🚨 Fully revert atomic checkpointing 🚨

[muellerzr]

What does this PR do?

As discussed offline, while on paper having atomic checkpointing was a good idea, it's unmaintainable by the core maintainers and has resulted in an influx of issues from users considering the wide variety of storage options available to them. (and untestable)

As the feature request, while reasonable and from a good mindset, was not due to an outright bug, we are reverting it entirely.

Related PRs:

#27820, #28364

Did revert based on this commit: fe8d130

Related issues:

• Save model checkpoint error when multi-gpu training #27925
• 4.36 transformers got wrong _save_checkpoint with deepspeed. work with previous versions #28027
• Permission Denied Error When Saving Checkpoint During Training #29382

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.

@ArthurZucker @LysandreJik

[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.

[ArthurZucker]

Thank you. That's a good learning experience for me, saving files is tricky and a lot of users have a lot of different hardwares!

[amyeroberts]

Thanks for fixing! Agreed removing this feature is a good idea.

Let's add a 🚨 prefix to the PR title so we know this is a "breaking" change - i.e. behaviour is changing between releases

[Chandler-Bing]

wonderful PR! hope 4.38.3 version release soon

[shiningliang]

Will this change be included in 4.38.3? Thanks

[amyeroberts]

@Chandler-Bing @shiningliang 4.38.2 -> 4.38.3 would represent a patch release, these are reserved for fixing regressions or unintended breaking changes from a minor release i.e. 4.37 -> 4.38.

We have minor releases roughly once a month, so the next release, 4.39.0 will be in a week or two.

[wxjiao]

Nice PR! Upgraded transformers to the latest version (i.e., main), now the checkpoints can be saved smoothly in mult-node setting.