Sharded state dicts save correctly when save_weights_only=True

[dimitri-voytan]

What does this PR do?

Fixes #19492 for FSDP sharded state_dicts. Optimizer states are default to an empty list if they are not in the state_dict, which can happen when the model checkpoint callback uses save_weights_only=True

Before submitting

• Was this discussed/agreed via a GitHub issue? (not for typos and docs)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests? (not for typos and docs)
• Did you verify new and existing tests pass locally with your changes?
• Did you list all the breaking changes introduced by this pull request?
• [ ]Did you update the CHANGELOG? (not for typos, docs, test updates, or minor internal changes/refactors)

PR review

Anyone in the community is welcome to review the PR.
Before you start reviewing, make sure you have read the review guidelines. In short, see the following bullet-list:

Reviewer checklist

• Is this pull request ready for review? (if not, please submit in draft mode)
• Check that all items from Before submitting are resolved
• Make sure the title is self-explanatory and the description concisely explains the PR
• Add labels and milestones (and optionally projects) to the PR so it can be classified

---

📚 Documentation preview 📚: https://pytorch-lightning--19524.org.readthedocs.build/en/19524/

[dimitri-voytan]

@awaelchli

This works now in that the test fails on master and passes with the corrections. It could improve with verification that the checkpoints are useable in the lightning ecosystem downstream. If this is of interest, can I ask the lightning team for some help as I don't have a lot of free time to work on this.

[awaelchli]

@dimitri-voytan Great fix, thanks a lot!

I integrated your test in an existing one we already had for weights_only=True, we just needed to add the parameterization for sharded checkpoints.