[Bug]: export静态图无报错但没有pdmodel文件

[cycao77]

软件环境

paddle2onnx                              2.0.0
paddlefsl                                1.1.0
paddlenlp                                3.0.0b2
paddlepaddle-gpu                         3.0.0rc1

重复问题

• I have searched the existing issues

错误描述

使用PaddleNLP/slm/examples/machine_reading_comprehension
/SQuAD训完后export静态图，没有报错但目录中没有pdmodel文件，export日志：
[2025-04-15 20:30:54,323] [ WARNING] - Detected that datasets module was imported before paddlenlp. This may cause PaddleNLP datasets to be unavalible in intranet. Please import paddlenlp before datasets module to avoid download issues
[2025-04-15 20:30:55,490] [    INFO] - Loading configuration file ./trained_models/mrc_squad2_all45_v2/model_172000/config.json
[2025-04-15 20:30:55,491] [    INFO] - Loading weights file ./trained_models/mrc_squad2_all45_v2/model_172000/model_state.pdparams
[2025-04-15 20:30:56,101] [    INFO] - Loaded weights file from disk, setting weights to model.
W0415 20:30:56.327435 59824 gpu_resources.cc:119] Please NOTE: device: 0, GPU Compute Capability: 8.0, Driver API Version: 12.4, Runtime API Version: 12.3
W0415 20:30:56.328116 59824 gpu_resources.cc:164] device: 0, cuDNN Version: 9.8.
[2025-04-15 20:30:56,989] [    INFO] - All model checkpoint weights were used when initializing ErnieForQuestionAnswering.

[2025-04-15 20:30:56,989] [    INFO] - All the weights of ErnieForQuestionAnswering were initialized from the model checkpoint at ./trained_models/mrc_squad2_all45_v2/model_172000.
If your task is similar to the task the model of the checkpoint was trained on, you can already use ErnieForQuestionAnswering for predictions without further training.
/ssd3/zhaol/anaconda3/envs/paddlenlp_env/lib/python3.10/site-packages/paddle/jit/dy2static/program_translator.py:780: UserWarning: full_graph=False don't support input_spec arguments. It will not produce any effect.
You can set full_graph=True, then you can assign input spec.

  warnings.warn(
/ssd3/zhaol/anaconda3/envs/paddlenlp_env/lib/python3.10/site-packages/paddlenlp/transformers/utils.py:204: UserWarning: The `forward` method of <class 'paddlenlp.transformers.ernie.modeling.ErnieModel'> is patched and the patch might be based on an old oversion which missing some arguments compared with the latest, such as ['output_hidden_states', 'output_attentions', 'return_dict']. We automatically add compatibility on the patch for these arguemnts, and maybe the patch should be updated.
  warnings.warn(
/ssd3/zhaol/anaconda3/envs/paddlenlp_env/lib/python3.10/site-packages/paddlenlp/transformers/utils.py:212: UserWarning: The `forward` method of <class 'paddle.nn.layer.transformer.TransformerEncoder'> is patched and the patch might be conflict with patches made by paddlenlp which seems have more arguments such as ['output_hidden_states', 'output_attentions', 'return_dict']. We automatically add compatibility on the patch for these arguemnts, and maybe the patch should be updated.
  warnings.warn(
[2025-04-15 20:30:58,691] [    INFO] - tokenizer config file saved in ./infer_model/tokenizer_config.json
[2025-04-15 20:30:58,691] [    INFO] - Special tokens file saved in ./infer_model/special_tokens_map.json

export之后的infer文件夹：
model.json  model.pdiparams  special_tokens_map.json  tokenizer_config.json  vocab.txt

运行python -u deploy/python/predict.py报错：
config = paddle.inference.Config(args.model_name_or_path + ".pdmodel", args.model_name_or_path + ".pdiparams")
RuntimeError: (NotFound) Cannot open file ./infer_model/model.pdmodel, please confirm whether the file is normal.
  [Hint: Expected paddle::inference::IsFileExists(prog_file_) == true, but received paddle::inference::IsFileExists(prog_file_):0 != true:1.] (at ../paddle/fluid/inference/api/analysis_config.cc:112)

稳定复现步骤 & 代码

我是按照https://github.com/PaddlePaddle/PaddleNLP/tree/develop/slm/examples/machine_reading_comprehension/SQuAD#%E9%9D%99%E6%80%81%E5%9B%BE%E9%A2%84%E6%B5%8B 的步骤进行的

[cycao77]

应该是新版本的bug，降低版本后可以export成功，降低后是：
paddlenlp 2.6.0
paddlepaddle-gpu 2.5.1.post102

paddle和paddlenlp版本兼容性太差了，用着真的让人不想用第二次，太烂了

[ZHUI]

你好，最新的Paddle推理格式进行了升级。有一些没适配好的地方，目前我们正在修。不好意思 #9763