Fix failing test on main

[muellerzr]

What does this PR do?

We should only save/load the sampler if it's an iterable dataset shard, and if we use our custom sampler. Otherwise pulling from the batch sampler and loading it is more nuanced. For the sake of the release, let's focus on just the new sampler. If users report that they can't reproduce trainings on batch samplers (or if we test and find we can't), then we can pivot after.

Fixes # (issue)

Failing test on main that caught this

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?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@LysandreJik @BenjaminBossan

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[BenjaminBossan]

LGTM, thanks for fixing the issue.

From my understanding, this is a -- potentially temporary -- fix that makes the saving and loading of the sampler more specific because the current code does not cover some cases. If that's right, could you elaborate where it goes wrong so that we have it written down for the future?

[muellerzr]

Where it goes askew is we need a better way to get the users batch sampler from our custom batch sampler. Such as dataloader.batch_sampler.batch_sampler. (Since our sampler stores that in the batch_sampler.

One potential fix: add dataloader.get_sampler() which will know to return either the batch_sampler or sampler of the original dataloader