_fp16 = const()[name = tensor("op_2560_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2561_cast_fp16 = add(x = var_2559_cast_fp16, y = var_2560_to_fp16)[name = tensor("op_2561_cast_fp16")]; + tensor denom_67_epsilon_0_to_fp16 = const()[name = tensor("denom_67_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_67_cast_fp16 = rsqrt(epsilon = denom_67_epsilon_0_to_fp16, x = var_2561_cast_fp16)[name = tensor("denom_67_cast_fp16")]; + tensor out_67_cast_fp16 = mul(x = zero_mean_67_cast_fp16, y = denom_67_cast_fp16)[name = tensor("out_67_cast_fp16")]; + tensor obj_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(711139136)))]; + 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(711141760)))]; + 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 var_2576 = const()[name = tensor("op_2576"), val = tensor([1, 1])]; + tensor var_2578 = const()[name = tensor("op_2578"), val = tensor([1, 1])]; + tensor query_45_pad_type_0 = const()[name = tensor("query_45_pad_type_0"), val = tensor("custom")]; + tensor query_45_pad_0 = const()[name = tensor("query_45_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(711144384)))]; + 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(714421248)))]; + tensor query_45_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_bias_to_fp16, dilations = var_2578, groups = var_2541, pad = query_45_pad_0, pad_type = query_45_pad_type_0, strides = var_2576, weight = layers_11_self_attn_q_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("query_45_cast_fp16")]; + tensor var_2582 = const()[name = tensor("op_2582"), val = tensor([1, 1])]; + tensor var_2584 = const()[name = tensor("op_2584"), val = tensor([1, 1])]; + tensor current_key_23_pad_type_0 = const()[name = tensor("current_key_23_pad_type_0"), val = tensor("custom")]; + tensor current_key_23_pad_0 = const()[name = tensor("current_key_23_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(714423872)))]; + tensor current_key_23_cast_fp16 = conv(dilations = var_2584, groups = var_2541, pad = current_key_23_pad_0, pad_type = current_key_23_pad_type_0, strides = var_2582, weight = layers_11_self_attn_k_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("current_key_23_cast_fp16")]; + tensor var_2589 = const()[name = tensor("op_2589"), val = tensor([1, 1])]; + tensor var_2591 = const()[name = tensor("op_2591"), val = tensor([1, 1])]; + tensor current_value_23_pad_type_0 = const()[name = tensor("current_value_23_pad_type_0"), val = tensor("custom")]; + tensor current_value_23_pad_0 = const()[name = tensor("current_value_23_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(717700736)))]; + 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(720977600)))]; + tensor current_value_23_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_bias_to_fp16, dilations = var_2591, groups = var_2541, pad = current_value_23_pad_0, pad_type = current_value_23_pad_type_0, strides = var_2589, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("current_value_23_cast_fp16")]; + tensor var_2598_cast_fp16 = mul(x = current_key_23_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_2598_cast_fp16")]; + tensor var_2600_cast_fp16 = mul(x = var_103_cast_fp16_11, y = var_241_cast_fp16)[name = tensor("op_2600_cast_fp16")]; + tensor key_45_cast_fp16 = add(x = var_2598_cast_fp16, y = var_2600_cast_fp16)[name = tensor("key_45_cast_fp16")]; + tensor var_2602_cast_fp16 = mul(x = current_value_23_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_2602_cast_fp16")]; + tensor var_2604_cast_fp16 = mul(x = var_138_cast_fp16_11, y = var_241_cast_fp16)[name = tensor("op_2604_cast_fp16")]; + tensor value_45_cast_fp16 = add(x = var_2602_cast_fp16, y = var_2604_cast_fp16)[name = tensor("value_45_cast_fp16")]; + tensor var_2607 = const()[name = tensor("op_2607"), val = tensor([1, 20, 64, -1])]; + tensor var_2608_cast_fp16 = reshape(shape = var_2607, x = query_45_cast_fp16)[name = tensor("op_2608_cast_fp16")]; + tensor var_2609_to_fp16 = const()[name = tensor("op_2609_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2610_cast_fp16 = mul(x = var_2608_cast_fp16, y = var_2609_to_fp16)[name = tensor("op_2610_cast_fp16")]; + tensor var_2611 = const()[name = tensor("op_2611"), val = tensor([1, 20, 64, -1])]; + tensor var_2612_cast_fp16 = reshape(shape = var_2611, x = key_45_cast_fp16)[name = tensor("op_2612_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_2610_cast_fp16, y = var_2612_cast_fp16)[name = tensor("mh_w_67_cast_fp16")]; + tensor mh_w_69_cast_fp16 = add(x = mh_w_67_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_69_cast_fp16")]; + tensor var_2620_cast_fp16 = softmax(axis = var_2534, x = mh_w_69_cast_fp16)[name = tensor("op_2620_cast_fp16")]; + tensor var_2621 = const()[name = tensor("op_2621"), val = tensor([1, 20, 64, -1])]; + tensor var_2622_cast_fp16 = reshape(shape = var_2621, x = value_45_cast_fp16)[name = tensor("op_2622_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_2622_cast_fp16, y = var_2620_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_2625 = const()[name = tensor("op_2625"), val = tensor([1, 1280, 1, -1])]; + tensor input_111_cast_fp16 = reshape(shape = var_2625, x = attn_45_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor var_2629 = const()[name = tensor("op_2629"), val = tensor([1, 1])]; + tensor var_2631 = const()[name = tensor("op_2631"), val = tensor([1, 1])]; + tensor obj_161_pad_type_0 = const()[name = tensor("obj_161_pad_type_0"), val = tensor("custom")]; + tensor obj_161_pad_0 = const()[name = tensor("obj_161_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(720980224)))]; + 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(724257088)))]; + tensor obj_161_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_bias_to_fp16, dilations = var_2631, groups = var_2541, pad = obj_161_pad_0, pad_type = obj_161_pad_type_0, strides = var_2629, 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 var_2641 = const()[name = tensor("op_2641"), val = tensor([1])]; + tensor channels_mean_69_cast_fp16 = reduce_mean(axes = var_2641, keep_dims = var_2542, x = inputs_69_cast_fp16)[name = tensor("channels_mean_69_cast_fp16")]; + tensor zero_mean_69_cast_fp16 = sub(x = inputs_69_cast_fp16, y = channels_mean_69_cast_fp16)[name = tensor("zero_mean_69_cast_fp16")]; + tensor zero_mean_sq_69_cast_fp16 = mul(x = zero_mean_69_cast_fp16, y = zero_mean_69_cast_fp16)[name = tensor("zero_mean_sq_69_cast_fp16")]; + tensor var_2645 = const()[name = tensor("op_2645"), val = tensor([1])]; + tensor var_2646_cast_fp16 = reduce_mean(axes = var_2645, keep_dims = var_2542, x = zero_mean_sq_69_cast_fp16)[name = tensor("op_2646_cast_fp16")]; + tensor var_2647_to_fp16 = const()[name = tensor("op_2647_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2648_cast_fp16 = add(x = var_2646_cast_fp16, y = var_2647_to_fp16)[name = tensor("op_2648_cast_fp16")]; + tensor denom_69_epsilon_0_to_fp16 = const()[name = tensor("denom_69_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_69_cast_fp16 = rsqrt(epsilon = denom_69_epsilon_0_to_fp16, x = var_2648_cast_fp16)[name = tensor("denom_69_cast_fp16")]; + tensor out_69_cast_fp16 = mul(x = zero_mean_69_cast_fp16, y = denom_69_cast_fp16)[name = tensor("out_69_cast_fp16")]; + tensor obj_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(724259712)))]; + 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(724262336)))]; + 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 var_2663 = const()[name = tensor("op_2663"), val = tensor([1, 1])]; + tensor var_2665 = const()[name = tensor("op_2665"), val = tensor([1, 1])]; + tensor query_47_pad_type_0 = const()[name = tensor("query_47_pad_type_0"), val = tensor("custom")]; + tensor query_47_pad_0 = const()[name = tensor("query_47_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(724264960)))]; + 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(727541824)))]; + tensor query_47_cast_fp16 = conv(bias = layers_11_encoder_attn_q_proj_bias_to_fp16, dilations = var_2665, groups = var_2541, pad = query_47_pad_0, pad_type = query_47_pad_type_0, strides = var_2663, weight = layers_11_encoder_attn_q_proj_weight_to_fp16, x = obj_163_cast_fp16)[name = tensor("query_47_cast_fp16")]; + tensor var_2669 = const()[name = tensor("op_2669"), val = tensor([1, 1])]; + tensor var_2671 = const()[name = tensor("op_2671"), val = tensor([1, 1])]; + tensor key_47_pad_type_0 = const()[name = tensor("key_47_pad_type_0"), val = tensor("custom")]; + tensor key_47_pad_0 = const()[name = tensor("key_47_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(727544448)))]; + tensor key_47_cast_fp16 = conv(dilations = var_2671, groups = var_2541, pad = key_47_pad_0, pad_type = key_47_pad_type_0, strides = var_2669, weight = layers_11_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_47_cast_fp16")]; + tensor var_2676 = const()[name = tensor("op_2676"), val = tensor([1, 1])]; + tensor var_2678 = const()[name = tensor("op_2678"), val = tensor([1, 1])]; + tensor value_47_pad_type_0 = const()[name = tensor("value_47_pad_type_0"), val = tensor("custom")]; + tensor value_47_pad_0 = const()[name = tensor("value_47_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(730821312)))]; + 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(734098176)))]; + tensor value_47_cast_fp16 = conv(bias = layers_11_encoder_attn_v_proj_bias_to_fp16, dilations = var_2678, groups = var_2541, pad = value_47_pad_0, pad_type = value_47_pad_type_0, strides = var_2676, weight = layers_11_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_47_cast_fp16")]; + tensor var_2682 = const()[name = tensor("op_2682"), val = tensor([1, 20, 64, -1])]; + tensor var_2683_cast_fp16 = reshape(shape = var_2682, x = query_47_cast_fp16)[name = tensor("op_2683_cast_fp16")]; + tensor var_2684_to_fp16 = const()[name = tensor("op_2684_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2685_cast_fp16 = mul(x = var_2683_cast_fp16, y = var_2684_to_fp16)[name = tensor("op_2685_cast_fp16")]; + tensor var_2686 = const()[name = tensor("op_2686"), val = tensor([1, 20, 64, -1])]; + tensor var_2687_cast_fp16 = reshape(shape = var_2686, x = key_47_cast_fp16)[name = tensor("op_2687_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_2685_cast_fp16, y = var_2687_cast_fp16)[name = tensor("mh_w_71_cast_fp16")]; + tensor obj_167_cast_fp16 = softmax(axis = var_2534, x = mh_w_71_cast_fp16)[name = tensor("obj_167_cast_fp16")]; + tensor var_2691 = const()[name = tensor("op_2691"), val = tensor([1, 20, 64, -1])]; + tensor var_2692_cast_fp16 = reshape(shape = var_2691, x = value_47_cast_fp16)[name = tensor("op_2692_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_2692_cast_fp16, y = obj_167_cast_fp16)[name = tensor("attn_47_cast_fp16")]; + tensor var_2695 = const()[name = tensor("op_2695"), val = tensor([1, 1280, 1, -1])]; + tensor input_113_cast_fp16 = reshape(shape = var_2695, x = attn_47_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor var_2699 = const()[name = tensor("op_2699"), val = tensor([1, 1])]; + tensor var_2701 = const()[name = tensor("op_2701"), val = tensor([1, 1])]; + tensor obj_165_pad_type_0 = const()[name = tensor("obj_165_pad_type_0"), val = tensor("custom")]; + tensor obj_165_pad_0 = const()[name = tensor("obj_165_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(734100800)))]; + 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(737377664)))]; + tensor obj_165_cast_fp16 = conv(bias = layers_11_encoder_attn_o_proj_bias_to_fp16, dilations = var_2701, groups = var_2541, pad = obj_165_pad_0, pad_type = obj_165_pad_type_0, strides = var_2699, 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 var_2707 = const()[name = tensor("op_2707"), val = tensor([1])]; + tensor channels_mean_71_cast_fp16 = reduce_mean(axes = var_2707, keep_dims = var_2542, x = inputs_71_cast_fp16)[name = tensor("channels_mean_71_cast_fp16")]; + tensor zero_mean_71_cast_fp16 = sub(x = inputs_71_cast_fp16, y = channels_mean_71_cast_fp16)[name = tensor("zero_mean_71_cast_fp16")]; + tensor zero_mean_sq_71_cast_fp16 = mul(x = zero_mean_71_cast_fp16, y = zero_mean_71_cast_fp16)[name = tensor("zero_mean_sq_71_cast_fp16")]; + tensor var_2711 = const()[name = tensor("op_2711"), val = tensor([1])]; + tensor var_2712_cast_fp16 = reduce_mean(axes = var_2711, keep_dims = var_2542, x = zero_mean_sq_71_cast_fp16)[name = tensor("op_2712_cast_fp16")]; + tensor var_2713_to_fp16 = const()[name = tensor("op_2713_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2714_cast_fp16 = add(x = var_2712_cast_fp16, y = var_2713_to_fp16)[name = tensor("op_2714_cast_fp16")]; + tensor denom_71_epsilon_0_to_fp16 = const()[name = tensor("denom_71_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_71_cast_fp16 = rsqrt(epsilon = denom_71_epsilon_0_to_fp16, x = var_2714_cast_fp16)[name = tensor("denom_71_cast_fp16")]; + tensor out_71_cast_fp16 = mul(x = zero_mean_71_cast_fp16, y = denom_71_cast_fp16)[name = tensor("out_71_cast_fp16")]; + tensor input_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(737380288)))]; + 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(737382912)))]; + 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 var_2725 = const()[name = tensor("op_2725"), val = tensor([1, 1])]; + tensor var_2727 = const()[name = tensor("op_2727"), val = tensor([1, 1])]; + tensor input_117_pad_type_0 = const()[name = tensor("input_117_pad_type_0"), val = tensor("custom")]; + tensor input_117_pad_0 = const()[name = tensor("input_117_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(737385536)))]; + 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(750492800)))]; + tensor input_117_cast_fp16 = conv(bias = layers_11_fc1_bias_to_fp16, dilations = var_2727, groups = var_2541, pad = input_117_pad_0, pad_type = input_117_pad_type_0, strides = var_2725, 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 var_2733 = const()[name = tensor("op_2733"), val = tensor([1, 1])]; + tensor var_2735 = const()[name = tensor("op_2735"), val = tensor([1, 1])]; + tensor hidden_states_25_pad_type_0 = const()[name = tensor("hidden_states_25_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_25_pad_0 = const()[name = tensor("hidden_states_25_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(750503104)))]; + 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(763610368)))]; + tensor hidden_states_25_cast_fp16 = conv(bias = layers_11_fc2_bias_to_fp16, dilations = var_2735, groups = var_2541, pad = hidden_states_25_pad_0, pad_type = hidden_states_25_pad_type_0, strides = var_2733, 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_2748 = const()[name = tensor("op_2748"), val = tensor(3)]; + tensor var_2755 = const()[name = tensor("op_2755"), val = tensor(1)]; + tensor var_2756 = const()[name = tensor("op_2756"), val = tensor(true)]; + tensor var_2768 = const()[name = tensor("op_2768"), val = tensor([1])]; + tensor channels_mean_73_cast_fp16 = reduce_mean(axes = var_2768, keep_dims = var_2756, x = inputs_73_cast_fp16)[name = tensor("channels_mean_73_cast_fp16")]; + tensor zero_mean_73_cast_fp16 = sub(x = inputs_73_cast_fp16, y = channels_mean_73_cast_fp16)[name = tensor("zero_mean_73_cast_fp16")]; + tensor zero_mean_sq_73_cast_fp16 = mul(x = zero_mean_73_cast_fp16, y = zero_mean_73_cast_fp16)[name = tensor("zero_mean_sq_73_cast_fp16")]; + tensor var_2772 = const()[name = tensor("op_2772"), val = tensor([1])]; + tensor var_2773_cast_fp16 = reduce_mean(axes = var_2772, keep_dims = var_2756, x = zero_mean_sq_73_cast_fp16)[name = tensor("op_2773_cast_fp16")]; + tensor var_2774_to_fp16 = const()[name = tensor("op_2774_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2775_cast_fp16 = add(x = var_2773_cast_fp16, y = var_2774_to_fp16)[name = tensor("op_2775_cast_fp16")]; + tensor denom_73_epsilon_0_to_fp16 = const()[name = tensor("denom_73_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_73_cast_fp16 = rsqrt(epsilon = denom_73_epsilon_0_to_fp16, x = var_2775_cast_fp16)[name = tensor("denom_73_cast_fp16")]; + tensor out_73_cast_fp16 = mul(x = zero_mean_73_cast_fp16, y = denom_73_cast_fp16)[name = tensor("out_73_cast_fp16")]; + tensor obj_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(763612992)))]; + 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(763615616)))]; + 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 var_2790 = const()[name = tensor("op_2790"), val = tensor([1, 1])]; + tensor var_2792 = const()[name = tensor("op_2792"), val = tensor([1, 1])]; + tensor query_49_pad_type_0 = const()[name = tensor("query_49_pad_type_0"), val = tensor("custom")]; + tensor query_49_pad_0 = const()[name = tensor("query_49_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(763618240)))]; + 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(766895104)))]; + tensor query_49_cast_fp16 = conv(bias = layers_12_self_attn_q_proj_bias_to_fp16, dilations = var_2792, groups = var_2755, pad = query_49_pad_0, pad_type = query_49_pad_type_0, strides = var_2790, weight = layers_12_self_attn_q_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("query_49_cast_fp16")]; + tensor var_2796 = const()[name = tensor("op_2796"), val = tensor([1, 1])]; + tensor var_2798 = const()[name = tensor("op_2798"), val = tensor([1, 1])]; + tensor current_key_25_pad_type_0 = const()[name = tensor("current_key_25_pad_type_0"), val = tensor("custom")]; + tensor current_key_25_pad_0 = const()[name = tensor("current_key_25_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(766897728)))]; + tensor current_key_25_cast_fp16 = conv(dilations = var_2798, groups = var_2755, pad = current_key_25_pad_0, pad_type = current_key_25_pad_type_0, strides = var_2796, weight = layers_12_self_attn_k_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("current_key_25_cast_fp16")]; + tensor var_2803 = const()[name = tensor("op_2803"), val = tensor([1, 1])]; + tensor var_2805 = const()[name = tensor("op_2805"), val = tensor([1, 1])]; + tensor current_value_25_pad_type_0 = const()[name = tensor("current_value_25_pad_type_0"), val = tensor("custom")]; + tensor current_value_25_pad_0 = const()[name = tensor("current_value_25_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(770174592)))]; + 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(773451456)))]; + tensor current_value_25_cast_fp16 = conv(bias = layers_12_self_attn_v_proj_bias_to_fp16, dilations = var_2805, groups = var_2755, pad = current_value_25_pad_0, pad_type = current_value_25_pad_type_0, strides = var_2803, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("current_value_25_cast_fp16")]; + tensor var_2812_cast_fp16 = mul(x = current_key_25_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_2812_cast_fp16")]; + tensor var_2814_cast_fp16 = mul(x = var_103_cast_fp16_12, y = var_241_cast_fp16)[name = tensor("op_2814_cast_fp16")]; + tensor key_49_cast_fp16 = add(x = var_2812_cast_fp16, y = var_2814_cast_fp16)[name = tensor("key_49_cast_fp16")]; + tensor var_2816_cast_fp16 = mul(x = current_value_25_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_2816_cast_fp16")]; + tensor var_2818_cast_fp16 = mul(x = var_138_cast_fp16_12, y = var_241_cast_fp16)[name = tensor("op_2818_cast_fp16")]; + tensor value_49_cast_fp16 = add(x = var_2816_cast_fp16, y = var_2818_cast_fp16)[name = tensor("value_49_cast_fp16")]; + tensor var_2821 = const()[name = tensor("op_2821"), val = tensor([1, 20, 64, -1])]; + tensor var_2822_cast_fp16 = reshape(shape = var_2821, x = query_49_cast_fp16)[name = tensor("op_2822_cast_fp16")]; + tensor var_2823_to_fp16 = const()[name = tensor("op_2823_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2824_cast_fp16 = mul(x = var_2822_cast_fp16, y = var_2823_to_fp16)[name = tensor("op_2824_cast_fp16")]; + tensor var_2825 = const()[name = tensor("op_2825"), val = tensor([1, 20, 64, -1])]; + tensor var_2826_cast_fp16 = reshape(shape = var_2825, x = key_49_cast_fp16)[name = tensor("op_2826_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_2824_cast_fp16, y = var_2826_cast_fp16)[name = tensor("mh_w_73_cast_fp16")]; + tensor mh_w_75_cast_fp16 = add(x = mh_w_73_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_75_cast_fp16")]; + tensor var_2834_cast_fp16 = softmax(axis = var_2748, x = mh_w_75_cast_fp16)[name = tensor("op_2834_cast_fp16")]; + tensor var_2835 = const()[name = tensor("op_2835"), val = tensor([1, 20, 64, -1])]; + tensor var_2836_cast_fp16 = reshape(shape = var_2835, x = value_49_cast_fp16)[name = tensor("op_2836_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_2836_cast_fp16, y = var_2834_cast_fp16)[name = tensor("attn_49_cast_fp16")]; + tensor var_2839 = const()[name = tensor("op_2839"), val = tensor([1, 1280, 1, -1])]; + tensor input_121_cast_fp16 = reshape(shape = var_2839, x = attn_49_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor var_2843 = const()[name = tensor("op_2843"), val = tensor([1, 1])]; + tensor var_2845 = const()[name = tensor("op_2845"), val = tensor([1, 1])]; + tensor obj_175_pad_type_0 = const()[name = tensor("obj_175_pad_type_0"), val = tensor("custom")]; + tensor obj_175_pad_0 = const()[name = tensor("obj_175_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(773454080)))]; + 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(776730944)))]; + tensor obj_175_cast_fp16 = conv(bias = layers_12_self_attn_o_proj_bias_to_fp16, dilations = var_2845, groups = var_2755, pad = obj_175_pad_0, pad_type = obj_175_pad_type_0, strides = var_2843, 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 var_2855 = const()[name = tensor("op_2855"), val = tensor([1])]; + tensor channels_mean_75_cast_fp16 = reduce_mean(axes = var_2855, keep_dims = var_2756, x = inputs_75_cast_fp16)[name = tensor("channels_mean_75_cast_fp16")]; + tensor zero_mean_75_cast_fp16 = sub(x = inputs_75_cast_fp16, y = channels_mean_75_cast_fp16)[name = tensor("zero_mean_75_cast_fp16")]; + tensor zero_mean_sq_75_cast_fp16 = mul(x = zero_mean_75_cast_fp16, y = zero_mean_75_cast_fp16)[name = tensor("zero_mean_sq_75_cast_fp16")]; + tensor var_2859 = const()[name = tensor("op_2859"), val = tensor([1])]; + tensor var_2860_cast_fp16 = reduce_mean(axes = var_2859, keep_dims = var_2756, x = zero_mean_sq_75_cast_fp16)[name = tensor("op_2860_cast_fp16")]; + tensor var_2861_to_fp16 = const()[name = tensor("op_2861_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2862_cast_fp16 = add(x = var_2860_cast_fp16, y = var_2861_to_fp16)[name = tensor("op_2862_cast_fp16")]; + tensor denom_75_epsilon_0_to_fp16 = const()[name = tensor("denom_75_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_75_cast_fp16 = rsqrt(epsilon = denom_75_epsilon_0_to_fp16, x = var_2862_cast_fp16)[name = tensor("denom_75_cast_fp16")]; + tensor out_75_cast_fp16 = mul(x = zero_mean_75_cast_fp16, y = denom_75_cast_fp16)[name = tensor("out_75_cast_fp16")]; + tensor obj_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(776733568)))]; + 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(776736192)))]; + 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 var_2877 = const()[name = tensor("op_2877"), val = tensor([1, 1])]; + tensor var_2879 = const()[name = tensor("op_2879"), val = tensor([1, 1])]; + tensor query_51_pad_type_0 = const()[name = tensor("query_51_pad_type_0"), val = tensor("custom")]; + tensor query_51_pad_0 = const()[name = tensor("query_51_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(776738816)))]; + 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(780015680)))]; + tensor query_51_cast_fp16 = conv(bias = layers_12_encoder_attn_q_proj_bias_to_fp16, dilations = var_2879, groups = var_2755, pad = query_51_pad_0, pad_type = query_51_pad_type_0, strides = var_2877, weight = layers_12_encoder_attn_q_proj_weight_to_fp16, x = obj_177_cast_fp16)[name = tensor("query_51_cast_fp16")]; + tensor var_2883 = const()[name = tensor("op_2883"), val = tensor([1, 1])]; + tensor var_2885 = const()[name = tensor("op_2885"), val = tensor([1, 1])]; + tensor key_51_pad_type_0 = const()[name = tensor("key_51_pad_type_0"), val = tensor("custom")]; + tensor key_51_pad_0 = const()[name = tensor("key_51_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(780018304)))]; + tensor key_51_cast_fp16 = conv(dilations = var_2885, groups = var_2755, pad = key_51_pad_0, pad_type = key_51_pad_type_0, strides = var_2883, weight = layers_12_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_51_cast_fp16")]; + tensor var_2890 = const()[name = tensor("op_2890"), val = tensor([1, 1])]; + tensor var_2892 = const()[name = tensor("op_2892"), val = tensor([1, 1])]; + tensor value_51_pad_type_0 = const()[name = tensor("value_51_pad_type_0"), val = tensor("custom")]; + tensor value_51_pad_0 = const()[name = tensor("value_51_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(783295168)))]; + 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(786572032)))]; + tensor value_51_cast_fp16 = conv(bias = layers_12_encoder_attn_v_proj_bias_to_fp16, dilations = var_2892, groups = var_2755, pad = value_51_pad_0, pad_type = value_51_pad_type_0, strides = var_2890, weight = layers_12_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_51_cast_fp16")]; + tensor var_2896 = const()[name = tensor("op_2896"), val = tensor([1, 20, 64, -1])]; + tensor var_2897_cast_fp16 = reshape(shape = var_2896, x = query_51_cast_fp16)[name = tensor("op_2897_cast_fp16")]; + tensor var_2898_to_fp16 = const()[name = tensor("op_2898_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2899_cast_fp16 = mul(x = var_2897_cast_fp16, y = var_2898_to_fp16)[name = tensor("op_2899_cast_fp16")]; + tensor var_2900 = const()[name = tensor("op_2900"), val = tensor([1, 20, 64, -1])]; + tensor var_2901_cast_fp16 = reshape(shape = var_2900, x = key_51_cast_fp16)[name = tensor("op_2901_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_2899_cast_fp16, y = var_2901_cast_fp16)[name = tensor("mh_w_77_cast_fp16")]; + tensor obj_181_cast_fp16 = softmax(axis = var_2748, x = mh_w_77_cast_fp16)[name = tensor("obj_181_cast_fp16")]; + tensor var_2905 = const()[name = tensor("op_2905"), val = tensor([1, 20, 64, -1])]; + tensor var_2906_cast_fp16 = reshape(shape = var_2905, x = value_51_cast_fp16)[name = tensor("op_2906_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_2906_cast_fp16, y = obj_181_cast_fp16)[name = tensor("attn_51_cast_fp16")]; + tensor var_2909 = const()[name = tensor("op_2909"), val = tensor([1, 1280, 1, -1])]; + tensor input_123_cast_fp16 = reshape(shape = var_2909, x = attn_51_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor var_2913 = const()[name = tensor("op_2913"), val = tensor([1, 1])]; + tensor var_2915 = const()[name = tensor("op_2915"), val = tensor([1, 1])]; + tensor obj_179_pad_type_0 = const()[name = tensor("obj_179_pad_type_0"), val = tensor("custom")]; + tensor obj_179_pad_0 = const()[name = tensor("obj_179_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(786574656)))]; + 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(789851520)))]; + tensor obj_179_cast_fp16 = conv(bias = layers_12_encoder_attn_o_proj_bias_to_fp16, dilations = var_2915, groups = var_2755, pad = obj_179_pad_0, pad_type = obj_179_pad_type_0, strides = var_2913, 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 var_2921 = const()[name = tensor("op_2921"), val = tensor([1])]; + tensor channels_mean_77_cast_fp16 = reduce_mean(axes = var_2921, keep_dims = var_2756, x = inputs_77_cast_fp16)[name = tensor("channels_mean_77_cast_fp16")]; + tensor zero_mean_77_cast_fp16 = sub(x = inputs_77_cast_fp16, y = channels_mean_77_cast_fp16)[name = tensor("zero_mean_77_cast_fp16")]; + tensor zero_mean_sq_77_cast_fp16 = mul(x = zero_mean_77_cast_fp16, y = zero_mean_77_cast_fp16)[name = tensor("zero_mean_sq_77_cast_fp16")]; + tensor var_2925 = const()[name = tensor("op_2925"), val = tensor([1])]; + tensor var_2926_cast_fp16 = reduce_mean(axes = var_2925, keep_dims = var_2756, x = zero_mean_sq_77_cast_fp16)[name = tensor("op_2926_cast_fp16")]; + tensor var_2927_to_fp16 = const()[name = tensor("op_2927_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2928_cast_fp16 = add(x = var_2926_cast_fp16, y = var_2927_to_fp16)[name = tensor("op_2928_cast_fp16")]; + tensor denom_77_epsilon_0_to_fp16 = const()[name = tensor("denom_77_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_77_cast_fp16 = rsqrt(epsilon = denom_77_epsilon_0_to_fp16, x = var_2928_cast_fp16)[name = tensor("denom_77_cast_fp16")]; + tensor out_77_cast_fp16 = mul(x = zero_mean_77_cast_fp16, y = denom_77_cast_fp16)[name = tensor("out_77_cast_fp16")]; + tensor input_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(789854144)))]; + 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(789856768)))]; + 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 var_2939 = const()[name = tensor("op_2939"), val = tensor([1, 1])]; + tensor var_2941 = const()[name = tensor("op_2941"), val = tensor([1, 1])]; + tensor input_127_pad_type_0 = const()[name = tensor("input_127_pad_type_0"), val = tensor("custom")]; + tensor input_127_pad_0 = const()[name = tensor("input_127_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(789859392)))]; + 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(802966656)))]; + tensor input_127_cast_fp16 = conv(bias = layers_12_fc1_bias_to_fp16, dilations = var_2941, groups = var_2755, pad = input_127_pad_0, pad_type = input_127_pad_type_0, strides = var_2939, 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 var_2947 = const()[name = tensor("op_2947"), val = tensor([1, 1])]; + tensor var_2949 = const()[name = tensor("op_2949"), val = tensor([1, 1])]; + tensor hidden_states_27_pad_type_0 = const()[name = tensor("hidden_states_27_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_27_pad_0 = const()[name = tensor("hidden_states_27_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(802976960)))]; + 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(816084224)))]; + tensor hidden_states_27_cast_fp16 = conv(bias = layers_12_fc2_bias_to_fp16, dilations = var_2949, groups = var_2755, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = var_2947, 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_2962 = const()[name = tensor("op_2962"), val = tensor(3)]; + tensor var_2969 = const()[name = tensor("op_2969"), val = tensor(1)]; + tensor var_2970 = const()[name = tensor("op_2970"), val = tensor(true)]; + tensor var_2982 = const()[name = tensor("op_2982"), val = tensor([1])]; + tensor channels_mean_79_cast_fp16 = reduce_mean(axes = var_2982, keep_dims = var_2970, x = inputs_79_cast_fp16)[name = tensor("channels_mean_79_cast_fp16")]; + tensor zero_mean_79_cast_fp16 = sub(x = inputs_79_cast_fp16, y = channels_mean_79_cast_fp16)[name = tensor("zero_mean_79_cast_fp16")]; + tensor zero_mean_sq_79_cast_fp16 = mul(x = zero_mean_79_cast_fp16, y = zero_mean_79_cast_fp16)[name = tensor("zero_mean_sq_79_cast_fp16")]; + tensor var_2986 = const()[name = tensor("op_2986"), val = tensor([1])]; + tensor var_2987_cast_fp16 = reduce_mean(axes = var_2986, keep_dims = var_2970, x = zero_mean_sq_79_cast_fp16)[name = tensor("op_2987_cast_fp16")]; + tensor var_2988_to_fp16 = const()[name = tensor("op_2988_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2989_cast_fp16 = add(x = var_2987_cast_fp16, y = var_2988_to_fp16)[name = tensor("op_2989_cast_fp16")]; + tensor denom_79_epsilon_0_to_fp16 = const()[name = tensor("denom_79_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_79_cast_fp16 = rsqrt(epsilon = denom_79_epsilon_0_to_fp16, x = var_2989_cast_fp16)[name = tensor("denom_79_cast_fp16")]; + tensor out_79_cast_fp16 = mul(x = zero_mean_79_cast_fp16, y = denom_79_cast_fp16)[name = tensor("out_79_cast_fp16")]; + tensor obj_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(816086848)))]; + 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(816089472)))]; + 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 var_3004 = const()[name = tensor("op_3004"), val = tensor([1, 1])]; + tensor var_3006 = const()[name = tensor("op_3006"), val = tensor([1, 1])]; + tensor query_53_pad_type_0 = const()[name = tensor("query_53_pad_type_0"), val = tensor("custom")]; + tensor query_53_pad_0 = const()[name = tensor("query_53_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(816092096)))]; + 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(819368960)))]; + tensor query_53_cast_fp16 = conv(bias = layers_13_self_attn_q_proj_bias_to_fp16, dilations = var_3006, groups = var_2969, pad = query_53_pad_0, pad_type = query_53_pad_type_0, strides = var_3004, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor("query_53_cast_fp16")]; + tensor var_3010 = const()[name = tensor("op_3010"), val = tensor([1, 1])]; + tensor var_3012 = const()[name = tensor("op_3012"), val = tensor([1, 1])]; + tensor current_key_27_pad_type_0 = const()[name = tensor("current_key_27_pad_type_0"), val = tensor("custom")]; + tensor current_key_27_pad_0 = const()[name = tensor("current_key_27_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(819371584)))]; + tensor current_key_27_cast_fp16 = conv(dilations = var_3012, groups = var_2969, pad = current_key_27_pad_0, pad_type = current_key_27_pad_type_0, strides = var_3010, weight = layers_13_self_attn_k_proj_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor("current_key_27_cast_fp16")]; + tensor var_3017 = const()[name = tensor("op_3017"), val = tensor([1, 1])]; + tensor var_3019 = const()[name = tensor("op_3019"), val = tensor([1, 1])]; + tensor current_value_27_pad_type_0 = const()[name = tensor("current_value_27_pad_type_0"), val = tensor("custom")]; + tensor current_value_27_pad_0 = const()[name = tensor("current_value_27_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(822648448)))]; + 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(825925312)))]; + tensor current_value_27_cast_fp16 = conv(bias = layers_13_self_attn_v_proj_bias_to_fp16, dilations = var_3019, groups = var_2969, pad = current_value_27_pad_0, pad_type = current_value_27_pad_type_0, strides = var_3017, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor("current_value_27_cast_fp16")]; + tensor var_3026_cast_fp16 = mul(x = current_key_27_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3026_cast_fp16")]; + tensor var_3028_cast_fp16 = mul(x = var_103_cast_fp16_13, y = var_241_cast_fp16)[name = tensor("op_3028_cast_fp16")]; + tensor key_53_cast_fp16 = add(x = var_3026_cast_fp16, y = var_3028_cast_fp16)[name = tensor("key_53_cast_fp16")]; + tensor var_3030_cast_fp16 = mul(x = current_value_27_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3030_cast_fp16")]; + tensor var_3032_cast_fp16 = mul(x = var_138_cast_fp16_13, y = var_241_cast_fp16)[name = tensor("op_3032_cast_fp16")]; + tensor value_53_cast_fp16 = add(x = var_3030_cast_fp16, y = var_3032_cast_fp16)[name = tensor("value_53_cast_fp16")]; + tensor var_3035 = const()[name = tensor("op_3035"), val = tensor([1, 20, 64, -1])]; + tensor var_3036_cast_fp16 = reshape(shape = var_3035, x = query_53_cast_fp16)[name = tensor("op_3036_cast_fp16")]; + tensor var_3037_to_fp16 = const()[name = tensor("op_3037_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3038_cast_fp16 = mul(x = var_3036_cast_fp16, y = var_3037_to_fp16)[name = tensor("op_3038_cast_fp16")]; + tensor var_3039 = const()[name = tensor("op_3039"), val = tensor([1, 20, 64, -1])]; + tensor var_3040_cast_fp16 = reshape(shape = var_3039, x = key_53_cast_fp16)[name = tensor("op_3040_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_3038_cast_fp16, y = var_3040_cast_fp16)[name = tensor("mh_w_79_cast_fp16")]; + tensor mh_w_81_cast_fp16 = add(x = mh_w_79_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_81_cast_fp16")]; + tensor var_3048_cast_fp16 = softmax(axis = var_2962, x = mh_w_81_cast_fp16)[name = tensor("op_3048_cast_fp16")]; + tensor var_3049 = const()[name = tensor("op_3049"), val = tensor([1, 20, 64, -1])]; + tensor var_3050_cast_fp16 = reshape(shape = var_3049, x = value_53_cast_fp16)[name = tensor("op_3050_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_3050_cast_fp16, y = var_3048_cast_fp16)[name = tensor("attn_53_cast_fp16")]; + tensor var_3053 = const()[name = tensor("op_3053"), val = tensor([1, 1280, 1, -1])]; + tensor input_131_cast_fp16 = reshape(shape = var_3053, x = attn_53_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor var_3057 = const()[name = tensor("op_3057"), val = tensor([1, 1])]; + tensor var_3059 = const()[name = tensor("op_3059"), val = tensor([1, 1])]; + tensor obj_189_pad_type_0 = const()[name = tensor("obj_189_pad_type_0"), val = tensor("custom")]; + tensor obj_189_pad_0 = const()[name = tensor("obj_189_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(825927936)))]; + 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(829204800)))]; + tensor obj_189_cast_fp16 = conv(bias = layers_13_self_attn_o_proj_bias_to_fp16, dilations = var_3059, groups = var_2969, pad = obj_189_pad_0, pad_type = obj_189_pad_type_0, strides = var_3057, 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 var_3069 = const()[name = tensor("op_3069"), val = tensor([1])]; + tensor channels_mean_81_cast_fp16 = reduce_mean(axes = var_3069, keep_dims = var_2970, x = inputs_81_cast_fp16)[name = tensor("channels_mean_81_cast_fp16")]; + tensor zero_mean_81_cast_fp16 = sub(x = inputs_81_cast_fp16, y = channels_mean_81_cast_fp16)[name = tensor("zero_mean_81_cast_fp16")]; + tensor zero_mean_sq_81_cast_fp16 = mul(x = zero_mean_81_cast_fp16, y = zero_mean_81_cast_fp16)[name = tensor("zero_mean_sq_81_cast_fp16")]; + tensor var_3073 = const()[name = tensor("op_3073"), val = tensor([1])]; + tensor var_3074_cast_fp16 = reduce_mean(axes = var_3073, keep_dims = var_2970, x = zero_mean_sq_81_cast_fp16)[name = tensor("op_3074_cast_fp16")]; + tensor var_3075_to_fp16 = const()[name = tensor("op_3075_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3076_cast_fp16 = add(x = var_3074_cast_fp16, y = var_3075_to_fp16)[name = tensor("op_3076_cast_fp16")]; + tensor denom_81_epsilon_0_to_fp16 = const()[name = tensor("denom_81_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_81_cast_fp16 = rsqrt(epsilon = denom_81_epsilon_0_to_fp16, x = var_3076_cast_fp16)[name = tensor("denom_81_cast_fp16")]; + tensor out_81_cast_fp16 = mul(x = zero_mean_81_cast_fp16, y = denom_81_cast_fp16)[name = tensor("out_81_cast_fp16")]; + tensor obj_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(829207424)))]; + 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(829210048)))]; + 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 var_3091 = const()[name = tensor("op_3091"), val = tensor([1, 1])]; + tensor var_3093 = const()[name = tensor("op_3093"), val = tensor([1, 1])]; + tensor query_55_pad_type_0 = const()[name = tensor("query_55_pad_type_0"), val = tensor("custom")]; + tensor query_55_pad_0 = const()[name = tensor("query_55_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(829212672)))]; + 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(832489536)))]; + tensor query_55_cast_fp16 = conv(bias = layers_13_encoder_attn_q_proj_bias_to_fp16, dilations = var_3093, groups = var_2969, pad = query_55_pad_0, pad_type = query_55_pad_type_0, strides = var_3091, weight = layers_13_encoder_attn_q_proj_weight_to_fp16, x = obj_191_cast_fp16)[name = tensor("query_55_cast_fp16")]; + tensor var_3097 = const()[name = tensor("op_3097"), val = tensor([1, 1])]; + tensor var_3099 = const()[name = tensor("op_3099"), val = tensor([1, 1])]; + tensor key_55_pad_type_0 = const()[name = tensor("key_55_pad_type_0"), val = tensor("custom")]; + tensor key_55_pad_0 = const()[name = tensor("key_55_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(832492160)))]; + tensor key_55_cast_fp16 = conv(dilations = var_3099, groups = var_2969, pad = key_55_pad_0, pad_type = key_55_pad_type_0, strides = var_3097, weight = layers_13_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_55_cast_fp16")]; + tensor var_3104 = const()[name = tensor("op_3104"), val = tensor([1, 1])]; + tensor var_3106 = const()[name = tensor("op_3106"), val = tensor([1, 1])]; + tensor value_55_pad_type_0 = const()[name = tensor("value_55_pad_type_0"), val = tensor("custom")]; + tensor value_55_pad_0 = const()[name = tensor("value_55_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(835769024)))]; + 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(839045888)))]; + tensor value_55_cast_fp16 = conv(bias = layers_13_encoder_attn_v_proj_bias_to_fp16, dilations = var_3106, groups = var_2969, pad = value_55_pad_0, pad_type = value_55_pad_type_0, strides = var_3104, weight = layers_13_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_55_cast_fp16")]; + tensor var_3110 = const()[name = tensor("op_3110"), val = tensor([1, 20, 64, -1])]; + tensor var_3111_cast_fp16 = reshape(shape = var_3110, x = query_55_cast_fp16)[name = tensor("op_3111_cast_fp16")]; + tensor var_3112_to_fp16 = const()[name = tensor("op_3112_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3113_cast_fp16 = mul(x = var_3111_cast_fp16, y = var_3112_to_fp16)[name = tensor("op_3113_cast_fp16")]; + tensor var_3114 = const()[name = tensor("op_3114"), val = tensor([1, 20, 64, -1])]; + tensor var_3115_cast_fp16 = reshape(shape = var_3114, x = key_55_cast_fp16)[name = tensor("op_3115_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_3113_cast_fp16, y = var_3115_cast_fp16)[name = tensor("mh_w_83_cast_fp16")]; + tensor obj_195_cast_fp16 = softmax(axis = var_2962, x = mh_w_83_cast_fp16)[name = tensor("obj_195_cast_fp16")]; + tensor var_3119 = const()[name = tensor("op_3119"), val = tensor([1, 20, 64, -1])]; + tensor var_3120_cast_fp16 = reshape(shape = var_3119, x = value_55_cast_fp16)[name = tensor("op_3120_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_3120_cast_fp16, y = obj_195_cast_fp16)[name = tensor("attn_55_cast_fp16")]; + tensor var_3123 = const()[name = tensor("op_3123"), val = tensor([1, 1280, 1, -1])]; + tensor input_133_cast_fp16 = reshape(shape = var_3123, x = attn_55_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor var_3127 = const()[name = tensor("op_3127"), val = tensor([1, 1])]; + tensor var_3129 = const()[name = tensor("op_3129"), val = tensor([1, 1])]; + tensor obj_193_pad_type_0 = const()[name = tensor("obj_193_pad_type_0"), val = tensor("custom")]; + tensor obj_193_pad_0 = const()[name = tensor("obj_193_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(839048512)))]; + 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(842325376)))]; + tensor obj_193_cast_fp16 = conv(bias = layers_13_encoder_attn_o_proj_bias_to_fp16, dilations = var_3129, groups = var_2969, pad = obj_193_pad_0, pad_type = obj_193_pad_type_0, strides = var_3127, 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 var_3138 = const()[name = tensor("op_3138"), val = tensor([1])]; + tensor channels_mean_83_cast_fp16 = reduce_mean(axes = var_3138, keep_dims = var_2970, x = inputs_83_cast_fp16)[name = tensor("channels_mean_83_cast_fp16")]; + tensor zero_mean_83_cast_fp16 = sub(x = inputs_83_cast_fp16, y = channels_mean_83_cast_fp16)[name = tensor("zero_mean_83_cast_fp16")]; + tensor zero_mean_sq_83_cast_fp16 = mul(x = zero_mean_83_cast_fp16, y = zero_mean_83_cast_fp16)[name = tensor("zero_mean_sq_83_cast_fp16")]; + tensor var_3142 = const()[name = tensor("op_3142"), val = tensor([1])]; + tensor var_3143_cast_fp16 = reduce_mean(axes = var_3142, keep_dims = var_2970, x = zero_mean_sq_83_cast_fp16)[name = tensor("op_3143_cast_fp16")]; + tensor var_3144_to_fp16 = const()[name = tensor("op_3144_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3145_cast_fp16 = add(x = var_3143_cast_fp16, y = var_3144_to_fp16)[name = tensor("op_3145_cast_fp16")]; + tensor denom_83_epsilon_0_to_fp16 = const()[name = tensor("denom_83_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_83_cast_fp16 = rsqrt(epsilon = denom_83_epsilon_0_to_fp16, x = var_3145_cast_fp16)[name = tensor("denom_83_cast_fp16")]; + tensor out_83_cast_fp16 = mul(x = zero_mean_83_cast_fp16, y = denom_83_cast_fp16)[name = tensor("out_83_cast_fp16")]; + tensor input_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(842328000)))]; + 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(842330624)))]; + 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 var_3156 = const()[name = tensor("op_3156"), val = tensor([1, 1])]; + tensor var_3158 = const()[name = tensor("op_3158"), val = tensor([1, 1])]; + tensor input_137_pad_type_0 = const()[name = tensor("input_137_pad_type_0"), val = tensor("custom")]; + tensor input_137_pad_0 = const()[name = tensor("input_137_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(842333248)))]; + 