onnx==1.18.0 is causing MONAI to fail to import on Windows #8455

Describe the bug

When installing onnx==1.18.0 alongside monai==1.4.0 on Windows, the import fails with no exception or log message.

To Reproduce

Steps to reproduce the behavior:

1. Create a fresh Python 3.9/3.10 conda environment (may affect other versions as well)

2. Install the following list of packages using pip install -r requirements.txt (note: we first install onnx==1.16.1 to showcase a working install):

   Details

   absl-py==2.2.2
   acres==0.3.0
   acvl_utils==0.2
   annotated-types==0.7.0
   astor==0.8.1
   batchgenerators==0.25.1
   bids-validator==1.14.7.post0
   bidsschematools==1.0.7
   certifi==2025.4.26
   charset-normalizer==3.4.2
   click==8.1.8
   colorama==0.4.6
   coloredlogs==15.0.1
   connected-components-3d==3.23.0
   contourpy==1.3.0
   coverage==7.8.0
   csv-diff==1.2
   cycler==0.12.1
   Cython==3.1.0
   deepdiff==8.5.0
   Deprecated==1.2.18
   dicom2nifti==2.6.0
   dictdiffer==0.9.0
   dipy==1.8.0
   docker-pycreds==0.4.0
   docopt==0.6.2
   dynamic_network_architectures==0.3.1
   einops==0.8.1
   eval_type_backport==0.2.2
   exceptiongroup==1.3.0
   filelock==3.18.0
   flatbuffers==25.2.10
   fonttools==4.58.0
   formulaic==0.3.4
   fsspec==2025.3.2
   future==1.0.0
   gitdb==4.0.12
   GitPython==3.1.44
   graphviz==0.20.3
   grpcio==1.71.0
   gryds==0.0.9
   h5py==3.13.0
   humanfriendly==10.0
   humanize==4.12.3
   idna==3.10
   imagecodecs==2024.12.30
   imageio==2.37.0
   importlib_metadata==8.7.0
   importlib_resources==6.5.2
   iniconfig==2.1.0
   interface-meta==1.3.0
   ivadomed==2.9.10
   Jinja2==3.1.6
   joblib==1.5.0
   kiwisolver==1.4.7
   lazy_loader==0.4
   linecache2==1.0.0
   loguru==0.7.3
   lxml==5.4.0
   Markdown==3.8
   markdown-it-py==3.0.0
   MarkupSafe==3.0.2
   matplotlib==3.9.4
   matplotlib-inline==0.1.7
   mdurl==0.1.2
   monai==1.4.0
   mpmath==1.3.0
   msvc-runtime==14.29.30133
   networkx==3.2.1
   neurite==0.2
   nibabel==5.3.2
   nilearn==0.10.4
   nnunetv2==2.3.1
   num2words==0.5.14
   numpy==1.26.4
   onnx==1.16.1
   onnxruntime==1.19.2
   orderly-set==5.4.1
   packaging==25.0
   pandas==1.5.3
   pillow==11.2.1
   platformdirs==4.3.8
   pluggy==1.6.0
   portalocker==3.1.1
   protobuf==6.31.0
   psutil==7.0.0
   pybids==0.15.5
   pydantic==2.11.4
   pydantic_core==2.33.2
   pydicom==2.4.4
   Pygments==2.19.1
   pyparsing==3.2.3
   PyQt5==5.12.3
   PyQt5-sip==12.12.2
   pyreadline3==3.5.4
   pystrum==0.4
   pytest==8.3.5
   pytest-console-scripts==1.4.1
   pytest-cov==6.1.1
   python-dateutil==2.9.0.post0
   python-gdcm==3.0.24.1
   pytz==2025.2
   pywin32==310
   PyYAML==6.0.2
   requests==2.32.3
   requirements-parser==0.11.0
   rich==14.0.0
   scikit-image==0.24.0
   scikit-learn==1.6.1
   scipy==1.13.1
   seaborn==0.13.2
   sentry-sdk==2.28.0
   setproctitle==1.3.6
   setuptools-scm==8.3.1
   shellingham==1.5.4
   simpleitk==2.5.0
   six==1.17.0
   smmap==5.0.2
   # Editable install with no version control (spinalcordtoolbox==7.1.dev0)
   -e c:\users\joshua\sct_7.1.dev0
   SQLAlchemy==1.3.24
   sympy==1.14.0
   tensorboard==2.19.0
   tensorboard-data-server==0.7.2
   threadpoolctl==3.6.0
   tifffile==2024.8.30
   tomli==2.2.1
   torch==2.2.2+cpu
   torchio==0.20.8
   torchvision==0.17.2+cpu
   totalspineseg==20250205
   tqdm==4.67.1
   traceback2==1.4.0
   traitlets==5.14.3
   transforms3d==0.4.2
   trx-python==0.3
   typer==0.15.4
   types-setuptools==80.7.0.20250516
   typing-inspection==0.4.0
   typing_extensions==4.13.2
   unittest2==1.1.0
   urllib3==2.4.0
   voxelmorph==0.2
   wandb==0.19.11
   Werkzeug==3.1.3
   win32_setctime==1.2.0
   wquantiles==0.6
   wrapt==1.17.2
   XlsxWriter==3.2.3
   xlwt==1.3.0
   yacs==0.1.8
   zipp==3.21.0
   

3. Run the command python -c "import monai; print('it worked!')", and the message will be printed.

4. Now, install onnx==1.18.0 into the environment instead.

5. Re-running the command python -c "import monai; print('it worked!')" will not print any message, or even an exception.

Additional context

onnx==1.18.0 and monai==1.4.0 are not enough to fully reproduce the issue. It seems some other package among my environment is contributing to this issue.

To better understand this issue, I tried comparing the import monai commands using the python -v option:

• onnx==1.18.0 + monai==1.4.0 + extra packages: Failure (output.txt)
• onnx==1.16.1 + monai==1.4.0 + extra packages: Success (output.txt)

Diffing these outputs, we can see that the outputs begin to diverge when importing onnx, specifically:

import 'onnx.serialization' # <_frozen_importlib_external.SourceFileLoader object at <0x00000200DD729520>
# extension module 'onnx.onnx_cpp2py_export' loaded from 'C:\\Users\\Joshua\\sct_7.1.dev0\\python\\envs\\venv_sct\\lib\\site-packages\\onnx\\onnx_cpp2py_export.cp39-win_amd64.pyd'

Also, the failing import halts specifically at these lines:

import 'onnx.model_container' # <_frozen_importlib_external.SourceFileLoader object at <0x00000200E84DF2E0>
# C:\Users\Joshua\sct_7.1.dev0\python\envs\venv_sct\lib\site-packages\onnx\reference\__pycache__\op_run.cpython-39.pyc matches C:\Users\Joshua\sct_7.1.dev0\python\envs\venv_sct\lib\site-packages\onnx\reference\op_run.py
# code object from 'C:\\Users\\Joshua\\sct_7.1.dev0\\python\\envs\\venv_sct\\lib\\site-packages\\onnx\\reference\\__pycache__\\op_run.cpython-39.pyc'

I am not sure why my combination of packages causes this difference, and I am also not sure why there is no failure for just onnx or just onnx+monai.

I am also not sure whether to report this issue on the monai repo or the onnx repo. But, since this issue doesn't seem to be affecting import onnx alone, I thought to report it here as well.

Given the mention of onnx_cpp2py_export specifically, I think it might be related to:

• [1.16.2/1.17?] ONNX build Windows onnx/onnx#6267