transformers - 💡(How to fix) Fix from_pretrained loads wrong custom module after save_pretrained [3 comments, 2 participants]

TL;DR

The issue can be fixed by ensuring that the custom_model.py file is not modified during the saving process, allowing the subsequent load to use the correct cached version.

Guidance

• The problem seems to be related to the way the custom_model.py file is being modified and cached during the saving process, causing the later load to use the local version instead of the cached one.
• To verify this, check the HF_MODULES_CACHE directory to see if the custom_model.py file is being updated correctly for each saved model.
• One possible workaround is to ensure that the custom_model.py file is not imported or modified during the saving process, allowing the subsequent load to use the correct cached version.
• Another possible solution is to use a different approach to save and load the models, avoiding the use of the custom_model.py file altogether.

Example

No code snippet is provided as the issue is more related to the workflow and caching mechanism rather than a specific code error.

Notes

The issue seems to be specific to the custom implementation of saving and loading models, and the provided information does not give a clear indication of the root cause. Further investigation into the caching mechanism and the saving process is needed to provide a more accurate solution.

Recommendation

Apply workaround: Avoid modifying the custom_model.py file during the saving process to ensure that the subsequent load uses the correct cached version. This can be done by reorganizing the code to separate the saving and loading processes, or by using a different approach to save and load the models.