transformers - 💡(How to fix) Fix Optimizer step being called 2 times when using deepspeed [2 comments, 2 participants]

TL;DR

The issue can be fixed by modifying the trainer.py in the transformers library to remove the redundant optimizer.step call after accelerator.backward.

Guidance

• Identify the line of code in trainer.py where optimizer.step is called after accelerator.backward and remove or comment it out to prevent the duplicate optimizer step.
• Verify that the modification fixes the issue by checking the number of optimizer steps performed during a single iteration.
• Review the code to ensure that the accelerator.backward call is correctly handling the gradient accumulation step without requiring an additional optimizer.step call.
• Test the modified code with a simple example to confirm that the fix does not introduce any new issues.

Example

No code snippet is provided as the issue does not include the specific code lines that need to be modified.

Notes

The fix assumes that the accelerator.backward call is correctly implemented and that the removal of the redundant optimizer.step call will not affect the overall functionality of the code.

Recommendation

Apply workaround: Modify the trainer.py file to remove the redundant optimizer.step call, as this is a specific fix for the identified issue and does not require upgrading to a new version.