Add support for distributed checkpointing of HF safetensors with DCP

[ankitageorge]

Context

What is the purpose of this PR? Is it to

• add a new feature
• fix a bug
• update tests and/or documentation
• other (please add here)

Please link to any issues this PR addresses.

Changelog

What are the changes made in this PR?

• Currently when saving and loading HF safetensors with DCP, we're still doing rank zero consolidation and then saving it. With this change we can save checkpoints in a distributed fashion and do the consolidation in DCP's finish step, and by-pass the rank-zero checkpointing which is the current approach. This PR adds the ability to use this functionality. This finish step consolidation is currently slow, but working on speeding it up. The API call from torchtune shouldn't change.
• Without the consolidation, on a 70B model that was taking 260s to save without DCP, is now taking 70s to save the sharded tensors.

Test plan

Please make sure to do each of the following if applicable to your PR. If you're unsure about any one of these just ask and we will happily help. We also have a contributing page for some guidance on contributing.

• [x ] run pre-commit hooks and linters (make sure you've first installed via pre-commit install)
• add unit tests for any new functionality
• update docstrings for any new or updated methods or classes
• [x ] run unit tests via pytest tests
• [x ] run recipe tests via pytest tests -m integration_test
• [ x] manually run any new or modified recipes with sufficient proof of correctness
• include relevant commands and any other artifacts in this summary (pastes of loss curves, eval results, etc.)

UX

If your function changed a public API, please add a dummy example of what the user experience will look like when calling it.
Here is a docstring example
and a tutorial example

• I did not change any public API
• [x ] I have added an example to docs or docstrings

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/torchtune/2851

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit b215e3a with merge base 3d73591 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.