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(855440512)))]; + tensor input_137_cast_fp16 = conv(bias = layers_13_fc1_bias_to_fp16, dilations = var_3158, groups = var_2969, pad = input_137_pad_0, pad_type = input_137_pad_type_0, strides = var_3156, 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 var_3164 = const()[name = tensor("op_3164"), val = tensor([1, 1])]; + tensor var_3166 = const()[name = tensor("op_3166"), val = tensor([1, 1])]; + tensor hidden_states_29_pad_type_0 = const()[name = tensor("hidden_states_29_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_29_pad_0 = const()[name = tensor("hidden_states_29_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(855450816)))]; + 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(868558080)))]; + tensor hidden_states_29_cast_fp16 = conv(bias = layers_13_fc2_bias_to_fp16, dilations = var_3166, groups = var_2969, pad = hidden_states_29_pad_0, pad_type = hidden_states_29_pad_type_0, strides = var_3164, 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_3180 = const()[name = tensor("op_3180"), val = tensor(3)]; + tensor var_3187 = const()[name = tensor("op_3187"), val = tensor(1)]; + tensor var_3188 = const()[name = tensor("op_3188"), val = tensor(true)]; + tensor var_3200 = const()[name = tensor("op_3200"), val = tensor([1])]; + tensor channels_mean_85_cast_fp16 = reduce_mean(axes = var_3200, keep_dims = var_3188, x = inputs_85_cast_fp16)[name = tensor("channels_mean_85_cast_fp16")]; + tensor zero_mean_85_cast_fp16 = sub(x = inputs_85_cast_fp16, y = channels_mean_85_cast_fp16)[name = tensor("zero_mean_85_cast_fp16")]; + tensor zero_mean_sq_85_cast_fp16 = mul(x = zero_mean_85_cast_fp16, y = zero_mean_85_cast_fp16)[name = tensor("zero_mean_sq_85_cast_fp16")]; + tensor var_3204 = const()[name = tensor("op_3204"), val = tensor([1])]; + tensor var_3205_cast_fp16 = reduce_mean(axes = var_3204, keep_dims = var_3188, x = zero_mean_sq_85_cast_fp16)[name = tensor("op_3205_cast_fp16")]; + tensor var_3206_to_fp16 = const()[name = tensor("op_3206_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3207_cast_fp16 = add(x = var_3205_cast_fp16, y = var_3206_to_fp16)[name = tensor("op_3207_cast_fp16")]; + tensor denom_85_epsilon_0_to_fp16 = const()[name = tensor("denom_85_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_85_cast_fp16 = rsqrt(epsilon = denom_85_epsilon_0_to_fp16, x = var_3207_cast_fp16)[name = tensor("denom_85_cast_fp16")]; + tensor out_85_cast_fp16 = mul(x = zero_mean_85_cast_fp16, y = denom_85_cast_fp16)[name = tensor("out_85_cast_fp16")]; + tensor obj_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(868560704)))]; + 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(868563328)))]; + 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 var_3222 = const()[name = tensor("op_3222"), val = tensor([1, 1])]; + tensor var_3224 = const()[name = tensor("op_3224"), val = tensor([1, 1])]; + tensor query_57_pad_type_0 = const()[name = tensor("query_57_pad_type_0"), val = tensor("custom")]; + tensor query_57_pad_0 = const()[name = tensor("query_57_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(868565952)))]; + 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(871842816)))]; + tensor query_57_cast_fp16 = conv(bias = layers_14_self_attn_q_proj_bias_to_fp16, dilations = var_3224, groups = var_3187, pad = query_57_pad_0, pad_type = query_57_pad_type_0, strides = var_3222, weight = layers_14_self_attn_q_proj_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor("query_57_cast_fp16")]; + tensor var_3228 = const()[name = tensor("op_3228"), val = tensor([1, 1])]; + tensor var_3230 = const()[name = tensor("op_3230"), val = tensor([1, 1])]; + tensor current_key_29_pad_type_0 = const()[name = tensor("current_key_29_pad_type_0"), val = tensor("custom")]; + tensor current_key_29_pad_0 = const()[name = tensor("current_key_29_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(871845440)))]; + tensor current_key_29_cast_fp16 = conv(dilations = var_3230, groups = var_3187, pad = current_key_29_pad_0, pad_type = current_key_29_pad_type_0, strides = var_3228, weight = layers_14_self_attn_k_proj_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor("current_key_29_cast_fp16")]; + tensor var_3235 = const()[name = tensor("op_3235"), val = tensor([1, 1])]; + tensor var_3237 = const()[name = tensor("op_3237"), val = tensor([1, 1])]; + tensor current_value_29_pad_type_0 = const()[name = tensor("current_value_29_pad_type_0"), val = tensor("custom")]; + tensor current_value_29_pad_0 = const()[name = tensor("current_value_29_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(875122304)))]; + 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(878399168)))]; + tensor current_value_29_cast_fp16 = conv(bias = layers_14_self_attn_v_proj_bias_to_fp16, dilations = var_3237, groups = var_3187, pad = current_value_29_pad_0, pad_type = current_value_29_pad_type_0, strides = var_3235, weight = layers_14_self_attn_v_proj_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor("current_value_29_cast_fp16")]; + tensor var_3244_cast_fp16 = mul(x = current_key_29_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3244_cast_fp16")]; + tensor var_3246_cast_fp16 = mul(x = var_103_cast_fp16_14, y = var_241_cast_fp16)[name = tensor("op_3246_cast_fp16")]; + tensor key_57_cast_fp16 = add(x = var_3244_cast_fp16, y = var_3246_cast_fp16)[name = tensor("key_57_cast_fp16")]; + tensor var_3248_cast_fp16 = mul(x = current_value_29_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3248_cast_fp16")]; + tensor var_3250_cast_fp16 = mul(x = var_138_cast_fp16_14, y = var_241_cast_fp16)[name = tensor("op_3250_cast_fp16")]; + tensor value_57_cast_fp16 = add(x = var_3248_cast_fp16, y = var_3250_cast_fp16)[name = tensor("value_57_cast_fp16")]; + tensor var_3253 = const()[name = tensor("op_3253"), val = tensor([1, 20, 64, -1])]; + tensor var_3254_cast_fp16 = reshape(shape = var_3253, x = query_57_cast_fp16)[name = tensor("op_3254_cast_fp16")]; + tensor var_3255_to_fp16 = const()[name = tensor("op_3255_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3256_cast_fp16 = mul(x = var_3254_cast_fp16, y = var_3255_to_fp16)[name = tensor("op_3256_cast_fp16")]; + tensor var_3257 = const()[name = tensor("op_3257"), val = tensor([1, 20, 64, -1])]; + tensor var_3258_cast_fp16 = reshape(shape = var_3257, x = key_57_cast_fp16)[name = tensor("op_3258_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_3256_cast_fp16, y = var_3258_cast_fp16)[name = tensor("mh_w_85_cast_fp16")]; + tensor mh_w_87_cast_fp16 = add(x = mh_w_85_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_87_cast_fp16")]; + tensor var_3266_cast_fp16 = softmax(axis = var_3180, x = mh_w_87_cast_fp16)[name = tensor("op_3266_cast_fp16")]; + tensor var_3267 = const()[name = tensor("op_3267"), val = tensor([1, 20, 64, -1])]; + tensor var_3268_cast_fp16 = reshape(shape = var_3267, x = value_57_cast_fp16)[name = tensor("op_3268_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_3268_cast_fp16, y = var_3266_cast_fp16)[name = tensor("attn_57_cast_fp16")]; + tensor var_3271 = const()[name = tensor("op_3271"), val = tensor([1, 1280, 1, -1])]; + tensor input_141_cast_fp16 = reshape(shape = var_3271, x = attn_57_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor var_3275 = const()[name = tensor("op_3275"), val = tensor([1, 1])]; + tensor var_3277 = const()[name = tensor("op_3277"), val = tensor([1, 1])]; + tensor obj_203_pad_type_0 = const()[name = tensor("obj_203_pad_type_0"), val = tensor("custom")]; + tensor obj_203_pad_0 = const()[name = tensor("obj_203_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(878401792)))]; + 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(881678656)))]; + tensor obj_203_cast_fp16 = conv(bias = layers_14_self_attn_o_proj_bias_to_fp16, dilations = var_3277, groups = var_3187, pad = obj_203_pad_0, pad_type = obj_203_pad_type_0, strides = var_3275, 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 var_3287 = const()[name = tensor("op_3287"), val = tensor([1])]; + tensor channels_mean_87_cast_fp16 = reduce_mean(axes = var_3287, keep_dims = var_3188, x = inputs_87_cast_fp16)[name = tensor("channels_mean_87_cast_fp16")]; + tensor zero_mean_87_cast_fp16 = sub(x = inputs_87_cast_fp16, y = channels_mean_87_cast_fp16)[name = tensor("zero_mean_87_cast_fp16")]; + tensor zero_mean_sq_87_cast_fp16 = mul(x = zero_mean_87_cast_fp16, y = zero_mean_87_cast_fp16)[name = tensor("zero_mean_sq_87_cast_fp16")]; + tensor var_3291 = const()[name = tensor("op_3291"), val = tensor([1])]; + tensor var_3292_cast_fp16 = reduce_mean(axes = var_3291, keep_dims = var_3188, x = zero_mean_sq_87_cast_fp16)[name = tensor("op_3292_cast_fp16")]; + tensor var_3293_to_fp16 = const()[name = tensor("op_3293_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3294_cast_fp16 = add(x = var_3292_cast_fp16, y = var_3293_to_fp16)[name = tensor("op_3294_cast_fp16")]; + tensor denom_87_epsilon_0_to_fp16 = const()[name = tensor("denom_87_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_87_cast_fp16 = rsqrt(epsilon = denom_87_epsilon_0_to_fp16, x = var_3294_cast_fp16)[name = tensor("denom_87_cast_fp16")]; + tensor out_87_cast_fp16 = mul(x = zero_mean_87_cast_fp16, y = denom_87_cast_fp16)[name = tensor("out_87_cast_fp16")]; + tensor obj_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(881681280)))]; + 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(881683904)))]; + 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 var_3309 = const()[name = tensor("op_3309"), val = tensor([1, 1])]; + tensor var_3311 = const()[name = tensor("op_3311"), val = tensor([1, 1])]; + tensor query_59_pad_type_0 = const()[name = tensor("query_59_pad_type_0"), val = tensor("custom")]; + tensor query_59_pad_0 = const()[name = tensor("query_59_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(881686528)))]; + 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(884963392)))]; + tensor query_59_cast_fp16 = conv(bias = layers_14_encoder_attn_q_proj_bias_to_fp16, dilations = var_3311, groups = var_3187, pad = query_59_pad_0, pad_type = query_59_pad_type_0, strides = var_3309, weight = layers_14_encoder_attn_q_proj_weight_to_fp16, x = obj_205_cast_fp16)[name = tensor("query_59_cast_fp16")]; + tensor var_3315 = const()[name = tensor("op_3315"), val = tensor([1, 1])]; + tensor var_3317 = const()[name = tensor("op_3317"), val = tensor([1, 1])]; + tensor key_59_pad_type_0 = const()[name = tensor("key_59_pad_type_0"), val = tensor("custom")]; + tensor key_59_pad_0 = const()[name = tensor("key_59_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(884966016)))]; + tensor key_59_cast_fp16 = conv(dilations = var_3317, groups = var_3187, pad = key_59_pad_0, pad_type = key_59_pad_type_0, strides = var_3315, weight = layers_14_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_59_cast_fp16")]; + tensor var_3322 = const()[name = tensor("op_3322"), val = tensor([1, 1])]; + tensor var_3324 = const()[name = tensor("op_3324"), val = tensor([1, 1])]; + tensor value_59_pad_type_0 = const()[name = tensor("value_59_pad_type_0"), val = tensor("custom")]; + tensor value_59_pad_0 = const()[name = tensor("value_59_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(888242880)))]; + 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(891519744)))]; + tensor value_59_cast_fp16 = conv(bias = layers_14_encoder_attn_v_proj_bias_to_fp16, dilations = var_3324, groups = var_3187, pad = value_59_pad_0, pad_type = value_59_pad_type_0, strides = var_3322, weight = layers_14_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_59_cast_fp16")]; + tensor var_3328 = const()[name = tensor("op_3328"), val = tensor([1, 20, 64, -1])]; + tensor var_3329_cast_fp16 = reshape(shape = var_3328, x = query_59_cast_fp16)[name = tensor("op_3329_cast_fp16")]; + tensor var_3330_to_fp16 = const()[name = tensor("op_3330_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3331_cast_fp16 = mul(x = var_3329_cast_fp16, y = var_3330_to_fp16)[name = tensor("op_3331_cast_fp16")]; + tensor var_3332 = const()[name = tensor("op_3332"), val = tensor([1, 20, 64, -1])]; + tensor var_3333_cast_fp16 = reshape(shape = var_3332, x = key_59_cast_fp16)[name = tensor("op_3333_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_3331_cast_fp16, y = var_3333_cast_fp16)[name = tensor("mh_w_89_cast_fp16")]; + tensor obj_209_cast_fp16 = softmax(axis = var_3180, x = mh_w_89_cast_fp16)[name = tensor("obj_209_cast_fp16")]; + tensor var_3337 = const()[name = tensor("op_3337"), val = tensor([1, 20, 64, -1])]; + tensor var_3338_cast_fp16 = reshape(shape = var_3337, x = value_59_cast_fp16)[name = tensor("op_3338_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_3338_cast_fp16, y = obj_209_cast_fp16)[name = tensor("attn_59_cast_fp16")]; + tensor var_3341 = const()[name = tensor("op_3341"), val = tensor([1, 1280, 1, -1])]; + tensor input_143_cast_fp16 = reshape(shape = var_3341, x = attn_59_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor var_3345 = const()[name = tensor("op_3345"), val = tensor([1, 1])]; + tensor var_3347 = const()[name = tensor("op_3347"), val = tensor([1, 1])]; + tensor obj_207_pad_type_0 = const()[name = tensor("obj_207_pad_type_0"), val = tensor("custom")]; + tensor obj_207_pad_0 = const()[name = tensor("obj_207_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(891522368)))]; + 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(894799232)))]; + tensor obj_207_cast_fp16 = conv(bias = layers_14_encoder_attn_o_proj_bias_to_fp16, dilations = var_3347, groups = var_3187, pad = obj_207_pad_0, pad_type = obj_207_pad_type_0, strides = var_3345, 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 var_3353 = const()[name = tensor("op_3353"), val = tensor([1])]; + tensor channels_mean_89_cast_fp16 = reduce_mean(axes = var_3353, keep_dims = var_3188, x = inputs_89_cast_fp16)[name = tensor("channels_mean_89_cast_fp16")]; + tensor zero_mean_89_cast_fp16 = sub(x = inputs_89_cast_fp16, y = channels_mean_89_cast_fp16)[name = tensor("zero_mean_89_cast_fp16")]; + tensor zero_mean_sq_89_cast_fp16 = mul(x = zero_mean_89_cast_fp16, y = zero_mean_89_cast_fp16)[name = tensor("zero_mean_sq_89_cast_fp16")]; + tensor var_3357 = const()[name = tensor("op_3357"), val = tensor([1])]; + tensor var_3358_cast_fp16 = reduce_mean(axes = var_3357, keep_dims = var_3188, x = zero_mean_sq_89_cast_fp16)[name = tensor("op_3358_cast_fp16")]; + tensor var_3359_to_fp16 = const()[name = tensor("op_3359_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3360_cast_fp16 = add(x = var_3358_cast_fp16, y = var_3359_to_fp16)[name = tensor("op_3360_cast_fp16")]; + tensor denom_89_epsilon_0_to_fp16 = const()[name = tensor("denom_89_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_89_cast_fp16 = rsqrt(epsilon = denom_89_epsilon_0_to_fp16, x = var_3360_cast_fp16)[name = tensor("denom_89_cast_fp16")]; + tensor out_89_cast_fp16 = mul(x = zero_mean_89_cast_fp16, y = denom_89_cast_fp16)[name = tensor("out_89_cast_fp16")]; + tensor input_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(894801856)))]; + 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(894804480)))]; + 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 var_3371 = const()[name = tensor("op_3371"), val = tensor([1, 1])]; + tensor var_3373 = const()[name = tensor("op_3373"), val = tensor([1, 1])]; + tensor input_147_pad_type_0 = const()[name = tensor("input_147_pad_type_0"), val = tensor("custom")]; + tensor input_147_pad_0 = const()[name = tensor("input_147_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(894807104)))]; + 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(907914368)))]; + tensor input_147_cast_fp16 = conv(bias = layers_14_fc1_bias_to_fp16, dilations = var_3373, groups = var_3187, pad = input_147_pad_0, pad_type = input_147_pad_type_0, strides = var_3371, 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 var_3379 = const()[name = tensor("op_3379"), val = tensor([1, 1])]; + tensor var_3381 = const()[name = tensor("op_3381"), val = tensor([1, 1])]; + tensor hidden_states_31_pad_type_0 = const()[name = tensor("hidden_states_31_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_31_pad_0 = const()[name = tensor("hidden_states_31_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(907924672)))]; + 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(921031936)))]; + tensor hidden_states_31_cast_fp16 = conv(bias = layers_14_fc2_bias_to_fp16, dilations = var_3381, groups = var_3187, pad = hidden_states_31_pad_0, pad_type = hidden_states_31_pad_type_0, strides = var_3379, 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_3394 = const()[name = tensor("op_3394"), val = tensor(3)]; + tensor var_3401 = const()[name = tensor("op_3401"), val = tensor(1)]; + tensor var_3402 = const()[name = tensor("op_3402"), val = tensor(true)]; + tensor var_3414 = const()[name = tensor("op_3414"), val = tensor([1])]; + tensor channels_mean_91_cast_fp16 = reduce_mean(axes = var_3414, keep_dims = var_3402, x = inputs_91_cast_fp16)[name = tensor("channels_mean_91_cast_fp16")]; + tensor zero_mean_91_cast_fp16 = sub(x = inputs_91_cast_fp16, y = channels_mean_91_cast_fp16)[name = tensor("zero_mean_91_cast_fp16")]; + tensor zero_mean_sq_91_cast_fp16 = mul(x = zero_mean_91_cast_fp16, y = zero_mean_91_cast_fp16)[name = tensor("zero_mean_sq_91_cast_fp16")]; + tensor var_3418 = const()[name = tensor("op_3418"), val = tensor([1])]; + tensor var_3419_cast_fp16 = reduce_mean(axes = var_3418, keep_dims = var_3402, x = zero_mean_sq_91_cast_fp16)[name = tensor("op_3419_cast_fp16")]; + tensor var_3420_to_fp16 = const()[name = tensor("op_3420_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3421_cast_fp16 = add(x = var_3419_cast_fp16, y = var_3420_to_fp16)[name = tensor("op_3421_cast_fp16")]; + tensor denom_91_epsilon_0_to_fp16 = const()[name = tensor("denom_91_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_91_cast_fp16 = rsqrt(epsilon = denom_91_epsilon_0_to_fp16, x = var_3421_cast_fp16)[name = tensor("denom_91_cast_fp16")]; + tensor out_91_cast_fp16 = mul(x = zero_mean_91_cast_fp16, y = denom_91_cast_fp16)[name = tensor("out_91_cast_fp16")]; + tensor obj_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(921034560)))]; + 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(921037184)))]; + 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 var_3436 = const()[name = tensor("op_3436"), val = tensor([1, 1])]; + tensor var_3438 = const()[name = tensor("op_3438"), val = tensor([1, 1])]; + tensor query_61_pad_type_0 = const()[name = tensor("query_61_pad_type_0"), val = tensor("custom")]; + tensor query_61_pad_0 = const()[name = tensor("query_61_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(921039808)))]; + 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(924316672)))]; + tensor query_61_cast_fp16 = conv(bias = layers_15_self_attn_q_proj_bias_to_fp16, dilations = var_3438, groups = var_3401, pad = query_61_pad_0, pad_type = query_61_pad_type_0, strides = var_3436, weight = layers_15_self_attn_q_proj_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor("query_61_cast_fp16")]; + tensor var_3442 = const()[name = tensor("op_3442"), val = tensor([1, 1])]; + tensor var_3444 = const()[name = tensor("op_3444"), val = tensor([1, 1])]; + tensor current_key_31_pad_type_0 = const()[name = tensor("current_key_31_pad_type_0"), val = tensor("custom")]; + tensor current_key_31_pad_0 = const()[name = tensor("current_key_31_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(924319296)))]; + tensor current_key_31_cast_fp16 = conv(dilations = var_3444, groups = var_3401, pad = current_key_31_pad_0, pad_type = current_key_31_pad_type_0, strides = var_3442, weight = layers_15_self_attn_k_proj_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor("current_key_31_cast_fp16")]; + tensor var_3449 = const()[name = tensor("op_3449"), val = tensor([1, 1])]; + tensor var_3451 = const()[name = tensor("op_3451"), val = tensor([1, 1])]; + tensor current_value_31_pad_type_0 = const()[name = tensor("current_value_31_pad_type_0"), val = tensor("custom")]; + tensor current_value_31_pad_0 = const()[name = tensor("current_value_31_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(927596160)))]; + 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(930873024)))]; + tensor current_value_31_cast_fp16 = conv(bias = layers_15_self_attn_v_proj_bias_to_fp16, dilations = var_3451, groups = var_3401, pad = current_value_31_pad_0, pad_type = current_value_31_pad_type_0, strides = var_3449, weight = layers_15_self_attn_v_proj_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor("current_value_31_cast_fp16")]; + tensor var_3458_cast_fp16 = mul(x = current_key_31_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3458_cast_fp16")]; + tensor var_3460_cast_fp16 = mul(x = var_103_cast_fp16_15, y = var_241_cast_fp16)[name = tensor("op_3460_cast_fp16")]; + tensor key_61_cast_fp16 = add(x = var_3458_cast_fp16, y = var_3460_cast_fp16)[name = tensor("key_61_cast_fp16")]; + tensor var_3462_cast_fp16 = mul(x = current_value_31_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3462_cast_fp16")]; + tensor var_3464_cast_fp16 = mul(x = var_138_cast_fp16_15, y = var_241_cast_fp16)[name = tensor("op_3464_cast_fp16")]; + tensor value_61_cast_fp16 = add(x = var_3462_cast_fp16, y = var_3464_cast_fp16)[name = tensor("value_61_cast_fp16")]; + tensor var_3467 = const()[name = tensor("op_3467"), val = tensor([1, 20, 64, -1])]; + tensor var_3468_cast_fp16 = reshape(shape = var_3467, x = query_61_cast_fp16)[name = tensor("op_3468_cast_fp16")]; + tensor var_3469_to_fp16 = const()[name = tensor("op_3469_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3470_cast_fp16 = mul(x = var_3468_cast_fp16, y = var_3469_to_fp16)[name = tensor("op_3470_cast_fp16")]; + tensor var_3471 = const()[name = tensor("op_3471"), val = tensor([1, 20, 64, -1])]; + tensor var_3472_cast_fp16 = reshape(shape = var_3471, x = key_61_cast_fp16)[name = tensor("op_3472_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_3470_cast_fp16, y = var_3472_cast_fp16)[name = tensor("mh_w_91_cast_fp16")]; + tensor mh_w_93_cast_fp16 = add(x = mh_w_91_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_93_cast_fp16")]; + tensor var_3480_cast_fp16 = softmax(axis = var_3394, x = mh_w_93_cast_fp16)[name = tensor("op_3480_cast_fp16")]; + tensor var_3481 = const()[name = tensor("op_3481"), val = tensor([1, 20, 64, -1])]; + tensor var_3482_cast_fp16 = reshape(shape = var_3481, x = value_61_cast_fp16)[name = tensor("op_3482_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_3482_cast_fp16, y = var_3480_cast_fp16)[name = tensor("attn_61_cast_fp16")]; + tensor var_3485 = const()[name = tensor("op_3485"), val = tensor([1, 1280, 1, -1])]; + tensor input_151_cast_fp16 = reshape(shape = var_3485, x = attn_61_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor var_3489 = const()[name = tensor("op_3489"), val = tensor([1, 1])]; + tensor var_3491 = const()[name = tensor("op_3491"), val = tensor([1, 1])]; + tensor obj_217_pad_type_0 = const()[name = tensor("obj_217_pad_type_0"), val = tensor("custom")]; + tensor obj_217_pad_0 = const()[name = tensor("obj_217_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(930875648)))]; + 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(934152512)))]; + tensor obj_217_cast_fp16 = conv(bias = layers_15_self_attn_o_proj_bias_to_fp16, dilations = var_3491, groups = var_3401, pad = obj_217_pad_0, pad_type = obj_217_pad_type_0, strides = var_3489, 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 var_3501 = const()[name = tensor("op_3501"), val = tensor([1])]; + tensor channels_mean_93_cast_fp16 = reduce_mean(axes = var_3501, keep_dims = var_3402, x = inputs_93_cast_fp16)[name = tensor("channels_mean_93_cast_fp16")]; + tensor zero_mean_93_cast_fp16 = sub(x = inputs_93_cast_fp16, y = channels_mean_93_cast_fp16)[name = tensor("zero_mean_93_cast_fp16")]; + tensor zero_mean_sq_93_cast_fp16 = mul(x = zero_mean_93_cast_fp16, y = zero_mean_93_cast_fp16)[name = tensor("zero_mean_sq_93_cast_fp16")]; + tensor var_3505 = const()[name = tensor("op_3505"), val = tensor([1])]; + tensor var_3506_cast_fp16 = reduce_mean(axes = var_3505, keep_dims = var_3402, x = zero_mean_sq_93_cast_fp16)[name = tensor("op_3506_cast_fp16")]; + tensor var_3507_to_fp16 = const()[name = tensor("op_3507_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3508_cast_fp16 = add(x = var_3506_cast_fp16, y = var_3507_to_fp16)[name = tensor("op_3508_cast_fp16")]; + tensor denom_93_epsilon_0_to_fp16 = const()[name = tensor("denom_93_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_93_cast_fp16 = rsqrt(epsilon = denom_93_epsilon_0_to_fp16, x = var_3508_cast_fp16)[name = tensor("denom_93_cast_fp16")]; + tensor out_93_cast_fp16 = mul(x = zero_mean_93_cast_fp16, y = denom_93_cast_fp16)[name = tensor("out_93_cast_fp16")]; + tensor obj_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(934155136)))]; + 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(934157760)))]; + 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 var_3523 = const()[name = tensor("op_3523"), val = tensor([1, 1])]; + tensor var_3525 = const()[name = tensor("op_3525"), val = tensor([1, 1])]; + tensor query_63_pad_type_0 = const()[name = tensor("query_63_pad_type_0"), val = tensor("custom")]; + tensor query_63_pad_0 = const()[name = tensor("query_63_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(934160384)))]; + 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(937437248)))]; + tensor query_63_cast_fp16 = conv(bias = layers_15_encoder_attn_q_proj_bias_to_fp16, dilations = var_3525, groups = var_3401, pad = query_63_pad_0, pad_type = query_63_pad_type_0, strides = var_3523, weight = layers_15_encoder_attn_q_proj_weight_to_fp16, x = obj_219_cast_fp16)[name = tensor("query_63_cast_fp16")]; + tensor var_3529 = const()[name = tensor("op_3529"), val = tensor([1, 1])]; + tensor var_3531 = const()[name = tensor("op_3531"), val = tensor([1, 1])]; + tensor key_63_pad_type_0 = const()[name = tensor("key_63_pad_type_0"), val = tensor("custom")]; + tensor key_63_pad_0 = const()[name = tensor("key_63_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(937439872)))]; + tensor key_63_cast_fp16 = conv(dilations = var_3531, groups = var_3401, pad = key_63_pad_0, pad_type = key_63_pad_type_0, strides = var_3529, weight = layers_15_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_63_cast_fp16")]; + tensor var_3536 = const()[name = tensor("op_3536"), val = tensor([1, 1])]; + tensor var_3538 = const()[name = tensor("op_3538"), val = tensor([1, 1])]; + tensor value_63_pad_type_0 = const()[name = tensor("value_63_pad_type_0"), val = tensor("custom")]; + tensor value_63_pad_0 = const()[name = tensor("value_63_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(940716736)))]; + 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(943993600)))]; + tensor value_63_cast_fp16 = conv(bias = layers_15_encoder_attn_v_proj_bias_to_fp16, dilations = var_3538, groups = var_3401, pad = value_63_pad_0, pad_type = value_63_pad_type_0, strides = var_3536, weight = layers_15_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_63_cast_fp16")]; + tensor var_3542 = const()[name = tensor("op_3542"), val = tensor([1, 20, 64, -1])]; + tensor var_3543_cast_fp16 = reshape(shape = var_3542, x = query_63_cast_fp16)[name = tensor("op_3543_cast_fp16")]; + tensor var_3544_to_fp16 = const()[name = tensor("op_3544_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3545_cast_fp16 = mul(x = var_3543_cast_fp16, y = var_3544_to_fp16)[name = tensor("op_3545_cast_fp16")]; + tensor var_3546 = const()[name = tensor("op_3546"), val = tensor([1, 20, 64, -1])]; + tensor var_3547_cast_fp16 = reshape(shape = var_3546, x = key_63_cast_fp16)[name = tensor("op_3547_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_3545_cast_fp16, y = var_3547_cast_fp16)[name = tensor("mh_w_95_cast_fp16")]; + tensor obj_223_cast_fp16 = softmax(axis = var_3394, x = mh_w_95_cast_fp16)[name = tensor("obj_223_cast_fp16")]; + tensor var_3551 = const()[name = tensor("op_3551"), val = tensor([1, 20, 64, -1])]; + tensor var_3552_cast_fp16 = reshape(shape = var_3551, x = value_63_cast_fp16)[name = tensor("op_3552_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_3552_cast_fp16, y = obj_223_cast_fp16)[name = tensor("attn_63_cast_fp16")]; + tensor var_3555 = const()[name = tensor("op_3555"), val = tensor([1, 1280, 1, -1])]; + tensor input_153_cast_fp16 = reshape(shape = var_3555, x = attn_63_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor var_3559 = const()[name = tensor("op_3559"), val = tensor([1, 1])]; + tensor var_3561 = const()[name = tensor("op_3561"), val = tensor([1, 1])]; + tensor obj_221_pad_type_0 = const()[name = tensor("obj_221_pad_type_0"), val = tensor("custom")]; + tensor obj_221_pad_0 = const()[name = tensor("obj_221_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(943996224)))]; + 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(947273088)))]; + tensor obj_221_cast_fp16 = conv(bias = layers_15_encoder_attn_o_proj_bias_to_fp16, dilations = var_3561, groups = var_3401, pad = obj_221_pad_0, pad_type = obj_221_pad_type_0, strides = var_3559, 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 var_3567 = const()[name = tensor("op_3567"), val = tensor([1])]; + tensor channels_mean_95_cast_fp16 = reduce_mean(axes = var_3567, keep_dims = var_3402, x = inputs_95_cast_fp16)[name = tensor("channels_mean_95_cast_fp16")]; + tensor zero_mean_95_cast_fp16 = sub(x = inputs_95_cast_fp16, y = channels_mean_95_cast_fp16)[name = tensor("zero_mean_95_cast_fp16")]; + tensor zero_mean_sq_95_cast_fp16 = mul(x = zero_mean_95_cast_fp16, y = zero_mean_95_cast_fp16)[name = tensor("zero_mean_sq_95_cast_fp16")]; + tensor var_3571 = const()[name = tensor("op_3571"), val = tensor([1])]; + tensor var_3572_cast_fp16 = reduce_mean(axes = var_3571, keep_dims = var_3402, x = zero_mean_sq_95_cast_fp16)[name = tensor("op_3572_cast_fp16")]; + tensor var_3573_to_fp16 = const()[name = tensor("op_3573_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3574_cast_fp16 = add(x = var_3572_cast_fp16, y = var_3573_to_fp16)[name = tensor("op_3574_cast_fp16")]; + tensor denom_95_epsilon_0_to_fp16 = const()[name = tensor("denom_95_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_95_cast_fp16 = rsqrt(epsilon = denom_95_epsilon_0_to_fp16, x = var_3574_cast_fp16)[name = tensor("denom_95_cast_fp16")]; + tensor out_95_cast_fp16 = mul(x = zero_mean_95_cast_fp16, y = denom_95_cast_fp16)[name = tensor("out_95_cast_fp16")]; + tensor input_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(947275712)))]; + 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(947278336)))]; + 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 var_3585 = const()[name = tensor("op_3585"), val = tensor([1, 1])]; + tensor var_3587 = const()[name = tensor("op_3587"), val = tensor([1, 1])]; + tensor input_157_pad_type_0 = const()[name = tensor("input_157_pad_type_0"), val = tensor("custom")]; + tensor input_157_pad_0 = const()[name = tensor("input_157_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(947280960)))]; + 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(960388224)))]; + tensor input_157_cast_fp16 = conv(bias = layers_15_fc1_bias_to_fp16, dilations = var_3587, groups = var_3401, pad = input_157_pad_0, pad_type = input_157_pad_type_0, strides = var_3585, 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 var_3593 = const()[name = tensor("op_3593"), val = tensor([1, 1])]; + tensor var_3595 = const()[name = tensor("op_3595"), val = tensor([1, 1])]; + tensor hidden_states_33_pad_type_0 = const()[name = tensor("hidden_states_33_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_33_pad_0 = const()[name = tensor("hidden_states_33_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(960398528)))]; + 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(973505792)))]; + tensor hidden_states_33_cast_fp16 = conv(bias = layers_15_fc2_bias_to_fp16, dilations = var_3595, groups = var_3401, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = var_3593, 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_3608 = const()[name = tensor("op_3608"), val = tensor(3)]; + tensor var_3615 = const()[name = tensor("op_3615"), val = tensor(1)]; + tensor var_3616 = const()[name = tensor("op_3616"), val = tensor(true)]; + tensor var_3628 = const()[name = tensor("op_3628"), val = tensor([1])]; + tensor channels_mean_97_cast_fp16 = reduce_mean(axes = var_3628, keep_dims = var_3616, x = inputs_97_cast_fp16)[name = tensor("channels_mean_97_cast_fp16")]; + tensor zero_mean_97_cast_fp16 = sub(x = inputs_97_cast_fp16, y = channels_mean_97_cast_fp16)[name = tensor("zero_mean_97_cast_fp16")]; + tensor zero_mean_sq_97_cast_fp16 = mul(x = zero_mean_97_cast_fp16, y = zero_mean_97_cast_fp16)[name = tensor("zero_mean_sq_97_cast_fp16")]; + tensor var_3632 = const()[name = tensor("op_3632"), val = tensor([1])]; + tensor var_3633_cast_fp16 = reduce_mean(axes = var_3632, keep_dims = var_3616, x = zero_mean_sq_97_cast_fp16)[name = tensor("op_3633_cast_fp16")]; + tensor var_3634_to_fp16 = const()[name = tensor("op_3634_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3635_cast_fp16 = add(x = var_3633_cast_fp16, y = var_3634_to_fp16)[name = tensor("op_3635_cast_fp16")]; + tensor denom_97_epsilon_0_to_fp16 = const()[name = tensor("denom_97_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_97_cast_fp16 = rsqrt(epsilon = denom_97_epsilon_0_to_fp16, x = var_3635_cast_fp16)[name = tensor("denom_97_cast_fp16")]; + tensor out_97_cast_fp16 = mul(x = zero_mean_97_cast_fp16, y = denom_97_cast_fp16)[name = tensor("out_97_cast_fp16")]; + tensor obj_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(973508416)))]; + 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(973511040)))]; + 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 var_3650 = const()[name = tensor("op_3650"), val = tensor([1, 1])]; + tensor var_3652 = const()[name = tensor("op_3652"), val = tensor([1, 1])]; + tensor query_65_pad_type_0 = const()[name = tensor("query_65_pad_type_0"), val = tensor("custom")]; + tensor query_65_pad_0 = const()[name = tensor("query_65_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(973513664)))]; + 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(976790528)))]; + tensor query_65_cast_fp16 = conv(bias = layers_16_self_attn_q_proj_bias_to_fp16, dilations = var_3652, groups = var_3615, pad = query_65_pad_0, pad_type = query_65_pad_type_0, strides = var_3650, weight = layers_16_self_attn_q_proj_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("query_65_cast_fp16")]; + tensor var_3656 = const()[name = tensor("op_3656"), val = tensor([1, 1])]; + tensor var_3658 = const()[name = tensor("op_3658"), val = tensor([1, 1])]; + tensor current_key_33_pad_type_0 = const()[name = tensor("current_key_33_pad_type_0"), val = tensor("custom")]; + tensor current_key_33_pad_0 = const()[name = tensor("current_key_33_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(976793152)))]; + tensor current_key_33_cast_fp16 = conv(dilations = var_3658, groups = var_3615, pad = current_key_33_pad_0, pad_type = current_key_33_pad_type_0, strides = var_3656, weight = layers_16_self_attn_k_proj_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("current_key_33_cast_fp16")]; + tensor var_3663 = const()[name = tensor("op_3663"), val = tensor([1, 1])]; + tensor var_3665 = const()[name = tensor("op_3665"), val = tensor([1, 1])]; + tensor current_value_33_pad_type_0 = const()[name = tensor("current_value_33_pad_type_0"), val = tensor("custom")]; + tensor current_value_33_pad_0 = const()[name = tensor("current_value_33_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(980070016)))]; + 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(983346880)))]; + tensor current_value_33_cast_fp16 = conv(bias = layers_16_self_attn_v_proj_bias_to_fp16, dilations = var_3665, groups = var_3615, pad = current_value_33_pad_0, pad_type = current_value_33_pad_type_0, strides = var_3663, weight = layers_16_self_attn_v_proj_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("current_value_33_cast_fp16")]; + tensor var_3672_cast_fp16 = mul(x = current_key_33_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3672_cast_fp16")]; + tensor var_3674_cast_fp16 = mul(x = var_103_cast_fp16_16, y = var_241_cast_fp16)[name = tensor("op_3674_cast_fp16")]; + tensor key_65_cast_fp16 = add(x = var_3672_cast_fp16, y = var_3674_cast_fp16)[name = tensor("key_65_cast_fp16")]; + tensor var_3676_cast_fp16 = mul(x = current_value_33_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3676_cast_fp16")]; + tensor var_3678_cast_fp16 = mul(x = var_138_cast_fp16_16, y = var_241_cast_fp16)[name = tensor("op_3678_cast_fp16")]; + tensor value_65_cast_fp16 = add(x = var_3676_cast_fp16, y = var_3678_cast_fp16)[name = tensor("value_65_cast_fp16")]; + tensor var_3681 = const()[name = tensor("op_3681"), val = tensor([1, 20, 64, -1])]; + tensor var_3682_cast_fp16 = reshape(shape = var_3681, x = query_65_cast_fp16)[name = tensor("op_3682_cast_fp16")]; + tensor var_3683_to_fp16 = const()[name = tensor("op_3683_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3684_cast_fp16 = mul(x = var_3682_cast_fp16, y = var_3683_to_fp16)[name = tensor("op_3684_cast_fp16")]; + tensor var_3685 = const()[name = tensor("op_3685"), val = tensor([1, 20, 64, -1])]; + tensor var_3686_cast_fp16 = reshape(shape = var_3685, x = key_65_cast_fp16)[name = tensor("op_3686_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_3684_cast_fp16, y = var_3686_cast_fp16)[name = tensor("mh_w_97_cast_fp16")]; + tensor mh_w_99_cast_fp16 = add(x = mh_w_97_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_99_cast_fp16")]; + tensor var_3694_cast_fp16 = softmax(axis = var_3608, x = mh_w_99_cast_fp16)[name = tensor("op_3694_cast_fp16")]; + tensor var_3695 = const()[name = tensor("op_3695"), val = tensor([1, 20, 64, -1])]; + tensor var_3696_cast_fp16 = reshape(shape = var_3695, x = value_65_cast_fp16)[name = tensor("op_3696_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_3696_cast_fp16, y = var_3694_cast_fp16)[name = tensor("attn_65_cast_fp16")]; + tensor var_3699 = const()[name = tensor("op_3699"), val = tensor([1, 1280, 1, -1])]; + tensor input_161_cast_fp16 = reshape(shape = var_3699, x = attn_65_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor var_3703 = const()[name = tensor("op_3703"), val = tensor([1, 1])]; + tensor var_3705 = const()[name = tensor("op_3705"), val = tensor([1, 1])]; + tensor obj_231_pad_type_0 = const()[name = tensor("obj_231_pad_type_0"), val = tensor("custom")]; + tensor obj_231_pad_0 = const()[name = tensor("obj_231_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(983349504)))]; + 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(986626368)))]; + tensor obj_231_cast_fp16 = conv(bias = layers_16_self_attn_o_proj_bias_to_fp16, dilations = var_3705, groups = var_3615, pad = obj_231_pad_0, pad_type = obj_231_pad_type_0, strides = var_3703, 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 var_3715 = const()[name = tensor("op_3715"), val = tensor([1])]; + tensor channels_mean_99_cast_fp16 = reduce_mean(axes = var_3715, keep_dims = var_3616, x = inputs_99_cast_fp16)[name = tensor("channels_mean_99_cast_fp16")]; + tensor zero_mean_99_cast_fp16 = sub(x = inputs_99_cast_fp16, y = channels_mean_99_cast_fp16)[name = tensor("zero_mean_99_cast_fp16")]; + tensor zero_mean_sq_99_cast_fp16 = mul(x = zero_mean_99_cast_fp16, y = zero_mean_99_cast_fp16)[name = tensor("zero_mean_sq_99_cast_fp16")]; + tensor var_3719 = const()[name = tensor("op_3719"), val = tensor([1])]; + tensor var_3720_cast_fp16 = reduce_mean(axes = var_3719, keep_dims = var_3616, x = zero_mean_sq_99_cast_fp16)[name = tensor("op_3720_cast_fp16")]; + tensor var_3721_to_fp16 = const()[name = tensor("op_3721_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3722_cast_fp16 = add(x = var_3720_cast_fp16, y = var_3721_to_fp16)[name = tensor("op_3722_cast_fp16")]; + tensor denom_99_epsilon_0_to_fp16 = const()[name = tensor("denom_99_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_99_cast_fp16 = rsqrt(epsilon = denom_99_epsilon_0_to_fp16, x = var_3722_cast_fp16)[name = tensor("denom_99_cast_fp16")]; + tensor out_99_cast_fp16 = mul(x = zero_mean_99_cast_fp16, y = denom_99_cast_fp16)[name = tensor("out_99_cast_fp16")]; + tensor obj_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(986628992)))]; + 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(986631616)))]; + 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 var_3737 = const()[name = tensor("op_3737"), val = tensor([1, 1])]; + tensor var_3739 = const()[name = tensor("op_3739"), val = tensor([1, 1])]; + tensor query_67_pad_type_0 = const()[name = tensor("query_67_pad_type_0"), val = tensor("custom")]; + tensor query_67_pad_0 = const()[name = tensor("query_67_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(986634240)))]; + 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(989911104)))]; + tensor query_67_cast_fp16 = conv(bias = layers_16_encoder_attn_q_proj_bias_to_fp16, dilations = var_3739, groups = var_3615, pad = query_67_pad_0, pad_type = query_67_pad_type_0, strides = var_3737, weight = layers_16_encoder_attn_q_proj_weight_to_fp16, x = obj_233_cast_fp16)[name = tensor("query_67_cast_fp16")]; + tensor var_3743 = const()[name = tensor("op_3743"), val = tensor([1, 1])]; + tensor var_3745 = const()[name = tensor("op_3745"), val = tensor([1, 1])]; + tensor key_67_pad_type_0 = const()[name = tensor("key_67_pad_type_0"), val = tensor("custom")]; + tensor key_67_pad_0 = const()[name = tensor("key_67_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(989913728)))]; + tensor key_67_cast_fp16 = conv(dilations = var_3745, groups = var_3615, pad = key_67_pad_0, pad_type = key_67_pad_type_0, strides = var_3743, weight = layers_16_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_67_cast_fp16")]; + tensor var_3750 = const()[name = tensor("op_3750"), val = tensor([1, 1])]; + tensor var_3752 = const()[name = tensor("op_3752"), val = tensor([1, 1])]; + tensor value_67_pad_type_0 = const()[name = tensor("value_67_pad_type_0"), val = tensor("custom")]; + tensor value_67_pad_0 = const()[name = tensor("value_67_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(993190592)))]; + 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(996467456)))]; + tensor value_67_cast_fp16 = conv(bias = layers_16_encoder_attn_v_proj_bias_to_fp16, dilations = var_3752, groups = var_3615, pad = value_67_pad_0, pad_type = value_67_pad_type_0, strides = var_3750, weight = layers_16_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_67_cast_fp16")]; + tensor var_3756 = const()[name = tensor("op_3756"), val = tensor([1, 20, 64, -1])]; + tensor var_3757_cast_fp16 = reshape(shape = var_3756, x = query_67_cast_fp16)[name = tensor("op_3757_cast_fp16")]; + tensor var_3758_to_fp16 = const()[name = tensor("op_3758_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3759_cast_fp16 = mul(x = var_3757_cast_fp16, y = var_3758_to_fp16)[name = tensor("op_3759_cast_fp16")]; + tensor var_3760 = const()[name = tensor("op_3760"), val = tensor([1, 20, 64, -1])]; + tensor var_3761_cast_fp16 = reshape(shape = var_3760, x = key_67_cast_fp16)[name = tensor("op_3761_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_3759_cast_fp16, y = var_3761_cast_fp16)[name = tensor("mh_w_101_cast_fp16")]; + tensor obj_237_cast_fp16 = softmax(axis = var_3608, x = mh_w_101_cast_fp16)[name = tensor("obj_237_cast_fp16")]; + tensor var_3765 = const()[name = tensor("op_3765"), val = tensor([1, 20, 64, -1])]; + tensor var_3766_cast_fp16 = reshape(shape = var_3765, x = value_67_cast_fp16)[name = tensor("op_3766_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_3766_cast_fp16, y = obj_237_cast_fp16)[name = tensor("attn_67_cast_fp16")]; + tensor var_3769 = const()[name = tensor("op_3769"), val = tensor([1, 1280, 1, -1])]; + tensor input_163_cast_fp16 = reshape(shape = var_3769, x = attn_67_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor var_3773 = const()[name = tensor("op_3773"), val = tensor([1, 1])]; + tensor var_3775 = const()[name = tensor("op_3775"), val = tensor([1, 1])]; + tensor obj_235_pad_type_0 = const()[name = tensor("obj_235_pad_type_0"), val = tensor("custom")]; + tensor obj_235_pad_0 = const()[name = tensor("obj_235_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(996470080)))]; + 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(999746944)))]; + tensor