Nougat OCR中cached_property导入错误与BARTDecoder参数异常问题

错误1：cached_property导入错误（ImportError） #

错误信息 #

ImportError: cannot import name 'cached_property' from 'nougat.utils' (/lfs/skampere1/0/emilyhyf/miniconda/lib/python3.12/site-packages/nougat/utils/__init__.py)
OCRing with base model failed on /lfs/skampere1/0/emilyhyf/massive-evaporation-4-math/data/debug_pdfs/dummy_folder1/The_IMO_Compendium.pdf... trying small model
Traceback (most recent call last):
  File "/lfs/skampere1/0/emilyhyf/miniconda/bin/nougat", line 5, in <module>
    from predict import main
  File "/lfs/skampere1/0/emilyhyf/miniconda/lib/python3.12/site-packages/predict.py", line 18, in <module>
    from nougat import NougatModel
  File "/lfs/skampere1/0/emilyhyf/miniconda/lib/python3.12/site-packages/nougat/__init__.py", line 1, in <module>
    from nougat.app import Nougat
  File "/lfs/skampere1/0/emilyhyf/miniconda/lib/python3.12/site-packages/nougat/app.py", line 5, in <module>
    from nougat.asgi import serve
  File "/lfs/skampere1/0/emilyhyf/miniconda/lib/python3.12/site-packages/nougat/asgi.py", line 6, in <module>
    from nougat.context.request import Request
  File "/lfs/skampere1/0/emilyhyf/miniconda/lib/python3.12/site-packages/nougat/context/__init__.py", line 1, in <module>
    from nougat.context.request import Request
  File "/lfs/skampere1/0/emilyhyf/miniconda/lib/python3.12/site-packages/nougat/context/request.py", line 8, in <module>
    from nougat.utils import cached_property, File
ImportError: cannot import name 'cached_property' from 'nougat.utils' (/lfs/skampere1/0/emilyhyf/miniconda/lib/python3.12/site-packages/nougat/utils/__init__.py)
OCRing with small model failed on /lfs/skampere1/0/emilyhyf/massive-evaporation-4-math/data/debug_pdfs/dummy_folder1/The_IMO_Compendium.pdf
The following files failed OCRing and need manual review:
/lfs/skampere1/0/emilyhyf/massive-evaporation-4-math/data/debug_pdfs/putnam_and_beyond.pdf
/lfs/skampere1/0/emilyhyf/massive-evaporation-4-math/data/debug_pdfs/dummy_folder1/The_IMO_Compendium.pdf
Time taken: 8.497612476348877 seconds, 0.1416268746058146 minutes, 0.00236044791009691 hours

成因 #

Nougat版本与Python 3.12不兼容：Python 3.8+标准库已内置functools.cached_property，部分旧版Nougat可能移除了内部实现的cached_property，但代码中仍保留从nougat.utils导入的逻辑；或者该模块未在nougat.utils.__init__.py中导出。

解决方法 #

1. 升级Nougat到兼容版本：执行命令

pip install --upgrade nougat-ocr

2. 临时修改导入逻辑：
   • 打开nougat/context/request.py文件，将from nougat.utils import cached_property替换为from functools import cached_property；
   • 或在nougat/utils/__init__.py中添加一行：from functools import cached_property。

错误2：BARTDecoder的cache_position参数错误（TypeError） #

错误信息 #

TypeError: BARTDecoder.prepare_inputs_for_inference() got an unexpected keyword argument 'cache_position'
-> Cannot close object, library is destroyed. This may cause a memory leak!
-> Cannot close object, library is destroyed. This may cause a memory leak!
-> Cannot close object, library is destroyed. This may cause a memory leak!
OCRing with base model failed on /lfs/skampere1/0/naveenkc/contest_scraper/test_nougat/731991.pdf... trying small model
/lfs/skampere1/0/naveenkc/miniconda/lib/python3.12/site-packages/torch/functional.py:512: UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at ../aten/src/ATen/native/TensorShape.cpp:3587.)
  return _VF.meshgrid(tensors, **kwargs) # type: ignore[attr-defined]
  0%|                                                 | 0/288 [00:24<?, ?it/s]
Traceback (most recent call last):
  File "/lfs/skampere1/0/naveenkc/miniconda/bin/nougat", line 8, in <module>
    sys.exit(main())
             ^^^^^^
  File "/lfs/skampere1/0/naveenkc/miniconda/lib/python3.12/site-packages/predict.py", line 167, in main
    model_output = model.inference(
                   ^^^^^^^^^^^^^^^^
  File "/lfs/skampere1/0/naveenkc/miniconda/lib/python3.12/site-packages/nougat/model.py", line 592, in inference
    decoder_output = self.decoder.model.generate(
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/lfs/skampere1/0/naveenkc/miniconda/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/lfs/skampere1/0/naveenkc/miniconda/lib/python3.12/site-packages/transformers/generation/utils.py", line 1758, in generate
    result = self._sample(
             ^^^^^^^^^^^^^
  File "/lfs/skampere1/0/naveenkc/miniconda/lib/python3.12/site-packages/transformers/generation/utils.py", line 2394, in _sample
    model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: BARTDecoder.prepare_inputs_for_inference() got an unexpected keyword argument 'cache_position'
-> Cannot close object, library is destroyed. This may cause a memory leak!
-> Cannot close object, library is destroyed. This may cause a memory leak!
-> Cannot close object, library is destroyed. This may cause a memory leak!
-> Cannot close object, library is destroyed. This may cause a memory leak!
-> Cannot close object, library is destroyed. This may cause a memory leak!
OCRing with small model failed on /lfs/skampere1/0/naveenkc/contest_scraper/test_nougat/731991.pdf
The following files failed OCRing and need manual review:
/lfs/skampere1/0/naveenkc/contest_scraper/test_nougat/731991.pdf
Time taken: 65.70022702217102 seconds, 1.0950037837028503 minutes, 0.018250063061714172 hours

成因 #

Transformers版本与Nougat不兼容：新版本Transformers在文本生成逻辑中新增了cache_position参数，但旧版Nougat的BARTDecoder.prepare_inputs_for_inference方法未处理该参数，导致传参时报错。同时Python 3.12环境下，部分依赖版本的兼容性问题会放大这类参数不匹配的错误。

解决方法 #

1. 安装Nougat指定的Transformers版本：查看Nougat官方requirements文件，安装对应版本，例如：

pip install transformers==4.33.2

2. 修改Nougat代码适配新版本Transformers：
   • 打开nougat/model.py，找到调用generate方法的位置，移除cache_position相关参数；
   • 或修改BARTDecoder的prepare_inputs_for_inference方法，添加**kwargs参数以接收多余的关键字参数。
3. 确保Torch与Transformers版本兼容：根据Transformers官方文档，安装与Python 3.12匹配的Torch版本，避免跨版本兼容性问题。

内容的提问来源于stack exchange，提问作者Charlie Parker