[data_loader] Optionally also propagate set_epoch to batch sampler

[tomaarsen]

What does this PR do?

When training with the transformers Trainer or related (Sentence Transformers, SpanMarker, SetFit, etc.), set_epoch is called on the dataloader. This is propagated down to the dataloader.batch_sampler.sampler if that has a set_epoch , but not to dataloader.batch_sampler.

This prevents epoch-specific generator seeding in custom batch samplers, such as the ones that are common in Sentence Transformers: https://sbert.net/docs/package_reference/sentence_transformer/sampler.html

See also UKPLab/sentence-transformers#3069 by @antigregory which showed that set_epoch in my Batch Samplers is not called like I expected them to be.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@muellerzr (p.s. get better soon!)

• Tom Aarsen

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[muellerzr]

Thanks for propagating! Can you confirm it fixes your issue? (If so, we either should add a test, or add SentenceTransformer test subsets into accelerate)

[tomaarsen]

Can you confirm it fixes your issue?

It does indeed. I think I can write a test to go alongside this PR - it doesn't need to be ST-specific.

• Tom Aarsen

[tomaarsen]

This test roughly mirrors my use case and should act as an effective test to make sure ST should work. Feel free to update it if e.g. you don't normally use Accelerator in the data_loader tests.

• Tom Aarsen

[muellerzr]

Thanks for adding!