obj_235_cast_fp16 = conv(bias = layers_16_encoder_attn_o_proj_bias_to_fp16, dilations = var_3775, groups = var_3615, pad = obj_235_pad_0, pad_type = obj_235_pad_type_0, strides = var_3773, 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 var_3784 = const()[name = tensor("op_3784"), val = tensor([1])]; + tensor channels_mean_101_cast_fp16 = reduce_mean(axes = var_3784, keep_dims = var_3616, x = inputs_101_cast_fp16)[name = tensor("channels_mean_101_cast_fp16")]; + tensor zero_mean_101_cast_fp16 = sub(x = inputs_101_cast_fp16, y = channels_mean_101_cast_fp16)[name = tensor("zero_mean_101_cast_fp16")]; + tensor zero_mean_sq_101_cast_fp16 = mul(x = zero_mean_101_cast_fp16, y = zero_mean_101_cast_fp16)[name = tensor("zero_mean_sq_101_cast_fp16")]; + tensor var_3788 = const()[name = tensor("op_3788"), val = tensor([1])]; + tensor var_3789_cast_fp16 = reduce_mean(axes = var_3788, keep_dims = var_3616, x = zero_mean_sq_101_cast_fp16)[name = tensor("op_3789_cast_fp16")]; + tensor var_3790_to_fp16 = const()[name = tensor("op_3790_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3791_cast_fp16 = add(x = var_3789_cast_fp16, y = var_3790_to_fp16)[name = tensor("op_3791_cast_fp16")]; + tensor denom_101_epsilon_0_to_fp16 = const()[name = tensor("denom_101_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_101_cast_fp16 = rsqrt(epsilon = denom_101_epsilon_0_to_fp16, x = var_3791_cast_fp16)[name = tensor("denom_101_cast_fp16")]; + tensor out_101_cast_fp16 = mul(x = zero_mean_101_cast_fp16, y = denom_101_cast_fp16)[name = tensor("out_101_cast_fp16")]; + tensor input_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(999749568)))]; + 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(999752192)))]; + 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 var_3802 = const()[name = tensor("op_3802"), val = tensor([1, 1])]; + tensor var_3804 = const()[name = tensor("op_3804"), val = tensor([1, 1])]; + tensor input_167_pad_type_0 = const()[name = tensor("input_167_pad_type_0"), val = tensor("custom")]; + tensor input_167_pad_0 = const()[name = tensor("input_167_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(999754816)))]; + 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(1012862080)))]; + tensor input_167_cast_fp16 = conv(bias = layers_16_fc1_bias_to_fp16, dilations = var_3804, groups = var_3615, pad = input_167_pad_0, pad_type = input_167_pad_type_0, strides = var_3802, 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 var_3810 = const()[name = tensor("op_3810"), val = tensor([1, 1])]; + tensor var_3812 = const()[name = tensor("op_3812"), val = tensor([1, 1])]; + tensor hidden_states_35_pad_type_0 = const()[name = tensor("hidden_states_35_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_35_pad_0 = const()[name = tensor("hidden_states_35_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1012872384)))]; + 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(1025979648)))]; + tensor hidden_states_35_cast_fp16 = conv(bias = layers_16_fc2_bias_to_fp16, dilations = var_3812, groups = var_3615, pad = hidden_states_35_pad_0, pad_type = hidden_states_35_pad_type_0, strides = var_3810, 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_3826 = const()[name = tensor("op_3826"), val = tensor(3)]; + tensor var_3833 = const()[name = tensor("op_3833"), val = tensor(1)]; + tensor var_3834 = const()[name = tensor("op_3834"), val = tensor(true)]; + tensor var_3846 = const()[name = tensor("op_3846"), val = tensor([1])]; + tensor channels_mean_103_cast_fp16 = reduce_mean(axes = var_3846, keep_dims = var_3834, x = inputs_103_cast_fp16)[name = tensor("channels_mean_103_cast_fp16")]; + tensor zero_mean_103_cast_fp16 = sub(x = inputs_103_cast_fp16, y = channels_mean_103_cast_fp16)[name = tensor("zero_mean_103_cast_fp16")]; + tensor zero_mean_sq_103_cast_fp16 = mul(x = zero_mean_103_cast_fp16, y = zero_mean_103_cast_fp16)[name = tensor("zero_mean_sq_103_cast_fp16")]; + tensor var_3850 = const()[name = tensor("op_3850"), val = tensor([1])]; + tensor var_3851_cast_fp16 = reduce_mean(axes = var_3850, keep_dims = var_3834, x = zero_mean_sq_103_cast_fp16)[name = tensor("op_3851_cast_fp16")]; + tensor var_3852_to_fp16 = const()[name = tensor("op_3852_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3853_cast_fp16 = add(x = var_3851_cast_fp16, y = var_3852_to_fp16)[name = tensor("op_3853_cast_fp16")]; + tensor denom_103_epsilon_0_to_fp16 = const()[name = tensor("denom_103_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_103_cast_fp16 = rsqrt(epsilon = denom_103_epsilon_0_to_fp16, x = var_3853_cast_fp16)[name = tensor("denom_103_cast_fp16")]; + tensor out_103_cast_fp16 = mul(x = zero_mean_103_cast_fp16, y = denom_103_cast_fp16)[name = tensor("out_103_cast_fp16")]; + tensor obj_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(1025982272)))]; + 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(1025984896)))]; + 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 var_3868 = const()[name = tensor("op_3868"), val = tensor([1, 1])]; + tensor var_3870 = const()[name = tensor("op_3870"), val = tensor([1, 1])]; + tensor query_69_pad_type_0 = const()[name = tensor("query_69_pad_type_0"), val = tensor("custom")]; + tensor query_69_pad_0 = const()[name = tensor("query_69_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1025987520)))]; + 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(1029264384)))]; + tensor query_69_cast_fp16 = conv(bias = layers_17_self_attn_q_proj_bias_to_fp16, dilations = var_3870, groups = var_3833, pad = query_69_pad_0, pad_type = query_69_pad_type_0, strides = var_3868, weight = layers_17_self_attn_q_proj_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor("query_69_cast_fp16")]; + tensor var_3874 = const()[name = tensor("op_3874"), val = tensor([1, 1])]; + tensor var_3876 = const()[name = tensor("op_3876"), val = tensor([1, 1])]; + tensor current_key_35_pad_type_0 = const()[name = tensor("current_key_35_pad_type_0"), val = tensor("custom")]; + tensor current_key_35_pad_0 = const()[name = tensor("current_key_35_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1029267008)))]; + tensor current_key_35_cast_fp16 = conv(dilations = var_3876, groups = var_3833, pad = current_key_35_pad_0, pad_type = current_key_35_pad_type_0, strides = var_3874, weight = layers_17_self_attn_k_proj_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor("current_key_35_cast_fp16")]; + tensor var_3881 = const()[name = tensor("op_3881"), val = tensor([1, 1])]; + tensor var_3883 = const()[name = tensor("op_3883"), val = tensor([1, 1])]; + tensor current_value_35_pad_type_0 = const()[name = tensor("current_value_35_pad_type_0"), val = tensor("custom")]; + tensor current_value_35_pad_0 = const()[name = tensor("current_value_35_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1032543872)))]; + 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(1035820736)))]; + tensor current_value_35_cast_fp16 = conv(bias = layers_17_self_attn_v_proj_bias_to_fp16, dilations = var_3883, groups = var_3833, pad = current_value_35_pad_0, pad_type = current_value_35_pad_type_0, strides = var_3881, weight = layers_17_self_attn_v_proj_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor("current_value_35_cast_fp16")]; + tensor var_3890_cast_fp16 = mul(x = current_key_35_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3890_cast_fp16")]; + tensor var_3892_cast_fp16 = mul(x = var_103_cast_fp16_17, y = var_241_cast_fp16)[name = tensor("op_3892_cast_fp16")]; + tensor key_69_cast_fp16 = add(x = var_3890_cast_fp16, y = var_3892_cast_fp16)[name = tensor("key_69_cast_fp16")]; + tensor var_3894_cast_fp16 = mul(x = current_value_35_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_3894_cast_fp16")]; + tensor var_3896_cast_fp16 = mul(x = var_138_cast_fp16_17, y = var_241_cast_fp16)[name = tensor("op_3896_cast_fp16")]; + tensor value_69_cast_fp16 = add(x = var_3894_cast_fp16, y = var_3896_cast_fp16)[name = tensor("value_69_cast_fp16")]; + tensor var_3899 = const()[name = tensor("op_3899"), val = tensor([1, 20, 64, -1])]; + tensor var_3900_cast_fp16 = reshape(shape = var_3899, x = query_69_cast_fp16)[name = tensor("op_3900_cast_fp16")]; + tensor var_3901_to_fp16 = const()[name = tensor("op_3901_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3902_cast_fp16 = mul(x = var_3900_cast_fp16, y = var_3901_to_fp16)[name = tensor("op_3902_cast_fp16")]; + tensor var_3903 = const()[name = tensor("op_3903"), val = tensor([1, 20, 64, -1])]; + tensor var_3904_cast_fp16 = reshape(shape = var_3903, x = key_69_cast_fp16)[name = tensor("op_3904_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_3902_cast_fp16, y = var_3904_cast_fp16)[name = tensor("mh_w_103_cast_fp16")]; + tensor mh_w_105_cast_fp16 = add(x = mh_w_103_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_105_cast_fp16")]; + tensor var_3912_cast_fp16 = softmax(axis = var_3826, x = mh_w_105_cast_fp16)[name = tensor("op_3912_cast_fp16")]; + tensor var_3913 = const()[name = tensor("op_3913"), val = tensor([1, 20, 64, -1])]; + tensor var_3914_cast_fp16 = reshape(shape = var_3913, x = value_69_cast_fp16)[name = tensor("op_3914_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_3914_cast_fp16, y = var_3912_cast_fp16)[name = tensor("attn_69_cast_fp16")]; + tensor var_3917 = const()[name = tensor("op_3917"), val = tensor([1, 1280, 1, -1])]; + tensor input_171_cast_fp16 = reshape(shape = var_3917, x = attn_69_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor var_3921 = const()[name = tensor("op_3921"), val = tensor([1, 1])]; + tensor var_3923 = const()[name = tensor("op_3923"), val = tensor([1, 1])]; + tensor obj_245_pad_type_0 = const()[name = tensor("obj_245_pad_type_0"), val = tensor("custom")]; + tensor obj_245_pad_0 = const()[name = tensor("obj_245_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1035823360)))]; + 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(1039100224)))]; + tensor obj_245_cast_fp16 = conv(bias = layers_17_self_attn_o_proj_bias_to_fp16, dilations = var_3923, groups = var_3833, pad = obj_245_pad_0, pad_type = obj_245_pad_type_0, strides = var_3921, 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 var_3933 = const()[name = tensor("op_3933"), val = tensor([1])]; + tensor channels_mean_105_cast_fp16 = reduce_mean(axes = var_3933, keep_dims = var_3834, x = inputs_105_cast_fp16)[name = tensor("channels_mean_105_cast_fp16")]; + tensor zero_mean_105_cast_fp16 = sub(x = inputs_105_cast_fp16, y = channels_mean_105_cast_fp16)[name = tensor("zero_mean_105_cast_fp16")]; + tensor zero_mean_sq_105_cast_fp16 = mul(x = zero_mean_105_cast_fp16, y = zero_mean_105_cast_fp16)[name = tensor("zero_mean_sq_105_cast_fp16")]; + tensor var_3937 = const()[name = tensor("op_3937"), val = tensor([1])]; + tensor var_3938_cast_fp16 = reduce_mean(axes = var_3937, keep_dims = var_3834, x = zero_mean_sq_105_cast_fp16)[name = tensor("op_3938_cast_fp16")]; + tensor var_3939_to_fp16 = const()[name = tensor("op_3939_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3940_cast_fp16 = add(x = var_3938_cast_fp16, y = var_3939_to_fp16)[name = tensor("op_3940_cast_fp16")]; + tensor denom_105_epsilon_0_to_fp16 = const()[name = tensor("denom_105_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_105_cast_fp16 = rsqrt(epsilon = denom_105_epsilon_0_to_fp16, x = var_3940_cast_fp16)[name = tensor("denom_105_cast_fp16")]; + tensor out_105_cast_fp16 = mul(x = zero_mean_105_cast_fp16, y = denom_105_cast_fp16)[name = tensor("out_105_cast_fp16")]; + tensor obj_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(1039102848)))]; + 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(1039105472)))]; + 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 var_3955 = const()[name = tensor("op_3955"), val = tensor([1, 1])]; + tensor var_3957 = const()[name = tensor("op_3957"), val = tensor([1, 1])]; + tensor query_71_pad_type_0 = const()[name = tensor("query_71_pad_type_0"), val = tensor("custom")]; + tensor query_71_pad_0 = const()[name = tensor("query_71_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1039108096)))]; + 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(1042384960)))]; + tensor query_71_cast_fp16 = conv(bias = layers_17_encoder_attn_q_proj_bias_to_fp16, dilations = var_3957, groups = var_3833, pad = query_71_pad_0, pad_type = query_71_pad_type_0, strides = var_3955, weight = layers_17_encoder_attn_q_proj_weight_to_fp16, x = obj_247_cast_fp16)[name = tensor("query_71_cast_fp16")]; + tensor var_3961 = const()[name = tensor("op_3961"), val = tensor([1, 1])]; + tensor var_3963 = const()[name = tensor("op_3963"), val = tensor([1, 1])]; + tensor key_71_pad_type_0 = const()[name = tensor("key_71_pad_type_0"), val = tensor("custom")]; + tensor key_71_pad_0 = const()[name = tensor("key_71_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1042387584)))]; + tensor key_71_cast_fp16 = conv(dilations = var_3963, groups = var_3833, pad = key_71_pad_0, pad_type = key_71_pad_type_0, strides = var_3961, weight = layers_17_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_71_cast_fp16")]; + tensor var_3968 = const()[name = tensor("op_3968"), val = tensor([1, 1])]; + tensor var_3970 = const()[name = tensor("op_3970"), val = tensor([1, 1])]; + tensor value_71_pad_type_0 = const()[name = tensor("value_71_pad_type_0"), val = tensor("custom")]; + tensor value_71_pad_0 = const()[name = tensor("value_71_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1045664448)))]; + 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(1048941312)))]; + tensor value_71_cast_fp16 = conv(bias = layers_17_encoder_attn_v_proj_bias_to_fp16, dilations = var_3970, groups = var_3833, pad = value_71_pad_0, pad_type = value_71_pad_type_0, strides = var_3968, weight = layers_17_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_71_cast_fp16")]; + tensor var_3974 = const()[name = tensor("op_3974"), val = tensor([1, 20, 64, -1])]; + tensor var_3975_cast_fp16 = reshape(shape = var_3974, x = query_71_cast_fp16)[name = tensor("op_3975_cast_fp16")]; + tensor var_3976_to_fp16 = const()[name = tensor("op_3976_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3977_cast_fp16 = mul(x = var_3975_cast_fp16, y = var_3976_to_fp16)[name = tensor("op_3977_cast_fp16")]; + tensor var_3978 = const()[name = tensor("op_3978"), val = tensor([1, 20, 64, -1])]; + tensor var_3979_cast_fp16 = reshape(shape = var_3978, x = key_71_cast_fp16)[name = tensor("op_3979_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_3977_cast_fp16, y = var_3979_cast_fp16)[name = tensor("mh_w_107_cast_fp16")]; + tensor obj_251_cast_fp16 = softmax(axis = var_3826, x = mh_w_107_cast_fp16)[name = tensor("obj_251_cast_fp16")]; + tensor var_3983 = const()[name = tensor("op_3983"), val = tensor([1, 20, 64, -1])]; + tensor var_3984_cast_fp16 = reshape(shape = var_3983, x = value_71_cast_fp16)[name = tensor("op_3984_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_3984_cast_fp16, y = obj_251_cast_fp16)[name = tensor("attn_71_cast_fp16")]; + tensor var_3987 = const()[name = tensor("op_3987"), val = tensor([1, 1280, 1, -1])]; + tensor input_173_cast_fp16 = reshape(shape = var_3987, x = attn_71_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor var_3991 = const()[name = tensor("op_3991"), val = tensor([1, 1])]; + tensor var_3993 = const()[name = tensor("op_3993"), val = tensor([1, 1])]; + tensor obj_249_pad_type_0 = const()[name = tensor("obj_249_pad_type_0"), val = tensor("custom")]; + tensor obj_249_pad_0 = const()[name = tensor("obj_249_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1048943936)))]; + 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(1052220800)))]; + tensor obj_249_cast_fp16 = conv(bias = layers_17_encoder_attn_o_proj_bias_to_fp16, dilations = var_3993, groups = var_3833, pad = obj_249_pad_0, pad_type = obj_249_pad_type_0, strides = var_3991, 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 var_4002 = const()[name = tensor("op_4002"), val = tensor([1])]; + tensor channels_mean_107_cast_fp16 = reduce_mean(axes = var_4002, keep_dims = var_3834, x = inputs_107_cast_fp16)[name = tensor("channels_mean_107_cast_fp16")]; + tensor zero_mean_107_cast_fp16 = sub(x = inputs_107_cast_fp16, y = channels_mean_107_cast_fp16)[name = tensor("zero_mean_107_cast_fp16")]; + tensor zero_mean_sq_107_cast_fp16 = mul(x = zero_mean_107_cast_fp16, y = zero_mean_107_cast_fp16)[name = tensor("zero_mean_sq_107_cast_fp16")]; + tensor var_4006 = const()[name = tensor("op_4006"), val = tensor([1])]; + tensor var_4007_cast_fp16 = reduce_mean(axes = var_4006, keep_dims = var_3834, x = zero_mean_sq_107_cast_fp16)[name = tensor("op_4007_cast_fp16")]; + tensor var_4008_to_fp16 = const()[name = tensor("op_4008_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4009_cast_fp16 = add(x = var_4007_cast_fp16, y = var_4008_to_fp16)[name = tensor("op_4009_cast_fp16")]; + tensor denom_107_epsilon_0_to_fp16 = const()[name = tensor("denom_107_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_107_cast_fp16 = rsqrt(epsilon = denom_107_epsilon_0_to_fp16, x = var_4009_cast_fp16)[name = tensor("denom_107_cast_fp16")]; + tensor out_107_cast_fp16 = mul(x = zero_mean_107_cast_fp16, y = denom_107_cast_fp16)[name = tensor("out_107_cast_fp16")]; + tensor input_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(1052223424)))]; + 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(1052226048)))]; + 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 var_4020 = const()[name = tensor("op_4020"), val = tensor([1, 1])]; + tensor var_4022 = const()[name = tensor("op_4022"), val = tensor([1, 1])]; + tensor input_177_pad_type_0 = const()[name = tensor("input_177_pad_type_0"), val = tensor("custom")]; + tensor input_177_pad_0 = const()[name = tensor("input_177_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1052228672)))]; + 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(1065335936)))]; + tensor input_177_cast_fp16 = conv(bias = layers_17_fc1_bias_to_fp16, dilations = var_4022, groups = var_3833, pad = input_177_pad_0, pad_type = input_177_pad_type_0, strides = var_4020, 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 var_4028 = const()[name = tensor("op_4028"), val = tensor([1, 1])]; + tensor var_4030 = const()[name = tensor("op_4030"), val = tensor([1, 1])]; + tensor hidden_states_37_pad_type_0 = const()[name = tensor("hidden_states_37_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_37_pad_0 = const()[name = tensor("hidden_states_37_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1065346240)))]; + 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(1078453504)))]; + tensor hidden_states_37_cast_fp16 = conv(bias = layers_17_fc2_bias_to_fp16, dilations = var_4030, groups = var_3833, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = var_4028, 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_4044 = const()[name = tensor("op_4044"), val = tensor(3)]; + tensor var_4051 = const()[name = tensor("op_4051"), val = tensor(1)]; + tensor var_4052 = const()[name = tensor("op_4052"), val = tensor(true)]; + tensor var_4064 = const()[name = tensor("op_4064"), val = tensor([1])]; + tensor channels_mean_109_cast_fp16 = reduce_mean(axes = var_4064, keep_dims = var_4052, x = inputs_109_cast_fp16)[name = tensor("channels_mean_109_cast_fp16")]; + tensor zero_mean_109_cast_fp16 = sub(x = inputs_109_cast_fp16, y = channels_mean_109_cast_fp16)[name = tensor("zero_mean_109_cast_fp16")]; + tensor zero_mean_sq_109_cast_fp16 = mul(x = zero_mean_109_cast_fp16, y = zero_mean_109_cast_fp16)[name = tensor("zero_mean_sq_109_cast_fp16")]; + tensor var_4068 = const()[name = tensor("op_4068"), val = tensor([1])]; + tensor var_4069_cast_fp16 = reduce_mean(axes = var_4068, keep_dims = var_4052, x = zero_mean_sq_109_cast_fp16)[name = tensor("op_4069_cast_fp16")]; + tensor var_4070_to_fp16 = const()[name = tensor("op_4070_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4071_cast_fp16 = add(x = var_4069_cast_fp16, y = var_4070_to_fp16)[name = tensor("op_4071_cast_fp16")]; + tensor denom_109_epsilon_0_to_fp16 = const()[name = tensor("denom_109_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_109_cast_fp16 = rsqrt(epsilon = denom_109_epsilon_0_to_fp16, x = var_4071_cast_fp16)[name = tensor("denom_109_cast_fp16")]; + tensor out_109_cast_fp16 = mul(x = zero_mean_109_cast_fp16, y = denom_109_cast_fp16)[name = tensor("out_109_cast_fp16")]; + tensor obj_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(1078456128)))]; + 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(1078458752)))]; + 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 var_4086 = const()[name = tensor("op_4086"), val = tensor([1, 1])]; + tensor var_4088 = const()[name = tensor("op_4088"), val = tensor([1, 1])]; + tensor query_73_pad_type_0 = const()[name = tensor("query_73_pad_type_0"), val = tensor("custom")]; + tensor query_73_pad_0 = const()[name = tensor("query_73_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1078461376)))]; + 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(1081738240)))]; + tensor query_73_cast_fp16 = conv(bias = layers_18_self_attn_q_proj_bias_to_fp16, dilations = var_4088, groups = var_4051, pad = query_73_pad_0, pad_type = query_73_pad_type_0, strides = var_4086, weight = layers_18_self_attn_q_proj_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("query_73_cast_fp16")]; + tensor var_4092 = const()[name = tensor("op_4092"), val = tensor([1, 1])]; + tensor var_4094 = const()[name = tensor("op_4094"), val = tensor([1, 1])]; + tensor current_key_37_pad_type_0 = const()[name = tensor("current_key_37_pad_type_0"), val = tensor("custom")]; + tensor current_key_37_pad_0 = const()[name = tensor("current_key_37_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1081740864)))]; + tensor current_key_37_cast_fp16 = conv(dilations = var_4094, groups = var_4051, pad = current_key_37_pad_0, pad_type = current_key_37_pad_type_0, strides = var_4092, weight = layers_18_self_attn_k_proj_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("current_key_37_cast_fp16")]; + tensor var_4099 = const()[name = tensor("op_4099"), val = tensor([1, 1])]; + tensor var_4101 = const()[name = tensor("op_4101"), val = tensor([1, 1])]; + tensor current_value_37_pad_type_0 = const()[name = tensor("current_value_37_pad_type_0"), val = tensor("custom")]; + tensor current_value_37_pad_0 = const()[name = tensor("current_value_37_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1085017728)))]; + 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(1088294592)))]; + tensor current_value_37_cast_fp16 = conv(bias = layers_18_self_attn_v_proj_bias_to_fp16, dilations = var_4101, groups = var_4051, pad = current_value_37_pad_0, pad_type = current_value_37_pad_type_0, strides = var_4099, weight = layers_18_self_attn_v_proj_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("current_value_37_cast_fp16")]; + tensor var_4108_cast_fp16 = mul(x = current_key_37_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4108_cast_fp16")]; + tensor var_4110_cast_fp16 = mul(x = var_103_cast_fp16_18, y = var_241_cast_fp16)[name = tensor("op_4110_cast_fp16")]; + tensor key_73_cast_fp16 = add(x = var_4108_cast_fp16, y = var_4110_cast_fp16)[name = tensor("key_73_cast_fp16")]; + tensor var_4112_cast_fp16 = mul(x = current_value_37_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4112_cast_fp16")]; + tensor var_4114_cast_fp16 = mul(x = var_138_cast_fp16_18, y = var_241_cast_fp16)[name = tensor("op_4114_cast_fp16")]; + tensor value_73_cast_fp16 = add(x = var_4112_cast_fp16, y = var_4114_cast_fp16)[name = tensor("value_73_cast_fp16")]; + tensor var_4117 = const()[name = tensor("op_4117"), val = tensor([1, 20, 64, -1])]; + tensor var_4118_cast_fp16 = reshape(shape = var_4117, x = query_73_cast_fp16)[name = tensor("op_4118_cast_fp16")]; + tensor var_4119_to_fp16 = const()[name = tensor("op_4119_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4120_cast_fp16 = mul(x = var_4118_cast_fp16, y = var_4119_to_fp16)[name = tensor("op_4120_cast_fp16")]; + tensor var_4121 = const()[name = tensor("op_4121"), val = tensor([1, 20, 64, -1])]; + tensor var_4122_cast_fp16 = reshape(shape = var_4121, x = key_73_cast_fp16)[name = tensor("op_4122_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_4120_cast_fp16, y = var_4122_cast_fp16)[name = tensor("mh_w_109_cast_fp16")]; + tensor mh_w_111_cast_fp16 = add(x = mh_w_109_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_111_cast_fp16")]; + tensor var_4130_cast_fp16 = softmax(axis = var_4044, x = mh_w_111_cast_fp16)[name = tensor("op_4130_cast_fp16")]; + tensor var_4131 = const()[name = tensor("op_4131"), val = tensor([1, 20, 64, -1])]; + tensor var_4132_cast_fp16 = reshape(shape = var_4131, x = value_73_cast_fp16)[name = tensor("op_4132_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_4132_cast_fp16, y = var_4130_cast_fp16)[name = tensor("attn_73_cast_fp16")]; + tensor var_4135 = const()[name = tensor("op_4135"), val = tensor([1, 1280, 1, -1])]; + tensor input_181_cast_fp16 = reshape(shape = var_4135, x = attn_73_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor var_4139 = const()[name = tensor("op_4139"), val = tensor([1, 1])]; + tensor var_4141 = const()[name = tensor("op_4141"), val = tensor([1, 1])]; + tensor obj_259_pad_type_0 = const()[name = tensor("obj_259_pad_type_0"), val = tensor("custom")]; + tensor obj_259_pad_0 = const()[name = tensor("obj_259_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1088297216)))]; + 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(1091574080)))]; + tensor obj_259_cast_fp16 = conv(bias = layers_18_self_attn_o_proj_bias_to_fp16, dilations = var_4141, groups = var_4051, pad = obj_259_pad_0, pad_type = obj_259_pad_type_0, strides = var_4139, 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 var_4151 = const()[name = tensor("op_4151"), val = tensor([1])]; + tensor channels_mean_111_cast_fp16 = reduce_mean(axes = var_4151, keep_dims = var_4052, x = inputs_111_cast_fp16)[name = tensor("channels_mean_111_cast_fp16")]; + tensor zero_mean_111_cast_fp16 = sub(x = inputs_111_cast_fp16, y = channels_mean_111_cast_fp16)[name = tensor("zero_mean_111_cast_fp16")]; + tensor zero_mean_sq_111_cast_fp16 = mul(x = zero_mean_111_cast_fp16, y = zero_mean_111_cast_fp16)[name = tensor("zero_mean_sq_111_cast_fp16")]; + tensor var_4155 = const()[name = tensor("op_4155"), val = tensor([1])]; + tensor var_4156_cast_fp16 = reduce_mean(axes = var_4155, keep_dims = var_4052, x = zero_mean_sq_111_cast_fp16)[name = tensor("op_4156_cast_fp16")]; + tensor var_4157_to_fp16 = const()[name = tensor("op_4157_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4158_cast_fp16 = add(x = var_4156_cast_fp16, y = var_4157_to_fp16)[name = tensor("op_4158_cast_fp16")]; + tensor denom_111_epsilon_0_to_fp16 = const()[name = tensor("denom_111_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_111_cast_fp16 = rsqrt(epsilon = denom_111_epsilon_0_to_fp16, x = var_4158_cast_fp16)[name = tensor("denom_111_cast_fp16")]; + tensor out_111_cast_fp16 = mul(x = zero_mean_111_cast_fp16, y = denom_111_cast_fp16)[name = tensor("out_111_cast_fp16")]; + tensor obj_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(1091576704)))]; + 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(1091579328)))]; + 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 var_4173 = const()[name = tensor("op_4173"), val = tensor([1, 1])]; + tensor var_4175 = const()[name = tensor("op_4175"), val = tensor([1, 1])]; + tensor query_75_pad_type_0 = const()[name = tensor("query_75_pad_type_0"), val = tensor("custom")]; + tensor query_75_pad_0 = const()[name = tensor("query_75_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1091581952)))]; + 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(1094858816)))]; + tensor query_75_cast_fp16 = conv(bias = layers_18_encoder_attn_q_proj_bias_to_fp16, dilations = var_4175, groups = var_4051, pad = query_75_pad_0, pad_type = query_75_pad_type_0, strides = var_4173, weight = layers_18_encoder_attn_q_proj_weight_to_fp16, x = obj_261_cast_fp16)[name = tensor("query_75_cast_fp16")]; + tensor var_4179 = const()[name = tensor("op_4179"), val = tensor([1, 1])]; + tensor var_4181 = const()[name = tensor("op_4181"), val = tensor([1, 1])]; + tensor key_75_pad_type_0 = const()[name = tensor("key_75_pad_type_0"), val = tensor("custom")]; + tensor key_75_pad_0 = const()[name = tensor("key_75_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1094861440)))]; + tensor key_75_cast_fp16 = conv(dilations = var_4181, groups = var_4051, pad = key_75_pad_0, pad_type = key_75_pad_type_0, strides = var_4179, weight = layers_18_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_75_cast_fp16")]; + tensor var_4186 = const()[name = tensor("op_4186"), val = tensor([1, 1])]; + tensor var_4188 = const()[name = tensor("op_4188"), val = tensor([1, 1])]; + tensor value_75_pad_type_0 = const()[name = tensor("value_75_pad_type_0"), val = tensor("custom")]; + tensor value_75_pad_0 = const()[name = tensor("value_75_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1098138304)))]; + 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(1101415168)))]; + tensor value_75_cast_fp16 = conv(bias = layers_18_encoder_attn_v_proj_bias_to_fp16, dilations = var_4188, groups = var_4051, pad = value_75_pad_0, pad_type = value_75_pad_type_0, strides = var_4186, weight = layers_18_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_75_cast_fp16")]; + tensor var_4192 = const()[name = tensor("op_4192"), val = tensor([1, 20, 64, -1])]; + tensor var_4193_cast_fp16 = reshape(shape = var_4192, x = query_75_cast_fp16)[name = tensor("op_4193_cast_fp16")]; + tensor var_4194_to_fp16 = const()[name = tensor("op_4194_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4195_cast_fp16 = mul(x = var_4193_cast_fp16, y = var_4194_to_fp16)[name = tensor("op_4195_cast_fp16")]; + tensor var_4196 = const()[name = tensor("op_4196"), val = tensor([1, 20, 64, -1])]; + tensor var_4197_cast_fp16 = reshape(shape = var_4196, x = key_75_cast_fp16)[name = tensor("op_4197_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_4195_cast_fp16, y = var_4197_cast_fp16)[name = tensor("mh_w_113_cast_fp16")]; + tensor obj_265_cast_fp16 = softmax(axis = var_4044, x = mh_w_113_cast_fp16)[name = tensor("obj_265_cast_fp16")]; + tensor var_4201 = const()[name = tensor("op_4201"), val = tensor([1, 20, 64, -1])]; + tensor var_4202_cast_fp16 = reshape(shape = var_4201, x = value_75_cast_fp16)[name = tensor("op_4202_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_4202_cast_fp16, y = obj_265_cast_fp16)[name = tensor("attn_75_cast_fp16")]; + tensor var_4205 = const()[name = tensor("op_4205"), val = tensor([1, 1280, 1, -1])]; + tensor input_183_cast_fp16 = reshape(shape = var_4205, x = attn_75_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor var_4209 = const()[name = tensor("op_4209"), val = tensor([1, 1])]; + tensor var_4211 = const()[name = tensor("op_4211"), val = tensor([1, 1])]; + tensor obj_263_pad_type_0 = const()[name = tensor("obj_263_pad_type_0"), val = tensor("custom")]; + tensor obj_263_pad_0 = const()[name = tensor("obj_263_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1101417792)))]; + 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(1104694656)))]; + tensor obj_263_cast_fp16 = conv(bias = layers_18_encoder_attn_o_proj_bias_to_fp16, dilations = var_4211, groups = var_4051, pad = obj_263_pad_0, pad_type = obj_263_pad_type_0, strides = var_4209, 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 var_4220 = const()[name = tensor("op_4220"), val = tensor([1])]; + tensor channels_mean_113_cast_fp16 = reduce_mean(axes = var_4220, keep_dims = var_4052, x = inputs_113_cast_fp16)[name = tensor("channels_mean_113_cast_fp16")]; + tensor zero_mean_113_cast_fp16 = sub(x = inputs_113_cast_fp16, y = channels_mean_113_cast_fp16)[name = tensor("zero_mean_113_cast_fp16")]; + tensor zero_mean_sq_113_cast_fp16 = mul(x = zero_mean_113_cast_fp16, y = zero_mean_113_cast_fp16)[name = tensor("zero_mean_sq_113_cast_fp16")]; + tensor var_4224 = const()[name = tensor("op_4224"), val = tensor([1])]; + tensor var_4225_cast_fp16 = reduce_mean(axes = var_4224, keep_dims = var_4052, x = zero_mean_sq_113_cast_fp16)[name = tensor("op_4225_cast_fp16")]; + tensor var_4226_to_fp16 = const()[name = tensor("op_4226_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4227_cast_fp16 = add(x = var_4225_cast_fp16, y = var_4226_to_fp16)[name = tensor("op_4227_cast_fp16")]; + tensor denom_113_epsilon_0_to_fp16 = const()[name = tensor("denom_113_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_113_cast_fp16 = rsqrt(epsilon = denom_113_epsilon_0_to_fp16, x = var_4227_cast_fp16)[name = tensor("denom_113_cast_fp16")]; + tensor out_113_cast_fp16 = mul(x = zero_mean_113_cast_fp16, y = denom_113_cast_fp16)[name = tensor("out_113_cast_fp16")]; + tensor input_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(1104697280)))]; + 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(1104699904)))]; + 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 var_4238 = const()[name = tensor("op_4238"), val = tensor([1, 1])]; + tensor var_4240 = const()[name = tensor("op_4240"), val = tensor([1, 1])]; + tensor input_187_pad_type_0 = const()[name = tensor("input_187_pad_type_0"), val = tensor("custom")]; + tensor input_187_pad_0 = const()[name = tensor("input_187_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1104702528)))]; + 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(1117809792)))]; + tensor input_187_cast_fp16 = conv(bias = layers_18_fc1_bias_to_fp16, dilations = var_4240, groups = var_4051, pad = input_187_pad_0, pad_type = input_187_pad_type_0, strides = var_4238, 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 var_4246 = const()[name = tensor("op_4246"), val = tensor([1, 1])]; + tensor var_4248 = const()[name = tensor("op_4248"), val = tensor([1, 1])]; + tensor hidden_states_39_pad_type_0 = const()[name = tensor("hidden_states_39_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_39_pad_0 = const()[name = tensor("hidden_states_39_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1117820096)))]; + 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(1130927360)))]; + tensor hidden_states_39_cast_fp16 = conv(bias = layers_18_fc2_bias_to_fp16, dilations = var_4248, groups = var_4051, pad = hidden_states_39_pad_0, pad_type = hidden_states_39_pad_type_0, strides = var_4246, 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_4262 = const()[name = tensor("op_4262"), val = tensor(3)]; + tensor var_4269 = const()[name = tensor("op_4269"), val = tensor(1)]; + tensor var_4270 = const()[name = tensor("op_4270"), val = tensor(true)]; + tensor var_4282 = const()[name = tensor("op_4282"), val = tensor([1])]; + tensor channels_mean_115_cast_fp16 = reduce_mean(axes = var_4282, keep_dims = var_4270, x = inputs_115_cast_fp16)[name = tensor("channels_mean_115_cast_fp16")]; + tensor zero_mean_115_cast_fp16 = sub(x = inputs_115_cast_fp16, y = channels_mean_115_cast_fp16)[name = tensor("zero_mean_115_cast_fp16")]; + tensor zero_mean_sq_115_cast_fp16 = mul(x = zero_mean_115_cast_fp16, y = zero_mean_115_cast_fp16)[name = tensor("zero_mean_sq_115_cast_fp16")]; + tensor var_4286 = const()[name = tensor("op_4286"), val = tensor([1])]; + tensor var_4287_cast_fp16 = reduce_mean(axes = var_4286, keep_dims = var_4270, x = zero_mean_sq_115_cast_fp16)[name = tensor("op_4287_cast_fp16")]; + tensor var_4288_to_fp16 = const()[name = tensor("op_4288_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4289_cast_fp16 = add(x = var_4287_cast_fp16, y = var_4288_to_fp16)[name = tensor("op_4289_cast_fp16")]; + tensor denom_115_epsilon_0_to_fp16 = const()[name = tensor("denom_115_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_115_cast_fp16 = rsqrt(epsilon = denom_115_epsilon_0_to_fp16, x = var_4289_cast_fp16)[name = tensor("denom_115_cast_fp16")]; + tensor out_115_cast_fp16 = mul(x = zero_mean_115_cast_fp16, y = denom_115_cast_fp16)[name = tensor("out_115_cast_fp16")]; + tensor obj_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(1130929984)))]; + 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(1130932608)))]; + 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 var_4304 = const()[name = tensor("op_4304"), val = tensor([1, 1])]; + tensor var_4306 = const()[name = tensor("op_4306"), val = tensor([1, 1])]; + tensor query_77_pad_type_0 = const()[name = tensor("query_77_pad_type_0"), val = tensor("custom")]; + tensor query_77_pad_0 = const()[name = tensor("query_77_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1130935232)))]; + 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(1134212096)))]; + tensor query_77_cast_fp16 = conv(bias = layers_19_self_attn_q_proj_bias_to_fp16, dilations = var_4306, groups = var_4269, pad = query_77_pad_0, pad_type = query_77_pad_type_0, strides = var_4304, weight = layers_19_self_attn_q_proj_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor("query_77_cast_fp16")]; + tensor var_4310 = const()[name = tensor("op_4310"), val = tensor([1, 1])]; + tensor var_4312 = const()[name = tensor("op_4312"), val = tensor([1, 1])]; + tensor current_key_39_pad_type_0 = const()[name = tensor("current_key_39_pad_type_0"), val = tensor("custom")]; + tensor current_key_39_pad_0 = const()[name = tensor("current_key_39_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1134214720)))]; + tensor current_key_39_cast_fp16 = conv(dilations = var_4312, groups = var_4269, pad = current_key_39_pad_0, pad_type = current_key_39_pad_type_0, strides = var_4310, weight = layers_19_self_attn_k_proj_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor("current_key_39_cast_fp16")]; + tensor var_4317 = const()[name = tensor("op_4317"), val = tensor([1, 1])]; + tensor var_4319 = const()[name = tensor("op_4319"), val = tensor([1, 1])]; + tensor current_value_39_pad_type_0 = const()[name = tensor("current_value_39_pad_type_0"), val = tensor("custom")]; + tensor current_value_39_pad_0 = const()[name = tensor("current_value_39_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1137491584)))]; + 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(1140768448)))]; + tensor current_value_39_cast_fp16 = conv(bias = layers_19_self_attn_v_proj_bias_to_fp16, dilations = var_4319, groups = var_4269, pad = current_value_39_pad_0, pad_type = current_value_39_pad_type_0, strides = var_4317, weight = layers_19_self_attn_v_proj_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor("current_value_39_cast_fp16")]; + tensor var_4326_cast_fp16 = mul(x = current_key_39_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4326_cast_fp16")]; + tensor var_4328_cast_fp16 = mul(x = var_103_cast_fp16_19, y = var_241_cast_fp16)[name = tensor("op_4328_cast_fp16")]; + tensor key_77_cast_fp16 = add(x = var_4326_cast_fp16, y = var_4328_cast_fp16)[name = tensor("key_77_cast_fp16")]; + tensor var_4330_cast_fp16 = mul(x = current_value_39_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4330_cast_fp16")]; + tensor var_4332_cast_fp16 = mul(x = var_138_cast_fp16_19, y = var_241_cast_fp16)[name = tensor("op_4332_cast_fp16")]; + tensor value_77_cast_fp16 = add(x = var_4330_cast_fp16, y = var_4332_cast_fp16)[name = tensor("value_77_cast_fp16")]; + tensor var_4335 = const()[name = tensor("op_4335"), val = tensor([1, 20, 64, -1])]; + tensor var_4336_cast_fp16 = reshape(shape = var_4335, x = query_77_cast_fp16)[name = tensor("op_4336_cast_fp16")]; + tensor var_4337_to_fp16 = const()[name = tensor("op_4337_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4338_cast_fp16 = mul(x = var_4336_cast_fp16, y = var_4337_to_fp16)[name = tensor("op_4338_cast_fp16")]; + tensor var_4339 = const()[name = tensor("op_4339"), val = tensor([1, 20, 64, -1])]; + tensor var_4340_cast_fp16 = reshape(shape = var_4339, x = key_77_cast_fp16)[name = tensor("op_4340_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_4338_cast_fp16, y = var_4340_cast_fp16)[name = tensor("mh_w_115_cast_fp16")]; + tensor mh_w_117_cast_fp16 = add(x = mh_w_115_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_117_cast_fp16")]; + tensor var_4348_cast_fp16 = softmax(axis = var_4262, x = mh_w_117_cast_fp16)[name = tensor("op_4348_cast_fp16")]; + tensor var_4349 = const()[name = tensor("op_4349"), val = tensor([1, 20, 64, -1])]; + tensor var_4350_cast_fp16 = reshape(shape = var_4349, x = value_77_cast_fp16)[name = tensor("op_4350_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_4350_cast_fp16, y = var_4348_cast_fp16)[name = tensor("attn_77_cast_fp16")]; + tensor var_4353 = const()[name = tensor("op_4353"), val = tensor([1, 1280, 1, -1])]; + tensor input_191_cast_fp16 = reshape(shape = var_4353, x = attn_77_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor var_4357 = const()[name = tensor("op_4357"), val = tensor([1, 1])]; + tensor var_4359 = const()[name = tensor("op_4359"), val = tensor([1, 1])]; + tensor obj_273_pad_type_0 = const()[name = tensor("obj_273_pad_type_0"), val = tensor("custom")]; + tensor obj_273_pad_0 = const()[name = tensor("obj_273_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1140771072)))]; + 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(1144047936)))]; + tensor obj_273_cast_fp16 = conv(bias = layers_19_self_attn_o_proj_bias_to_fp16, dilations = var_4359, groups = var_4269, pad = obj_273_pad_0, pad_type = obj_273_pad_type_0, strides = var_4357, 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 var_4369 = const()[name = tensor("op_4369"), val = tensor([1])]; + tensor channels_mean_117_cast_fp16 = reduce_mean(axes = var_4369, keep_dims = var_4270, x = inputs_117_cast_fp16)[name = tensor("channels_mean_117_cast_fp16")]; + tensor zero_mean_117_cast_fp16 = sub(x = inputs_117_cast_fp16, y = channels_mean_117_cast_fp16)[name = tensor("zero_mean_117_cast_fp16")]; + tensor zero_mean_sq_117_cast_fp16 = mul(x = zero_mean_117_cast_fp16, y = zero_mean_117_cast_fp16)[name = tensor("zero_mean_sq_117_cast_fp16")]; + tensor var_4373 = const()[name = tensor("op_4373"), val = tensor([1])]; + tensor var_4374_cast_fp16 = reduce_mean(axes = var_4373, keep_dims = var_4270, x = zero_mean_sq_117_cast_fp16)[name = tensor("op_4374_cast_fp16")]; + tensor var_4375_to_fp16 = const()[name = tensor("op_4375_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4376_cast_fp16 = add(x = var_4374_cast_fp16, y = var_4375_to_fp16)[name = tensor("op_4376_cast_fp16")]; + tensor denom_117_epsilon_0_to_fp16 = const()[name = tensor("denom_117_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_117_cast_fp16 = rsqrt(epsilon = denom_117_epsilon_0_to_fp16, x = var_4376_cast_fp16)[name = tensor("denom_117_cast_fp16")]; + tensor out_117_cast_fp16 = mul(x = zero_mean_117_cast_fp16, y = denom_117_cast_fp16)[name = tensor("out_117_cast_fp16")]; + tensor obj_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(1144050560)))]; + 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(1144053184)))]; + 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 var_4391 = const()[name = tensor("op_4391"), val = tensor([1, 1])]; + tensor var_4393 = const()[name = tensor("op_4393"), val = tensor([1, 1])]; + tensor query_79_pad_type_0 = const()[name = tensor("query_79_pad_type_0"), val = tensor("custom")]; + tensor query_79_pad_0 = const()[name = tensor("query_79_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1144055808)))]; + 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(1147332672)))]; + tensor query_79_cast_fp16 = conv(bias = layers_19_encoder_attn_q_proj_bias_to_fp16, dilations = var_4393, groups = var_4269, pad = query_79_pad_0, pad_type = query_79_pad_type_0, strides = var_4391, weight = layers_19_encoder_attn_q_proj_weight_to_fp16, x = obj_275_cast_fp16)[name = tensor("query_79_cast_fp16")]; + tensor var_4397 = const()[name = tensor("op_4397"), val = tensor([1, 1])]; + tensor var_4399 = const()[name = tensor("op_4399"), val = tensor([1, 1])]; + tensor key_79_pad_type_0 = const()[name = tensor("key_79_pad_type_0"), val = tensor("custom")]; + tensor key_79_pad_0 = const()[name = tensor("key_79_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1147335296)))]; + tensor key_79_cast_fp16 = conv(dilations = var_4399, groups = var_4269, pad = key_79_pad_0, pad_type = key_79_pad_type_0, strides = var_4397, weight = layers_19_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_79_cast_fp16")]; + tensor var_4404 = const()[name = tensor("op_4404"), val = tensor([1, 1])]; + tensor var_4406 = const()[name = tensor("op_4406"), val = tensor([1, 1])]; + tensor value_79_pad_type_0 = const()[name = tensor("value_79_pad_type_0"), val = tensor("custom")]; + tensor value_79_pad_0 = const()[name = tensor("value_79_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1150612160)))]; + 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(1153889024)))]; + tensor value_79_cast_fp16 = conv(bias = layers_19_encoder_attn_v_proj_bias_to_fp16, dilations = var_4406, groups = var_4269, pad = value_79_pad_0, pad_type = value_79_pad_type_0, strides = var_4404, weight = layers_19_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_79_cast_fp16")]; + tensor var_4410 = const()[name = tensor("op_4410"), val = tensor([1, 20, 64, -1])]; + tensor var_4411_cast_fp16 = reshape(shape = var_4410, x = query_79_cast_fp16)[name = tensor("op_4411_cast_fp16")]; + tensor var_4412_to_fp16 = const()[name = tensor("op_4412_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4413_cast_fp16 = mul(x = var_4411_cast_fp16, y = var_4412_to_fp16)[name = tensor("op_4413_cast_fp16")]; + tensor var_4414 = const()[name = tensor("op_4414"), val = tensor([1, 20, 64, -1])]; + tensor var_4415_cast_fp16 = reshape(shape = var_4414, x = key_79_cast_fp16)[name = tensor("op_4415_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_4413_cast_fp16, y = var_4415_cast_fp16)[name = tensor("mh_w_119_cast_fp16")]; + tensor obj_279_cast_fp16 = softmax(axis = var_4262, x = mh_w_119_cast_fp16)[name = tensor("obj_279_cast_fp16")]; + tensor var_4419 = const()[name = tensor("op_4419"), val = tensor([1, 20, 64, -1])]; + tensor var_4420_cast_fp16 = reshape(shape = var_4419, x = value_79_cast_fp16)[name = tensor("op_4420_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_4420_cast_fp16, y = obj_279_cast_fp16)[name = tensor("attn_79_cast_fp16")]; + tensor var_4423 = const()[name = tensor("op_4423"), val = tensor([1, 1280, 1, -1])]; + tensor input_193_cast_fp16 = reshape(shape = var_4423, x = attn_79_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor var_4427 = const()[name = tensor("op_4427"), val = tensor([1, 1])]; + tensor var_4429 = const()[name = tensor("op_4429"), val = tensor([1, 1])]; + tensor obj_277_pad_type_0 = const()[name = tensor("obj_277_pad_type_0"), val = tensor("custom")]; + tensor obj_277_pad_0 = const()[name = tensor("obj_277_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1153891648)))]; + 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(1157168512)))]; + tensor obj_277_cast_fp16 = conv(bias = layers_19_encoder_attn_o_proj_bias_to_fp16, dilations = var_4429, groups = var_4269, pad = obj_277_pad_0, pad_type = obj_277_pad_type_0, strides = var_4427, 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 var_4438 = const()[name = tensor("op_4438"), val = tensor([1])]; + tensor channels_mean_119_cast_fp16 = reduce_mean(axes = var_4438, keep_dims = var_4270, x = inputs_119_cast_fp16)[name = tensor("channels_mean_119_cast_fp16")]; + tensor zero_mean_119_cast_fp16 = sub(x = inputs_119_cast_fp16, y = channels_mean_119_cast_fp16)[name = tensor("zero_mean_119_cast_fp16")]; + tensor zero_mean_sq_119_cast_fp16 = mul(x = zero_mean_119_cast_fp16, y = zero_mean_119_cast_fp16)[name = tensor("zero_mean_sq_119_cast_fp16")]; + tensor var_4442 = const()[name = tensor("op_4442"), val = tensor([1])]; + tensor var_4443_cast_fp16 = reduce_mean(axes = var_4442, keep_dims = var_4270, x = zero_mean_sq_119_cast_fp16)[name = tensor("op_4443_cast_fp16")]; + tensor var_4444_to_fp16 = const()[name = tensor("op_4444_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4445_cast_fp16 = add(x = var_4443_cast_fp16, y = var_4444_to_fp16)[name = tensor("op_4445_cast_fp16")]; + tensor denom_119_epsilon_0_to_fp16 = const()[name = tensor("denom_119_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_119_cast_fp16 = rsqrt(epsilon = denom_119_epsilon_0_to_fp16, x = var_4445_cast_fp16)[name = tensor("denom_119_cast_fp16")]; + tensor out_119_cast_fp16 = mul(x = zero_mean_119_cast_fp16, y = denom_119_cast_fp16)[name = tensor("out_119_cast_fp16")]; + tensor input_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(1157171136)))]; + 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(1157173760)))]; + 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 var_4456 = const()[name = tensor("op_4456"), val = tensor([1, 1])]; + tensor var_4458 = const()[name = tensor("op_4458"), val = tensor([1, 1])]; + tensor input_197_pad_type_0 = const()[name = tensor("input_197_pad_type_0"), val = tensor("custom")]; + tensor input_197_pad_0 = const()[name = tensor("input_197_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1157176384)))]; + 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(1170283648)))]; + tensor input_197_cast_fp16 = conv(bias = layers_19_fc1_bias_to_fp16, dilations = var_4458, groups = var_4269, pad = input_197_pad_0, pad_type = input_197_pad_type_0, strides = var_4456, 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 var_4464 = const()[name = tensor("op_4464"), val = tensor([1, 1])]; + tensor var_4466 = const()[name = tensor("op_4466"), val = tensor([1, 1])]; + tensor hidden_states_41_pad_type_0 = const()[name = tensor("hidden_states_41_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_41_pad_0 = const()[name = tensor("hidden_states_41_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1170293952)))]; + 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(1183401216)))]; + tensor hidden_states_41_cast_fp16 = conv(bias = layers_19_fc2_bias_to_fp16, dilations = var_4466, groups = var_4269, pad = hidden_states_41_pad_0, pad_type = hidden_states_41_pad_type_0, strides = var_4464, 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_4480 = const()[name = tensor("op_4480"), val = tensor(3)]; + tensor var_4487 = const()[name = tensor("op_4487"), val = tensor(1)]; + tensor var_4488 = const()[name = tensor("op_4488"), val = tensor(true)]; + tensor var_4500 = const()[name = tensor("op_4500"), val = tensor([1])]; + tensor channels_mean_121_cast_fp16 = reduce_mean(axes = var_4500, keep_dims = var_4488, x = inputs_121_cast_fp16)[name = tensor("channels_mean_121_cast_fp16")]; + tensor zero_mean_121_cast_fp16 = sub(x = inputs_121_cast_fp16, y = channels_mean_121_cast_fp16)[name = tensor("zero_mean_121_cast_fp16")]; + tensor zero_mean_sq_121_cast_fp16 = mul(x = zero_mean_121_cast_fp16, y = zero_mean_121_cast_fp16)[name = tensor("zero_mean_sq_121_cast_fp16")]; + tensor var_4504 = const()[name = tensor("op_4504"), val = tensor([1])]; + tensor var_4505_cast_fp16 = reduce_mean(axes = var_4504, keep_dims = var_4488, x = zero_mean_sq_121_cast_fp16)[name = tensor("op_4505_cast_fp16")]; + tensor var_4506_to_fp16 = const()[name = tensor("op_4506_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4507_cast_fp16 = add(x = var_4505_cast_fp16, y = var_4506_to_fp16)[name = tensor("op_4507_cast_fp16")]; + tensor denom_121_epsilon_0_to_fp16 = const()[name = tensor("denom_121_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_121_cast_fp16 = rsqrt(epsilon = denom_121_epsilon_0_to_fp16, x = var_4507_cast_fp16)[name = tensor("denom_121_cast_fp16")]; + tensor out_121_cast_fp16 = mul(x = zero_mean_121_cast_fp16, y = denom_121_cast_fp16)[name = tensor("out_121_cast_fp16")]; + tensor obj_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(1183403840)))]; + 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(1183406464)))]; + 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 var_4522 = const()[name = tensor("op_4522"), val = tensor([1, 1])]; + tensor var_4524 = const()[name = tensor("op_4524"), val = tensor([1, 1])]; + tensor query_81_pad_type_0 = const()[name = tensor("query_81_pad_type_0"), val = tensor("custom")]; + tensor query_81_pad_0 = const()[name = tensor("query_81_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1183409088)))]; + 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(1186685952)))]; + tensor query_81_cast_fp16 = conv(bias = layers_20_self_attn_q_proj_bias_to_fp16, dilations = var_4524, groups = var_4487, pad = query_81_pad_0, pad_type = query_81_pad_type_0, strides = var_4522, weight = layers_20_self_attn_q_proj_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor("query_81_cast_fp16")]; + tensor var_4528 = const()[name = tensor("op_4528"), val = tensor([1, 1])]; + tensor var_4530 = const()[name = tensor("op_4530"), val = tensor([1, 1])]; + tensor current_key_41_pad_type_0 = const()[name = tensor("current_key_41_pad_type_0"), val = tensor("custom")]; + tensor current_key_41_pad_0 = const()[name = tensor("current_key_41_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1186688576)))]; + tensor current_key_41_cast_fp16 = conv(dilations = var_4530, groups = var_4487, pad = current_key_41_pad_0, pad_type = current_key_41_pad_type_0, strides = var_4528, weight = layers_20_self_attn_k_proj_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor("current_key_41_cast_fp16")]; + tensor var_4535 = const()[name = tensor("op_4535"), val = tensor([1, 1])]; + tensor var_4537 = const()[name = tensor("op_4537"), val = tensor([1, 1])]; + tensor current_value_41_pad_type_0 = const()[name = tensor("current_value_41_pad_type_0"), val = tensor("custom")]; + tensor current_value_41_pad_0 = const()[name = tensor("current_value_41_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1189965440)))]; + 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(1193242304)))]; + tensor current_value_41_cast_fp16 = conv(bias = layers_20_self_attn_v_proj_bias_to_fp16, dilations = var_4537, groups = var_4487, pad = current_value_41_pad_0, pad_type = current_value_41_pad_type_0, strides = var_4535, weight = layers_20_self_attn_v_proj_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor("current_value_41_cast_fp16")]; + tensor var_4544_cast_fp16 = mul(x = current_key_41_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4544_cast_fp16")]; + tensor var_4546_cast_fp16 = mul(x = var_103_cast_fp16_20, y = var_241_cast_fp16)[name = tensor("op_4546_cast_fp16")]; + tensor key_81_cast_fp16 = add(x = var_4544_cast_fp16, y = var_4546_cast_fp16)[name = tensor("key_81_cast_fp16")]; + tensor var_4548_cast_fp16 = mul(x = current_value_41_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4548_cast_fp16")]; + tensor var_4550_cast_fp16 = mul(x = var_138_cast_fp16_20, y = var_241_cast_fp16)[name = tensor("op_4550_cast_fp16")]; + tensor value_81_cast_fp16 = add(x = var_4548_cast_fp16, y = var_4550_cast_fp16)[name = tensor("value_81_cast_fp16")]; + tensor var_4553 = const()[name = tensor("op_4553"), val = tensor([1, 20, 64, -1])]; + tensor var_4554_cast_fp16 = reshape(shape = var_4553, x = query_81_cast_fp16)[name = tensor("op_4554_cast_fp16")]; + tensor var_4555_to_fp16 = const()[name = tensor("op_4555_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4556_cast_fp16 = mul(x = var_4554_cast_fp16, y = var_4555_to_fp16)[name = tensor("op_4556_cast_fp16")]; + tensor var_4557 = const()[name = tensor("op_4557"), val = tensor([1, 20, 64, -1])]; + tensor var_4558_cast_fp16 = reshape(shape = var_4557, x = key_81_cast_fp16)[name = tensor("op_4558_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_4556_cast_fp16, y = var_4558_cast_fp16)[name = tensor("mh_w_121_cast_fp16")]; + tensor mh_w_123_cast_fp16 = add(x = mh_w_121_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_123_cast_fp16")]; + tensor var_4566_cast_fp16 = softmax(axis = var_4480, x = mh_w_123_cast_fp16)[name = tensor("op_4566_cast_fp16")]; + tensor var_4567 = const()[name = tensor("op_4567"), val = tensor([1, 20, 64, -1])]; + tensor var_4568_cast_fp16 = reshape(shape = var_4567, x = value_81_cast_fp16)[name = tensor("op_4568_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_4568_cast_fp16, y = var_4566_cast_fp16)[name = tensor("attn_81_cast_fp16")]; + tensor var_4571 = const()[name = tensor("op_4571"), val = tensor([1, 1280, 1, -1])]; + tensor input_201_cast_fp16 = reshape(shape = var_4571, x = attn_81_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor var_4575 = const()[name = tensor("op_4575"), val = tensor([1, 1])]; + tensor var_4577 = const()[name = tensor("op_4577"), val = tensor([1, 1])]; + tensor obj_287_pad_type_0 = const()[name = tensor("obj_287_pad_type_0"), val = tensor("custom")]; + tensor obj_287_pad_0 = const()[name = tensor("obj_287_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1193244928)))]; + 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(1196521792)))]; + tensor obj_287_cast_fp16 = conv(bias = layers_20_self_attn_o_proj_bias_to_fp16, dilations = var_4577, groups = var_4487, pad = obj_287_pad_0, pad_type = obj_287_pad_type_0, strides = var_4575, 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 var_4587 = const()[name = tensor("op_4587"), val = tensor([1])]; + tensor channels_mean_123_cast_fp16 = reduce_mean(axes = var_4587, keep_dims = var_4488, x = inputs_123_cast_fp16)[name = tensor("channels_mean_123_cast_fp16")]; + tensor zero_mean_123_cast_fp16 = sub(x = inputs_123_cast_fp16, y = channels_mean_123_cast_fp16)[name = tensor("zero_mean_123_cast_fp16")]; + tensor zero_mean_sq_123_cast_fp16 = mul(x = zero_mean_123_cast_fp16, y = zero_mean_123_cast_fp16)[name = tensor("zero_mean_sq_123_cast_fp16")]; + tensor var_4591 = const()[name = tensor("op_4591"), val = tensor([1])]; + tensor var_4592_cast_fp16 = reduce_mean(axes = var_4591, keep_dims = var_4488, x = zero_mean_sq_123_cast_fp16)[name = tensor("op_4592_cast_fp16")]; + tensor var_4593_to_fp16 = const()[name = tensor("op_4593_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4594_cast_fp16 = add(x = var_4592_cast_fp16, y = var_4593_to_fp16)[name = tensor("op_4594_cast_fp16")]; + tensor denom_123_epsilon_0_to_fp16 = const()[name = tensor("denom_123_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_123_cast_fp16 = rsqrt(epsilon = denom_123_epsilon_0_to_fp16, x = var_4594_cast_fp16)[name = tensor("denom_123_cast_fp16")]; + tensor out_123_cast_fp16 = mul(x = zero_mean_123_cast_fp16, y = denom_123_cast_fp16)[name = tensor("out_123_cast_fp16")]; + tensor obj_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(1196524416)))]; + 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(1196527040)))]; + 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 var_4609 = const()[name = tensor("op_4609"), val = tensor([1, 1])]; + tensor var_4611 = const()[name = tensor("op_4611"), val = tensor([1, 1])]; + tensor query_83_pad_type_0 = const()[name = tensor("query_83_pad_type_0"), val = tensor("custom")]; + tensor query_83_pad_0 = const()[name = tensor("query_83_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1196529664)))]; + 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(1199806528)))]; + tensor query_83_cast_fp16 = conv(bias = layers_20_encoder_attn_q_proj_bias_to_fp16, dilations = var_4611, groups = var_4487, pad = query_83_pad_0, pad_type = query_83_pad_type_0, strides = var_4609, weight = layers_20_encoder_attn_q_proj_weight_to_fp16, x = obj_289_cast_fp16)[name = tensor("query_83_cast_fp16")]; + tensor var_4615 = const()[name = tensor("op_4615"), val = tensor([1, 1])]; + tensor var_4617 = const()[name = tensor("op_4617"), val = tensor([1, 1])]; + tensor key_83_pad_type_0 = const()[name = tensor("key_83_pad_type_0"), val = tensor("custom")]; + tensor key_83_pad_0 = const()[name = tensor("key_83_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1199809152)))]; + tensor key_83_cast_fp16 = conv(dilations = var_4617, groups = var_4487, pad = key_83_pad_0, pad_type = key_83_pad_type_0, strides = var_4615, weight = layers_20_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_83_cast_fp16")]; + tensor var_4622 = const()[name = tensor("op_4622"), val = tensor([1, 1])]; + tensor var_4624 = const()[name = tensor("op_4624"), val = tensor([1, 1])]; + tensor value_83_pad_type_0 = const()[name = tensor("value_83_pad_type_0"), val = tensor("custom")]; + tensor value_83_pad_0 = const()[name = tensor("value_83_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1203086016)))]; + 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(1206362880)))]; + tensor value_83_cast_fp16 = conv(bias = layers_20_encoder_attn_v_proj_bias_to_fp16, dilations = var_4624, groups = var_4487, pad = value_83_pad_0, pad_type = value_83_pad_type_0, strides = var_4622, weight = layers_20_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_83_cast_fp16")]; + tensor var_4628 = const()[name = tensor("op_4628"), val = tensor([1, 20, 64, -1])]; + tensor var_4629_cast_fp16 = reshape(shape = var_4628, x = query_83_cast_fp16)[name = tensor("op_4629_cast_fp16")]; + tensor var_4630_to_fp16 = const()[name = tensor("op_4630_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4631_cast_fp16 = mul(x = var_4629_cast_fp16, y = var_4630_to_fp16)[name = tensor("op_4631_cast_fp16")]; + tensor var_4632 = const()[name = tensor("op_4632"), val = tensor([1, 20, 64, -1])]; + tensor var_4633_cast_fp16 = reshape(shape = var_4632, x = key_83_cast_fp16)[name = tensor("op_4633_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_4631_cast_fp16, y = var_4633_cast_fp16)[name = tensor("mh_w_125_cast_fp16")]; + tensor obj_293_cast_fp16 = softmax(axis = var_4480, x = mh_w_125_cast_fp16)[name = tensor("obj_293_cast_fp16")]; + tensor var_4637 = const()[name = tensor("op_4637"), val = tensor([1, 20, 64, -1])]; + tensor var_4638_cast_fp16 = reshape(shape = var_4637, x = value_83_cast_fp16)[name = tensor("op_4638_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_4638_cast_fp16, y = obj_293_cast_fp16)[name = tensor("attn_83_cast_fp16")]; + tensor var_4641 = const()[name = tensor("op_4641"), val = tensor([1, 1280, 1, -1])]; + tensor input_203_cast_fp16 = reshape(shape = var_4641, x = attn_83_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor var_4645 = const()[name = tensor("op_4645"), val = tensor([1, 1])]; + tensor var_4647 = const()[name = tensor("op_4647"), val = tensor([1, 1])]; + tensor obj_291_pad_type_0 = const()[name = tensor("obj_291_pad_type_0"), val = tensor("custom")]; + tensor obj_291_pad_0 = const()[name = tensor("obj_291_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1206365504)))]; + 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(1209642368)))]; + tensor obj_291_cast_fp16 = conv(bias = layers_20_encoder_attn_o_proj_bias_to_fp16, dilations = var_4647, groups = var_4487, pad = obj_291_pad_0, pad_type = obj_291_pad_type_0, strides = var_4645, 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 var_4653 = const()[name = tensor("op_4653"), val = tensor([1])]; + tensor channels_mean_125_cast_fp16 = reduce_mean(axes = var_4653, keep_dims = var_4488, x = inputs_125_cast_fp16)[name = tensor("channels_mean_125_cast_fp16")]; + tensor zero_mean_125_cast_fp16 = sub(x = inputs_125_cast_fp16, y = channels_mean_125_cast_fp16)[name = tensor("zero_mean_125_cast_fp16")]; + tensor zero_mean_sq_125_cast_fp16 = mul(x = zero_mean_125_cast_fp16, y = zero_mean_125_cast_fp16)[name = tensor("zero_mean_sq_125_cast_fp16")]; + tensor var_4657 = const()[name = tensor("op_4657"), val = tensor([1])]; + tensor var_4658_cast_fp16 = reduce_mean(axes = var_4657, keep_dims = var_4488, x = zero_mean_sq_125_cast_fp16)[name = tensor("op_4658_cast_fp16")]; + tensor var_4659_to_fp16 = const()[name = tensor("op_4659_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4660_cast_fp16 = add(x = var_4658_cast_fp16, y = var_4659_to_fp16)[name = tensor("op_4660_cast_fp16")]; + tensor denom_125_epsilon_0_to_fp16 = const()[name = tensor("denom_125_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_125_cast_fp16 = rsqrt(epsilon = denom_125_epsilon_0_to_fp16, x = var_4660_cast_fp16)[name = tensor("denom_125_cast_fp16")]; + tensor out_125_cast_fp16 = mul(x = zero_mean_125_cast_fp16, y = denom_125_cast_fp16)[name = tensor("out_125_cast_fp16")]; + tensor input_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(1209644992)))]; + 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(1209647616)))]; + 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 var_4671 = const()[name = tensor("op_4671"), val = tensor([1, 1])]; + tensor var_4673 = const()[name = tensor("op_4673"), val = tensor([1, 1])]; + tensor input_207_pad_type_0 = const()[name = tensor("input_207_pad_type_0"), val = tensor("custom")]; + tensor input_207_pad_0 = const()[name = tensor("input_207_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1209650240)))]; + 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(1222757504)))]; + tensor input_207_cast_fp16 = conv(bias = layers_20_fc1_bias_to_fp16, dilations = var_4673, groups = var_4487, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = var_4671, 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 var_4679 = const()[name = tensor("op_4679"), val = tensor([1, 1])]; + tensor var_4681 = const()[name = tensor("op_4681"), val = tensor([1, 1])]; + tensor hidden_states_43_pad_type_0 = const()[name = tensor("hidden_states_43_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_43_pad_0 = const()[name = tensor("hidden_states_43_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1222767808)))]; + 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(1235875072)))]; + tensor hidden_states_43_cast_fp16 = conv(bias = layers_20_fc2_bias_to_fp16, dilations = var_4681, groups = var_4487, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = var_4679, 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_4694 = const()[name = tensor("op_4694"), val = tensor(3)]; + tensor var_4701 = const()[name = tensor("op_4701"), val = tensor(1)]; + tensor var_4702 = const()[name = tensor("op_4702"), val = tensor(true)]; + tensor var_4714 = const()[name = tensor("op_4714"), val = tensor([1])]; + tensor channels_mean_127_cast_fp16 = reduce_mean(axes = var_4714, keep_dims = var_4702, x = inputs_127_cast_fp16)[name = tensor("channels_mean_127_cast_fp16")]; + tensor zero_mean_127_cast_fp16 = sub(x = inputs_127_cast_fp16, y = channels_mean_127_cast_fp16)[name = tensor("zero_mean_127_cast_fp16")]; + tensor zero_mean_sq_127_cast_fp16 = mul(x = zero_mean_127_cast_fp16, y = zero_mean_127_cast_fp16)[name = tensor("zero_mean_sq_127_cast_fp16")]; + tensor var_4718 = const()[name = tensor("op_4718"), val = tensor([1])]; + tensor var_4719_cast_fp16 = reduce_mean(axes = var_4718, keep_dims = var_4702, x = zero_mean_sq_127_cast_fp16)[name = tensor("op_4719_cast_fp16")]; + tensor var_4720_to_fp16 = const()[name = tensor("op_4720_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4721_cast_fp16 = add(x = var_4719_cast_fp16, y = var_4720_to_fp16)[name = tensor("op_4721_cast_fp16")]; + tensor denom_127_epsilon_0_to_fp16 = const()[name = tensor("denom_127_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_127_cast_fp16 = rsqrt(epsilon = denom_127_epsilon_0_to_fp16, x = var_4721_cast_fp16)[name = tensor("denom_127_cast_fp16")]; + tensor out_127_cast_fp16 = mul(x = zero_mean_127_cast_fp16, y = denom_127_cast_fp16)[name = tensor("out_127_cast_fp16")]; + tensor obj_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(1235877696)))]; + 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(1235880320)))]; + 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 var_4736 = const()[name = tensor("op_4736"), val = tensor([1, 1])]; + tensor var_4738 = const()[name = tensor("op_4738"), val = tensor([1, 1])]; + tensor query_85_pad_type_0 = const()[name = tensor("query_85_pad_type_0"), val = tensor("custom")]; + tensor query_85_pad_0 = const()[name = tensor("query_85_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1235882944)))]; + 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(1239159808)))]; + tensor query_85_cast_fp16 = conv(bias = layers_21_self_attn_q_proj_bias_to_fp16, dilations = var_4738, groups = var_4701, pad = query_85_pad_0, pad_type = query_85_pad_type_0, strides = var_4736, weight = layers_21_self_attn_q_proj_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor("query_85_cast_fp16")]; + tensor var_4742 = const()[name = tensor("op_4742"), val = tensor([1, 1])]; + tensor var_4744 = const()[name = tensor("op_4744"), val = tensor([1, 1])]; + tensor current_key_43_pad_type_0 = const()[name = tensor("current_key_43_pad_type_0"), val = tensor("custom")]; + tensor current_key_43_pad_0 = const()[name = tensor("current_key_43_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1239162432)))]; + tensor current_key_43_cast_fp16 = conv(dilations = var_4744, groups = var_4701, pad = current_key_43_pad_0, pad_type = current_key_43_pad_type_0, strides = var_4742, weight = layers_21_self_attn_k_proj_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor("current_key_43_cast_fp16")]; + tensor var_4749 = const()[name = tensor("op_4749"), val = tensor([1, 1])]; + tensor var_4751 = const()[name = tensor("op_4751"), val = tensor([1, 1])]; + tensor current_value_43_pad_type_0 = const()[name = tensor("current_value_43_pad_type_0"), val = tensor("custom")]; + tensor current_value_43_pad_0 = const()[name = tensor("current_value_43_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1242439296)))]; + 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(1245716160)))]; + tensor current_value_43_cast_fp16 = conv(bias = layers_21_self_attn_v_proj_bias_to_fp16, dilations = var_4751, groups = var_4701, pad = current_value_43_pad_0, pad_type = current_value_43_pad_type_0, strides = var_4749, weight = layers_21_self_attn_v_proj_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor("current_value_43_cast_fp16")]; + tensor var_4758_cast_fp16 = mul(x = current_key_43_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4758_cast_fp16")]; + tensor var_4760_cast_fp16 = mul(x = var_103_cast_fp16_21, y = var_241_cast_fp16)[name = tensor("op_4760_cast_fp16")]; + tensor key_85_cast_fp16 = add(x = var_4758_cast_fp16, y = var_4760_cast_fp16)[name = tensor("key_85_cast_fp16")]; + tensor var_4762_cast_fp16 = mul(x = current_value_43_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4762_cast_fp16")]; + tensor var_4764_cast_fp16 = mul(x = var_138_cast_fp16_21, y = var_241_cast_fp16)[name = tensor("op_4764_cast_fp16")]; + tensor value_85_cast_fp16 = add(x = var_4762_cast_fp16, y = var_4764_cast_fp16)[name = tensor("value_85_cast_fp16")]; + tensor var_4767 = const()[name = tensor("op_4767"), val = tensor([1, 20, 64, -1])]; + tensor var_4768_cast_fp16 = reshape(shape = var_4767, x = query_85_cast_fp16)[name = tensor("op_4768_cast_fp16")]; + tensor var_4769_to_fp16 = const()[name = tensor("op_4769_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4770_cast_fp16 = mul(x = var_4768_cast_fp16, y = var_4769_to_fp16)[name = tensor("op_4770_cast_fp16")]; + tensor var_4771 = const()[name = tensor("op_4771"), val = tensor([1, 20, 64, -1])]; + tensor var_4772_cast_fp16 = reshape(shape = var_4771, x = key_85_cast_fp16)[name = tensor("op_4772_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_4770_cast_fp16, y = var_4772_cast_fp16)[name = tensor("mh_w_127_cast_fp16")]; + tensor mh_w_129_cast_fp16 = add(x = mh_w_127_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_129_cast_fp16")]; + tensor var_4780_cast_fp16 = softmax(axis = var_4694, x = mh_w_129_cast_fp16)[name = tensor("op_4780_cast_fp16")]; + tensor var_4781 = const()[name = tensor("op_4781"), val = tensor([1, 20, 64, -1])]; + tensor var_4782_cast_fp16 = reshape(shape = var_4781, x = value_85_cast_fp16)[name = tensor("op_4782_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_4782_cast_fp16, y = var_4780_cast_fp16)[name = tensor("attn_85_cast_fp16")]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([1, 1280, 1, -1])]; + tensor input_211_cast_fp16 = reshape(shape = var_4785, x = attn_85_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor var_4789 = const()[name = tensor("op_4789"), val = tensor([1, 1])]; + tensor var_4791 = const()[name = tensor("op_4791"), val = tensor([1, 1])]; + tensor obj_301_pad_type_0 = const()[name = tensor("obj_301_pad_type_0"), val = tensor("custom")]; + tensor obj_301_pad_0 = const()[name = tensor("obj_301_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1245718784)))]; + 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(1248995648)))]; + tensor obj_301_cast_fp16 = conv(bias = layers_21_self_attn_o_proj_bias_to_fp16, dilations = var_4791, groups = var_4701, pad = obj_301_pad_0, pad_type = obj_301_pad_type_0, strides = var_4789, 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 var_4801 = const()[name = tensor("op_4801"), val = tensor([1])]; + tensor channels_mean_129_cast_fp16 = reduce_mean(axes = var_4801, keep_dims = var_4702, x = inputs_129_cast_fp16)[name = tensor("channels_mean_129_cast_fp16")]; + tensor zero_mean_129_cast_fp16 = sub(x = inputs_129_cast_fp16, y = channels_mean_129_cast_fp16)[name = tensor("zero_mean_129_cast_fp16")]; + tensor zero_mean_sq_129_cast_fp16 = mul(x = zero_mean_129_cast_fp16, y = zero_mean_129_cast_fp16)[name = tensor("zero_mean_sq_129_cast_fp16")]; + tensor var_4805 = const()[name = tensor("op_4805"), val = tensor([1])]; + tensor var_4806_cast_fp16 = reduce_mean(axes = var_4805, keep_dims = var_4702, x = zero_mean_sq_129_cast_fp16)[name = tensor("op_4806_cast_fp16")]; + tensor var_4807_to_fp16 = const()[name = tensor("op_4807_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4808_cast_fp16 = add(x = var_4806_cast_fp16, y = var_4807_to_fp16)[name = tensor("op_4808_cast_fp16")]; + tensor denom_129_epsilon_0_to_fp16 = const()[name = tensor("denom_129_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_129_cast_fp16 = rsqrt(epsilon = denom_129_epsilon_0_to_fp16, x = var_4808_cast_fp16)[name = tensor("denom_129_cast_fp16")]; + tensor out_129_cast_fp16 = mul(x = zero_mean_129_cast_fp16, y = denom_129_cast_fp16)[name = tensor("out_129_cast_fp16")]; + tensor obj_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(1248998272)))]; + 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(1249000896)))]; + 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 var_4823 = const()[name = tensor("op_4823"), val = tensor([1, 1])]; + tensor var_4825 = const()[name = tensor("op_4825"), val = tensor([1, 1])]; + tensor query_87_pad_type_0 = const()[name = tensor("query_87_pad_type_0"), val = tensor("custom")]; + tensor query_87_pad_0 = const()[name = tensor("query_87_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1249003520)))]; + 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(1252280384)))]; + tensor query_87_cast_fp16 = conv(bias = layers_21_encoder_attn_q_proj_bias_to_fp16, dilations = var_4825, groups = var_4701, pad = query_87_pad_0, pad_type = query_87_pad_type_0, strides = var_4823, weight = layers_21_encoder_attn_q_proj_weight_to_fp16, x = obj_303_cast_fp16)[name = tensor("query_87_cast_fp16")]; + tensor var_4829 = const()[name = tensor("op_4829"), val = tensor([1, 1])]; + tensor var_4831 = const()[name = tensor("op_4831"), val = tensor([1, 1])]; + tensor key_87_pad_type_0 = const()[name = tensor("key_87_pad_type_0"), val = tensor("custom")]; + tensor key_87_pad_0 = const()[name = tensor("key_87_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1252283008)))]; + tensor key_87_cast_fp16 = conv(dilations = var_4831, groups = var_4701, pad = key_87_pad_0, pad_type = key_87_pad_type_0, strides = var_4829, weight = layers_21_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_87_cast_fp16")]; + tensor var_4836 = const()[name = tensor("op_4836"), val = tensor([1, 1])]; + tensor var_4838 = const()[name = tensor("op_4838"), val = tensor([1, 1])]; + tensor value_87_pad_type_0 = const()[name = tensor("value_87_pad_type_0"), val = tensor("custom")]; + tensor value_87_pad_0 = const()[name = tensor("value_87_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1255559872)))]; + 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(1258836736)))]; + tensor value_87_cast_fp16 = conv(bias = layers_21_encoder_attn_v_proj_bias_to_fp16, dilations = var_4838, groups = var_4701, pad = value_87_pad_0, pad_type = value_87_pad_type_0, strides = var_4836, weight = layers_21_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_87_cast_fp16")]; + tensor var_4842 = const()[name = tensor("op_4842"), val = tensor([1, 20, 64, -1])]; + tensor var_4843_cast_fp16 = reshape(shape = var_4842, x = query_87_cast_fp16)[name = tensor("op_4843_cast_fp16")]; + tensor var_4844_to_fp16 = const()[name = tensor("op_4844_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4845_cast_fp16 = mul(x = var_4843_cast_fp16, y = var_4844_to_fp16)[name = tensor("op_4845_cast_fp16")]; + tensor var_4846 = const()[name = tensor("op_4846"), val = tensor([1, 20, 64, -1])]; + tensor var_4847_cast_fp16 = reshape(shape = var_4846, x = key_87_cast_fp16)[name = tensor("op_4847_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_4845_cast_fp16, y = var_4847_cast_fp16)[name = tensor("mh_w_131_cast_fp16")]; + tensor obj_307_cast_fp16 = softmax(axis = var_4694, x = mh_w_131_cast_fp16)[name = tensor("obj_307_cast_fp16")]; + tensor var_4851 = const()[name = tensor("op_4851"), val = tensor([1, 20, 64, -1])]; + tensor var_4852_cast_fp16 = reshape(shape = var_4851, x = value_87_cast_fp16)[name = tensor("op_4852_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_4852_cast_fp16, y = obj_307_cast_fp16)[name = tensor("attn_87_cast_fp16")]; + tensor var_4855 = const()[name = tensor("op_4855"), val = tensor([1, 1280, 1, -1])]; + tensor input_213_cast_fp16 = reshape(shape = var_4855, x = attn_87_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor var_4859 = const()[name = tensor("op_4859"), val = tensor([1, 1])]; + tensor var_4861 = const()[name = tensor("op_4861"), val = tensor([1, 1])]; + tensor obj_305_pad_type_0 = const()[name = tensor("obj_305_pad_type_0"), val = tensor("custom")]; + tensor obj_305_pad_0 = const()[name = tensor("obj_305_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1258839360)))]; + 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(1262116224)))]; + tensor obj_305_cast_fp16 = conv(bias = layers_21_encoder_attn_o_proj_bias_to_fp16, dilations = var_4861, groups = var_4701, pad = obj_305_pad_0, pad_type = obj_305_pad_type_0, strides = var_4859, 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 var_4870 = const()[name = tensor("op_4870"), val = tensor([1])]; + tensor channels_mean_131_cast_fp16 = reduce_mean(axes = var_4870, keep_dims = var_4702, x = inputs_131_cast_fp16)[name = tensor("channels_mean_131_cast_fp16")]; + tensor zero_mean_131_cast_fp16 = sub(x = inputs_131_cast_fp16, y = channels_mean_131_cast_fp16)[name = tensor("zero_mean_131_cast_fp16")]; + tensor zero_mean_sq_131_cast_fp16 = mul(x = zero_mean_131_cast_fp16, y = zero_mean_131_cast_fp16)[name = tensor("zero_mean_sq_131_cast_fp16")]; + tensor var_4874 = const()[name = tensor("op_4874"), val = tensor([1])]; + tensor var_4875_cast_fp16 = reduce_mean(axes = var_4874, keep_dims = var_4702, x = zero_mean_sq_131_cast_fp16)[name = tensor("op_4875_cast_fp16")]; + tensor var_4876_to_fp16 = const()[name = tensor("op_4876_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4877_cast_fp16 = add(x = var_4875_cast_fp16, y = var_4876_to_fp16)[name = tensor("op_4877_cast_fp16")]; + tensor denom_131_epsilon_0_to_fp16 = const()[name = tensor("denom_131_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_131_cast_fp16 = rsqrt(epsilon = denom_131_epsilon_0_to_fp16, x = var_4877_cast_fp16)[name = tensor("denom_131_cast_fp16")]; + tensor out_131_cast_fp16 = mul(x = zero_mean_131_cast_fp16, y = denom_131_cast_fp16)[name = tensor("out_131_cast_fp16")]; + tensor input_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(1262118848)))]; + 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(1262121472)))]; + 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 var_4888 = const()[name = tensor("op_4888"), val = tensor([1, 1])]; + tensor var_4890 = const()[name = tensor("op_4890"), val = tensor([1, 1])]; + tensor input_217_pad_type_0 = const()[name = tensor("input_217_pad_type_0"), val = tensor("custom")]; + tensor input_217_pad_0 = const()[name = tensor("input_217_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1262124096)))]; + 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(1275231360)))]; + tensor input_217_cast_fp16 = conv(bias = layers_21_fc1_bias_to_fp16, dilations = var_4890, groups = var_4701, pad = input_217_pad_0, pad_type = input_217_pad_type_0, strides = var_4888, 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 var_4896 = const()[name = tensor("op_4896"), val = tensor([1, 1])]; + tensor var_4898 = const()[name = tensor("op_4898"), val = tensor([1, 1])]; + tensor hidden_states_45_pad_type_0 = const()[name = tensor("hidden_states_45_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_45_pad_0 = const()[name = tensor("hidden_states_45_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1275241664)))]; + 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(1288348928)))]; + tensor hidden_states_45_cast_fp16 = conv(bias = layers_21_fc2_bias_to_fp16, dilations = var_4898, groups = var_4701, pad = hidden_states_45_pad_0, pad_type = hidden_states_45_pad_type_0, strides = var_4896, 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_4912 = const()[name = tensor("op_4912"), val = tensor(3)]; + tensor var_4919 = const()[name = tensor("op_4919"), val = tensor(1)]; + tensor var_4920 = const()[name = tensor("op_4920"), val = tensor(true)]; + tensor var_4932 = const()[name = tensor("op_4932"), val = tensor([1])]; + tensor channels_mean_133_cast_fp16 = reduce_mean(axes = var_4932, keep_dims = var_4920, x = inputs_133_cast_fp16)[name = tensor("channels_mean_133_cast_fp16")]; + tensor zero_mean_133_cast_fp16 = sub(x = inputs_133_cast_fp16, y = channels_mean_133_cast_fp16)[name = tensor("zero_mean_133_cast_fp16")]; + tensor zero_mean_sq_133_cast_fp16 = mul(x = zero_mean_133_cast_fp16, y = zero_mean_133_cast_fp16)[name = tensor("zero_mean_sq_133_cast_fp16")]; + tensor var_4936 = const()[name = tensor("op_4936"), val = tensor([1])]; + tensor var_4937_cast_fp16 = reduce_mean(axes = var_4936, keep_dims = var_4920, x = zero_mean_sq_133_cast_fp16)[name = tensor("op_4937_cast_fp16")]; + tensor var_4938_to_fp16 = const()[name = tensor("op_4938_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4939_cast_fp16 = add(x = var_4937_cast_fp16, y = var_4938_to_fp16)[name = tensor("op_4939_cast_fp16")]; + tensor denom_133_epsilon_0_to_fp16 = const()[name = tensor("denom_133_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_133_cast_fp16 = rsqrt(epsilon = denom_133_epsilon_0_to_fp16, x = var_4939_cast_fp16)[name = tensor("denom_133_cast_fp16")]; + tensor out_133_cast_fp16 = mul(x = zero_mean_133_cast_fp16, y = denom_133_cast_fp16)[name = tensor("out_133_cast_fp16")]; + tensor obj_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(1288351552)))]; + 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(1288354176)))]; + 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 var_4954 = const()[name = tensor("op_4954"), val = tensor([1, 1])]; + tensor var_4956 = const()[name = tensor("op_4956"), val = tensor([1, 1])]; + tensor query_89_pad_type_0 = const()[name = tensor("query_89_pad_type_0"), val = tensor("custom")]; + tensor query_89_pad_0 = const()[name = tensor("query_89_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1288356800)))]; + 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(1291633664)))]; + tensor query_89_cast_fp16 = conv(bias = layers_22_self_attn_q_proj_bias_to_fp16, dilations = var_4956, groups = var_4919, pad = query_89_pad_0, pad_type = query_89_pad_type_0, strides = var_4954, weight = layers_22_self_attn_q_proj_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("query_89_cast_fp16")]; + tensor var_4960 = const()[name = tensor("op_4960"), val = tensor([1, 1])]; + tensor var_4962 = const()[name = tensor("op_4962"), val = tensor([1, 1])]; + tensor current_key_45_pad_type_0 = const()[name = tensor("current_key_45_pad_type_0"), val = tensor("custom")]; + tensor current_key_45_pad_0 = const()[name = tensor("current_key_45_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1291636288)))]; + tensor current_key_45_cast_fp16 = conv(dilations = var_4962, groups = var_4919, pad = current_key_45_pad_0, pad_type = current_key_45_pad_type_0, strides = var_4960, weight = layers_22_self_attn_k_proj_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("current_key_45_cast_fp16")]; + tensor var_4967 = const()[name = tensor("op_4967"), val = tensor([1, 1])]; + tensor var_4969 = const()[name = tensor("op_4969"), val = tensor([1, 1])]; + tensor current_value_45_pad_type_0 = const()[name = tensor("current_value_45_pad_type_0"), val = tensor("custom")]; + tensor current_value_45_pad_0 = const()[name = tensor("current_value_45_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1294913152)))]; + 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(1298190016)))]; + tensor current_value_45_cast_fp16 = conv(bias = layers_22_self_attn_v_proj_bias_to_fp16, dilations = var_4969, groups = var_4919, pad = current_value_45_pad_0, pad_type = current_value_45_pad_type_0, strides = var_4967, weight = layers_22_self_attn_v_proj_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("current_value_45_cast_fp16")]; + tensor var_4976_cast_fp16 = mul(x = current_key_45_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4976_cast_fp16")]; + tensor var_4978_cast_fp16 = mul(x = var_103_cast_fp16_22, y = var_241_cast_fp16)[name = tensor("op_4978_cast_fp16")]; + tensor key_89_cast_fp16 = add(x = var_4976_cast_fp16, y = var_4978_cast_fp16)[name = tensor("key_89_cast_fp16")]; + tensor var_4980_cast_fp16 = mul(x = current_value_45_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_4980_cast_fp16")]; + tensor var_4982_cast_fp16 = mul(x = var_138_cast_fp16_22, y = var_241_cast_fp16)[name = tensor("op_4982_cast_fp16")]; + tensor value_89_cast_fp16 = add(x = var_4980_cast_fp16, y = var_4982_cast_fp16)[name = tensor("value_89_cast_fp16")]; + tensor var_4985 = const()[name = tensor("op_4985"), val = tensor([1, 20, 64, -1])]; + tensor var_4986_cast_fp16 = reshape(shape = var_4985, x = query_89_cast_fp16)[name = tensor("op_4986_cast_fp16")]; + tensor var_4987_to_fp16 = const()[name = tensor("op_4987_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4988_cast_fp16 = mul(x = var_4986_cast_fp16, y = var_4987_to_fp16)[name = tensor("op_4988_cast_fp16")]; + tensor var_4989 = const()[name = tensor("op_4989"), val = tensor([1, 20, 64, -1])]; + tensor var_4990_cast_fp16 = reshape(shape = var_4989, x = key_89_cast_fp16)[name = tensor("op_4990_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_4988_cast_fp16, y = var_4990_cast_fp16)[name = tensor("mh_w_133_cast_fp16")]; + tensor mh_w_135_cast_fp16 = add(x = mh_w_133_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_135_cast_fp16")]; + tensor var_4998_cast_fp16 = softmax(axis = var_4912, x = mh_w_135_cast_fp16)[name = tensor("op_4998_cast_fp16")]; + tensor var_4999 = const()[name = tensor("op_4999"), val = tensor([1, 20, 64, -1])]; + tensor var_5000_cast_fp16 = reshape(shape = var_4999, x = value_89_cast_fp16)[name = tensor("op_5000_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_5000_cast_fp16, y = var_4998_cast_fp16)[name = tensor("attn_89_cast_fp16")]; + tensor var_5003 = const()[name = tensor("op_5003"), val = tensor([1, 1280, 1, -1])]; + tensor input_221_cast_fp16 = reshape(shape = var_5003, x = attn_89_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor var_5007 = const()[name = tensor("op_5007"), val = tensor([1, 1])]; + tensor var_5009 = const()[name = tensor("op_5009"), val = tensor([1, 1])]; + tensor obj_315_pad_type_0 = const()[name = tensor("obj_315_pad_type_0"), val = tensor("custom")]; + tensor obj_315_pad_0 = const()[name = tensor("obj_315_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1298192640)))]; + 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(1301469504)))]; + tensor obj_315_cast_fp16 = conv(bias = layers_22_self_attn_o_proj_bias_to_fp16, dilations = var_5009, groups = var_4919, pad = obj_315_pad_0, pad_type = obj_315_pad_type_0, strides = var_5007, 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 var_5019 = const()[name = tensor("op_5019"), val = tensor([1])]; + tensor channels_mean_135_cast_fp16 = reduce_mean(axes = var_5019, keep_dims = var_4920, x = inputs_135_cast_fp16)[name = tensor("channels_mean_135_cast_fp16")]; + tensor zero_mean_135_cast_fp16 = sub(x = inputs_135_cast_fp16, y = channels_mean_135_cast_fp16)[name = tensor("zero_mean_135_cast_fp16")]; + tensor zero_mean_sq_135_cast_fp16 = mul(x = zero_mean_135_cast_fp16, y = zero_mean_135_cast_fp16)[name = tensor("zero_mean_sq_135_cast_fp16")]; + tensor var_5023 = const()[name = tensor("op_5023"), val = tensor([1])]; + tensor var_5024_cast_fp16 = reduce_mean(axes = var_5023, keep_dims = var_4920, x = zero_mean_sq_135_cast_fp16)[name = tensor("op_5024_cast_fp16")]; + tensor var_5025_to_fp16 = const()[name = tensor("op_5025_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5026_cast_fp16 = add(x = var_5024_cast_fp16, y = var_5025_to_fp16)[name = tensor("op_5026_cast_fp16")]; + tensor denom_135_epsilon_0_to_fp16 = const()[name = tensor("denom_135_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_135_cast_fp16 = rsqrt(epsilon = denom_135_epsilon_0_to_fp16, x = var_5026_cast_fp16)[name = tensor("denom_135_cast_fp16")]; + tensor out_135_cast_fp16 = mul(x = zero_mean_135_cast_fp16, y = denom_135_cast_fp16)[name = tensor("out_135_cast_fp16")]; + tensor obj_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(1301472128)))]; + 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(1301474752)))]; + 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 var_5041 = const()[name = tensor("op_5041"), val = tensor([1, 1])]; + tensor var_5043 = const()[name = tensor("op_5043"), val = tensor([1, 1])]; + tensor query_91_pad_type_0 = const()[name = tensor("query_91_pad_type_0"), val = tensor("custom")]; + tensor query_91_pad_0 = const()[name = tensor("query_91_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1301477376)))]; + 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(1304754240)))]; + tensor query_91_cast_fp16 = conv(bias = layers_22_encoder_attn_q_proj_bias_to_fp16, dilations = var_5043, groups = var_4919, pad = query_91_pad_0, pad_type = query_91_pad_type_0, strides = var_5041, weight = layers_22_encoder_attn_q_proj_weight_to_fp16, x = obj_317_cast_fp16)[name = tensor("query_91_cast_fp16")]; + tensor var_5047 = const()[name = tensor("op_5047"), val = tensor([1, 1])]; + tensor var_5049 = const()[name = tensor("op_5049"), val = tensor([1, 1])]; + tensor key_91_pad_type_0 = const()[name = tensor("key_91_pad_type_0"), val = tensor("custom")]; + tensor key_91_pad_0 = const()[name = tensor("key_91_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1304756864)))]; + tensor key_91_cast_fp16 = conv(dilations = var_5049, groups = var_4919, pad = key_91_pad_0, pad_type = key_91_pad_type_0, strides = var_5047, weight = layers_22_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_91_cast_fp16")]; + tensor var_5054 = const()[name = tensor("op_5054"), val = tensor([1, 1])]; + tensor var_5056 = const()[name = tensor("op_5056"), val = tensor([1, 1])]; + tensor value_91_pad_type_0 = const()[name = tensor("value_91_pad_type_0"), val = tensor("custom")]; + tensor value_91_pad_0 = const()[name = tensor("value_91_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1308033728)))]; + 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(1311310592)))]; + tensor value_91_cast_fp16 = conv(bias = layers_22_encoder_attn_v_proj_bias_to_fp16, dilations = var_5056, groups = var_4919, pad = value_91_pad_0, pad_type = value_91_pad_type_0, strides = var_5054, weight = layers_22_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_91_cast_fp16")]; + tensor var_5060 = const()[name = tensor("op_5060"), val = tensor([1, 20, 64, -1])]; + tensor var_5061_cast_fp16 = reshape(shape = var_5060, x = query_91_cast_fp16)[name = tensor("op_5061_cast_fp16")]; + tensor var_5062_to_fp16 = const()[name = tensor("op_5062_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5063_cast_fp16 = mul(x = var_5061_cast_fp16, y = var_5062_to_fp16)[name = tensor("op_5063_cast_fp16")]; + tensor var_5064 = const()[name = tensor("op_5064"), val = tensor([1, 20, 64, -1])]; + tensor var_5065_cast_fp16 = reshape(shape = var_5064, x = key_91_cast_fp16)[name = tensor("op_5065_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_5063_cast_fp16, y = var_5065_cast_fp16)[name = tensor("mh_w_137_cast_fp16")]; + tensor obj_321_cast_fp16 = softmax(axis = var_4912, x = mh_w_137_cast_fp16)[name = tensor("obj_321_cast_fp16")]; + tensor var_5069 = const()[name = tensor("op_5069"), val = tensor([1, 20, 64, -1])]; + tensor var_5070_cast_fp16 = reshape(shape = var_5069, x = value_91_cast_fp16)[name = tensor("op_5070_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_5070_cast_fp16, y = obj_321_cast_fp16)[name = tensor("attn_91_cast_fp16")]; + tensor var_5073 = const()[name = tensor("op_5073"), val = tensor([1, 1280, 1, -1])]; + tensor input_223_cast_fp16 = reshape(shape = var_5073, x = attn_91_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor var_5077 = const()[name = tensor("op_5077"), val = tensor([1, 1])]; + tensor var_5079 = const()[name = tensor("op_5079"), val = tensor([1, 1])]; + tensor obj_319_pad_type_0 = const()[name = tensor("obj_319_pad_type_0"), val = tensor("custom")]; + tensor obj_319_pad_0 = const()[name = tensor("obj_319_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1311313216)))]; + 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(1314590080)))]; + tensor obj_319_cast_fp16 = conv(bias = layers_22_encoder_attn_o_proj_bias_to_fp16, dilations = var_5079, groups = var_4919, pad = obj_319_pad_0, pad_type = obj_319_pad_type_0, strides = var_5077, 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 var_5088 = const()[name = tensor("op_5088"), val = tensor([1])]; + tensor channels_mean_137_cast_fp16 = reduce_mean(axes = var_5088, keep_dims = var_4920, x = inputs_137_cast_fp16)[name = tensor("channels_mean_137_cast_fp16")]; + tensor zero_mean_137_cast_fp16 = sub(x = inputs_137_cast_fp16, y = channels_mean_137_cast_fp16)[name = tensor("zero_mean_137_cast_fp16")]; + tensor zero_mean_sq_137_cast_fp16 = mul(x = zero_mean_137_cast_fp16, y = zero_mean_137_cast_fp16)[name = tensor("zero_mean_sq_137_cast_fp16")]; + tensor var_5092 = const()[name = tensor("op_5092"), val = tensor([1])]; + tensor var_5093_cast_fp16 = reduce_mean(axes = var_5092, keep_dims = var_4920, x = zero_mean_sq_137_cast_fp16)[name = tensor("op_5093_cast_fp16")]; + tensor var_5094_to_fp16 = const()[name = tensor("op_5094_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5095_cast_fp16 = add(x = var_5093_cast_fp16, y = var_5094_to_fp16)[name = tensor("op_5095_cast_fp16")]; + tensor denom_137_epsilon_0_to_fp16 = const()[name = tensor("denom_137_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_137_cast_fp16 = rsqrt(epsilon = denom_137_epsilon_0_to_fp16, x = var_5095_cast_fp16)[name = tensor("denom_137_cast_fp16")]; + tensor out_137_cast_fp16 = mul(x = zero_mean_137_cast_fp16, y = denom_137_cast_fp16)[name = tensor("out_137_cast_fp16")]; + tensor input_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(1314592704)))]; + 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(1314595328)))]; + 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 var_5106 = const()[name = tensor("op_5106"), val = tensor([1, 1])]; + tensor var_5108 = const()[name = tensor("op_5108"), val = tensor([1, 1])]; + tensor input_227_pad_type_0 = const()[name = tensor("input_227_pad_type_0"), val = tensor("custom")]; + tensor input_227_pad_0 = const()[name = tensor("input_227_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1314597952)))]; + 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(1327705216)))]; + tensor input_227_cast_fp16 = conv(bias = layers_22_fc1_bias_to_fp16, dilations = var_5108, groups = var_4919, pad = input_227_pad_0, pad_type = input_227_pad_type_0, strides = var_5106, 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 var_5114 = const()[name = tensor("op_5114"), val = tensor([1, 1])]; + tensor var_5116 = const()[name = tensor("op_5116"), val = tensor([1, 1])]; + tensor hidden_states_47_pad_type_0 = const()[name = tensor("hidden_states_47_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_47_pad_0 = const()[name = tensor("hidden_states_47_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1327715520)))]; + 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(1340822784)))]; + tensor hidden_states_47_cast_fp16 = conv(bias = layers_22_fc2_bias_to_fp16, dilations = var_5116, groups = var_4919, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = var_5114, 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_5130 = const()[name = tensor("op_5130"), val = tensor(3)]; + tensor var_5137 = const()[name = tensor("op_5137"), val = tensor(1)]; + tensor var_5138 = const()[name = tensor("op_5138"), val = tensor(true)]; + tensor var_5150 = const()[name = tensor("op_5150"), val = tensor([1])]; + tensor channels_mean_139_cast_fp16 = reduce_mean(axes = var_5150, keep_dims = var_5138, x = inputs_139_cast_fp16)[name = tensor("channels_mean_139_cast_fp16")]; + tensor zero_mean_139_cast_fp16 = sub(x = inputs_139_cast_fp16, y = channels_mean_139_cast_fp16)[name = tensor("zero_mean_139_cast_fp16")]; + tensor zero_mean_sq_139_cast_fp16 = mul(x = zero_mean_139_cast_fp16, y = zero_mean_139_cast_fp16)[name = tensor("zero_mean_sq_139_cast_fp16")]; + tensor var_5154 = const()[name = tensor("op_5154"), val = tensor([1])]; + tensor var_5155_cast_fp16 = reduce_mean(axes = var_5154, keep_dims = var_5138, x = zero_mean_sq_139_cast_fp16)[name = tensor("op_5155_cast_fp16")]; + tensor var_5156_to_fp16 = const()[name = tensor("op_5156_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5157_cast_fp16 = add(x = var_5155_cast_fp16, y = var_5156_to_fp16)[name = tensor("op_5157_cast_fp16")]; + tensor denom_139_epsilon_0_to_fp16 = const()[name = tensor("denom_139_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_139_cast_fp16 = rsqrt(epsilon = denom_139_epsilon_0_to_fp16, x = var_5157_cast_fp16)[name = tensor("denom_139_cast_fp16")]; + tensor out_139_cast_fp16 = mul(x = zero_mean_139_cast_fp16, y = denom_139_cast_fp16)[name = tensor("out_139_cast_fp16")]; + tensor obj_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(1340825408)))]; + 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(1340828032)))]; + 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 var_5172 = const()[name = tensor("op_5172"), val = tensor([1, 1])]; + tensor var_5174 = const()[name = tensor("op_5174"), val = tensor([1, 1])]; + tensor query_93_pad_type_0 = const()[name = tensor("query_93_pad_type_0"), val = tensor("custom")]; + tensor query_93_pad_0 = const()[name = tensor("query_93_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1340830656)))]; + 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(1344107520)))]; + tensor query_93_cast_fp16 = conv(bias = layers_23_self_attn_q_proj_bias_to_fp16, dilations = var_5174, groups = var_5137, pad = query_93_pad_0, pad_type = query_93_pad_type_0, strides = var_5172, weight = layers_23_self_attn_q_proj_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor("query_93_cast_fp16")]; + tensor var_5178 = const()[name = tensor("op_5178"), val = tensor([1, 1])]; + tensor var_5180 = const()[name = tensor("op_5180"), val = tensor([1, 1])]; + tensor current_key_47_pad_type_0 = const()[name = tensor("current_key_47_pad_type_0"), val = tensor("custom")]; + tensor current_key_47_pad_0 = const()[name = tensor("current_key_47_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1344110144)))]; + tensor current_key_47_cast_fp16 = conv(dilations = var_5180, groups = var_5137, pad = current_key_47_pad_0, pad_type = current_key_47_pad_type_0, strides = var_5178, weight = layers_23_self_attn_k_proj_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor("current_key_47_cast_fp16")]; + tensor var_5185 = const()[name = tensor("op_5185"), val = tensor([1, 1])]; + tensor var_5187 = const()[name = tensor("op_5187"), val = tensor([1, 1])]; + tensor current_value_47_pad_type_0 = const()[name = tensor("current_value_47_pad_type_0"), val = tensor("custom")]; + tensor current_value_47_pad_0 = const()[name = tensor("current_value_47_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1347387008)))]; + 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(1350663872)))]; + tensor current_value_47_cast_fp16 = conv(bias = layers_23_self_attn_v_proj_bias_to_fp16, dilations = var_5187, groups = var_5137, pad = current_value_47_pad_0, pad_type = current_value_47_pad_type_0, strides = var_5185, weight = layers_23_self_attn_v_proj_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor("current_value_47_cast_fp16")]; + tensor var_5194_cast_fp16 = mul(x = current_key_47_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_5194_cast_fp16")]; + tensor var_5196_cast_fp16 = mul(x = var_103_cast_fp16_23, y = var_241_cast_fp16)[name = tensor("op_5196_cast_fp16")]; + tensor key_93_cast_fp16 = add(x = var_5194_cast_fp16, y = var_5196_cast_fp16)[name = tensor("key_93_cast_fp16")]; + tensor var_5198_cast_fp16 = mul(x = current_value_47_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_5198_cast_fp16")]; + tensor var_5200_cast_fp16 = mul(x = var_138_cast_fp16_23, y = var_241_cast_fp16)[name = tensor("op_5200_cast_fp16")]; + tensor value_93_cast_fp16 = add(x = var_5198_cast_fp16, y = var_5200_cast_fp16)[name = tensor("value_93_cast_fp16")]; + tensor var_5203 = const()[name = tensor("op_5203"), val = tensor([1, 20, 64, -1])]; + tensor var_5204_cast_fp16 = reshape(shape = var_5203, x = query_93_cast_fp16)[name = tensor("op_5204_cast_fp16")]; + tensor var_5205_to_fp16 = const()[name = tensor("op_5205_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5206_cast_fp16 = mul(x = var_5204_cast_fp16, y = var_5205_to_fp16)[name = tensor("op_5206_cast_fp16")]; + tensor var_5207 = const()[name = tensor("op_5207"), val = tensor([1, 20, 64, -1])]; + tensor var_5208_cast_fp16 = reshape(shape = var_5207, x = key_93_cast_fp16)[name = tensor("op_5208_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_5206_cast_fp16, y = var_5208_cast_fp16)[name = tensor("mh_w_139_cast_fp16")]; + tensor mh_w_141_cast_fp16 = add(x = mh_w_139_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_141_cast_fp16")]; + tensor var_5216_cast_fp16 = softmax(axis = var_5130, x = mh_w_141_cast_fp16)[name = tensor("op_5216_cast_fp16")]; + tensor var_5217 = const()[name = tensor("op_5217"), val = tensor([1, 20, 64, -1])]; + tensor var_5218_cast_fp16 = reshape(shape = var_5217, x = value_93_cast_fp16)[name = tensor("op_5218_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_5218_cast_fp16, y = var_5216_cast_fp16)[name = tensor("attn_93_cast_fp16")]; + tensor var_5221 = const()[name = tensor("op_5221"), val = tensor([1, 1280, 1, -1])]; + tensor input_231_cast_fp16 = reshape(shape = var_5221, x = attn_93_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor var_5225 = const()[name = tensor("op_5225"), val = tensor([1, 1])]; + tensor var_5227 = const()[name = tensor("op_5227"), val = tensor([1, 1])]; + tensor obj_329_pad_type_0 = const()[name = tensor("obj_329_pad_type_0"), val = tensor("custom")]; + tensor obj_329_pad_0 = const()[name = tensor("obj_329_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1350666496)))]; + 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(1353943360)))]; + tensor obj_329_cast_fp16 = conv(bias = layers_23_self_attn_o_proj_bias_to_fp16, dilations = var_5227, groups = var_5137, pad = obj_329_pad_0, pad_type = obj_329_pad_type_0, strides = var_5225, 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 var_5237 = const()[name = tensor("op_5237"), val = tensor([1])]; + tensor channels_mean_141_cast_fp16 = reduce_mean(axes = var_5237, keep_dims = var_5138, x = inputs_141_cast_fp16)[name = tensor("channels_mean_141_cast_fp16")]; + tensor zero_mean_141_cast_fp16 = sub(x = inputs_141_cast_fp16, y = channels_mean_141_cast_fp16)[name = tensor("zero_mean_141_cast_fp16")]; + tensor zero_mean_sq_141_cast_fp16 = mul(x = zero_mean_141_cast_fp16, y = zero_mean_141_cast_fp16)[name = tensor("zero_mean_sq_141_cast_fp16")]; + tensor var_5241 = const()[name = tensor("op_5241"), val = tensor([1])]; + tensor var_5242_cast_fp16 = reduce_mean(axes = var_5241, keep_dims = var_5138, x = zero_mean_sq_141_cast_fp16)[name = tensor("op_5242_cast_fp16")]; + tensor var_5243_to_fp16 = const()[name = tensor("op_5243_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5244_cast_fp16 = add(x = var_5242_cast_fp16, y = var_5243_to_fp16)[name = tensor("op_5244_cast_fp16")]; + tensor denom_141_epsilon_0_to_fp16 = const()[name = tensor("denom_141_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_141_cast_fp16 = rsqrt(epsilon = denom_141_epsilon_0_to_fp16, x = var_5244_cast_fp16)[name = tensor("denom_141_cast_fp16")]; + tensor out_141_cast_fp16 = mul(x = zero_mean_141_cast_fp16, y = denom_141_cast_fp16)[name = tensor("out_141_cast_fp16")]; + tensor obj_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(1353945984)))]; + 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(1353948608)))]; + 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 var_5259 = const()[name = tensor("op_5259"), val = tensor([1, 1])]; + tensor var_5261 = const()[name = tensor("op_5261"), val = tensor([1, 1])]; + tensor query_95_pad_type_0 = const()[name = tensor("query_95_pad_type_0"), val = tensor("custom")]; + tensor query_95_pad_0 = const()[name = tensor("query_95_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1353951232)))]; + 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(1357228096)))]; + tensor query_95_cast_fp16 = conv(bias = layers_23_encoder_attn_q_proj_bias_to_fp16, dilations = var_5261, groups = var_5137, pad = query_95_pad_0, pad_type = query_95_pad_type_0, strides = var_5259, weight = layers_23_encoder_attn_q_proj_weight_to_fp16, x = obj_331_cast_fp16)[name = tensor("query_95_cast_fp16")]; + tensor var_5265 = const()[name = tensor("op_5265"), val = tensor([1, 1])]; + tensor var_5267 = const()[name = tensor("op_5267"), val = tensor([1, 1])]; + tensor key_95_pad_type_0 = const()[name = tensor("key_95_pad_type_0"), val = tensor("custom")]; + tensor key_95_pad_0 = const()[name = tensor("key_95_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1357230720)))]; + tensor key_95_cast_fp16 = conv(dilations = var_5267, groups = var_5137, pad = key_95_pad_0, pad_type = key_95_pad_type_0, strides = var_5265, weight = layers_23_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_95_cast_fp16")]; + tensor var_5272 = const()[name = tensor("op_5272"), val = tensor([1, 1])]; + tensor var_5274 = const()[name = tensor("op_5274"), val = tensor([1, 1])]; + tensor value_95_pad_type_0 = const()[name = tensor("value_95_pad_type_0"), val = tensor("custom")]; + tensor value_95_pad_0 = const()[name = tensor("value_95_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1360507584)))]; + 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(1363784448)))]; + tensor value_95_cast_fp16 = conv(bias = layers_23_encoder_attn_v_proj_bias_to_fp16, dilations = var_5274, groups = var_5137, pad = value_95_pad_0, pad_type = value_95_pad_type_0, strides = var_5272, weight = layers_23_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_95_cast_fp16")]; + tensor var_5278 = const()[name = tensor("op_5278"), val = tensor([1, 20, 64, -1])]; + tensor var_5279_cast_fp16 = reshape(shape = var_5278, x = query_95_cast_fp16)[name = tensor("op_5279_cast_fp16")]; + tensor var_5280_to_fp16 = const()[name = tensor("op_5280_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5281_cast_fp16 = mul(x = var_5279_cast_fp16, y = var_5280_to_fp16)[name = tensor("op_5281_cast_fp16")]; + tensor var_5282 = const()[name = tensor("op_5282"), val = tensor([1, 20, 64, -1])]; + tensor var_5283_cast_fp16 = reshape(shape = var_5282, x = key_95_cast_fp16)[name = tensor("op_5283_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_5281_cast_fp16, y = var_5283_cast_fp16)[name = tensor("mh_w_143_cast_fp16")]; + tensor obj_335_cast_fp16 = softmax(axis = var_5130, x = mh_w_143_cast_fp16)[name = tensor("obj_335_cast_fp16")]; + tensor var_5287 = const()[name = tensor("op_5287"), val = tensor([1, 20, 64, -1])]; + tensor var_5288_cast_fp16 = reshape(shape = var_5287, x = value_95_cast_fp16)[name = tensor("op_5288_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_5288_cast_fp16, y = obj_335_cast_fp16)[name = tensor("attn_95_cast_fp16")]; + tensor var_5291 = const()[name = tensor("op_5291"), val = tensor([1, 1280, 1, -1])]; + tensor input_233_cast_fp16 = reshape(shape = var_5291, x = attn_95_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor var_5295 = const()[name = tensor("op_5295"), val = tensor([1, 1])]; + tensor var_5297 = const()[name = tensor("op_5297"), val = tensor([1, 1])]; + tensor obj_333_pad_type_0 = const()[name = tensor("obj_333_pad_type_0"), val = tensor("custom")]; + tensor obj_333_pad_0 = const()[name = tensor("obj_333_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1363787072)))]; + 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(1367063936)))]; + tensor obj_333_cast_fp16 = conv(bias = layers_23_encoder_attn_o_proj_bias_to_fp16, dilations = var_5297, groups = var_5137, pad = obj_333_pad_0, pad_type = obj_333_pad_type_0, strides = var_5295, 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 var_5306 = const()[name = tensor("op_5306"), val = tensor([1])]; + tensor channels_mean_143_cast_fp16 = reduce_mean(axes = var_5306, keep_dims = var_5138, x = inputs_143_cast_fp16)[name = tensor("channels_mean_143_cast_fp16")]; + tensor zero_mean_143_cast_fp16 = sub(x = inputs_143_cast_fp16, y = channels_mean_143_cast_fp16)[name = tensor("zero_mean_143_cast_fp16")]; + tensor zero_mean_sq_143_cast_fp16 = mul(x = zero_mean_143_cast_fp16, y = zero_mean_143_cast_fp16)[name = tensor("zero_mean_sq_143_cast_fp16")]; + tensor var_5310 = const()[name = tensor("op_5310"), val = tensor([1])]; + tensor var_5311_cast_fp16 = reduce_mean(axes = var_5310, keep_dims = var_5138, x = zero_mean_sq_143_cast_fp16)[name = tensor("op_5311_cast_fp16")]; + tensor var_5312_to_fp16 = const()[name = tensor("op_5312_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5313_cast_fp16 = add(x = var_5311_cast_fp16, y = var_5312_to_fp16)[name = tensor("op_5313_cast_fp16")]; + tensor denom_143_epsilon_0_to_fp16 = const()[name = tensor("denom_143_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_143_cast_fp16 = rsqrt(epsilon = denom_143_epsilon_0_to_fp16, x = var_5313_cast_fp16)[name = tensor("denom_143_cast_fp16")]; + tensor out_143_cast_fp16 = mul(x = zero_mean_143_cast_fp16, y = denom_143_cast_fp16)[name = tensor("out_143_cast_fp16")]; + tensor input_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(1367066560)))]; + 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(1367069184)))]; + 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 var_5324 = const()[name = tensor("op_5324"), val = tensor([1, 1])]; + tensor var_5326 = const()[name = tensor("op_5326"), val = tensor([1, 1])]; + tensor input_237_pad_type_0 = const()[name = tensor("input_237_pad_type_0"), val = tensor("custom")]; + tensor input_237_pad_0 = const()[name = tensor("input_237_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1367071808)))]; + 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(1380179072)))]; + tensor input_237_cast_fp16 = conv(bias = layers_23_fc1_bias_to_fp16, dilations = var_5326, groups = var_5137, pad = input_237_pad_0, pad_type = input_237_pad_type_0, strides = var_5324, 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 var_5332 = const()[name = tensor("op_5332"), val = tensor([1, 1])]; + tensor var_5334 = const()[name = tensor("op_5334"), val = tensor([1, 1])]; + tensor hidden_states_49_pad_type_0 = const()[name = tensor("hidden_states_49_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_49_pad_0 = const()[name = tensor("hidden_states_49_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1380189376)))]; + 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(1393296640)))]; + tensor hidden_states_49_cast_fp16 = conv(bias = layers_23_fc2_bias_to_fp16, dilations = var_5334, groups = var_5137, pad = hidden_states_49_pad_0, pad_type = hidden_states_49_pad_type_0, strides = var_5332, 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_5348 = const()[name = tensor("op_5348"), val = tensor(3)]; + tensor var_5355 = const()[name = tensor("op_5355"), val = tensor(1)]; + tensor var_5356 = const()[name = tensor("op_5356"), val = tensor(true)]; + tensor var_5368 = const()[name = tensor("op_5368"), val = tensor([1])]; + tensor channels_mean_145_cast_fp16 = reduce_mean(axes = var_5368, keep_dims = var_5356, x = inputs_145_cast_fp16)[name = tensor("channels_mean_145_cast_fp16")]; + tensor zero_mean_145_cast_fp16 = sub(x = inputs_145_cast_fp16, y = channels_mean_145_cast_fp16)[name = tensor("zero_mean_145_cast_fp16")]; + tensor zero_mean_sq_145_cast_fp16 = mul(x = zero_mean_145_cast_fp16, y = zero_mean_145_cast_fp16)[name = tensor("zero_mean_sq_145_cast_fp16")]; + tensor var_5372 = const()[name = tensor("op_5372"), val = tensor([1])]; + tensor var_5373_cast_fp16 = reduce_mean(axes = var_5372, keep_dims = var_5356, x = zero_mean_sq_145_cast_fp16)[name = tensor("op_5373_cast_fp16")]; + tensor var_5374_to_fp16 = const()[name = tensor("op_5374_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5375_cast_fp16 = add(x = var_5373_cast_fp16, y = var_5374_to_fp16)[name = tensor("op_5375_cast_fp16")]; + tensor denom_145_epsilon_0_to_fp16 = const()[name = tensor("denom_145_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_145_cast_fp16 = rsqrt(epsilon = denom_145_epsilon_0_to_fp16, x = var_5375_cast_fp16)[name = tensor("denom_145_cast_fp16")]; + tensor out_145_cast_fp16 = mul(x = zero_mean_145_cast_fp16, y = denom_145_cast_fp16)[name = tensor("out_145_cast_fp16")]; + tensor obj_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(1393299264)))]; + 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(1393301888)))]; + 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 var_5390 = const()[name = tensor("op_5390"), val = tensor([1, 1])]; + tensor var_5392 = const()[name = tensor("op_5392"), val = tensor([1, 1])]; + tensor query_97_pad_type_0 = const()[name = tensor("query_97_pad_type_0"), val = tensor("custom")]; + tensor query_97_pad_0 = const()[name = tensor("query_97_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1393304512)))]; + 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(1396581376)))]; + tensor query_97_cast_fp16 = conv(bias = layers_24_self_attn_q_proj_bias_to_fp16, dilations = var_5392, groups = var_5355, pad = query_97_pad_0, pad_type = query_97_pad_type_0, strides = var_5390, weight = layers_24_self_attn_q_proj_weight_to_fp16, x = obj_337_cast_fp16)[name = tensor("query_97_cast_fp16")]; + tensor var_5396 = const()[name = tensor("op_5396"), val = tensor([1, 1])]; + tensor var_5398 = const()[name = tensor("op_5398"), val = tensor([1, 1])]; + tensor current_key_49_pad_type_0 = const()[name = tensor("current_key_49_pad_type_0"), val = tensor("custom")]; + tensor current_key_49_pad_0 = const()[name = tensor("current_key_49_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1396584000)))]; + tensor current_key_49_cast_fp16 = conv(dilations = var_5398, groups = var_5355, pad = current_key_49_pad_0, pad_type = current_key_49_pad_type_0, strides = var_5396, weight = layers_24_self_attn_k_proj_weight_to_fp16, x = obj_337_cast_fp16)[name = tensor("current_key_49_cast_fp16")]; + tensor var_5403 = const()[name = tensor("op_5403"), val = tensor([1, 1])]; + tensor var_5405 = const()[name = tensor("op_5405"), val = tensor([1, 1])]; + tensor current_value_49_pad_type_0 = const()[name = tensor("current_value_49_pad_type_0"), val = tensor("custom")]; + tensor current_value_49_pad_0 = const()[name = tensor("current_value_49_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1399860864)))]; + 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(1403137728)))]; + tensor current_value_49_cast_fp16 = conv(bias = layers_24_self_attn_v_proj_bias_to_fp16, dilations = var_5405, groups = var_5355, pad = current_value_49_pad_0, pad_type = current_value_49_pad_type_0, strides = var_5403, weight = layers_24_self_attn_v_proj_weight_to_fp16, x = obj_337_cast_fp16)[name = tensor("current_value_49_cast_fp16")]; + tensor var_5412_cast_fp16 = mul(x = current_key_49_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_5412_cast_fp16")]; + tensor var_5414_cast_fp16 = mul(x = var_103_cast_fp16_24, y = var_241_cast_fp16)[name = tensor("op_5414_cast_fp16")]; + tensor key_97_cast_fp16 = add(x = var_5412_cast_fp16, y = var_5414_cast_fp16)[name = tensor("key_97_cast_fp16")]; + tensor var_5416_cast_fp16 = mul(x = current_value_49_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_5416_cast_fp16")]; + tensor var_5418_cast_fp16 = mul(x = var_138_cast_fp16_24, y = var_241_cast_fp16)[name = tensor("op_5418_cast_fp16")]; + tensor value_97_cast_fp16 = add(x = var_5416_cast_fp16, y = var_5418_cast_fp16)[name = tensor("value_97_cast_fp16")]; + tensor var_5421 = const()[name = tensor("op_5421"), val = tensor([1, 20, 64, -1])]; + tensor var_5422_cast_fp16 = reshape(shape = var_5421, x = query_97_cast_fp16)[name = tensor("op_5422_cast_fp16")]; + tensor var_5423_to_fp16 = const()[name = tensor("op_5423_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5424_cast_fp16 = mul(x = var_5422_cast_fp16, y = var_5423_to_fp16)[name = tensor("op_5424_cast_fp16")]; + tensor var_5425 = const()[name = tensor("op_5425"), val = tensor([1, 20, 64, -1])]; + tensor var_5426_cast_fp16 = reshape(shape = var_5425, x = key_97_cast_fp16)[name = tensor("op_5426_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_5424_cast_fp16, y = var_5426_cast_fp16)[name = tensor("mh_w_145_cast_fp16")]; + tensor mh_w_147_cast_fp16 = add(x = mh_w_145_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_147_cast_fp16")]; + tensor var_5434_cast_fp16 = softmax(axis = var_5348, x = mh_w_147_cast_fp16)[name = tensor("op_5434_cast_fp16")]; + tensor var_5435 = const()[name = tensor("op_5435"), val = tensor([1, 20, 64, -1])]; + tensor var_5436_cast_fp16 = reshape(shape = var_5435, x = value_97_cast_fp16)[name = tensor("op_5436_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_5436_cast_fp16, y = var_5434_cast_fp16)[name = tensor("attn_97_cast_fp16")]; + tensor var_5439 = const()[name = tensor("op_5439"), val = tensor([1, 1280, 1, -1])]; + tensor input_241_cast_fp16 = reshape(shape = var_5439, x = attn_97_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor var_5443 = const()[name = tensor("op_5443"), val = tensor([1, 1])]; + tensor var_5445 = const()[name = tensor("op_5445"), val = tensor([1, 1])]; + tensor obj_343_pad_type_0 = const()[name = tensor("obj_343_pad_type_0"), val = tensor("custom")]; + tensor obj_343_pad_0 = const()[name = tensor("obj_343_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1403140352)))]; + 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(1406417216)))]; + tensor obj_343_cast_fp16 = conv(bias = layers_24_self_attn_o_proj_bias_to_fp16, dilations = var_5445, groups = var_5355, pad = obj_343_pad_0, pad_type = obj_343_pad_type_0, strides = var_5443, 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 var_5455 = const()[name = tensor("op_5455"), val = tensor([1])]; + tensor channels_mean_147_cast_fp16 = reduce_mean(axes = var_5455, keep_dims = var_5356, x = inputs_147_cast_fp16)[name = tensor("channels_mean_147_cast_fp16")]; + tensor zero_mean_147_cast_fp16 = sub(x = inputs_147_cast_fp16, y = channels_mean_147_cast_fp16)[name = tensor("zero_mean_147_cast_fp16")]; + tensor zero_mean_sq_147_cast_fp16 = mul(x = zero_mean_147_cast_fp16, y = zero_mean_147_cast_fp16)[name = tensor("zero_mean_sq_147_cast_fp16")]; + tensor var_5459 = const()[name = tensor("op_5459"), val = tensor([1])]; + tensor var_5460_cast_fp16 = reduce_mean(axes = var_5459, keep_dims = var_5356, x = zero_mean_sq_147_cast_fp16)[name = tensor("op_5460_cast_fp16")]; + tensor var_5461_to_fp16 = const()[name = tensor("op_5461_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5462_cast_fp16 = add(x = var_5460_cast_fp16, y = var_5461_to_fp16)[name = tensor("op_5462_cast_fp16")]; + tensor denom_147_epsilon_0_to_fp16 = const()[name = tensor("denom_147_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_147_cast_fp16 = rsqrt(epsilon = denom_147_epsilon_0_to_fp16, x = var_5462_cast_fp16)[name = tensor("denom_147_cast_fp16")]; + tensor out_147_cast_fp16 = mul(x = zero_mean_147_cast_fp16, y = denom_147_cast_fp16)[name = tensor("out_147_cast_fp16")]; + tensor obj_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(1406419840)))]; + 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(1406422464)))]; + 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 var_5477 = const()[name = tensor("op_5477"), val = tensor([1, 1])]; + tensor var_5479 = const()[name = tensor("op_5479"), val = tensor([1, 1])]; + tensor query_99_pad_type_0 = const()[name = tensor("query_99_pad_type_0"), val = tensor("custom")]; + tensor query_99_pad_0 = const()[name = tensor("query_99_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1406425088)))]; + 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(1409701952)))]; + tensor query_99_cast_fp16 = conv(bias = layers_24_encoder_attn_q_proj_bias_to_fp16, dilations = var_5479, groups = var_5355, pad = query_99_pad_0, pad_type = query_99_pad_type_0, strides = var_5477, weight = layers_24_encoder_attn_q_proj_weight_to_fp16, x = obj_345_cast_fp16)[name = tensor("query_99_cast_fp16")]; + tensor var_5483 = const()[name = tensor("op_5483"), val = tensor([1, 1])]; + tensor var_5485 = const()[name = tensor("op_5485"), val = tensor([1, 1])]; + tensor key_99_pad_type_0 = const()[name = tensor("key_99_pad_type_0"), val = tensor("custom")]; + tensor key_99_pad_0 = const()[name = tensor("key_99_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1409704576)))]; + tensor key_99_cast_fp16 = conv(dilations = var_5485, groups = var_5355, pad = key_99_pad_0, pad_type = key_99_pad_type_0, strides = var_5483, weight = layers_24_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_99_cast_fp16")]; + tensor var_5490 = const()[name = tensor("op_5490"), val = tensor([1, 1])]; + tensor var_5492 = const()[name = tensor("op_5492"), val = tensor([1, 1])]; + tensor value_99_pad_type_0 = const()[name = tensor("value_99_pad_type_0"), val = tensor("custom")]; + tensor value_99_pad_0 = const()[name = tensor("value_99_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1412981440)))]; + 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(1416258304)))]; + tensor value_99_cast_fp16 = conv(bias = layers_24_encoder_attn_v_proj_bias_to_fp16, dilations = var_5492, groups = var_5355, pad = value_99_pad_0, pad_type = value_99_pad_type_0, strides = var_5490, weight = layers_24_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_99_cast_fp16")]; + tensor var_5496 = const()[name = tensor("op_5496"), val = tensor([1, 20, 64, -1])]; + tensor var_5497_cast_fp16 = reshape(shape = var_5496, x = query_99_cast_fp16)[name = tensor("op_5497_cast_fp16")]; + tensor var_5498_to_fp16 = const()[name = tensor("op_5498_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5499_cast_fp16 = mul(x = var_5497_cast_fp16, y = var_5498_to_fp16)[name = tensor("op_5499_cast_fp16")]; + tensor var_5500 = const()[name = tensor("op_5500"), val = tensor([1, 20, 64, -1])]; + tensor var_5501_cast_fp16 = reshape(shape = var_5500, x = key_99_cast_fp16)[name = tensor("op_5501_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_5499_cast_fp16, y = var_5501_cast_fp16)[name = tensor("mh_w_149_cast_fp16")]; + tensor obj_349_cast_fp16 = softmax(axis = var_5348, x = mh_w_149_cast_fp16)[name = tensor("obj_349_cast_fp16")]; + tensor var_5505 = const()[name = tensor("op_5505"), val = tensor([1, 20, 64, -1])]; + tensor var_5506_cast_fp16 = reshape(shape = var_5505, x = value_99_cast_fp16)[name = tensor("op_5506_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_5506_cast_fp16, y = obj_349_cast_fp16)[name = tensor("attn_99_cast_fp16")]; + tensor var_5509 = const()[name = tensor("op_5509"), val = tensor([1, 1280, 1, -1])]; + tensor input_243_cast_fp16 = reshape(shape = var_5509, x = attn_99_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor var_5513 = const()[name = tensor("op_5513"), val = tensor([1, 1])]; + tensor var_5515 = const()[name = tensor("op_5515"), val = tensor([1, 1])]; + tensor obj_347_pad_type_0 = const()[name = tensor("obj_347_pad_type_0"), val = tensor("custom")]; + tensor obj_347_pad_0 = const()[name = tensor("obj_347_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1416260928)))]; + 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(1419537792)))]; + tensor obj_347_cast_fp16 = conv(bias = layers_24_encoder_attn_o_proj_bias_to_fp16, dilations = var_5515, groups = var_5355, pad = obj_347_pad_0, pad_type = obj_347_pad_type_0, strides = var_5513, 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 var_5521 = const()[name = tensor("op_5521"), val = tensor([1])]; + tensor channels_mean_149_cast_fp16 = reduce_mean(axes = var_5521, keep_dims = var_5356, x = inputs_149_cast_fp16)[name = tensor("channels_mean_149_cast_fp16")]; + tensor zero_mean_149_cast_fp16 = sub(x = inputs_149_cast_fp16, y = channels_mean_149_cast_fp16)[name = tensor("zero_mean_149_cast_fp16")]; + tensor zero_mean_sq_149_cast_fp16 = mul(x = zero_mean_149_cast_fp16, y = zero_mean_149_cast_fp16)[name = tensor("zero_mean_sq_149_cast_fp16")]; + tensor var_5525 = const()[name = tensor("op_5525"), val = tensor([1])]; + tensor var_5526_cast_fp16 = reduce_mean(axes = var_5525, keep_dims = var_5356, x = zero_mean_sq_149_cast_fp16)[name = tensor("op_5526_cast_fp16")]; + tensor var_5527_to_fp16 = const()[name = tensor("op_5527_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5528_cast_fp16 = add(x = var_5526_cast_fp16, y = var_5527_to_fp16)[name = tensor("op_5528_cast_fp16")]; + tensor denom_149_epsilon_0_to_fp16 = const()[name = tensor("denom_149_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_149_cast_fp16 = rsqrt(epsilon = denom_149_epsilon_0_to_fp16, x = var_5528_cast_fp16)[name = tensor("denom_149_cast_fp16")]; + tensor out_149_cast_fp16 = mul(x = zero_mean_149_cast_fp16, y = denom_149_cast_fp16)[name = tensor("out_149_cast_fp16")]; + tensor input_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(1419540416)))]; + 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(1419543040)))]; + 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 var_5539 = const()[name = tensor("op_5539"), val = tensor([1, 1])]; + tensor var_5541 = const()[name = tensor("op_5541"), val = tensor([1, 1])]; + tensor input_247_pad_type_0 = const()[name = tensor("input_247_pad_type_0"), val = tensor("custom")]; + tensor input_247_pad_0 = const()[name = tensor("input_247_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1419545664)))]; + 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(1432652928)))]; + tensor input_247_cast_fp16 = conv(bias = layers_24_fc1_bias_to_fp16, dilations = var_5541, groups = var_5355, pad = input_247_pad_0, pad_type = input_247_pad_type_0, strides = var_5539, 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 var_5547 = const()[name = tensor("op_5547"), val = tensor([1, 1])]; + tensor var_5549 = const()[name = tensor("op_5549"), val = tensor([1, 1])]; + tensor hidden_states_51_pad_type_0 = const()[name = tensor("hidden_states_51_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_51_pad_0 = const()[name = tensor("hidden_states_51_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1432663232)))]; + 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(1445770496)))]; + tensor hidden_states_51_cast_fp16 = conv(bias = layers_24_fc2_bias_to_fp16, dilations = var_5549, groups = var_5355, pad = hidden_states_51_pad_0, pad_type = hidden_states_51_pad_type_0, strides = var_5547, 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_5562 = const()[name = tensor("op_5562"), val = tensor(3)]; + tensor var_5569 = const()[name = tensor("op_5569"), val = tensor(1)]; + tensor var_5570 = const()[name = tensor("op_5570"), val = tensor(true)]; + tensor var_5582 = const()[name = tensor("op_5582"), val = tensor([1])]; + tensor channels_mean_151_cast_fp16 = reduce_mean(axes = var_5582, keep_dims = var_5570, x = inputs_151_cast_fp16)[name = tensor("channels_mean_151_cast_fp16")]; + tensor zero_mean_151_cast_fp16 = sub(x = inputs_151_cast_fp16, y = channels_mean_151_cast_fp16)[name = tensor("zero_mean_151_cast_fp16")]; + tensor zero_mean_sq_151_cast_fp16 = mul(x = zero_mean_151_cast_fp16, y = zero_mean_151_cast_fp16)[name = tensor("zero_mean_sq_151_cast_fp16")]; + tensor var_5586 = const()[name = tensor("op_5586"), val = tensor([1])]; + tensor var_5587_cast_fp16 = reduce_mean(axes = var_5586, keep_dims = var_5570, x = zero_mean_sq_151_cast_fp16)[name = tensor("op_5587_cast_fp16")]; + tensor var_5588_to_fp16 = const()[name = tensor("op_5588_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5589_cast_fp16 = add(x = var_5587_cast_fp16, y = var_5588_to_fp16)[name = tensor("op_5589_cast_fp16")]; + tensor denom_151_epsilon_0_to_fp16 = const()[name = tensor("denom_151_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_151_cast_fp16 = rsqrt(epsilon = denom_151_epsilon_0_to_fp16, x = var_5589_cast_fp16)[name = tensor("denom_151_cast_fp16")]; + tensor out_151_cast_fp16 = mul(x = zero_mean_151_cast_fp16, y = denom_151_cast_fp16)[name = tensor("out_151_cast_fp16")]; + tensor obj_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(1445773120)))]; + 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(1445775744)))]; + 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 var_5604 = const()[name = tensor("op_5604"), val = tensor([1, 1])]; + tensor var_5606 = const()[name = tensor("op_5606"), val = tensor([1, 1])]; + tensor query_101_pad_type_0 = const()[name = tensor("query_101_pad_type_0"), val = tensor("custom")]; + tensor query_101_pad_0 = const()[name = tensor("query_101_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1445778368)))]; + 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(1449055232)))]; + tensor query_101_cast_fp16 = conv(bias = layers_25_self_attn_q_proj_bias_to_fp16, dilations = var_5606, groups = var_5569, pad = query_101_pad_0, pad_type = query_101_pad_type_0, strides = var_5604, weight = layers_25_self_attn_q_proj_weight_to_fp16, x = obj_351_cast_fp16)[name = tensor("query_101_cast_fp16")]; + tensor var_5610 = const()[name = tensor("op_5610"), val = tensor([1, 1])]; + tensor var_5612 = const()[name = tensor("op_5612"), val = tensor([1, 1])]; + tensor current_key_51_pad_type_0 = const()[name = tensor("current_key_51_pad_type_0"), val = tensor("custom")]; + tensor current_key_51_pad_0 = const()[name = tensor("current_key_51_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1449057856)))]; + tensor current_key_51_cast_fp16 = conv(dilations = var_5612, groups = var_5569, pad = current_key_51_pad_0, pad_type = current_key_51_pad_type_0, strides = var_5610, weight = layers_25_self_attn_k_proj_weight_to_fp16, x = obj_351_cast_fp16)[name = tensor("current_key_51_cast_fp16")]; + tensor var_5617 = const()[name = tensor("op_5617"), val = tensor([1, 1])]; + tensor var_5619 = const()[name = tensor("op_5619"), val = tensor([1, 1])]; + tensor current_value_51_pad_type_0 = const()[name = tensor("current_value_51_pad_type_0"), val = tensor("custom")]; + tensor current_value_51_pad_0 = const()[name = tensor("current_value_51_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1452334720)))]; + 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(1455611584)))]; + tensor current_value_51_cast_fp16 = conv(bias = layers_25_self_attn_v_proj_bias_to_fp16, dilations = var_5619, groups = var_5569, pad = current_value_51_pad_0, pad_type = current_value_51_pad_type_0, strides = var_5617, weight = layers_25_self_attn_v_proj_weight_to_fp16, x = obj_351_cast_fp16)[name = tensor("current_value_51_cast_fp16")]; + tensor var_5626_cast_fp16 = mul(x = current_key_51_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_5626_cast_fp16")]; + tensor var_5628_cast_fp16 = mul(x = var_103_cast_fp16_25, y = var_241_cast_fp16)[name = tensor("op_5628_cast_fp16")]; + tensor key_101_cast_fp16 = add(x = var_5626_cast_fp16, y = var_5628_cast_fp16)[name = tensor("key_101_cast_fp16")]; + tensor var_5630_cast_fp16 = mul(x = current_value_51_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_5630_cast_fp16")]; + tensor var_5632_cast_fp16 = mul(x = var_138_cast_fp16_25, y = var_241_cast_fp16)[name = tensor("op_5632_cast_fp16")]; + tensor value_101_cast_fp16 = add(x = var_5630_cast_fp16, y = var_5632_cast_fp16)[name = tensor("value_101_cast_fp16")]; + tensor var_5635 = const()[name = tensor("op_5635"), val = tensor([1, 20, 64, -1])]; + tensor var_5636_cast_fp16 = reshape(shape = var_5635, x = query_101_cast_fp16)[name = tensor("op_5636_cast_fp16")]; + tensor var_5637_to_fp16 = const()[name = tensor("op_5637_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5638_cast_fp16 = mul(x = var_5636_cast_fp16, y = var_5637_to_fp16)[name = tensor("op_5638_cast_fp16")]; + tensor var_5639 = const()[name = tensor("op_5639"), val = tensor([1, 20, 64, -1])]; + tensor var_5640_cast_fp16 = reshape(shape = var_5639, x = key_101_cast_fp16)[name = tensor("op_5640_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_5638_cast_fp16, y = var_5640_cast_fp16)[name = tensor("mh_w_151_cast_fp16")]; + tensor mh_w_153_cast_fp16 = add(x = mh_w_151_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_153_cast_fp16")]; + tensor var_5648_cast_fp16 = softmax(axis = var_5562, x = mh_w_153_cast_fp16)[name = tensor("op_5648_cast_fp16")]; + tensor var_5649 = const()[name = tensor("op_5649"), val = tensor([1, 20, 64, -1])]; + tensor var_5650_cast_fp16 = reshape(shape = var_5649, x = value_101_cast_fp16)[name = tensor("op_5650_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_5650_cast_fp16, y = var_5648_cast_fp16)[name = tensor("attn_101_cast_fp16")]; + tensor var_5653 = const()[name = tensor("op_5653"), val = tensor([1, 1280, 1, -1])]; + tensor input_251_cast_fp16 = reshape(shape = var_5653, x = attn_101_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor var_5657 = const()[name = tensor("op_5657"), val = tensor([1, 1])]; + tensor var_5659 = const()[name = tensor("op_5659"), val = tensor([1, 1])]; + tensor obj_357_pad_type_0 = const()[name = tensor("obj_357_pad_type_0"), val = tensor("custom")]; + tensor obj_357_pad_0 = const()[name = tensor("obj_357_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1455614208)))]; + 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(1458891072)))]; + tensor obj_357_cast_fp16 = conv(bias = layers_25_self_attn_o_proj_bias_to_fp16, dilations = var_5659, groups = var_5569, pad = obj_357_pad_0, pad_type = obj_357_pad_type_0, strides = var_5657, 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 var_5669 = const()[name = tensor("op_5669"), val = tensor([1])]; + tensor channels_mean_153_cast_fp16 = reduce_mean(axes = var_5669, keep_dims = var_5570, x = inputs_153_cast_fp16)[name = tensor("channels_mean_153_cast_fp16")]; + tensor zero_mean_153_cast_fp16 = sub(x = inputs_153_cast_fp16, y = channels_mean_153_cast_fp16)[name = tensor("zero_mean_153_cast_fp16")]; + tensor zero_mean_sq_153_cast_fp16 = mul(x = zero_mean_153_cast_fp16, y = zero_mean_153_cast_fp16)[name = tensor("zero_mean_sq_153_cast_fp16")]; + tensor var_5673 = const()[name = tensor("op_5673"), val = tensor([1])]; + tensor var_5674_cast_fp16 = reduce_mean(axes = var_5673, keep_dims = var_5570, x = zero_mean_sq_153_cast_fp16)[name = tensor("op_5674_cast_fp16")]; + tensor var_5675_to_fp16 = const()[name = tensor("op_5675_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5676_cast_fp16 = add(x = var_5674_cast_fp16, y = var_5675_to_fp16)[name = tensor("op_5676_cast_fp16")]; + tensor denom_153_epsilon_0_to_fp16 = const()[name = tensor("denom_153_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_153_cast_fp16 = rsqrt(epsilon = denom_153_epsilon_0_to_fp16, x = var_5676_cast_fp16)[name = tensor("denom_153_cast_fp16")]; + tensor out_153_cast_fp16 = mul(x = zero_mean_153_cast_fp16, y = denom_153_cast_fp16)[name = tensor("out_153_cast_fp16")]; + tensor obj_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(1458893696)))]; + 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(1458896320)))]; + 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 var_5691 = const()[name = tensor("op_5691"), val = tensor([1, 1])]; + tensor var_5693 = const()[name = tensor("op_5693"), val = tensor([1, 1])]; + tensor query_103_pad_type_0 = const()[name = tensor("query_103_pad_type_0"), val = tensor("custom")]; + tensor query_103_pad_0 = const()[name = tensor("query_103_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1458898944)))]; + 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(1462175808)))]; + tensor query_103_cast_fp16 = conv(bias = layers_25_encoder_attn_q_proj_bias_to_fp16, dilations = var_5693, groups = var_5569, pad = query_103_pad_0, pad_type = query_103_pad_type_0, strides = var_5691, weight = layers_25_encoder_attn_q_proj_weight_to_fp16, x = obj_359_cast_fp16)[name = tensor("query_103_cast_fp16")]; + tensor var_5697 = const()[name = tensor("op_5697"), val = tensor([1, 1])]; + tensor var_5699 = const()[name = tensor("op_5699"), val = tensor([1, 1])]; + tensor key_103_pad_type_0 = const()[name = tensor("key_103_pad_type_0"), val = tensor("custom")]; + tensor key_103_pad_0 = const()[name = tensor("key_103_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1462178432)))]; + tensor key_103_cast_fp16 = conv(dilations = var_5699, groups = var_5569, pad = key_103_pad_0, pad_type = key_103_pad_type_0, strides = var_5697, weight = layers_25_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_103_cast_fp16")]; + tensor var_5704 = const()[name = tensor("op_5704"), val = tensor([1, 1])]; + tensor var_5706 = const()[name = tensor("op_5706"), val = tensor([1, 1])]; + tensor value_103_pad_type_0 = const()[name = tensor("value_103_pad_type_0"), val = tensor("custom")]; + tensor value_103_pad_0 = const()[name = tensor("value_103_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1465455296)))]; + 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(1468732160)))]; + tensor value_103_cast_fp16 = conv(bias = layers_25_encoder_attn_v_proj_bias_to_fp16, dilations = var_5706, groups = var_5569, pad = value_103_pad_0, pad_type = value_103_pad_type_0, strides = var_5704, weight = layers_25_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_103_cast_fp16")]; + tensor var_5710 = const()[name = tensor("op_5710"), val = tensor([1, 20, 64, -1])]; + tensor var_5711_cast_fp16 = reshape(shape = var_5710, x = query_103_cast_fp16)[name = tensor("op_5711_cast_fp16")]; + tensor var_5712_to_fp16 = const()[name = tensor("op_5712_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5713_cast_fp16 = mul(x = var_5711_cast_fp16, y = var_5712_to_fp16)[name = tensor("op_5713_cast_fp16")]; + tensor var_5714 = const()[name = tensor("op_5714"), val = tensor([1, 20, 64, -1])]; + tensor var_5715_cast_fp16 = reshape(shape = var_5714, x = key_103_cast_fp16)[name = tensor("op_5715_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_5713_cast_fp16, y = var_5715_cast_fp16)[name = tensor("mh_w_155_cast_fp16")]; + tensor obj_363_cast_fp16 = softmax(axis = var_5562, x = mh_w_155_cast_fp16)[name = tensor("obj_363_cast_fp16")]; + tensor var_5719 = const()[name = tensor("op_5719"), val = tensor([1, 20, 64, -1])]; + tensor var_5720_cast_fp16 = reshape(shape = var_5719, x = value_103_cast_fp16)[name = tensor("op_5720_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_5720_cast_fp16, y = obj_363_cast_fp16)[name = tensor("attn_103_cast_fp16")]; + tensor var_5723 = const()[name = tensor("op_5723"), val = tensor([1, 1280, 1, -1])]; + tensor input_253_cast_fp16 = reshape(shape = var_5723, x = attn_103_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor var_5727 = const()[name = tensor("op_5727"), val = tensor([1, 1])]; + tensor var_5729 = const()[name = tensor("op_5729"), val = tensor([1, 1])]; + tensor obj_361_pad_type_0 = const()[name = tensor("obj_361_pad_type_0"), val = tensor("custom")]; + tensor obj_361_pad_0 = const()[name = tensor("obj_361_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1468734784)))]; + 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(1472011648)))]; + tensor obj_361_cast_fp16 = conv(bias = layers_25_encoder_attn_o_proj_bias_to_fp16, dilations = var_5729, groups = var_5569, pad = obj_361_pad_0, pad_type = obj_361_pad_type_0, strides = var_5727, 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 var_5738 = const()[name = tensor("op_5738"), val = tensor([1])]; + tensor channels_mean_155_cast_fp16 = reduce_mean(axes = var_5738, keep_dims = var_5570, x = inputs_155_cast_fp16)[name = tensor("channels_mean_155_cast_fp16")]; + tensor zero_mean_155_cast_fp16 = sub(x = inputs_155_cast_fp16, y = channels_mean_155_cast_fp16)[name = tensor("zero_mean_155_cast_fp16")]; + tensor zero_mean_sq_155_cast_fp16 = mul(x = zero_mean_155_cast_fp16, y = zero_mean_155_cast_fp16)[name = tensor("zero_mean_sq_155_cast_fp16")]; + tensor var_5742 = const()[name = tensor("op_5742"), val = tensor([1])]; + tensor var_5743_cast_fp16 = reduce_mean(axes = var_5742, keep_dims = var_5570, x = zero_mean_sq_155_cast_fp16)[name = tensor("op_5743_cast_fp16")]; + tensor var_5744_to_fp16 = const()[name = tensor("op_5744_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5745_cast_fp16 = add(x = var_5743_cast_fp16, y = var_5744_to_fp16)[name = tensor("op_5745_cast_fp16")]; + tensor denom_155_epsilon_0_to_fp16 = const()[name = tensor("denom_155_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_155_cast_fp16 = rsqrt(epsilon = denom_155_epsilon_0_to_fp16, x = var_5745_cast_fp16)[name = tensor("denom_155_cast_fp16")]; + tensor out_155_cast_fp16 = mul(x = zero_mean_155_cast_fp16, y = denom_155_cast_fp16)[name = tensor("out_155_cast_fp16")]; + tensor input_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(1472014272)))]; + 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(1472016896)))]; + 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 var_5756 = const()[name = tensor("op_5756"), val = tensor([1, 1])]; + tensor var_5758 = const()[name = tensor("op_5758"), val = tensor([1, 1])]; + tensor input_257_pad_type_0 = const()[name = tensor("input_257_pad_type_0"), val = tensor("custom")]; + tensor input_257_pad_0 = const()[name = tensor("input_257_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1472019520)))]; + 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(1485126784)))]; + tensor input_257_cast_fp16 = conv(bias = layers_25_fc1_bias_to_fp16, dilations = var_5758, groups = var_5569, pad = input_257_pad_0, pad_type = input_257_pad_type_0, strides = var_5756, 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 var_5764 = const()[name = tensor("op_5764"), val = tensor([1, 1])]; + tensor var_5766 = const()[name = tensor("op_5766"), val = tensor([1, 1])]; + tensor hidden_states_53_pad_type_0 = const()[name = tensor("hidden_states_53_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_53_pad_0 = const()[name = tensor("hidden_states_53_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1485137088)))]; + 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(1498244352)))]; + tensor hidden_states_53_cast_fp16 = conv(bias = layers_25_fc2_bias_to_fp16, dilations = var_5766, groups = var_5569, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = var_5764, 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_5780 = const()[name = tensor("op_5780"), val = tensor(3)]; + tensor var_5787 = const()[name = tensor("op_5787"), val = tensor(1)]; + tensor var_5788 = const()[name = tensor("op_5788"), val = tensor(true)]; + tensor var_5800 = const()[name = tensor("op_5800"), val = tensor([1])]; + tensor channels_mean_157_cast_fp16 = reduce_mean(axes = var_5800, keep_dims = var_5788, x = inputs_157_cast_fp16)[name = tensor("channels_mean_157_cast_fp16")]; + tensor zero_mean_157_cast_fp16 = sub(x = inputs_157_cast_fp16, y = channels_mean_157_cast_fp16)[name = tensor("zero_mean_157_cast_fp16")]; + tensor zero_mean_sq_157_cast_fp16 = mul(x = zero_mean_157_cast_fp16, y = zero_mean_157_cast_fp16)[name = tensor("zero_mean_sq_157_cast_fp16")]; + tensor var_5804 = const()[name = tensor("op_5804"), val = tensor([1])]; + tensor var_5805_cast_fp16 = reduce_mean(axes = var_5804, keep_dims = var_5788, x = zero_mean_sq_157_cast_fp16)[name = tensor("op_5805_cast_fp16")]; + tensor var_5806_to_fp16 = const()[name = tensor("op_5806_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5807_cast_fp16 = add(x = var_5805_cast_fp16, y = var_5806_to_fp16)[name = tensor("op_5807_cast_fp16")]; + tensor denom_157_epsilon_0_to_fp16 = const()[name = tensor("denom_157_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_157_cast_fp16 = rsqrt(epsilon = denom_157_epsilon_0_to_fp16, x = var_5807_cast_fp16)[name = tensor("denom_157_cast_fp16")]; + tensor out_157_cast_fp16 = mul(x = zero_mean_157_cast_fp16, y = denom_157_cast_fp16)[name = tensor("out_157_cast_fp16")]; + tensor obj_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(1498246976)))]; + 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(1498249600)))]; + 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 var_5822 = const()[name = tensor("op_5822"), val = tensor([1, 1])]; + tensor var_5824 = const()[name = tensor("op_5824"), val = tensor([1, 1])]; + tensor query_105_pad_type_0 = const()[name = tensor("query_105_pad_type_0"), val = tensor("custom")]; + tensor query_105_pad_0 = const()[name = tensor("query_105_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1498252224)))]; + 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(1501529088)))]; + tensor query_105_cast_fp16 = conv(bias = layers_26_self_attn_q_proj_bias_to_fp16, dilations = var_5824, groups = var_5787, pad = query_105_pad_0, pad_type = query_105_pad_type_0, strides = var_5822, weight = layers_26_self_attn_q_proj_weight_to_fp16, x = obj_365_cast_fp16)[name = tensor("query_105_cast_fp16")]; + tensor var_5828 = const()[name = tensor("op_5828"), val = tensor([1, 1])]; + tensor var_5830 = const()[name = tensor("op_5830"), val = tensor([1, 1])]; + tensor current_key_53_pad_type_0 = const()[name = tensor("current_key_53_pad_type_0"), val = tensor("custom")]; + tensor current_key_53_pad_0 = const()[name = tensor("current_key_53_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1501531712)))]; + tensor current_key_53_cast_fp16 = conv(dilations = var_5830, groups = var_5787, pad = current_key_53_pad_0, pad_type = current_key_53_pad_type_0, strides = var_5828, weight = layers_26_self_attn_k_proj_weight_to_fp16, x = obj_365_cast_fp16)[name = tensor("current_key_53_cast_fp16")]; + tensor var_5835 = const()[name = tensor("op_5835"), val = tensor([1, 1])]; + tensor var_5837 = const()[name = tensor("op_5837"), val = tensor([1, 1])]; + tensor current_value_53_pad_type_0 = const()[name = tensor("current_value_53_pad_type_0"), val = tensor("custom")]; + tensor current_value_53_pad_0 = const()[name = tensor("current_value_53_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1504808576)))]; + 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(1508085440)))]; + tensor current_value_53_cast_fp16 = conv(bias = layers_26_self_attn_v_proj_bias_to_fp16, dilations = var_5837, groups = var_5787, pad = current_value_53_pad_0, pad_type = current_value_53_pad_type_0, strides = var_5835, weight = layers_26_self_attn_v_proj_weight_to_fp16, x = obj_365_cast_fp16)[name = tensor("current_value_53_cast_fp16")]; + tensor var_5844_cast_fp16 = mul(x = current_key_53_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_5844_cast_fp16")]; + tensor var_5846_cast_fp16 = mul(x = var_103_cast_fp16_26, y = var_241_cast_fp16)[name = tensor("op_5846_cast_fp16")]; + tensor key_105_cast_fp16 = add(x = var_5844_cast_fp16, y = var_5846_cast_fp16)[name = tensor("key_105_cast_fp16")]; + tensor var_5848_cast_fp16 = mul(x = current_value_53_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_5848_cast_fp16")]; + tensor var_5850_cast_fp16 = mul(x = var_138_cast_fp16_26, y = var_241_cast_fp16)[name = tensor("op_5850_cast_fp16")]; + tensor value_105_cast_fp16 = add(x = var_5848_cast_fp16, y = var_5850_cast_fp16)[name = tensor("value_105_cast_fp16")]; + tensor var_5853 = const()[name = tensor("op_5853"), val = tensor([1, 20, 64, -1])]; + tensor var_5854_cast_fp16 = reshape(shape = var_5853, x = query_105_cast_fp16)[name = tensor("op_5854_cast_fp16")]; + tensor var_5855_to_fp16 = const()[name = tensor("op_5855_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5856_cast_fp16 = mul(x = var_5854_cast_fp16, y = var_5855_to_fp16)[name = tensor("op_5856_cast_fp16")]; + tensor var_5857 = const()[name = tensor("op_5857"), val = tensor([1, 20, 64, -1])]; + tensor var_5858_cast_fp16 = reshape(shape = var_5857, x = key_105_cast_fp16)[name = tensor("op_5858_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_5856_cast_fp16, y = var_5858_cast_fp16)[name = tensor("mh_w_157_cast_fp16")]; + tensor mh_w_159_cast_fp16 = add(x = mh_w_157_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_159_cast_fp16")]; + tensor var_5866_cast_fp16 = softmax(axis = var_5780, x = mh_w_159_cast_fp16)[name = tensor("op_5866_cast_fp16")]; + tensor var_5867 = const()[name = tensor("op_5867"), val = tensor([1, 20, 64, -1])]; + tensor var_5868_cast_fp16 = reshape(shape = var_5867, x = value_105_cast_fp16)[name = tensor("op_5868_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_5868_cast_fp16, y = var_5866_cast_fp16)[name = tensor("attn_105_cast_fp16")]; + tensor var_5871 = const()[name = tensor("op_5871"), val = tensor([1, 1280, 1, -1])]; + tensor input_261_cast_fp16 = reshape(shape = var_5871, x = attn_105_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor var_5875 = const()[name = tensor("op_5875"), val = tensor([1, 1])]; + tensor var_5877 = const()[name = tensor("op_5877"), val = tensor([1, 1])]; + tensor obj_371_pad_type_0 = const()[name = tensor("obj_371_pad_type_0"), val = tensor("custom")]; + tensor obj_371_pad_0 = const()[name = tensor("obj_371_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1508088064)))]; + 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(1511364928)))]; + tensor obj_371_cast_fp16 = conv(bias = layers_26_self_attn_o_proj_bias_to_fp16, dilations = var_5877, groups = var_5787, pad = obj_371_pad_0, pad_type = obj_371_pad_type_0, strides = var_5875, 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 var_5887 = const()[name = tensor("op_5887"), val = tensor([1])]; + tensor channels_mean_159_cast_fp16 = reduce_mean(axes = var_5887, keep_dims = var_5788, x = inputs_159_cast_fp16)[name = tensor("channels_mean_159_cast_fp16")]; + tensor zero_mean_159_cast_fp16 = sub(x = inputs_159_cast_fp16, y = channels_mean_159_cast_fp16)[name = tensor("zero_mean_159_cast_fp16")]; + tensor zero_mean_sq_159_cast_fp16 = mul(x = zero_mean_159_cast_fp16, y = zero_mean_159_cast_fp16)[name = tensor("zero_mean_sq_159_cast_fp16")]; + tensor var_5891 = const()[name = tensor("op_5891"), val = tensor([1])]; + tensor var_5892_cast_fp16 = reduce_mean(axes = var_5891, keep_dims = var_5788, x = zero_mean_sq_159_cast_fp16)[name = tensor("op_5892_cast_fp16")]; + tensor var_5893_to_fp16 = const()[name = tensor("op_5893_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5894_cast_fp16 = add(x = var_5892_cast_fp16, y = var_5893_to_fp16)[name = tensor("op_5894_cast_fp16")]; + tensor denom_159_epsilon_0_to_fp16 = const()[name = tensor("denom_159_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_159_cast_fp16 = rsqrt(epsilon = denom_159_epsilon_0_to_fp16, x = var_5894_cast_fp16)[name = tensor("denom_159_cast_fp16")]; + tensor out_159_cast_fp16 = mul(x = zero_mean_159_cast_fp16, y = denom_159_cast_fp16)[name = tensor("out_159_cast_fp16")]; + tensor obj_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(1511367552)))]; + 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(1511370176)))]; + 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 var_5909 = const()[name = tensor("op_5909"), val = tensor([1, 1])]; + tensor var_5911 = const()[name = tensor("op_5911"), val = tensor([1, 1])]; + tensor query_107_pad_type_0 = const()[name = tensor("query_107_pad_type_0"), val = tensor("custom")]; + tensor query_107_pad_0 = const()[name = tensor("query_107_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1511372800)))]; + 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(1514649664)))]; + tensor query_107_cast_fp16 = conv(bias = layers_26_encoder_attn_q_proj_bias_to_fp16, dilations = var_5911, groups = var_5787, pad = query_107_pad_0, pad_type = query_107_pad_type_0, strides = var_5909, weight = layers_26_encoder_attn_q_proj_weight_to_fp16, x = obj_373_cast_fp16)[name = tensor("query_107_cast_fp16")]; + tensor var_5915 = const()[name = tensor("op_5915"), val = tensor([1, 1])]; + tensor var_5917 = const()[name = tensor("op_5917"), val = tensor([1, 1])]; + tensor key_107_pad_type_0 = const()[name = tensor("key_107_pad_type_0"), val = tensor("custom")]; + tensor key_107_pad_0 = const()[name = tensor("key_107_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1514652288)))]; + tensor key_107_cast_fp16 = conv(dilations = var_5917, groups = var_5787, pad = key_107_pad_0, pad_type = key_107_pad_type_0, strides = var_5915, weight = layers_26_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_107_cast_fp16")]; + tensor var_5922 = const()[name = tensor("op_5922"), val = tensor([1, 1])]; + tensor var_5924 = const()[name = tensor("op_5924"), val = tensor([1, 1])]; + tensor value_107_pad_type_0 = const()[name = tensor("value_107_pad_type_0"), val = tensor("custom")]; + tensor value_107_pad_0 = const()[name = tensor("value_107_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1517929152)))]; + 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(1521206016)))]; + tensor value_107_cast_fp16 = conv(bias = layers_26_encoder_attn_v_proj_bias_to_fp16, dilations = var_5924, groups = var_5787, pad = value_107_pad_0, pad_type = value_107_pad_type_0, strides = var_5922, weight = layers_26_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_107_cast_fp16")]; + tensor var_5928 = const()[name = tensor("op_5928"), val = tensor([1, 20, 64, -1])]; + tensor var_5929_cast_fp16 = reshape(shape = var_5928, x = query_107_cast_fp16)[name = tensor("op_5929_cast_fp16")]; + tensor var_5930_to_fp16 = const()[name = tensor("op_5930_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5931_cast_fp16 = mul(x = var_5929_cast_fp16, y = var_5930_to_fp16)[name = tensor("op_5931_cast_fp16")]; + tensor var_5932 = const()[name = tensor("op_5932"), val = tensor([1, 20, 64, -1])]; + tensor var_5933_cast_fp16 = reshape(shape = var_5932, x = key_107_cast_fp16)[name = tensor("op_5933_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_5931_cast_fp16, y = var_5933_cast_fp16)[name = tensor("mh_w_161_cast_fp16")]; + tensor obj_377_cast_fp16 = softmax(axis = var_5780, x = mh_w_161_cast_fp16)[name = tensor("obj_377_cast_fp16")]; + tensor var_5937 = const()[name = tensor("op_5937"), val = tensor([1, 20, 64, -1])]; + tensor var_5938_cast_fp16 = reshape(shape = var_5937, x = value_107_cast_fp16)[name = tensor("op_5938_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_5938_cast_fp16, y = obj_377_cast_fp16)[name = tensor("attn_107_cast_fp16")]; + tensor var_5941 = const()[name = tensor("op_5941"), val = tensor([1, 1280, 1, -1])]; + tensor input_263_cast_fp16 = reshape(shape = var_5941, x = attn_107_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor var_5945 = const()[name = tensor("op_5945"), val = tensor([1, 1])]; + tensor var_5947 = const()[name = tensor("op_5947"), val = tensor([1, 1])]; + tensor obj_375_pad_type_0 = const()[name = tensor("obj_375_pad_type_0"), val = tensor("custom")]; + tensor obj_375_pad_0 = const()[name = tensor("obj_375_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1521208640)))]; + 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(1524485504)))]; + tensor obj_375_cast_fp16 = conv(bias = layers_26_encoder_attn_o_proj_bias_to_fp16, dilations = var_5947, groups = var_5787, pad = obj_375_pad_0, pad_type = obj_375_pad_type_0, strides = var_5945, 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 var_5956 = const()[name = tensor("op_5956"), val = tensor([1])]; + tensor channels_mean_161_cast_fp16 = reduce_mean(axes = var_5956, keep_dims = var_5788, x = inputs_161_cast_fp16)[name = tensor("channels_mean_161_cast_fp16")]; + tensor zero_mean_161_cast_fp16 = sub(x = inputs_161_cast_fp16, y = channels_mean_161_cast_fp16)[name = tensor("zero_mean_161_cast_fp16")]; + tensor zero_mean_sq_161_cast_fp16 = mul(x = zero_mean_161_cast_fp16, y = zero_mean_161_cast_fp16)[name = tensor("zero_mean_sq_161_cast_fp16")]; + tensor var_5960 = const()[name = tensor("op_5960"), val = tensor([1])]; + tensor var_5961_cast_fp16 = reduce_mean(axes = var_5960, keep_dims = var_5788, x = zero_mean_sq_161_cast_fp16)[name = tensor("op_5961_cast_fp16")]; + tensor var_5962_to_fp16 = const()[name = tensor("op_5962_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5963_cast_fp16 = add(x = var_5961_cast_fp16, y = var_5962_to_fp16)[name = tensor("op_5963_cast_fp16")]; + tensor denom_161_epsilon_0_to_fp16 = const()[name = tensor("denom_161_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_161_cast_fp16 = rsqrt(epsilon = denom_161_epsilon_0_to_fp16, x = var_5963_cast_fp16)[name = tensor("denom_161_cast_fp16")]; + tensor out_161_cast_fp16 = mul(x = zero_mean_161_cast_fp16, y = denom_161_cast_fp16)[name = tensor("out_161_cast_fp16")]; + tensor input_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(1524488128)))]; + 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(1524490752)))]; + 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 var_5974 = const()[name = tensor("op_5974"), val = tensor([1, 1])]; + tensor var_5976 = const()[name = tensor("op_5976"), val = tensor([1, 1])]; + tensor input_267_pad_type_0 = const()[name = tensor("input_267_pad_type_0"), val = tensor("custom")]; + tensor input_267_pad_0 = const()[name = tensor("input_267_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1524493376)))]; + 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(1537600640)))]; + tensor input_267_cast_fp16 = conv(bias = layers_26_fc1_bias_to_fp16, dilations = var_5976, groups = var_5787, pad = input_267_pad_0, pad_type = input_267_pad_type_0, strides = var_5974, 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 var_5982 = const()[name = tensor("op_5982"), val = tensor([1, 1])]; + tensor var_5984 = const()[name = tensor("op_5984"), val = tensor([1, 1])]; + tensor hidden_states_55_pad_type_0 = const()[name = tensor("hidden_states_55_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_55_pad_0 = const()[name = tensor("hidden_states_55_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1537610944)))]; + 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(1550718208)))]; + tensor hidden_states_55_cast_fp16 = conv(bias = layers_26_fc2_bias_to_fp16, dilations = var_5984, groups = var_5787, pad = hidden_states_55_pad_0, pad_type = hidden_states_55_pad_type_0, strides = var_5982, 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_5998 = const()[name = tensor("op_5998"), val = tensor(3)]; + tensor var_6005 = const()[name = tensor("op_6005"), val = tensor(1)]; + tensor var_6006 = const()[name = tensor("op_6006"), val = tensor(true)]; + tensor var_6018 = const()[name = tensor("op_6018"), val = tensor([1])]; + tensor channels_mean_163_cast_fp16 = reduce_mean(axes = var_6018, keep_dims = var_6006, x = inputs_163_cast_fp16)[name = tensor("channels_mean_163_cast_fp16")]; + tensor zero_mean_163_cast_fp16 = sub(x = inputs_163_cast_fp16, y = channels_mean_163_cast_fp16)[name = tensor("zero_mean_163_cast_fp16")]; + tensor zero_mean_sq_163_cast_fp16 = mul(x = zero_mean_163_cast_fp16, y = zero_mean_163_cast_fp16)[name = tensor("zero_mean_sq_163_cast_fp16")]; + tensor var_6022 = const()[name = tensor("op_6022"), val = tensor([1])]; + tensor var_6023_cast_fp16 = reduce_mean(axes = var_6022, keep_dims = var_6006, x = zero_mean_sq_163_cast_fp16)[name = tensor("op_6023_cast_fp16")]; + tensor var_6024_to_fp16 = const()[name = tensor("op_6024_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6025_cast_fp16 = add(x = var_6023_cast_fp16, y = var_6024_to_fp16)[name = tensor("op_6025_cast_fp16")]; + tensor denom_163_epsilon_0_to_fp16 = const()[name = tensor("denom_163_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_163_cast_fp16 = rsqrt(epsilon = denom_163_epsilon_0_to_fp16, x = var_6025_cast_fp16)[name = tensor("denom_163_cast_fp16")]; + tensor out_163_cast_fp16 = mul(x = zero_mean_163_cast_fp16, y = denom_163_cast_fp16)[name = tensor("out_163_cast_fp16")]; + tensor obj_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(1550720832)))]; + 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(1550723456)))]; + 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 var_6040 = const()[name = tensor("op_6040"), val = tensor([1, 1])]; + tensor var_6042 = const()[name = tensor("op_6042"), val = tensor([1, 1])]; + tensor query_109_pad_type_0 = const()[name = tensor("query_109_pad_type_0"), val = tensor("custom")]; + tensor query_109_pad_0 = const()[name = tensor("query_109_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1550726080)))]; + 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(1554002944)))]; + tensor query_109_cast_fp16 = conv(bias = layers_27_self_attn_q_proj_bias_to_fp16, dilations = var_6042, groups = var_6005, pad = query_109_pad_0, pad_type = query_109_pad_type_0, strides = var_6040, weight = layers_27_self_attn_q_proj_weight_to_fp16, x = obj_379_cast_fp16)[name = tensor("query_109_cast_fp16")]; + tensor var_6046 = const()[name = tensor("op_6046"), val = tensor([1, 1])]; + tensor var_6048 = const()[name = tensor("op_6048"), val = tensor([1, 1])]; + tensor current_key_55_pad_type_0 = const()[name = tensor("current_key_55_pad_type_0"), val = tensor("custom")]; + tensor current_key_55_pad_0 = const()[name = tensor("current_key_55_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1554005568)))]; + tensor current_key_55_cast_fp16 = conv(dilations = var_6048, groups = var_6005, pad = current_key_55_pad_0, pad_type = current_key_55_pad_type_0, strides = var_6046, weight = layers_27_self_attn_k_proj_weight_to_fp16, x = obj_379_cast_fp16)[name = tensor("current_key_55_cast_fp16")]; + tensor var_6053 = const()[name = tensor("op_6053"), val = tensor([1, 1])]; + tensor var_6055 = const()[name = tensor("op_6055"), val = tensor([1, 1])]; + tensor current_value_55_pad_type_0 = const()[name = tensor("current_value_55_pad_type_0"), val = tensor("custom")]; + tensor current_value_55_pad_0 = const()[name = tensor("current_value_55_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1557282432)))]; + 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(1560559296)))]; + tensor current_value_55_cast_fp16 = conv(bias = layers_27_self_attn_v_proj_bias_to_fp16, dilations = var_6055, groups = var_6005, pad = current_value_55_pad_0, pad_type = current_value_55_pad_type_0, strides = var_6053, weight = layers_27_self_attn_v_proj_weight_to_fp16, x = obj_379_cast_fp16)[name = tensor("current_value_55_cast_fp16")]; + tensor var_6062_cast_fp16 = mul(x = current_key_55_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6062_cast_fp16")]; + tensor var_6064_cast_fp16 = mul(x = var_103_cast_fp16_27, y = var_241_cast_fp16)[name = tensor("op_6064_cast_fp16")]; + tensor key_109_cast_fp16 = add(x = var_6062_cast_fp16, y = var_6064_cast_fp16)[name = tensor("key_109_cast_fp16")]; + tensor var_6066_cast_fp16 = mul(x = current_value_55_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6066_cast_fp16")]; + tensor var_6068_cast_fp16 = mul(x = var_138_cast_fp16_27, y = var_241_cast_fp16)[name = tensor("op_6068_cast_fp16")]; + tensor value_109_cast_fp16 = add(x = var_6066_cast_fp16, y = var_6068_cast_fp16)[name = tensor("value_109_cast_fp16")]; + tensor var_6071 = const()[name = tensor("op_6071"), val = tensor([1, 20, 64, -1])]; + tensor var_6072_cast_fp16 = reshape(shape = var_6071, x = query_109_cast_fp16)[name = tensor("op_6072_cast_fp16")]; + tensor var_6073_to_fp16 = const()[name = tensor("op_6073_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6074_cast_fp16 = mul(x = var_6072_cast_fp16, y = var_6073_to_fp16)[name = tensor("op_6074_cast_fp16")]; + tensor var_6075 = const()[name = tensor("op_6075"), val = tensor([1, 20, 64, -1])]; + tensor var_6076_cast_fp16 = reshape(shape = var_6075, x = key_109_cast_fp16)[name = tensor("op_6076_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_6074_cast_fp16, y = var_6076_cast_fp16)[name = tensor("mh_w_163_cast_fp16")]; + tensor mh_w_165_cast_fp16 = add(x = mh_w_163_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_165_cast_fp16")]; + tensor var_6084_cast_fp16 = softmax(axis = var_5998, x = mh_w_165_cast_fp16)[name = tensor("op_6084_cast_fp16")]; + tensor var_6085 = const()[name = tensor("op_6085"), val = tensor([1, 20, 64, -1])]; + tensor var_6086_cast_fp16 = reshape(shape = var_6085, x = value_109_cast_fp16)[name = tensor("op_6086_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_6086_cast_fp16, y = var_6084_cast_fp16)[name = tensor("attn_109_cast_fp16")]; + tensor var_6089 = const()[name = tensor("op_6089"), val = tensor([1, 1280, 1, -1])]; + tensor input_271_cast_fp16 = reshape(shape = var_6089, x = attn_109_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor var_6093 = const()[name = tensor("op_6093"), val = tensor([1, 1])]; + tensor var_6095 = const()[name = tensor("op_6095"), val = tensor([1, 1])]; + tensor obj_385_pad_type_0 = const()[name = tensor("obj_385_pad_type_0"), val = tensor("custom")]; + tensor obj_385_pad_0 = const()[name = tensor("obj_385_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1560561920)))]; + 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(1563838784)))]; + tensor obj_385_cast_fp16 = conv(bias = layers_27_self_attn_o_proj_bias_to_fp16, dilations = var_6095, groups = var_6005, pad = obj_385_pad_0, pad_type = obj_385_pad_type_0, strides = var_6093, 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 var_6105 = const()[name = tensor("op_6105"), val = tensor([1])]; + tensor channels_mean_165_cast_fp16 = reduce_mean(axes = var_6105, keep_dims = var_6006, x = inputs_165_cast_fp16)[name = tensor("channels_mean_165_cast_fp16")]; + tensor zero_mean_165_cast_fp16 = sub(x = inputs_165_cast_fp16, y = channels_mean_165_cast_fp16)[name = tensor("zero_mean_165_cast_fp16")]; + tensor zero_mean_sq_165_cast_fp16 = mul(x = zero_mean_165_cast_fp16, y = zero_mean_165_cast_fp16)[name = tensor("zero_mean_sq_165_cast_fp16")]; + tensor var_6109 = const()[name = tensor("op_6109"), val = tensor([1])]; + tensor var_6110_cast_fp16 = reduce_mean(axes = var_6109, keep_dims = var_6006, x = zero_mean_sq_165_cast_fp16)[name = tensor("op_6110_cast_fp16")]; + tensor var_6111_to_fp16 = const()[name = tensor("op_6111_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6112_cast_fp16 = add(x = var_6110_cast_fp16, y = var_6111_to_fp16)[name = tensor("op_6112_cast_fp16")]; + tensor denom_165_epsilon_0_to_fp16 = const()[name = tensor("denom_165_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_165_cast_fp16 = rsqrt(epsilon = denom_165_epsilon_0_to_fp16, x = var_6112_cast_fp16)[name = tensor("denom_165_cast_fp16")]; + tensor out_165_cast_fp16 = mul(x = zero_mean_165_cast_fp16, y = denom_165_cast_fp16)[name = tensor("out_165_cast_fp16")]; + tensor obj_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(1563841408)))]; + 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(1563844032)))]; + 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 var_6127 = const()[name = tensor("op_6127"), val = tensor([1, 1])]; + tensor var_6129 = const()[name = tensor("op_6129"), val = tensor([1, 1])]; + tensor query_111_pad_type_0 = const()[name = tensor("query_111_pad_type_0"), val = tensor("custom")]; + tensor query_111_pad_0 = const()[name = tensor("query_111_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1563846656)))]; + 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(1567123520)))]; + tensor query_111_cast_fp16 = conv(bias = layers_27_encoder_attn_q_proj_bias_to_fp16, dilations = var_6129, groups = var_6005, pad = query_111_pad_0, pad_type = query_111_pad_type_0, strides = var_6127, weight = layers_27_encoder_attn_q_proj_weight_to_fp16, x = obj_387_cast_fp16)[name = tensor("query_111_cast_fp16")]; + tensor var_6133 = const()[name = tensor("op_6133"), val = tensor([1, 1])]; + tensor var_6135 = const()[name = tensor("op_6135"), val = tensor([1, 1])]; + tensor key_111_pad_type_0 = const()[name = tensor("key_111_pad_type_0"), val = tensor("custom")]; + tensor key_111_pad_0 = const()[name = tensor("key_111_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1567126144)))]; + tensor key_111_cast_fp16 = conv(dilations = var_6135, groups = var_6005, pad = key_111_pad_0, pad_type = key_111_pad_type_0, strides = var_6133, weight = layers_27_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_111_cast_fp16")]; + tensor var_6140 = const()[name = tensor("op_6140"), val = tensor([1, 1])]; + tensor var_6142 = const()[name = tensor("op_6142"), val = tensor([1, 1])]; + tensor value_111_pad_type_0 = const()[name = tensor("value_111_pad_type_0"), val = tensor("custom")]; + tensor value_111_pad_0 = const()[name = tensor("value_111_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1570403008)))]; + 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(1573679872)))]; + tensor value_111_cast_fp16 = conv(bias = layers_27_encoder_attn_v_proj_bias_to_fp16, dilations = var_6142, groups = var_6005, pad = value_111_pad_0, pad_type = value_111_pad_type_0, strides = var_6140, weight = layers_27_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_111_cast_fp16")]; + tensor var_6146 = const()[name = tensor("op_6146"), val = tensor([1, 20, 64, -1])]; + tensor var_6147_cast_fp16 = reshape(shape = var_6146, x = query_111_cast_fp16)[name = tensor("op_6147_cast_fp16")]; + tensor var_6148_to_fp16 = const()[name = tensor("op_6148_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6149_cast_fp16 = mul(x = var_6147_cast_fp16, y = var_6148_to_fp16)[name = tensor("op_6149_cast_fp16")]; + tensor var_6150 = const()[name = tensor("op_6150"), val = tensor([1, 20, 64, -1])]; + tensor var_6151_cast_fp16 = reshape(shape = var_6150, x = key_111_cast_fp16)[name = tensor("op_6151_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_6149_cast_fp16, y = var_6151_cast_fp16)[name = tensor("mh_w_167_cast_fp16")]; + tensor obj_391_cast_fp16 = softmax(axis = var_5998, x = mh_w_167_cast_fp16)[name = tensor("obj_391_cast_fp16")]; + tensor var_6155 = const()[name = tensor("op_6155"), val = tensor([1, 20, 64, -1])]; + tensor var_6156_cast_fp16 = reshape(shape = var_6155, x = value_111_cast_fp16)[name = tensor("op_6156_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_6156_cast_fp16, y = obj_391_cast_fp16)[name = tensor("attn_111_cast_fp16")]; + tensor var_6159 = const()[name = tensor("op_6159"), val = tensor([1, 1280, 1, -1])]; + tensor input_273_cast_fp16 = reshape(shape = var_6159, x = attn_111_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor var_6163 = const()[name = tensor("op_6163"), val = tensor([1, 1])]; + tensor var_6165 = const()[name = tensor("op_6165"), val = tensor([1, 1])]; + tensor obj_389_pad_type_0 = const()[name = tensor("obj_389_pad_type_0"), val = tensor("custom")]; + tensor obj_389_pad_0 = const()[name = tensor("obj_389_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1573682496)))]; + 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(1576959360)))]; + tensor obj_389_cast_fp16 = conv(bias = layers_27_encoder_attn_o_proj_bias_to_fp16, dilations = var_6165, groups = var_6005, pad = obj_389_pad_0, pad_type = obj_389_pad_type_0, strides = var_6163, 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 var_6174 = const()[name = tensor("op_6174"), val = tensor([1])]; + tensor channels_mean_167_cast_fp16 = reduce_mean(axes = var_6174, keep_dims = var_6006, x = inputs_167_cast_fp16)[name = tensor("channels_mean_167_cast_fp16")]; + tensor zero_mean_167_cast_fp16 = sub(x = inputs_167_cast_fp16, y = channels_mean_167_cast_fp16)[name = tensor("zero_mean_167_cast_fp16")]; + tensor zero_mean_sq_167_cast_fp16 = mul(x = zero_mean_167_cast_fp16, y = zero_mean_167_cast_fp16)[name = tensor("zero_mean_sq_167_cast_fp16")]; + tensor var_6178 = const()[name = tensor("op_6178"), val = tensor([1])]; + tensor var_6179_cast_fp16 = reduce_mean(axes = var_6178, keep_dims = var_6006, x = zero_mean_sq_167_cast_fp16)[name = tensor("op_6179_cast_fp16")]; + tensor var_6180_to_fp16 = const()[name = tensor("op_6180_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6181_cast_fp16 = add(x = var_6179_cast_fp16, y = var_6180_to_fp16)[name = tensor("op_6181_cast_fp16")]; + tensor denom_167_epsilon_0_to_fp16 = const()[name = tensor("denom_167_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_167_cast_fp16 = rsqrt(epsilon = denom_167_epsilon_0_to_fp16, x = var_6181_cast_fp16)[name = tensor("denom_167_cast_fp16")]; + tensor out_167_cast_fp16 = mul(x = zero_mean_167_cast_fp16, y = denom_167_cast_fp16)[name = tensor("out_167_cast_fp16")]; + tensor input_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(1576961984)))]; + 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(1576964608)))]; + 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 var_6192 = const()[name = tensor("op_6192"), val = tensor([1, 1])]; + tensor var_6194 = const()[name = tensor("op_6194"), val = tensor([1, 1])]; + tensor input_277_pad_type_0 = const()[name = tensor("input_277_pad_type_0"), val = tensor("custom")]; + tensor input_277_pad_0 = const()[name = tensor("input_277_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1576967232)))]; + 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(1590074496)))]; + tensor input_277_cast_fp16 = conv(bias = layers_27_fc1_bias_to_fp16, dilations = var_6194, groups = var_6005, pad = input_277_pad_0, pad_type = input_277_pad_type_0, strides = var_6192, 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 var_6200 = const()[name = tensor("op_6200"), val = tensor([1, 1])]; + tensor var_6202 = const()[name = tensor("op_6202"), val = tensor([1, 1])]; + tensor hidden_states_57_pad_type_0 = const()[name = tensor("hidden_states_57_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_57_pad_0 = const()[name = tensor("hidden_states_57_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1590084800)))]; + 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(1603192064)))]; + tensor hidden_states_57_cast_fp16 = conv(bias = layers_27_fc2_bias_to_fp16, dilations = var_6202, groups = var_6005, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = var_6200, 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_6216 = const()[name = tensor("op_6216"), val = tensor(3)]; + tensor var_6223 = const()[name = tensor("op_6223"), val = tensor(1)]; + tensor var_6224 = const()[name = tensor("op_6224"), val = tensor(true)]; + tensor var_6236 = const()[name = tensor("op_6236"), val = tensor([1])]; + tensor channels_mean_169_cast_fp16 = reduce_mean(axes = var_6236, keep_dims = var_6224, x = inputs_169_cast_fp16)[name = tensor("channels_mean_169_cast_fp16")]; + tensor zero_mean_169_cast_fp16 = sub(x = inputs_169_cast_fp16, y = channels_mean_169_cast_fp16)[name = tensor("zero_mean_169_cast_fp16")]; + tensor zero_mean_sq_169_cast_fp16 = mul(x = zero_mean_169_cast_fp16, y = zero_mean_169_cast_fp16)[name = tensor("zero_mean_sq_169_cast_fp16")]; + tensor var_6240 = const()[name = tensor("op_6240"), val = tensor([1])]; + tensor var_6241_cast_fp16 = reduce_mean(axes = var_6240, keep_dims = var_6224, x = zero_mean_sq_169_cast_fp16)[name = tensor("op_6241_cast_fp16")]; + tensor var_6242_to_fp16 = const()[name = tensor("op_6242_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6243_cast_fp16 = add(x = var_6241_cast_fp16, y = var_6242_to_fp16)[name = tensor("op_6243_cast_fp16")]; + tensor denom_169_epsilon_0_to_fp16 = const()[name = tensor("denom_169_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_169_cast_fp16 = rsqrt(epsilon = denom_169_epsilon_0_to_fp16, x = var_6243_cast_fp16)[name = tensor("denom_169_cast_fp16")]; + tensor out_169_cast_fp16 = mul(x = zero_mean_169_cast_fp16, y = denom_169_cast_fp16)[name = tensor("out_169_cast_fp16")]; + tensor obj_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(1603194688)))]; + 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(1603197312)))]; + 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 var_6258 = const()[name = tensor("op_6258"), val = tensor([1, 1])]; + tensor var_6260 = const()[name = tensor("op_6260"), val = tensor([1, 1])]; + tensor query_113_pad_type_0 = const()[name = tensor("query_113_pad_type_0"), val = tensor("custom")]; + tensor query_113_pad_0 = const()[name = tensor("query_113_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1603199936)))]; + 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(1606476800)))]; + tensor query_113_cast_fp16 = conv(bias = layers_28_self_attn_q_proj_bias_to_fp16, dilations = var_6260, groups = var_6223, pad = query_113_pad_0, pad_type = query_113_pad_type_0, strides = var_6258, weight = layers_28_self_attn_q_proj_weight_to_fp16, x = obj_393_cast_fp16)[name = tensor("query_113_cast_fp16")]; + tensor var_6264 = const()[name = tensor("op_6264"), val = tensor([1, 1])]; + tensor var_6266 = const()[name = tensor("op_6266"), val = tensor([1, 1])]; + tensor current_key_57_pad_type_0 = const()[name = tensor("current_key_57_pad_type_0"), val = tensor("custom")]; + tensor current_key_57_pad_0 = const()[name = tensor("current_key_57_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1606479424)))]; + tensor current_key_57_cast_fp16 = conv(dilations = var_6266, groups = var_6223, pad = current_key_57_pad_0, pad_type = current_key_57_pad_type_0, strides = var_6264, weight = layers_28_self_attn_k_proj_weight_to_fp16, x = obj_393_cast_fp16)[name = tensor("current_key_57_cast_fp16")]; + tensor var_6271 = const()[name = tensor("op_6271"), val = tensor([1, 1])]; + tensor var_6273 = const()[name = tensor("op_6273"), val = tensor([1, 1])]; + tensor current_value_57_pad_type_0 = const()[name = tensor("current_value_57_pad_type_0"), val = tensor("custom")]; + tensor current_value_57_pad_0 = const()[name = tensor("current_value_57_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1609756288)))]; + 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(1613033152)))]; + tensor current_value_57_cast_fp16 = conv(bias = layers_28_self_attn_v_proj_bias_to_fp16, dilations = var_6273, groups = var_6223, pad = current_value_57_pad_0, pad_type = current_value_57_pad_type_0, strides = var_6271, weight = layers_28_self_attn_v_proj_weight_to_fp16, x = obj_393_cast_fp16)[name = tensor("current_value_57_cast_fp16")]; + tensor var_6280_cast_fp16 = mul(x = current_key_57_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6280_cast_fp16")]; + tensor var_6282_cast_fp16 = mul(x = var_103_cast_fp16_28, y = var_241_cast_fp16)[name = tensor("op_6282_cast_fp16")]; + tensor key_113_cast_fp16 = add(x = var_6280_cast_fp16, y = var_6282_cast_fp16)[name = tensor("key_113_cast_fp16")]; + tensor var_6284_cast_fp16 = mul(x = current_value_57_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6284_cast_fp16")]; + tensor var_6286_cast_fp16 = mul(x = var_138_cast_fp16_28, y = var_241_cast_fp16)[name = tensor("op_6286_cast_fp16")]; + tensor value_113_cast_fp16 = add(x = var_6284_cast_fp16, y = var_6286_cast_fp16)[name = tensor("value_113_cast_fp16")]; + tensor var_6289 = const()[name = tensor("op_6289"), val = tensor([1, 20, 64, -1])]; + tensor var_6290_cast_fp16 = reshape(shape = var_6289, x = query_113_cast_fp16)[name = tensor("op_6290_cast_fp16")]; + tensor var_6291_to_fp16 = const()[name = tensor("op_6291_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6292_cast_fp16 = mul(x = var_6290_cast_fp16, y = var_6291_to_fp16)[name = tensor("op_6292_cast_fp16")]; + tensor var_6293 = const()[name = tensor("op_6293"), val = tensor([1, 20, 64, -1])]; + tensor var_6294_cast_fp16 = reshape(shape = var_6293, x = key_113_cast_fp16)[name = tensor("op_6294_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_6292_cast_fp16, y = var_6294_cast_fp16)[name = tensor("mh_w_169_cast_fp16")]; + tensor mh_w_171_cast_fp16 = add(x = mh_w_169_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_171_cast_fp16")]; + tensor var_6302_cast_fp16 = softmax(axis = var_6216, x = mh_w_171_cast_fp16)[name = tensor("op_6302_cast_fp16")]; + tensor var_6303 = const()[name = tensor("op_6303"), val = tensor([1, 20, 64, -1])]; + tensor var_6304_cast_fp16 = reshape(shape = var_6303, x = value_113_cast_fp16)[name = tensor("op_6304_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_6304_cast_fp16, y = var_6302_cast_fp16)[name = tensor("attn_113_cast_fp16")]; + tensor var_6307 = const()[name = tensor("op_6307"), val = tensor([1, 1280, 1, -1])]; + tensor input_281_cast_fp16 = reshape(shape = var_6307, x = attn_113_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor var_6311 = const()[name = tensor("op_6311"), val = tensor([1, 1])]; + tensor var_6313 = const()[name = tensor("op_6313"), val = tensor([1, 1])]; + tensor obj_399_pad_type_0 = const()[name = tensor("obj_399_pad_type_0"), val = tensor("custom")]; + tensor obj_399_pad_0 = const()[name = tensor("obj_399_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1613035776)))]; + 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(1616312640)))]; + tensor obj_399_cast_fp16 = conv(bias = layers_28_self_attn_o_proj_bias_to_fp16, dilations = var_6313, groups = var_6223, pad = obj_399_pad_0, pad_type = obj_399_pad_type_0, strides = var_6311, 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 var_6323 = const()[name = tensor("op_6323"), val = tensor([1])]; + tensor channels_mean_171_cast_fp16 = reduce_mean(axes = var_6323, keep_dims = var_6224, x = inputs_171_cast_fp16)[name = tensor("channels_mean_171_cast_fp16")]; + tensor zero_mean_171_cast_fp16 = sub(x = inputs_171_cast_fp16, y = channels_mean_171_cast_fp16)[name = tensor("zero_mean_171_cast_fp16")]; + tensor zero_mean_sq_171_cast_fp16 = mul(x = zero_mean_171_cast_fp16, y = zero_mean_171_cast_fp16)[name = tensor("zero_mean_sq_171_cast_fp16")]; + tensor var_6327 = const()[name = tensor("op_6327"), val = tensor([1])]; + tensor var_6328_cast_fp16 = reduce_mean(axes = var_6327, keep_dims = var_6224, x = zero_mean_sq_171_cast_fp16)[name = tensor("op_6328_cast_fp16")]; + tensor var_6329_to_fp16 = const()[name = tensor("op_6329_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6330_cast_fp16 = add(x = var_6328_cast_fp16, y = var_6329_to_fp16)[name = tensor("op_6330_cast_fp16")]; + tensor denom_171_epsilon_0_to_fp16 = const()[name = tensor("denom_171_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_171_cast_fp16 = rsqrt(epsilon = denom_171_epsilon_0_to_fp16, x = var_6330_cast_fp16)[name = tensor("denom_171_cast_fp16")]; + tensor out_171_cast_fp16 = mul(x = zero_mean_171_cast_fp16, y = denom_171_cast_fp16)[name = tensor("out_171_cast_fp16")]; + tensor obj_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(1616315264)))]; + 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(1616317888)))]; + 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 var_6345 = const()[name = tensor("op_6345"), val = tensor([1, 1])]; + tensor var_6347 = const()[name = tensor("op_6347"), val = tensor([1, 1])]; + tensor query_115_pad_type_0 = const()[name = tensor("query_115_pad_type_0"), val = tensor("custom")]; + tensor query_115_pad_0 = const()[name = tensor("query_115_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1616320512)))]; + 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(1619597376)))]; + tensor query_115_cast_fp16 = conv(bias = layers_28_encoder_attn_q_proj_bias_to_fp16, dilations = var_6347, groups = var_6223, pad = query_115_pad_0, pad_type = query_115_pad_type_0, strides = var_6345, weight = layers_28_encoder_attn_q_proj_weight_to_fp16, x = obj_401_cast_fp16)[name = tensor("query_115_cast_fp16")]; + tensor var_6351 = const()[name = tensor("op_6351"), val = tensor([1, 1])]; + tensor var_6353 = const()[name = tensor("op_6353"), val = tensor([1, 1])]; + tensor key_115_pad_type_0 = const()[name = tensor("key_115_pad_type_0"), val = tensor("custom")]; + tensor key_115_pad_0 = const()[name = tensor("key_115_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1619600000)))]; + tensor key_115_cast_fp16 = conv(dilations = var_6353, groups = var_6223, pad = key_115_pad_0, pad_type = key_115_pad_type_0, strides = var_6351, weight = layers_28_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_115_cast_fp16")]; + tensor var_6358 = const()[name = tensor("op_6358"), val = tensor([1, 1])]; + tensor var_6360 = const()[name = tensor("op_6360"), val = tensor([1, 1])]; + tensor value_115_pad_type_0 = const()[name = tensor("value_115_pad_type_0"), val = tensor("custom")]; + tensor value_115_pad_0 = const()[name = tensor("value_115_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1622876864)))]; + 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(1626153728)))]; + tensor value_115_cast_fp16 = conv(bias = layers_28_encoder_attn_v_proj_bias_to_fp16, dilations = var_6360, groups = var_6223, pad = value_115_pad_0, pad_type = value_115_pad_type_0, strides = var_6358, weight = layers_28_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_115_cast_fp16")]; + tensor var_6364 = const()[name = tensor("op_6364"), val = tensor([1, 20, 64, -1])]; + tensor var_6365_cast_fp16 = reshape(shape = var_6364, x = query_115_cast_fp16)[name = tensor("op_6365_cast_fp16")]; + tensor var_6366_to_fp16 = const()[name = tensor("op_6366_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6367_cast_fp16 = mul(x = var_6365_cast_fp16, y = var_6366_to_fp16)[name = tensor("op_6367_cast_fp16")]; + tensor var_6368 = const()[name = tensor("op_6368"), val = tensor([1, 20, 64, -1])]; + tensor var_6369_cast_fp16 = reshape(shape = var_6368, x = key_115_cast_fp16)[name = tensor("op_6369_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_6367_cast_fp16, y = var_6369_cast_fp16)[name = tensor("mh_w_173_cast_fp16")]; + tensor obj_405_cast_fp16 = softmax(axis = var_6216, x = mh_w_173_cast_fp16)[name = tensor("obj_405_cast_fp16")]; + tensor var_6373 = const()[name = tensor("op_6373"), val = tensor([1, 20, 64, -1])]; + tensor var_6374_cast_fp16 = reshape(shape = var_6373, x = value_115_cast_fp16)[name = tensor("op_6374_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_6374_cast_fp16, y = obj_405_cast_fp16)[name = tensor("attn_115_cast_fp16")]; + tensor var_6377 = const()[name = tensor("op_6377"), val = tensor([1, 1280, 1, -1])]; + tensor input_283_cast_fp16 = reshape(shape = var_6377, x = attn_115_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor var_6381 = const()[name = tensor("op_6381"), val = tensor([1, 1])]; + tensor var_6383 = const()[name = tensor("op_6383"), val = tensor([1, 1])]; + tensor obj_403_pad_type_0 = const()[name = tensor("obj_403_pad_type_0"), val = tensor("custom")]; + tensor obj_403_pad_0 = const()[name = tensor("obj_403_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1626156352)))]; + 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(1629433216)))]; + tensor obj_403_cast_fp16 = conv(bias = layers_28_encoder_attn_o_proj_bias_to_fp16, dilations = var_6383, groups = var_6223, pad = obj_403_pad_0, pad_type = obj_403_pad_type_0, strides = var_6381, 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 var_6389 = const()[name = tensor("op_6389"), val = tensor([1])]; + tensor channels_mean_173_cast_fp16 = reduce_mean(axes = var_6389, keep_dims = var_6224, x = inputs_173_cast_fp16)[name = tensor("channels_mean_173_cast_fp16")]; + tensor zero_mean_173_cast_fp16 = sub(x = inputs_173_cast_fp16, y = channels_mean_173_cast_fp16)[name = tensor("zero_mean_173_cast_fp16")]; + tensor zero_mean_sq_173_cast_fp16 = mul(x = zero_mean_173_cast_fp16, y = zero_mean_173_cast_fp16)[name = tensor("zero_mean_sq_173_cast_fp16")]; + tensor var_6393 = const()[name = tensor("op_6393"), val = tensor([1])]; + tensor var_6394_cast_fp16 = reduce_mean(axes = var_6393, keep_dims = var_6224, x = zero_mean_sq_173_cast_fp16)[name = tensor("op_6394_cast_fp16")]; + tensor var_6395_to_fp16 = const()[name = tensor("op_6395_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6396_cast_fp16 = add(x = var_6394_cast_fp16, y = var_6395_to_fp16)[name = tensor("op_6396_cast_fp16")]; + tensor denom_173_epsilon_0_to_fp16 = const()[name = tensor("denom_173_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_173_cast_fp16 = rsqrt(epsilon = denom_173_epsilon_0_to_fp16, x = var_6396_cast_fp16)[name = tensor("denom_173_cast_fp16")]; + tensor out_173_cast_fp16 = mul(x = zero_mean_173_cast_fp16, y = denom_173_cast_fp16)[name = tensor("out_173_cast_fp16")]; + tensor input_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(1629435840)))]; + 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(1629438464)))]; + 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 var_6407 = const()[name = tensor("op_6407"), val = tensor([1, 1])]; + tensor var_6409 = const()[name = tensor("op_6409"), val = tensor([1, 1])]; + tensor input_287_pad_type_0 = const()[name = tensor("input_287_pad_type_0"), val = tensor("custom")]; + tensor input_287_pad_0 = const()[name = tensor("input_287_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1629441088)))]; + 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(1642548352)))]; + tensor input_287_cast_fp16 = conv(bias = layers_28_fc1_bias_to_fp16, dilations = var_6409, groups = var_6223, pad = input_287_pad_0, pad_type = input_287_pad_type_0, strides = var_6407, 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 var_6415 = const()[name = tensor("op_6415"), val = tensor([1, 1])]; + tensor var_6417 = const()[name = tensor("op_6417"), val = tensor([1, 1])]; + tensor hidden_states_59_pad_type_0 = const()[name = tensor("hidden_states_59_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_59_pad_0 = const()[name = tensor("hidden_states_59_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1642558656)))]; + 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(1655665920)))]; + tensor hidden_states_59_cast_fp16 = conv(bias = layers_28_fc2_bias_to_fp16, dilations = var_6417, groups = var_6223, pad = hidden_states_59_pad_0, pad_type = hidden_states_59_pad_type_0, strides = var_6415, 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_6430 = const()[name = tensor("op_6430"), val = tensor(3)]; + tensor var_6437 = const()[name = tensor("op_6437"), val = tensor(1)]; + tensor var_6438 = const()[name = tensor("op_6438"), val = tensor(true)]; + tensor var_6450 = const()[name = tensor("op_6450"), val = tensor([1])]; + tensor channels_mean_175_cast_fp16 = reduce_mean(axes = var_6450, keep_dims = var_6438, x = inputs_175_cast_fp16)[name = tensor("channels_mean_175_cast_fp16")]; + tensor zero_mean_175_cast_fp16 = sub(x = inputs_175_cast_fp16, y = channels_mean_175_cast_fp16)[name = tensor("zero_mean_175_cast_fp16")]; + tensor zero_mean_sq_175_cast_fp16 = mul(x = zero_mean_175_cast_fp16, y = zero_mean_175_cast_fp16)[name = tensor("zero_mean_sq_175_cast_fp16")]; + tensor var_6454 = const()[name = tensor("op_6454"), val = tensor([1])]; + tensor var_6455_cast_fp16 = reduce_mean(axes = var_6454, keep_dims = var_6438, x = zero_mean_sq_175_cast_fp16)[name = tensor("op_6455_cast_fp16")]; + tensor var_6456_to_fp16 = const()[name = tensor("op_6456_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6457_cast_fp16 = add(x = var_6455_cast_fp16, y = var_6456_to_fp16)[name = tensor("op_6457_cast_fp16")]; + tensor denom_175_epsilon_0_to_fp16 = const()[name = tensor("denom_175_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_175_cast_fp16 = rsqrt(epsilon = denom_175_epsilon_0_to_fp16, x = var_6457_cast_fp16)[name = tensor("denom_175_cast_fp16")]; + tensor out_175_cast_fp16 = mul(x = zero_mean_175_cast_fp16, y = denom_175_cast_fp16)[name = tensor("out_175_cast_fp16")]; + tensor obj_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(1655668544)))]; + 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(1655671168)))]; + 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 var_6472 = const()[name = tensor("op_6472"), val = tensor([1, 1])]; + tensor var_6474 = const()[name = tensor("op_6474"), val = tensor([1, 1])]; + tensor query_117_pad_type_0 = const()[name = tensor("query_117_pad_type_0"), val = tensor("custom")]; + tensor query_117_pad_0 = const()[name = tensor("query_117_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1655673792)))]; + 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(1658950656)))]; + tensor query_117_cast_fp16 = conv(bias = layers_29_self_attn_q_proj_bias_to_fp16, dilations = var_6474, groups = var_6437, pad = query_117_pad_0, pad_type = query_117_pad_type_0, strides = var_6472, weight = layers_29_self_attn_q_proj_weight_to_fp16, x = obj_407_cast_fp16)[name = tensor("query_117_cast_fp16")]; + tensor var_6478 = const()[name = tensor("op_6478"), val = tensor([1, 1])]; + tensor var_6480 = const()[name = tensor("op_6480"), val = tensor([1, 1])]; + tensor current_key_59_pad_type_0 = const()[name = tensor("current_key_59_pad_type_0"), val = tensor("custom")]; + tensor current_key_59_pad_0 = const()[name = tensor("current_key_59_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1658953280)))]; + tensor current_key_59_cast_fp16 = conv(dilations = var_6480, groups = var_6437, pad = current_key_59_pad_0, pad_type = current_key_59_pad_type_0, strides = var_6478, weight = layers_29_self_attn_k_proj_weight_to_fp16, x = obj_407_cast_fp16)[name = tensor("current_key_59_cast_fp16")]; + tensor var_6485 = const()[name = tensor("op_6485"), val = tensor([1, 1])]; + tensor var_6487 = const()[name = tensor("op_6487"), val = tensor([1, 1])]; + tensor current_value_59_pad_type_0 = const()[name = tensor("current_value_59_pad_type_0"), val = tensor("custom")]; + tensor current_value_59_pad_0 = const()[name = tensor("current_value_59_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1662230144)))]; + 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(1665507008)))]; + tensor current_value_59_cast_fp16 = conv(bias = layers_29_self_attn_v_proj_bias_to_fp16, dilations = var_6487, groups = var_6437, pad = current_value_59_pad_0, pad_type = current_value_59_pad_type_0, strides = var_6485, weight = layers_29_self_attn_v_proj_weight_to_fp16, x = obj_407_cast_fp16)[name = tensor("current_value_59_cast_fp16")]; + tensor var_6494_cast_fp16 = mul(x = current_key_59_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6494_cast_fp16")]; + tensor var_6496_cast_fp16 = mul(x = var_103_cast_fp16_29, y = var_241_cast_fp16)[name = tensor("op_6496_cast_fp16")]; + tensor key_117_cast_fp16 = add(x = var_6494_cast_fp16, y = var_6496_cast_fp16)[name = tensor("key_117_cast_fp16")]; + tensor var_6498_cast_fp16 = mul(x = current_value_59_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6498_cast_fp16")]; + tensor var_6500_cast_fp16 = mul(x = var_138_cast_fp16_29, y = var_241_cast_fp16)[name = tensor("op_6500_cast_fp16")]; + tensor value_117_cast_fp16 = add(x = var_6498_cast_fp16, y = var_6500_cast_fp16)[name = tensor("value_117_cast_fp16")]; + tensor var_6503 = const()[name = tensor("op_6503"), val = tensor([1, 20, 64, -1])]; + tensor var_6504_cast_fp16 = reshape(shape = var_6503, x = query_117_cast_fp16)[name = tensor("op_6504_cast_fp16")]; + tensor var_6505_to_fp16 = const()[name = tensor("op_6505_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6506_cast_fp16 = mul(x = var_6504_cast_fp16, y = var_6505_to_fp16)[name = tensor("op_6506_cast_fp16")]; + tensor var_6507 = const()[name = tensor("op_6507"), val = tensor([1, 20, 64, -1])]; + tensor var_6508_cast_fp16 = reshape(shape = var_6507, x = key_117_cast_fp16)[name = tensor("op_6508_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_6506_cast_fp16, y = var_6508_cast_fp16)[name = tensor("mh_w_175_cast_fp16")]; + tensor mh_w_177_cast_fp16 = add(x = mh_w_175_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_177_cast_fp16")]; + tensor var_6516_cast_fp16 = softmax(axis = var_6430, x = mh_w_177_cast_fp16)[name = tensor("op_6516_cast_fp16")]; + tensor var_6517 = const()[name = tensor("op_6517"), val = tensor([1, 20, 64, -1])]; + tensor var_6518_cast_fp16 = reshape(shape = var_6517, x = value_117_cast_fp16)[name = tensor("op_6518_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_6518_cast_fp16, y = var_6516_cast_fp16)[name = tensor("attn_117_cast_fp16")]; + tensor var_6521 = const()[name = tensor("op_6521"), val = tensor([1, 1280, 1, -1])]; + tensor input_291_cast_fp16 = reshape(shape = var_6521, x = attn_117_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor var_6525 = const()[name = tensor("op_6525"), val = tensor([1, 1])]; + tensor var_6527 = const()[name = tensor("op_6527"), val = tensor([1, 1])]; + tensor obj_413_pad_type_0 = const()[name = tensor("obj_413_pad_type_0"), val = tensor("custom")]; + tensor obj_413_pad_0 = const()[name = tensor("obj_413_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1665509632)))]; + 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(1668786496)))]; + tensor obj_413_cast_fp16 = conv(bias = layers_29_self_attn_o_proj_bias_to_fp16, dilations = var_6527, groups = var_6437, pad = obj_413_pad_0, pad_type = obj_413_pad_type_0, strides = var_6525, 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 var_6537 = const()[name = tensor("op_6537"), val = tensor([1])]; + tensor channels_mean_177_cast_fp16 = reduce_mean(axes = var_6537, keep_dims = var_6438, x = inputs_177_cast_fp16)[name = tensor("channels_mean_177_cast_fp16")]; + tensor zero_mean_177_cast_fp16 = sub(x = inputs_177_cast_fp16, y = channels_mean_177_cast_fp16)[name = tensor("zero_mean_177_cast_fp16")]; + tensor zero_mean_sq_177_cast_fp16 = mul(x = zero_mean_177_cast_fp16, y = zero_mean_177_cast_fp16)[name = tensor("zero_mean_sq_177_cast_fp16")]; + tensor var_6541 = const()[name = tensor("op_6541"), val = tensor([1])]; + tensor var_6542_cast_fp16 = reduce_mean(axes = var_6541, keep_dims = var_6438, x = zero_mean_sq_177_cast_fp16)[name = tensor("op_6542_cast_fp16")]; + tensor var_6543_to_fp16 = const()[name = tensor("op_6543_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6544_cast_fp16 = add(x = var_6542_cast_fp16, y = var_6543_to_fp16)[name = tensor("op_6544_cast_fp16")]; + tensor denom_177_epsilon_0_to_fp16 = const()[name = tensor("denom_177_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_177_cast_fp16 = rsqrt(epsilon = denom_177_epsilon_0_to_fp16, x = var_6544_cast_fp16)[name = tensor("denom_177_cast_fp16")]; + tensor out_177_cast_fp16 = mul(x = zero_mean_177_cast_fp16, y = denom_177_cast_fp16)[name = tensor("out_177_cast_fp16")]; + tensor obj_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(1668789120)))]; + 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(1668791744)))]; + 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 var_6559 = const()[name = tensor("op_6559"), val = tensor([1, 1])]; + tensor var_6561 = const()[name = tensor("op_6561"), val = tensor([1, 1])]; + tensor query_119_pad_type_0 = const()[name = tensor("query_119_pad_type_0"), val = tensor("custom")]; + tensor query_119_pad_0 = const()[name = tensor("query_119_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1668794368)))]; + 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(1672071232)))]; + tensor query_119_cast_fp16 = conv(bias = layers_29_encoder_attn_q_proj_bias_to_fp16, dilations = var_6561, groups = var_6437, pad = query_119_pad_0, pad_type = query_119_pad_type_0, strides = var_6559, weight = layers_29_encoder_attn_q_proj_weight_to_fp16, x = obj_415_cast_fp16)[name = tensor("query_119_cast_fp16")]; + tensor var_6565 = const()[name = tensor("op_6565"), val = tensor([1, 1])]; + tensor var_6567 = const()[name = tensor("op_6567"), val = tensor([1, 1])]; + tensor key_119_pad_type_0 = const()[name = tensor("key_119_pad_type_0"), val = tensor("custom")]; + tensor key_119_pad_0 = const()[name = tensor("key_119_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1672073856)))]; + tensor key_119_cast_fp16 = conv(dilations = var_6567, groups = var_6437, pad = key_119_pad_0, pad_type = key_119_pad_type_0, strides = var_6565, weight = layers_29_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_119_cast_fp16")]; + tensor var_6572 = const()[name = tensor("op_6572"), val = tensor([1, 1])]; + tensor var_6574 = const()[name = tensor("op_6574"), val = tensor([1, 1])]; + tensor value_119_pad_type_0 = const()[name = tensor("value_119_pad_type_0"), val = tensor("custom")]; + tensor value_119_pad_0 = const()[name = tensor("value_119_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1675350720)))]; + 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(1678627584)))]; + tensor value_119_cast_fp16 = conv(bias = layers_29_encoder_attn_v_proj_bias_to_fp16, dilations = var_6574, groups = var_6437, pad = value_119_pad_0, pad_type = value_119_pad_type_0, strides = var_6572, weight = layers_29_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_119_cast_fp16")]; + tensor var_6578 = const()[name = tensor("op_6578"), val = tensor([1, 20, 64, -1])]; + tensor var_6579_cast_fp16 = reshape(shape = var_6578, x = query_119_cast_fp16)[name = tensor("op_6579_cast_fp16")]; + tensor var_6580_to_fp16 = const()[name = tensor("op_6580_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6581_cast_fp16 = mul(x = var_6579_cast_fp16, y = var_6580_to_fp16)[name = tensor("op_6581_cast_fp16")]; + tensor var_6582 = const()[name = tensor("op_6582"), val = tensor([1, 20, 64, -1])]; + tensor var_6583_cast_fp16 = reshape(shape = var_6582, x = key_119_cast_fp16)[name = tensor("op_6583_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_6581_cast_fp16, y = var_6583_cast_fp16)[name = tensor("mh_w_179_cast_fp16")]; + tensor obj_419_cast_fp16 = softmax(axis = var_6430, x = mh_w_179_cast_fp16)[name = tensor("obj_419_cast_fp16")]; + tensor var_6587 = const()[name = tensor("op_6587"), val = tensor([1, 20, 64, -1])]; + tensor var_6588_cast_fp16 = reshape(shape = var_6587, x = value_119_cast_fp16)[name = tensor("op_6588_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_6588_cast_fp16, y = obj_419_cast_fp16)[name = tensor("attn_119_cast_fp16")]; + tensor var_6591 = const()[name = tensor("op_6591"), val = tensor([1, 1280, 1, -1])]; + tensor input_293_cast_fp16 = reshape(shape = var_6591, x = attn_119_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor var_6595 = const()[name = tensor("op_6595"), val = tensor([1, 1])]; + tensor var_6597 = const()[name = tensor("op_6597"), val = tensor([1, 1])]; + tensor obj_417_pad_type_0 = const()[name = tensor("obj_417_pad_type_0"), val = tensor("custom")]; + tensor obj_417_pad_0 = const()[name = tensor("obj_417_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1678630208)))]; + 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(1681907072)))]; + tensor obj_417_cast_fp16 = conv(bias = layers_29_encoder_attn_o_proj_bias_to_fp16, dilations = var_6597, groups = var_6437, pad = obj_417_pad_0, pad_type = obj_417_pad_type_0, strides = var_6595, 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 var_6603 = const()[name = tensor("op_6603"), val = tensor([1])]; + tensor channels_mean_179_cast_fp16 = reduce_mean(axes = var_6603, keep_dims = var_6438, x = inputs_179_cast_fp16)[name = tensor("channels_mean_179_cast_fp16")]; + tensor zero_mean_179_cast_fp16 = sub(x = inputs_179_cast_fp16, y = channels_mean_179_cast_fp16)[name = tensor("zero_mean_179_cast_fp16")]; + tensor zero_mean_sq_179_cast_fp16 = mul(x = zero_mean_179_cast_fp16, y = zero_mean_179_cast_fp16)[name = tensor("zero_mean_sq_179_cast_fp16")]; + tensor var_6607 = const()[name = tensor("op_6607"), val = tensor([1])]; + tensor var_6608_cast_fp16 = reduce_mean(axes = var_6607, keep_dims = var_6438, x = zero_mean_sq_179_cast_fp16)[name = tensor("op_6608_cast_fp16")]; + tensor var_6609_to_fp16 = const()[name = tensor("op_6609_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6610_cast_fp16 = add(x = var_6608_cast_fp16, y = var_6609_to_fp16)[name = tensor("op_6610_cast_fp16")]; + tensor denom_179_epsilon_0_to_fp16 = const()[name = tensor("denom_179_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_179_cast_fp16 = rsqrt(epsilon = denom_179_epsilon_0_to_fp16, x = var_6610_cast_fp16)[name = tensor("denom_179_cast_fp16")]; + tensor out_179_cast_fp16 = mul(x = zero_mean_179_cast_fp16, y = denom_179_cast_fp16)[name = tensor("out_179_cast_fp16")]; + tensor input_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(1681909696)))]; + 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(1681912320)))]; + 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 var_6621 = const()[name = tensor("op_6621"), val = tensor([1, 1])]; + tensor var_6623 = const()[name = tensor("op_6623"), val = tensor([1, 1])]; + tensor input_297_pad_type_0 = const()[name = tensor("input_297_pad_type_0"), val = tensor("custom")]; + tensor input_297_pad_0 = const()[name = tensor("input_297_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1681914944)))]; + 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(1695022208)))]; + tensor input_297_cast_fp16 = conv(bias = layers_29_fc1_bias_to_fp16, dilations = var_6623, groups = var_6437, pad = input_297_pad_0, pad_type = input_297_pad_type_0, strides = var_6621, 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 var_6629 = const()[name = tensor("op_6629"), val = tensor([1, 1])]; + tensor var_6631 = const()[name = tensor("op_6631"), val = tensor([1, 1])]; + tensor hidden_states_61_pad_type_0 = const()[name = tensor("hidden_states_61_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_61_pad_0 = const()[name = tensor("hidden_states_61_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1695032512)))]; + 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(1708139776)))]; + tensor hidden_states_61_cast_fp16 = conv(bias = layers_29_fc2_bias_to_fp16, dilations = var_6631, groups = var_6437, pad = hidden_states_61_pad_0, pad_type = hidden_states_61_pad_type_0, strides = var_6629, 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_6644 = const()[name = tensor("op_6644"), val = tensor(3)]; + tensor var_6651 = const()[name = tensor("op_6651"), val = tensor(1)]; + tensor var_6652 = const()[name = tensor("op_6652"), val = tensor(true)]; + tensor var_6664 = const()[name = tensor("op_6664"), val = tensor([1])]; + tensor channels_mean_181_cast_fp16 = reduce_mean(axes = var_6664, keep_dims = var_6652, x = inputs_181_cast_fp16)[name = tensor("channels_mean_181_cast_fp16")]; + tensor zero_mean_181_cast_fp16 = sub(x = inputs_181_cast_fp16, y = channels_mean_181_cast_fp16)[name = tensor("zero_mean_181_cast_fp16")]; + tensor zero_mean_sq_181_cast_fp16 = mul(x = zero_mean_181_cast_fp16, y = zero_mean_181_cast_fp16)[name = tensor("zero_mean_sq_181_cast_fp16")]; + tensor var_6668 = const()[name = tensor("op_6668"), val = tensor([1])]; + tensor var_6669_cast_fp16 = reduce_mean(axes = var_6668, keep_dims = var_6652, x = zero_mean_sq_181_cast_fp16)[name = tensor("op_6669_cast_fp16")]; + tensor var_6670_to_fp16 = const()[name = tensor("op_6670_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6671_cast_fp16 = add(x = var_6669_cast_fp16, y = var_6670_to_fp16)[name = tensor("op_6671_cast_fp16")]; + tensor denom_181_epsilon_0_to_fp16 = const()[name = tensor("denom_181_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_181_cast_fp16 = rsqrt(epsilon = denom_181_epsilon_0_to_fp16, x = var_6671_cast_fp16)[name = tensor("denom_181_cast_fp16")]; + tensor out_181_cast_fp16 = mul(x = zero_mean_181_cast_fp16, y = denom_181_cast_fp16)[name = tensor("out_181_cast_fp16")]; + tensor obj_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(1708142400)))]; + 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(1708145024)))]; + 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 var_6686 = const()[name = tensor("op_6686"), val = tensor([1, 1])]; + tensor var_6688 = const()[name = tensor("op_6688"), val = tensor([1, 1])]; + tensor query_121_pad_type_0 = const()[name = tensor("query_121_pad_type_0"), val = tensor("custom")]; + tensor query_121_pad_0 = const()[name = tensor("query_121_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1708147648)))]; + 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(1711424512)))]; + tensor query_121_cast_fp16 = conv(bias = layers_30_self_attn_q_proj_bias_to_fp16, dilations = var_6688, groups = var_6651, pad = query_121_pad_0, pad_type = query_121_pad_type_0, strides = var_6686, weight = layers_30_self_attn_q_proj_weight_to_fp16, x = obj_421_cast_fp16)[name = tensor("query_121_cast_fp16")]; + tensor var_6692 = const()[name = tensor("op_6692"), val = tensor([1, 1])]; + tensor var_6694 = const()[name = tensor("op_6694"), val = tensor([1, 1])]; + tensor current_key_61_pad_type_0 = const()[name = tensor("current_key_61_pad_type_0"), val = tensor("custom")]; + tensor current_key_61_pad_0 = const()[name = tensor("current_key_61_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1711427136)))]; + tensor current_key_61_cast_fp16 = conv(dilations = var_6694, groups = var_6651, pad = current_key_61_pad_0, pad_type = current_key_61_pad_type_0, strides = var_6692, weight = layers_30_self_attn_k_proj_weight_to_fp16, x = obj_421_cast_fp16)[name = tensor("current_key_61_cast_fp16")]; + tensor var_6699 = const()[name = tensor("op_6699"), val = tensor([1, 1])]; + tensor var_6701 = const()[name = tensor("op_6701"), val = tensor([1, 1])]; + tensor current_value_61_pad_type_0 = const()[name = tensor("current_value_61_pad_type_0"), val = tensor("custom")]; + tensor current_value_61_pad_0 = const()[name = tensor("current_value_61_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1714704000)))]; + 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(1717980864)))]; + tensor current_value_61_cast_fp16 = conv(bias = layers_30_self_attn_v_proj_bias_to_fp16, dilations = var_6701, groups = var_6651, pad = current_value_61_pad_0, pad_type = current_value_61_pad_type_0, strides = var_6699, weight = layers_30_self_attn_v_proj_weight_to_fp16, x = obj_421_cast_fp16)[name = tensor("current_value_61_cast_fp16")]; + tensor var_6708_cast_fp16 = mul(x = current_key_61_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6708_cast_fp16")]; + tensor var_6710_cast_fp16 = mul(x = var_103_cast_fp16_30, y = var_241_cast_fp16)[name = tensor("op_6710_cast_fp16")]; + tensor key_121_cast_fp16 = add(x = var_6708_cast_fp16, y = var_6710_cast_fp16)[name = tensor("key_121_cast_fp16")]; + tensor var_6712_cast_fp16 = mul(x = current_value_61_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6712_cast_fp16")]; + tensor var_6714_cast_fp16 = mul(x = var_138_cast_fp16_30, y = var_241_cast_fp16)[name = tensor("op_6714_cast_fp16")]; + tensor value_121_cast_fp16 = add(x = var_6712_cast_fp16, y = var_6714_cast_fp16)[name = tensor("value_121_cast_fp16")]; + tensor var_6717 = const()[name = tensor("op_6717"), val = tensor([1, 20, 64, -1])]; + tensor var_6718_cast_fp16 = reshape(shape = var_6717, x = query_121_cast_fp16)[name = tensor("op_6718_cast_fp16")]; + tensor var_6719_to_fp16 = const()[name = tensor("op_6719_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6720_cast_fp16 = mul(x = var_6718_cast_fp16, y = var_6719_to_fp16)[name = tensor("op_6720_cast_fp16")]; + tensor var_6721 = const()[name = tensor("op_6721"), val = tensor([1, 20, 64, -1])]; + tensor var_6722_cast_fp16 = reshape(shape = var_6721, x = key_121_cast_fp16)[name = tensor("op_6722_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_6720_cast_fp16, y = var_6722_cast_fp16)[name = tensor("mh_w_181_cast_fp16")]; + tensor mh_w_183_cast_fp16 = add(x = mh_w_181_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_183_cast_fp16")]; + tensor var_6730_cast_fp16 = softmax(axis = var_6644, x = mh_w_183_cast_fp16)[name = tensor("op_6730_cast_fp16")]; + tensor var_6731 = const()[name = tensor("op_6731"), val = tensor([1, 20, 64, -1])]; + tensor var_6732_cast_fp16 = reshape(shape = var_6731, x = value_121_cast_fp16)[name = tensor("op_6732_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_6732_cast_fp16, y = var_6730_cast_fp16)[name = tensor("attn_121_cast_fp16")]; + tensor var_6735 = const()[name = tensor("op_6735"), val = tensor([1, 1280, 1, -1])]; + tensor input_301_cast_fp16 = reshape(shape = var_6735, x = attn_121_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor var_6739 = const()[name = tensor("op_6739"), val = tensor([1, 1])]; + tensor var_6741 = const()[name = tensor("op_6741"), val = tensor([1, 1])]; + tensor obj_427_pad_type_0 = const()[name = tensor("obj_427_pad_type_0"), val = tensor("custom")]; + tensor obj_427_pad_0 = const()[name = tensor("obj_427_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1717983488)))]; + 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(1721260352)))]; + tensor obj_427_cast_fp16 = conv(bias = layers_30_self_attn_o_proj_bias_to_fp16, dilations = var_6741, groups = var_6651, pad = obj_427_pad_0, pad_type = obj_427_pad_type_0, strides = var_6739, 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 var_6751 = const()[name = tensor("op_6751"), val = tensor([1])]; + tensor channels_mean_183_cast_fp16 = reduce_mean(axes = var_6751, keep_dims = var_6652, x = inputs_183_cast_fp16)[name = tensor("channels_mean_183_cast_fp16")]; + tensor zero_mean_183_cast_fp16 = sub(x = inputs_183_cast_fp16, y = channels_mean_183_cast_fp16)[name = tensor("zero_mean_183_cast_fp16")]; + tensor zero_mean_sq_183_cast_fp16 = mul(x = zero_mean_183_cast_fp16, y = zero_mean_183_cast_fp16)[name = tensor("zero_mean_sq_183_cast_fp16")]; + tensor var_6755 = const()[name = tensor("op_6755"), val = tensor([1])]; + tensor var_6756_cast_fp16 = reduce_mean(axes = var_6755, keep_dims = var_6652, x = zero_mean_sq_183_cast_fp16)[name = tensor("op_6756_cast_fp16")]; + tensor var_6757_to_fp16 = const()[name = tensor("op_6757_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6758_cast_fp16 = add(x = var_6756_cast_fp16, y = var_6757_to_fp16)[name = tensor("op_6758_cast_fp16")]; + tensor denom_183_epsilon_0_to_fp16 = const()[name = tensor("denom_183_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_183_cast_fp16 = rsqrt(epsilon = denom_183_epsilon_0_to_fp16, x = var_6758_cast_fp16)[name = tensor("denom_183_cast_fp16")]; + tensor out_183_cast_fp16 = mul(x = zero_mean_183_cast_fp16, y = denom_183_cast_fp16)[name = tensor("out_183_cast_fp16")]; + tensor obj_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(1721262976)))]; + 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(1721265600)))]; + 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 var_6773 = const()[name = tensor("op_6773"), val = tensor([1, 1])]; + tensor var_6775 = const()[name = tensor("op_6775"), val = tensor([1, 1])]; + tensor query_123_pad_type_0 = const()[name = tensor("query_123_pad_type_0"), val = tensor("custom")]; + tensor query_123_pad_0 = const()[name = tensor("query_123_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1721268224)))]; + 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(1724545088)))]; + tensor query_123_cast_fp16 = conv(bias = layers_30_encoder_attn_q_proj_bias_to_fp16, dilations = var_6775, groups = var_6651, pad = query_123_pad_0, pad_type = query_123_pad_type_0, strides = var_6773, weight = layers_30_encoder_attn_q_proj_weight_to_fp16, x = obj_429_cast_fp16)[name = tensor("query_123_cast_fp16")]; + tensor var_6779 = const()[name = tensor("op_6779"), val = tensor([1, 1])]; + tensor var_6781 = const()[name = tensor("op_6781"), val = tensor([1, 1])]; + tensor key_123_pad_type_0 = const()[name = tensor("key_123_pad_type_0"), val = tensor("custom")]; + tensor key_123_pad_0 = const()[name = tensor("key_123_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1724547712)))]; + tensor key_123_cast_fp16 = conv(dilations = var_6781, groups = var_6651, pad = key_123_pad_0, pad_type = key_123_pad_type_0, strides = var_6779, weight = layers_30_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_123_cast_fp16")]; + tensor var_6786 = const()[name = tensor("op_6786"), val = tensor([1, 1])]; + tensor var_6788 = const()[name = tensor("op_6788"), val = tensor([1, 1])]; + tensor value_123_pad_type_0 = const()[name = tensor("value_123_pad_type_0"), val = tensor("custom")]; + tensor value_123_pad_0 = const()[name = tensor("value_123_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1727824576)))]; + 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(1731101440)))]; + tensor value_123_cast_fp16 = conv(bias = layers_30_encoder_attn_v_proj_bias_to_fp16, dilations = var_6788, groups = var_6651, pad = value_123_pad_0, pad_type = value_123_pad_type_0, strides = var_6786, weight = layers_30_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_123_cast_fp16")]; + tensor var_6792 = const()[name = tensor("op_6792"), val = tensor([1, 20, 64, -1])]; + tensor var_6793_cast_fp16 = reshape(shape = var_6792, x = query_123_cast_fp16)[name = tensor("op_6793_cast_fp16")]; + tensor var_6794_to_fp16 = const()[name = tensor("op_6794_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6795_cast_fp16 = mul(x = var_6793_cast_fp16, y = var_6794_to_fp16)[name = tensor("op_6795_cast_fp16")]; + tensor var_6796 = const()[name = tensor("op_6796"), val = tensor([1, 20, 64, -1])]; + tensor var_6797_cast_fp16 = reshape(shape = var_6796, x = key_123_cast_fp16)[name = tensor("op_6797_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_6795_cast_fp16, y = var_6797_cast_fp16)[name = tensor("mh_w_185_cast_fp16")]; + tensor obj_433_cast_fp16 = softmax(axis = var_6644, x = mh_w_185_cast_fp16)[name = tensor("obj_433_cast_fp16")]; + tensor var_6801 = const()[name = tensor("op_6801"), val = tensor([1, 20, 64, -1])]; + tensor var_6802_cast_fp16 = reshape(shape = var_6801, x = value_123_cast_fp16)[name = tensor("op_6802_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_6802_cast_fp16, y = obj_433_cast_fp16)[name = tensor("attn_123_cast_fp16")]; + tensor var_6805 = const()[name = tensor("op_6805"), val = tensor([1, 1280, 1, -1])]; + tensor input_303_cast_fp16 = reshape(shape = var_6805, x = attn_123_cast_fp16)[name = tensor("input_303_cast_fp16")]; + tensor var_6809 = const()[name = tensor("op_6809"), val = tensor([1, 1])]; + tensor var_6811 = const()[name = tensor("op_6811"), val = tensor([1, 1])]; + tensor obj_431_pad_type_0 = const()[name = tensor("obj_431_pad_type_0"), val = tensor("custom")]; + tensor obj_431_pad_0 = const()[name = tensor("obj_431_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1731104064)))]; + 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(1734380928)))]; + tensor obj_431_cast_fp16 = conv(bias = layers_30_encoder_attn_o_proj_bias_to_fp16, dilations = var_6811, groups = var_6651, pad = obj_431_pad_0, pad_type = obj_431_pad_type_0, strides = var_6809, 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 var_6817 = const()[name = tensor("op_6817"), val = tensor([1])]; + tensor channels_mean_185_cast_fp16 = reduce_mean(axes = var_6817, keep_dims = var_6652, x = inputs_185_cast_fp16)[name = tensor("channels_mean_185_cast_fp16")]; + tensor zero_mean_185_cast_fp16 = sub(x = inputs_185_cast_fp16, y = channels_mean_185_cast_fp16)[name = tensor("zero_mean_185_cast_fp16")]; + tensor zero_mean_sq_185_cast_fp16 = mul(x = zero_mean_185_cast_fp16, y = zero_mean_185_cast_fp16)[name = tensor("zero_mean_sq_185_cast_fp16")]; + tensor var_6821 = const()[name = tensor("op_6821"), val = tensor([1])]; + tensor var_6822_cast_fp16 = reduce_mean(axes = var_6821, keep_dims = var_6652, x = zero_mean_sq_185_cast_fp16)[name = tensor("op_6822_cast_fp16")]; + tensor var_6823_to_fp16 = const()[name = tensor("op_6823_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6824_cast_fp16 = add(x = var_6822_cast_fp16, y = var_6823_to_fp16)[name = tensor("op_6824_cast_fp16")]; + tensor denom_185_epsilon_0_to_fp16 = const()[name = tensor("denom_185_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_185_cast_fp16 = rsqrt(epsilon = denom_185_epsilon_0_to_fp16, x = var_6824_cast_fp16)[name = tensor("denom_185_cast_fp16")]; + tensor out_185_cast_fp16 = mul(x = zero_mean_185_cast_fp16, y = denom_185_cast_fp16)[name = tensor("out_185_cast_fp16")]; + tensor input_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(1734383552)))]; + 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(1734386176)))]; + 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 var_6835 = const()[name = tensor("op_6835"), val = tensor([1, 1])]; + tensor var_6837 = const()[name = tensor("op_6837"), val = tensor([1, 1])]; + tensor input_307_pad_type_0 = const()[name = tensor("input_307_pad_type_0"), val = tensor("custom")]; + tensor input_307_pad_0 = const()[name = tensor("input_307_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1734388800)))]; + 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(1747496064)))]; + tensor input_307_cast_fp16 = conv(bias = layers_30_fc1_bias_to_fp16, dilations = var_6837, groups = var_6651, pad = input_307_pad_0, pad_type = input_307_pad_type_0, strides = var_6835, 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 var_6843 = const()[name = tensor("op_6843"), val = tensor([1, 1])]; + tensor var_6845 = const()[name = tensor("op_6845"), val = tensor([1, 1])]; + tensor hidden_states_63_pad_type_0 = const()[name = tensor("hidden_states_63_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_63_pad_0 = const()[name = tensor("hidden_states_63_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1747506368)))]; + 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(1760613632)))]; + tensor hidden_states_63_cast_fp16 = conv(bias = layers_30_fc2_bias_to_fp16, dilations = var_6845, groups = var_6651, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = var_6843, 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_6858 = const()[name = tensor("op_6858"), val = tensor(3)]; + tensor var_6865 = const()[name = tensor("op_6865"), val = tensor(1)]; + tensor var_6866 = const()[name = tensor("op_6866"), val = tensor(true)]; + tensor var_6878 = const()[name = tensor("op_6878"), val = tensor([1])]; + tensor channels_mean_187_cast_fp16 = reduce_mean(axes = var_6878, keep_dims = var_6866, x = inputs_187_cast_fp16)[name = tensor("channels_mean_187_cast_fp16")]; + tensor zero_mean_187_cast_fp16 = sub(x = inputs_187_cast_fp16, y = channels_mean_187_cast_fp16)[name = tensor("zero_mean_187_cast_fp16")]; + tensor zero_mean_sq_187_cast_fp16 = mul(x = zero_mean_187_cast_fp16, y = zero_mean_187_cast_fp16)[name = tensor("zero_mean_sq_187_cast_fp16")]; + tensor var_6882 = const()[name = tensor("op_6882"), val = tensor([1])]; + tensor var_6883_cast_fp16 = reduce_mean(axes = var_6882, keep_dims = var_6866, x = zero_mean_sq_187_cast_fp16)[name = tensor("op_6883_cast_fp16")]; + tensor var_6884_to_fp16 = const()[name = tensor("op_6884_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6885_cast_fp16 = add(x = var_6883_cast_fp16, y = var_6884_to_fp16)[name = tensor("op_6885_cast_fp16")]; + tensor denom_187_epsilon_0_to_fp16 = const()[name = tensor("denom_187_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_187_cast_fp16 = rsqrt(epsilon = denom_187_epsilon_0_to_fp16, x = var_6885_cast_fp16)[name = tensor("denom_187_cast_fp16")]; + tensor out_187_cast_fp16 = mul(x = zero_mean_187_cast_fp16, y = denom_187_cast_fp16)[name = tensor("out_187_cast_fp16")]; + tensor obj_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(1760616256)))]; + 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(1760618880)))]; + 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 var_6900 = const()[name = tensor("op_6900"), val = tensor([1, 1])]; + tensor var_6902 = const()[name = tensor("op_6902"), val = tensor([1, 1])]; + tensor query_125_pad_type_0 = const()[name = tensor("query_125_pad_type_0"), val = tensor("custom")]; + tensor query_125_pad_0 = const()[name = tensor("query_125_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1760621504)))]; + 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(1763898368)))]; + tensor query_125_cast_fp16 = conv(bias = layers_31_self_attn_q_proj_bias_to_fp16, dilations = var_6902, groups = var_6865, pad = query_125_pad_0, pad_type = query_125_pad_type_0, strides = var_6900, weight = layers_31_self_attn_q_proj_weight_to_fp16, x = obj_435_cast_fp16)[name = tensor("query_125_cast_fp16")]; + tensor var_6906 = const()[name = tensor("op_6906"), val = tensor([1, 1])]; + tensor var_6908 = const()[name = tensor("op_6908"), val = tensor([1, 1])]; + tensor current_key_pad_type_0 = const()[name = tensor("current_key_pad_type_0"), val = tensor("custom")]; + tensor current_key_pad_0 = const()[name = tensor("current_key_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1763900992)))]; + tensor current_key_cast_fp16 = conv(dilations = var_6908, groups = var_6865, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = var_6906, weight = layers_31_self_attn_k_proj_weight_to_fp16, x = obj_435_cast_fp16)[name = tensor("current_key_cast_fp16")]; + tensor var_6913 = const()[name = tensor("op_6913"), val = tensor([1, 1])]; + tensor var_6915 = const()[name = tensor("op_6915"), val = tensor([1, 1])]; + tensor current_value_pad_type_0 = const()[name = tensor("current_value_pad_type_0"), val = tensor("custom")]; + tensor current_value_pad_0 = const()[name = tensor("current_value_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1767177856)))]; + 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(1770454720)))]; + tensor current_value_cast_fp16 = conv(bias = layers_31_self_attn_v_proj_bias_to_fp16, dilations = var_6915, groups = var_6865, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = var_6913, weight = layers_31_self_attn_v_proj_weight_to_fp16, x = obj_435_cast_fp16)[name = tensor("current_value_cast_fp16")]; + tensor var_6922_cast_fp16 = mul(x = current_key_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6922_cast_fp16")]; + tensor var_6924_cast_fp16 = mul(x = var_103_cast_fp16_31, y = var_241_cast_fp16)[name = tensor("op_6924_cast_fp16")]; + tensor key_125_cast_fp16 = add(x = var_6922_cast_fp16, y = var_6924_cast_fp16)[name = tensor("key_125_cast_fp16")]; + tensor var_6926_cast_fp16 = mul(x = current_value_cast_fp16, y = var_238_cast_fp16)[name = tensor("op_6926_cast_fp16")]; + tensor var_6928_cast_fp16 = mul(x = var_138_cast_fp16_31, y = var_241_cast_fp16)[name = tensor("op_6928_cast_fp16")]; + tensor value_125_cast_fp16 = add(x = var_6926_cast_fp16, y = var_6928_cast_fp16)[name = tensor("value_125_cast_fp16")]; + tensor var_6931 = const()[name = tensor("op_6931"), val = tensor([1, 20, 64, -1])]; + tensor var_6932_cast_fp16 = reshape(shape = var_6931, x = query_125_cast_fp16)[name = tensor("op_6932_cast_fp16")]; + tensor var_6933_to_fp16 = const()[name = tensor("op_6933_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6934_cast_fp16 = mul(x = var_6932_cast_fp16, y = var_6933_to_fp16)[name = tensor("op_6934_cast_fp16")]; + tensor var_6935 = const()[name = tensor("op_6935"), val = tensor([1, 20, 64, -1])]; + tensor var_6936_cast_fp16 = reshape(shape = var_6935, x = key_125_cast_fp16)[name = tensor("op_6936_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_6934_cast_fp16, y = var_6936_cast_fp16)[name = tensor("mh_w_187_cast_fp16")]; + tensor mh_w_189_cast_fp16 = add(x = mh_w_187_cast_fp16, y = var_259_cast_fp16)[name = tensor("mh_w_189_cast_fp16")]; + tensor var_6944_cast_fp16 = softmax(axis = var_6858, x = mh_w_189_cast_fp16)[name = tensor("op_6944_cast_fp16")]; + tensor var_6945 = const()[name = tensor("op_6945"), val = tensor([1, 20, 64, -1])]; + tensor var_6946_cast_fp16 = reshape(shape = var_6945, x = value_125_cast_fp16)[name = tensor("op_6946_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_6946_cast_fp16, y = var_6944_cast_fp16)[name = tensor("attn_125_cast_fp16")]; + tensor var_6949 = const()[name = tensor("op_6949"), val = tensor([1, 1280, 1, -1])]; + tensor input_311_cast_fp16 = reshape(shape = var_6949, x = attn_125_cast_fp16)[name = tensor("input_311_cast_fp16")]; + tensor var_6953 = const()[name = tensor("op_6953"), val = tensor([1, 1])]; + tensor var_6955 = const()[name = tensor("op_6955"), val = tensor([1, 1])]; + tensor obj_441_pad_type_0 = const()[name = tensor("obj_441_pad_type_0"), val = tensor("custom")]; + tensor obj_441_pad_0 = const()[name = tensor("obj_441_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1770457344)))]; + 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(1773734208)))]; + tensor obj_441_cast_fp16 = conv(bias = layers_31_self_attn_o_proj_bias_to_fp16, dilations = var_6955, groups = var_6865, pad = obj_441_pad_0, pad_type = obj_441_pad_type_0, strides = var_6953, 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 var_6965 = const()[name = tensor("op_6965"), val = tensor([1])]; + tensor channels_mean_189_cast_fp16 = reduce_mean(axes = var_6965, keep_dims = var_6866, x = inputs_189_cast_fp16)[name = tensor("channels_mean_189_cast_fp16")]; + tensor zero_mean_189_cast_fp16 = sub(x = inputs_189_cast_fp16, y = channels_mean_189_cast_fp16)[name = tensor("zero_mean_189_cast_fp16")]; + tensor zero_mean_sq_189_cast_fp16 = mul(x = zero_mean_189_cast_fp16, y = zero_mean_189_cast_fp16)[name = tensor("zero_mean_sq_189_cast_fp16")]; + tensor var_6969 = const()[name = tensor("op_6969"), val = tensor([1])]; + tensor var_6970_cast_fp16 = reduce_mean(axes = var_6969, keep_dims = var_6866, x = zero_mean_sq_189_cast_fp16)[name = tensor("op_6970_cast_fp16")]; + tensor var_6971_to_fp16 = const()[name = tensor("op_6971_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6972_cast_fp16 = add(x = var_6970_cast_fp16, y = var_6971_to_fp16)[name = tensor("op_6972_cast_fp16")]; + tensor denom_189_epsilon_0_to_fp16 = const()[name = tensor("denom_189_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_189_cast_fp16 = rsqrt(epsilon = denom_189_epsilon_0_to_fp16, x = var_6972_cast_fp16)[name = tensor("denom_189_cast_fp16")]; + tensor out_189_cast_fp16 = mul(x = zero_mean_189_cast_fp16, y = denom_189_cast_fp16)[name = tensor("out_189_cast_fp16")]; + tensor obj_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(1773736832)))]; + 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(1773739456)))]; + 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 var_6987 = const()[name = tensor("op_6987"), val = tensor([1, 1])]; + tensor var_6989 = const()[name = tensor("op_6989"), val = tensor([1, 1])]; + tensor query_pad_type_0 = const()[name = tensor("query_pad_type_0"), val = tensor("custom")]; + tensor query_pad_0 = const()[name = tensor("query_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1773742080)))]; + 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(1777018944)))]; + tensor query_cast_fp16 = conv(bias = layers_31_encoder_attn_q_proj_bias_to_fp16, dilations = var_6989, groups = var_6865, pad = query_pad_0, pad_type = query_pad_type_0, strides = var_6987, weight = layers_31_encoder_attn_q_proj_weight_to_fp16, x = obj_443_cast_fp16)[name = tensor("query_cast_fp16")]; + tensor var_6993 = const()[name = tensor("op_6993"), val = tensor([1, 1])]; + tensor var_6995 = const()[name = tensor("op_6995"), val = tensor([1, 1])]; + tensor key_pad_type_0 = const()[name = tensor("key_pad_type_0"), val = tensor("custom")]; + tensor key_pad_0 = const()[name = tensor("key_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1777021568)))]; + tensor key_cast_fp16 = conv(dilations = var_6995, groups = var_6865, pad = key_pad_0, pad_type = key_pad_type_0, strides = var_6993, weight = layers_31_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_cast_fp16")]; + tensor var_7000 = const()[name = tensor("op_7000"), val = tensor([1, 1])]; + tensor var_7002 = const()[name = tensor("op_7002"), val = tensor([1, 1])]; + tensor value_pad_type_0 = const()[name = tensor("value_pad_type_0"), val = tensor("custom")]; + tensor value_pad_0 = const()[name = tensor("value_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1780298432)))]; + 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(1783575296)))]; + tensor value_cast_fp16 = conv(bias = layers_31_encoder_attn_v_proj_bias_to_fp16, dilations = var_7002, groups = var_6865, pad = value_pad_0, pad_type = value_pad_type_0, strides = var_7000, weight = layers_31_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_cast_fp16")]; + tensor var_7006 = const()[name = tensor("op_7006"), val = tensor([1, 20, 64, -1])]; + tensor var_7007_cast_fp16 = reshape(shape = var_7006, x = query_cast_fp16)[name = tensor("op_7007_cast_fp16")]; + tensor var_7008_to_fp16 = const()[name = tensor("op_7008_to_fp16"), val = tensor(0x1p-3)]; + tensor var_7009_cast_fp16 = mul(x = var_7007_cast_fp16, y = var_7008_to_fp16)[name = tensor("op_7009_cast_fp16")]; + tensor var_7010 = const()[name = tensor("op_7010"), val = tensor([1, 20, 64, -1])]; + tensor var_7011_cast_fp16 = reshape(shape = var_7010, x = key_cast_fp16)[name = tensor("op_7011_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_7009_cast_fp16, y = var_7011_cast_fp16)[name = tensor("mh_w_cast_fp16")]; + tensor obj_447_cast_fp16 = softmax(axis = var_6858, x = mh_w_cast_fp16)[name = tensor("obj_447_cast_fp16")]; + tensor var_7015 = const()[name = tensor("op_7015"), val = tensor([1, 20, 64, -1])]; + tensor var_7016_cast_fp16 = reshape(shape = var_7015, x = value_cast_fp16)[name = tensor("op_7016_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_7016_cast_fp16, y = obj_447_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_7019 = const()[name = tensor("op_7019"), val = tensor([1, 1280, 1, -1])]; + tensor input_313_cast_fp16 = reshape(shape = var_7019, x = attn_cast_fp16)[name = tensor("input_313_cast_fp16")]; + tensor var_7023 = const()[name = tensor("op_7023"), val = tensor([1, 1])]; + tensor var_7025 = const()[name = tensor("op_7025"), val = tensor([1, 1])]; + tensor obj_445_pad_type_0 = const()[name = tensor("obj_445_pad_type_0"), val = tensor("custom")]; + tensor obj_445_pad_0 = const()[name = tensor("obj_445_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1783577920)))]; + 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(1786854784)))]; + tensor obj_445_cast_fp16 = conv(bias = layers_31_encoder_attn_o_proj_bias_to_fp16, dilations = var_7025, groups = var_6865, pad = obj_445_pad_0, pad_type = obj_445_pad_type_0, strides = var_7023, 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 var_7031 = const()[name = tensor("op_7031"), val = tensor([1])]; + tensor channels_mean_191_cast_fp16 = reduce_mean(axes = var_7031, keep_dims = var_6866, x = inputs_191_cast_fp16)[name = tensor("channels_mean_191_cast_fp16")]; + tensor zero_mean_191_cast_fp16 = sub(x = inputs_191_cast_fp16, y = channels_mean_191_cast_fp16)[name = tensor("zero_mean_191_cast_fp16")]; + tensor zero_mean_sq_191_cast_fp16 = mul(x = zero_mean_191_cast_fp16, y = zero_mean_191_cast_fp16)[name = tensor("zero_mean_sq_191_cast_fp16")]; + tensor var_7035 = const()[name = tensor("op_7035"), val = tensor([1])]; + tensor var_7036_cast_fp16 = reduce_mean(axes = var_7035, keep_dims = var_6866, x = zero_mean_sq_191_cast_fp16)[name = tensor("op_7036_cast_fp16")]; + tensor var_7037_to_fp16 = const()[name = tensor("op_7037_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_7038_cast_fp16 = add(x = var_7036_cast_fp16, y = var_7037_to_fp16)[name = tensor("op_7038_cast_fp16")]; + tensor denom_191_epsilon_0_to_fp16 = const()[name = tensor("denom_191_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_191_cast_fp16 = rsqrt(epsilon = denom_191_epsilon_0_to_fp16, x = var_7038_cast_fp16)[name = tensor("denom_191_cast_fp16")]; + tensor out_191_cast_fp16 = mul(x = zero_mean_191_cast_fp16, y = denom_191_cast_fp16)[name = tensor("out_191_cast_fp16")]; + tensor input_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(1786857408)))]; + 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(1786860032)))]; + 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 var_7049 = const()[name = tensor("op_7049"), val = tensor([1, 1])]; + tensor var_7051 = const()[name = tensor("op_7051"), val = tensor([1, 1])]; + tensor input_317_pad_type_0 = const()[name = tensor("input_317_pad_type_0"), val = tensor("custom")]; + tensor input_317_pad_0 = const()[name = tensor("input_317_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1786862656)))]; + 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(1799969920)))]; + tensor input_317_cast_fp16 = conv(bias = layers_31_fc1_bias_to_fp16, dilations = var_7051, groups = var_6865, pad = input_317_pad_0, pad_type = input_317_pad_type_0, strides = var_7049, 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 var_7057 = const()[name = tensor("op_7057"), val = tensor([1, 1])]; + tensor var_7059 = const()[name = tensor("op_7059"), val = tensor([1, 1])]; + tensor hidden_states_65_pad_type_0 = const()[name = tensor("hidden_states_65_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_65_pad_0 = const()[name = tensor("hidden_states_65_pad_0"), val = tensor([0, 0, 0, 0])]; + 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(1799980224)))]; + 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(1813087488)))]; + tensor hidden_states_65_cast_fp16 = conv(bias = layers_31_fc2_bias_to_fp16, dilations = var_7059, groups = var_6865, pad = hidden_states_65_pad_0, pad_type = hidden_states_65_pad_type_0, strides = var_7057, 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 var_7069 = const()[name = tensor("op_7069"), val = tensor(true)]; + tensor var_7073 = const()[name = tensor("op_7073"), val = tensor([1])]; + tensor channels_mean_cast_fp16 = reduce_mean(axes = var_7073, keep_dims = var_7069, x = inputs_cast_fp16)[name = tensor("channels_mean_cast_fp16")]; + tensor zero_mean_cast_fp16 = sub(x = inputs_cast_fp16, y = channels_mean_cast_fp16)[name = tensor("zero_mean_cast_fp16")]; + tensor zero_mean_sq_cast_fp16 = mul(x = zero_mean_cast_fp16, y = zero_mean_cast_fp16)[name = tensor("zero_mean_sq_cast_fp16")]; + tensor var_7077 = const()[name = tensor("op_7077"), val = tensor([1])]; + tensor var_7078_cast_fp16 = reduce_mean(axes = var_7077, keep_dims = var_7069, x = zero_mean_sq_cast_fp16)[name = tensor("op_7078_cast_fp16")]; + tensor var_7079_to_fp16 = const()[name = tensor("op_7079_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_7080_cast_fp16 = add(x = var_7078_cast_fp16, y = var_7079_to_fp16)[name = tensor("op_7080_cast_fp16")]; + tensor denom_epsilon_0_to_fp16 = const()[name = tensor("denom_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_cast_fp16 = rsqrt(epsilon = denom_epsilon_0_to_fp16, x = var_7080_cast_fp16)[name = tensor("denom_cast_fp16")]; + tensor out_cast_fp16 = mul(x = zero_mean_cast_fp16, y = denom_cast_fp16)[name = tensor("out_cast_fp16")]; + tensor hidden_states_gamma_0_to_fp16 = const()[name = tensor("hidden_states_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1813090112)))]; + 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(1813092736)))]; + 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_7090_axes_0 = const()[name = tensor("op_7090_axes_0"), val = tensor([2])]; + tensor var_7090_cast_fp16 = squeeze(axes = var_7090_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_7090_cast_fp16")]; + tensor var_7093_perm_0 = const()[name = tensor("op_7093_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(1813095360)))]; + tensor transpose_0 = transpose(perm = var_7093_perm_0, x = var_7090_cast_fp16)[name = tensor("transpose_0")]; + tensor logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = transpose_0)[name = tensor("linear_0_cast_fp16")]; + tensor var_7097 = const()[name = tensor("op_7097"), 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_7097, 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_7100 = const()[name = tensor("op_7100"), 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_7100, 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_7111_begin_0 = const()[name = tensor("op_7111_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_7111_end_0 = const()[name = tensor("op_7111_end_0"), val = tensor([1, 13, 1, 1500])]; + tensor var_7111_end_mask_0 = const()[name = tensor("op_7111_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7111_cast_fp16 = slice_by_index(begin = var_7111_begin_0, end = var_7111_end_0, end_mask = var_7111_end_mask_0, x = obj_153_cast_fp16)[name = tensor("op_7111_cast_fp16")]; + tensor var_7114_begin_0 = const()[name = tensor("op_7114_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7114_end_0 = const()[name = tensor("op_7114_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7114_end_mask_0 = const()[name = tensor("op_7114_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7114_squeeze_mask_0 = const()[name = tensor("op_7114_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7114_cast_fp16 = slice_by_index(begin = var_7114_begin_0, end = var_7114_end_0, end_mask = var_7114_end_mask_0, squeeze_mask = var_7114_squeeze_mask_0, x = var_7111_cast_fp16)[name = tensor("op_7114_cast_fp16")]; + tensor var_7129_begin_0 = const()[name = tensor("op_7129_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_7129_end_0 = const()[name = tensor("op_7129_end_0"), val = tensor([1, 18, 1, 1500])]; + tensor var_7129_end_mask_0 = const()[name = tensor("op_7129_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7129_cast_fp16 = slice_by_index(begin = var_7129_begin_0, end = var_7129_end_0, end_mask = var_7129_end_mask_0, x = obj_195_cast_fp16)[name = tensor("op_7129_cast_fp16")]; + tensor var_7132_begin_0 = const()[name = tensor("op_7132_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7132_end_0 = const()[name = tensor("op_7132_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7132_end_mask_0 = const()[name = tensor("op_7132_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7132_squeeze_mask_0 = const()[name = tensor("op_7132_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7132_cast_fp16 = slice_by_index(begin = var_7132_begin_0, end = var_7132_end_0, end_mask = var_7132_end_mask_0, squeeze_mask = var_7132_squeeze_mask_0, x = var_7129_cast_fp16)[name = tensor("op_7132_cast_fp16")]; + tensor var_7147_begin_0 = const()[name = tensor("op_7147_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_7147_end_0 = const()[name = tensor("op_7147_end_0"), val = tensor([1, 12, 1, 1500])]; + tensor var_7147_end_mask_0 = const()[name = tensor("op_7147_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7147_cast_fp16 = slice_by_index(begin = var_7147_begin_0, end = var_7147_end_0, end_mask = var_7147_end_mask_0, x = obj_237_cast_fp16)[name = tensor("op_7147_cast_fp16")]; + tensor var_7150_begin_0 = const()[name = tensor("op_7150_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7150_end_0 = const()[name = tensor("op_7150_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7150_end_mask_0 = const()[name = tensor("op_7150_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7150_squeeze_mask_0 = const()[name = tensor("op_7150_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7150_cast_fp16 = slice_by_index(begin = var_7150_begin_0, end = var_7150_end_0, end_mask = var_7150_end_mask_0, squeeze_mask = var_7150_squeeze_mask_0, x = var_7147_cast_fp16)[name = tensor("op_7150_cast_fp16")]; + tensor var_7165_begin_0 = const()[name = tensor("op_7165_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_7165_end_0 = const()[name = tensor("op_7165_end_0"), val = tensor([1, 13, 1, 1500])]; + tensor var_7165_end_mask_0 = const()[name = tensor("op_7165_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7165_cast_fp16 = slice_by_index(begin = var_7165_begin_0, end = var_7165_end_0, end_mask = var_7165_end_mask_0, x = obj_237_cast_fp16)[name = tensor("op_7165_cast_fp16")]; + tensor var_7168_begin_0 = const()[name = tensor("op_7168_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7168_end_0 = const()[name = tensor("op_7168_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7168_end_mask_0 = const()[name = tensor("op_7168_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7168_squeeze_mask_0 = const()[name = tensor("op_7168_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7168_cast_fp16 = slice_by_index(begin = var_7168_begin_0, end = var_7168_end_0, end_mask = var_7168_end_mask_0, squeeze_mask = var_7168_squeeze_mask_0, x = var_7165_cast_fp16)[name = tensor("op_7168_cast_fp16")]; + tensor var_7183_begin_0 = const()[name = tensor("op_7183_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_7183_end_0 = const()[name = tensor("op_7183_end_0"), val = tensor([1, 14, 1, 1500])]; + tensor var_7183_end_mask_0 = const()[name = tensor("op_7183_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7183_cast_fp16 = slice_by_index(begin = var_7183_begin_0, end = var_7183_end_0, end_mask = var_7183_end_mask_0, x = obj_237_cast_fp16)[name = tensor("op_7183_cast_fp16")]; + tensor var_7186_begin_0 = const()[name = tensor("op_7186_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7186_end_0 = const()[name = tensor("op_7186_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7186_end_mask_0 = const()[name = tensor("op_7186_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7186_squeeze_mask_0 = const()[name = tensor("op_7186_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7186_cast_fp16 = slice_by_index(begin = var_7186_begin_0, end = var_7186_end_0, end_mask = var_7186_end_mask_0, squeeze_mask = var_7186_squeeze_mask_0, x = var_7183_cast_fp16)[name = tensor("op_7186_cast_fp16")]; + tensor var_7201_begin_0 = const()[name = tensor("op_7201_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_7201_end_0 = const()[name = tensor("op_7201_end_0"), val = tensor([1, 16, 1, 1500])]; + tensor var_7201_end_mask_0 = const()[name = tensor("op_7201_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7201_cast_fp16 = slice_by_index(begin = var_7201_begin_0, end = var_7201_end_0, end_mask = var_7201_end_mask_0, x = obj_251_cast_fp16)[name = tensor("op_7201_cast_fp16")]; + tensor var_7204_begin_0 = const()[name = tensor("op_7204_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7204_end_0 = const()[name = tensor("op_7204_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7204_end_mask_0 = const()[name = tensor("op_7204_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7204_squeeze_mask_0 = const()[name = tensor("op_7204_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7204_cast_fp16 = slice_by_index(begin = var_7204_begin_0, end = var_7204_end_0, end_mask = var_7204_end_mask_0, squeeze_mask = var_7204_squeeze_mask_0, x = var_7201_cast_fp16)[name = tensor("op_7204_cast_fp16")]; + tensor var_7219_begin_0 = const()[name = tensor("op_7219_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_7219_end_0 = const()[name = tensor("op_7219_end_0"), val = tensor([1, 17, 1, 1500])]; + tensor var_7219_end_mask_0 = const()[name = tensor("op_7219_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7219_cast_fp16 = slice_by_index(begin = var_7219_begin_0, end = var_7219_end_0, end_mask = var_7219_end_mask_0, x = obj_251_cast_fp16)[name = tensor("op_7219_cast_fp16")]; + tensor var_7222_begin_0 = const()[name = tensor("op_7222_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7222_end_0 = const()[name = tensor("op_7222_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7222_end_mask_0 = const()[name = tensor("op_7222_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7222_squeeze_mask_0 = const()[name = tensor("op_7222_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7222_cast_fp16 = slice_by_index(begin = var_7222_begin_0, end = var_7222_end_0, end_mask = var_7222_end_mask_0, squeeze_mask = var_7222_squeeze_mask_0, x = var_7219_cast_fp16)[name = tensor("op_7222_cast_fp16")]; + tensor var_7237_begin_0 = const()[name = tensor("op_7237_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_7237_end_0 = const()[name = tensor("op_7237_end_0"), val = tensor([1, 5, 1, 1500])]; + tensor var_7237_end_mask_0 = const()[name = tensor("op_7237_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7237_cast_fp16 = slice_by_index(begin = var_7237_begin_0, end = var_7237_end_0, end_mask = var_7237_end_mask_0, x = obj_265_cast_fp16)[name = tensor("op_7237_cast_fp16")]; + tensor var_7240_begin_0 = const()[name = tensor("op_7240_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7240_end_0 = const()[name = tensor("op_7240_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7240_end_mask_0 = const()[name = tensor("op_7240_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7240_squeeze_mask_0 = const()[name = tensor("op_7240_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7240_cast_fp16 = slice_by_index(begin = var_7240_begin_0, end = var_7240_end_0, end_mask = var_7240_end_mask_0, squeeze_mask = var_7240_squeeze_mask_0, x = var_7237_cast_fp16)[name = tensor("op_7240_cast_fp16")]; + tensor var_7255_begin_0 = const()[name = tensor("op_7255_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_7255_end_0 = const()[name = tensor("op_7255_end_0"), val = tensor([1, 12, 1, 1500])]; + tensor var_7255_end_mask_0 = const()[name = tensor("op_7255_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7255_cast_fp16 = slice_by_index(begin = var_7255_begin_0, end = var_7255_end_0, end_mask = var_7255_end_mask_0, x = obj_265_cast_fp16)[name = tensor("op_7255_cast_fp16")]; + tensor var_7258_begin_0 = const()[name = tensor("op_7258_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7258_end_0 = const()[name = tensor("op_7258_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7258_end_mask_0 = const()[name = tensor("op_7258_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7258_squeeze_mask_0 = const()[name = tensor("op_7258_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7258_cast_fp16 = slice_by_index(begin = var_7258_begin_0, end = var_7258_end_0, end_mask = var_7258_end_mask_0, squeeze_mask = var_7258_squeeze_mask_0, x = var_7255_cast_fp16)[name = tensor("op_7258_cast_fp16")]; + tensor var_7273_begin_0 = const()[name = tensor("op_7273_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_7273_end_0 = const()[name = tensor("op_7273_end_0"), val = tensor([1, 20, 1, 1500])]; + tensor var_7273_end_mask_0 = const()[name = tensor("op_7273_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7273_cast_fp16 = slice_by_index(begin = var_7273_begin_0, end = var_7273_end_0, end_mask = var_7273_end_mask_0, x = obj_265_cast_fp16)[name = tensor("op_7273_cast_fp16")]; + tensor var_7276_begin_0 = const()[name = tensor("op_7276_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7276_end_0 = const()[name = tensor("op_7276_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7276_end_mask_0 = const()[name = tensor("op_7276_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7276_squeeze_mask_0 = const()[name = tensor("op_7276_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7276_cast_fp16 = slice_by_index(begin = var_7276_begin_0, end = var_7276_end_0, end_mask = var_7276_end_mask_0, squeeze_mask = var_7276_squeeze_mask_0, x = var_7273_cast_fp16)[name = tensor("op_7276_cast_fp16")]; + tensor var_7291_begin_0 = const()[name = tensor("op_7291_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_7291_end_0 = const()[name = tensor("op_7291_end_0"), val = tensor([1, 12, 1, 1500])]; + tensor var_7291_end_mask_0 = const()[name = tensor("op_7291_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7291_cast_fp16 = slice_by_index(begin = var_7291_begin_0, end = var_7291_end_0, end_mask = var_7291_end_mask_0, x = obj_279_cast_fp16)[name = tensor("op_7291_cast_fp16")]; + tensor var_7294_begin_0 = const()[name = tensor("op_7294_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7294_end_0 = const()[name = tensor("op_7294_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7294_end_mask_0 = const()[name = tensor("op_7294_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7294_squeeze_mask_0 = const()[name = tensor("op_7294_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7294_cast_fp16 = slice_by_index(begin = var_7294_begin_0, end = var_7294_end_0, end_mask = var_7294_end_mask_0, squeeze_mask = var_7294_squeeze_mask_0, x = var_7291_cast_fp16)[name = tensor("op_7294_cast_fp16")]; + tensor var_7309_begin_0 = const()[name = tensor("op_7309_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_7309_end_0 = const()[name = tensor("op_7309_end_0"), val = tensor([1, 3, 1, 1500])]; + tensor var_7309_end_mask_0 = const()[name = tensor("op_7309_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7309_cast_fp16 = slice_by_index(begin = var_7309_begin_0, end = var_7309_end_0, end_mask = var_7309_end_mask_0, x = obj_307_cast_fp16)[name = tensor("op_7309_cast_fp16")]; + tensor var_7312_begin_0 = const()[name = tensor("op_7312_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7312_end_0 = const()[name = tensor("op_7312_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7312_end_mask_0 = const()[name = tensor("op_7312_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7312_squeeze_mask_0 = const()[name = tensor("op_7312_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7312_cast_fp16 = slice_by_index(begin = var_7312_begin_0, end = var_7312_end_0, end_mask = var_7312_end_mask_0, squeeze_mask = var_7312_squeeze_mask_0, x = var_7309_cast_fp16)[name = tensor("op_7312_cast_fp16")]; + tensor var_7327_begin_0 = const()[name = tensor("op_7327_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_7327_end_0 = const()[name = tensor("op_7327_end_0"), val = tensor([1, 4, 1, 1500])]; + tensor var_7327_end_mask_0 = const()[name = tensor("op_7327_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7327_cast_fp16 = slice_by_index(begin = var_7327_begin_0, end = var_7327_end_0, end_mask = var_7327_end_mask_0, x = obj_307_cast_fp16)[name = tensor("op_7327_cast_fp16")]; + tensor var_7330_begin_0 = const()[name = tensor("op_7330_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7330_end_0 = const()[name = tensor("op_7330_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7330_end_mask_0 = const()[name = tensor("op_7330_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7330_squeeze_mask_0 = const()[name = tensor("op_7330_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7330_cast_fp16 = slice_by_index(begin = var_7330_begin_0, end = var_7330_end_0, end_mask = var_7330_end_mask_0, squeeze_mask = var_7330_squeeze_mask_0, x = var_7327_cast_fp16)[name = tensor("op_7330_cast_fp16")]; + tensor var_7345_begin_0 = const()[name = tensor("op_7345_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_7345_end_0 = const()[name = tensor("op_7345_end_0"), val = tensor([1, 4, 1, 1500])]; + tensor var_7345_end_mask_0 = const()[name = tensor("op_7345_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7345_cast_fp16 = slice_by_index(begin = var_7345_begin_0, end = var_7345_end_0, end_mask = var_7345_end_mask_0, x = obj_321_cast_fp16)[name = tensor("op_7345_cast_fp16")]; + tensor var_7348_begin_0 = const()[name = tensor("op_7348_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7348_end_0 = const()[name = tensor("op_7348_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7348_end_mask_0 = const()[name = tensor("op_7348_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7348_squeeze_mask_0 = const()[name = tensor("op_7348_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7348_cast_fp16 = slice_by_index(begin = var_7348_begin_0, end = var_7348_end_0, end_mask = var_7348_end_mask_0, squeeze_mask = var_7348_squeeze_mask_0, x = var_7345_cast_fp16)[name = tensor("op_7348_cast_fp16")]; + tensor var_7363_begin_0 = const()[name = tensor("op_7363_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_7363_end_0 = const()[name = tensor("op_7363_end_0"), val = tensor([1, 10, 1, 1500])]; + tensor var_7363_end_mask_0 = const()[name = tensor("op_7363_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7363_cast_fp16 = slice_by_index(begin = var_7363_begin_0, end = var_7363_end_0, end_mask = var_7363_end_mask_0, x = obj_321_cast_fp16)[name = tensor("op_7363_cast_fp16")]; + tensor var_7366_begin_0 = const()[name = tensor("op_7366_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7366_end_0 = const()[name = tensor("op_7366_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7366_end_mask_0 = const()[name = tensor("op_7366_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7366_squeeze_mask_0 = const()[name = tensor("op_7366_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7366_cast_fp16 = slice_by_index(begin = var_7366_begin_0, end = var_7366_end_0, end_mask = var_7366_end_mask_0, squeeze_mask = var_7366_squeeze_mask_0, x = var_7363_cast_fp16)[name = tensor("op_7366_cast_fp16")]; + tensor var_7381_begin_0 = const()[name = tensor("op_7381_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_7381_end_0 = const()[name = tensor("op_7381_end_0"), val = tensor([1, 13, 1, 1500])]; + tensor var_7381_end_mask_0 = const()[name = tensor("op_7381_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7381_cast_fp16 = slice_by_index(begin = var_7381_begin_0, end = var_7381_end_0, end_mask = var_7381_end_mask_0, x = obj_321_cast_fp16)[name = tensor("op_7381_cast_fp16")]; + tensor var_7384_begin_0 = const()[name = tensor("op_7384_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7384_end_0 = const()[name = tensor("op_7384_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7384_end_mask_0 = const()[name = tensor("op_7384_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7384_squeeze_mask_0 = const()[name = tensor("op_7384_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7384_cast_fp16 = slice_by_index(begin = var_7384_begin_0, end = var_7384_end_0, end_mask = var_7384_end_mask_0, squeeze_mask = var_7384_squeeze_mask_0, x = var_7381_cast_fp16)[name = tensor("op_7384_cast_fp16")]; + tensor var_7399_begin_0 = const()[name = tensor("op_7399_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_7399_end_0 = const()[name = tensor("op_7399_end_0"), val = tensor([1, 6, 1, 1500])]; + tensor var_7399_end_mask_0 = const()[name = tensor("op_7399_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7399_cast_fp16 = slice_by_index(begin = var_7399_begin_0, end = var_7399_end_0, end_mask = var_7399_end_mask_0, x = obj_335_cast_fp16)[name = tensor("op_7399_cast_fp16")]; + tensor var_7402_begin_0 = const()[name = tensor("op_7402_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7402_end_0 = const()[name = tensor("op_7402_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7402_end_mask_0 = const()[name = tensor("op_7402_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7402_squeeze_mask_0 = const()[name = tensor("op_7402_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7402_cast_fp16 = slice_by_index(begin = var_7402_begin_0, end = var_7402_end_0, end_mask = var_7402_end_mask_0, squeeze_mask = var_7402_squeeze_mask_0, x = var_7399_cast_fp16)[name = tensor("op_7402_cast_fp16")]; + tensor var_7417_begin_0 = const()[name = tensor("op_7417_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_7417_end_0 = const()[name = tensor("op_7417_end_0"), val = tensor([1, 8, 1, 1500])]; + tensor var_7417_end_mask_0 = const()[name = tensor("op_7417_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7417_cast_fp16 = slice_by_index(begin = var_7417_begin_0, end = var_7417_end_0, end_mask = var_7417_end_mask_0, x = obj_335_cast_fp16)[name = tensor("op_7417_cast_fp16")]; + tensor var_7420_begin_0 = const()[name = tensor("op_7420_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7420_end_0 = const()[name = tensor("op_7420_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7420_end_mask_0 = const()[name = tensor("op_7420_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7420_squeeze_mask_0 = const()[name = tensor("op_7420_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7420_cast_fp16 = slice_by_index(begin = var_7420_begin_0, end = var_7420_end_0, end_mask = var_7420_end_mask_0, squeeze_mask = var_7420_squeeze_mask_0, x = var_7417_cast_fp16)[name = tensor("op_7420_cast_fp16")]; + tensor var_7435_begin_0 = const()[name = tensor("op_7435_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_7435_end_0 = const()[name = tensor("op_7435_end_0"), val = tensor([1, 14, 1, 1500])]; + tensor var_7435_end_mask_0 = const()[name = tensor("op_7435_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7435_cast_fp16 = slice_by_index(begin = var_7435_begin_0, end = var_7435_end_0, end_mask = var_7435_end_mask_0, x = obj_335_cast_fp16)[name = tensor("op_7435_cast_fp16")]; + tensor var_7438_begin_0 = const()[name = tensor("op_7438_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7438_end_0 = const()[name = tensor("op_7438_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7438_end_mask_0 = const()[name = tensor("op_7438_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7438_squeeze_mask_0 = const()[name = tensor("op_7438_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7438_cast_fp16 = slice_by_index(begin = var_7438_begin_0, end = var_7438_end_0, end_mask = var_7438_end_mask_0, squeeze_mask = var_7438_squeeze_mask_0, x = var_7435_cast_fp16)[name = tensor("op_7438_cast_fp16")]; + tensor var_7453_begin_0 = const()[name = tensor("op_7453_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_7453_end_0 = const()[name = tensor("op_7453_end_0"), val = tensor([1, 6, 1, 1500])]; + tensor var_7453_end_mask_0 = const()[name = tensor("op_7453_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7453_cast_fp16 = slice_by_index(begin = var_7453_begin_0, end = var_7453_end_0, end_mask = var_7453_end_mask_0, x = obj_363_cast_fp16)[name = tensor("op_7453_cast_fp16")]; + tensor var_7456_begin_0 = const()[name = tensor("op_7456_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7456_end_0 = const()[name = tensor("op_7456_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7456_end_mask_0 = const()[name = tensor("op_7456_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7456_squeeze_mask_0 = const()[name = tensor("op_7456_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7456_cast_fp16 = slice_by_index(begin = var_7456_begin_0, end = var_7456_end_0, end_mask = var_7456_end_mask_0, squeeze_mask = var_7456_squeeze_mask_0, x = var_7453_cast_fp16)[name = tensor("op_7456_cast_fp16")]; + tensor var_7471_begin_0 = const()[name = tensor("op_7471_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_7471_end_0 = const()[name = tensor("op_7471_end_0"), val = tensor([1, 2, 1, 1500])]; + tensor var_7471_end_mask_0 = const()[name = tensor("op_7471_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7471_cast_fp16 = slice_by_index(begin = var_7471_begin_0, end = var_7471_end_0, end_mask = var_7471_end_mask_0, x = obj_377_cast_fp16)[name = tensor("op_7471_cast_fp16")]; + tensor var_7474_begin_0 = const()[name = tensor("op_7474_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7474_end_0 = const()[name = tensor("op_7474_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7474_end_mask_0 = const()[name = tensor("op_7474_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7474_squeeze_mask_0 = const()[name = tensor("op_7474_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7474_cast_fp16 = slice_by_index(begin = var_7474_begin_0, end = var_7474_end_0, end_mask = var_7474_end_mask_0, squeeze_mask = var_7474_squeeze_mask_0, x = var_7471_cast_fp16)[name = tensor("op_7474_cast_fp16")]; + tensor var_7489_begin_0 = const()[name = tensor("op_7489_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_7489_end_0 = const()[name = tensor("op_7489_end_0"), val = tensor([1, 13, 1, 1500])]; + tensor var_7489_end_mask_0 = const()[name = tensor("op_7489_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7489_cast_fp16 = slice_by_index(begin = var_7489_begin_0, end = var_7489_end_0, end_mask = var_7489_end_mask_0, x = obj_377_cast_fp16)[name = tensor("op_7489_cast_fp16")]; + tensor var_7492_begin_0 = const()[name = tensor("op_7492_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7492_end_0 = const()[name = tensor("op_7492_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7492_end_mask_0 = const()[name = tensor("op_7492_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7492_squeeze_mask_0 = const()[name = tensor("op_7492_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7492_cast_fp16 = slice_by_index(begin = var_7492_begin_0, end = var_7492_end_0, end_mask = var_7492_end_mask_0, squeeze_mask = var_7492_squeeze_mask_0, x = var_7489_cast_fp16)[name = tensor("op_7492_cast_fp16")]; + tensor var_7507_begin_0 = const()[name = tensor("op_7507_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_7507_end_0 = const()[name = tensor("op_7507_end_0"), val = tensor([1, 16, 1, 1500])]; + tensor var_7507_end_mask_0 = const()[name = tensor("op_7507_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_7507_cast_fp16 = slice_by_index(begin = var_7507_begin_0, end = var_7507_end_0, end_mask = var_7507_end_mask_0, x = obj_391_cast_fp16)[name = tensor("op_7507_cast_fp16")]; + tensor var_7510_begin_0 = const()[name = tensor("op_7510_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7510_end_0 = const()[name = tensor("op_7510_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_7510_end_mask_0 = const()[name = tensor("op_7510_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_7510_squeeze_mask_0 = const()[name = tensor("op_7510_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_7510_cast_fp16 = slice_by_index(begin = var_7510_begin_0, end = var_7510_end_0, end_mask = var_7510_end_mask_0, squeeze_mask = var_7510_squeeze_mask_0, x = var_7507_cast_fp16)[name = tensor("op_7510_cast_fp16")]; + tensor var_7517 = const()[name = tensor("op_7517"), val = tensor(1)]; + tensor var_7518_interleave_0 = const()[name = tensor("op_7518_interleave_0"), val = tensor(false)]; + tensor var_7518_cast_fp16 = concat(axis = var_7517, interleave = var_7518_interleave_0, values = (var_7114_cast_fp16, var_7132_cast_fp16, var_7150_cast_fp16, var_7168_cast_fp16, var_7186_cast_fp16, var_7204_cast_fp16, var_7222_cast_fp16, var_7240_cast_fp16, var_7258_cast_fp16, var_7276_cast_fp16, var_7294_cast_fp16, var_7312_cast_fp16, var_7330_cast_fp16, var_7348_cast_fp16, var_7366_cast_fp16, var_7384_cast_fp16, var_7402_cast_fp16, var_7420_cast_fp16, var_7438_cast_fp16, var_7456_cast_fp16, var_7474_cast_fp16, var_7492_cast_fp16, var_7510_cast_fp16))[name = tensor("op_7518_cast_fp16")]; + tensor var_7520 = const()[name = tensor("op_7520"), val = tensor([1])]; + tensor var_7521 = const()[name = tensor("op_7521"), val = tensor(false)]; + tensor alignment_heads_weights = reduce_mean(axes = var_7520, keep_dims = var_7521, x = var_7518_cast_fp16)[name = tensor("obj_cast_fp16")]; + } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); +} \ No newline at end of file