fix: allow loading of nested states in Fabric.load [wip]

[Markus28]

What does this PR do?

This PR aims to allow loading nested states via Fabric.load (fixes #20208). Nested states can add some structure when multiple models are trained simultaneously (e.g. in GANs or in reinforcement learning).

Currently, it is my understanding that the following methods need to be modified

• Fabric.load: In this method, some entries of the state dictionaries are replaced with those loaded from the checkpoint. Until now, only fabric-internal objects were not replaced (since the respective state dicts had been loaded previously). We need to apply this replacement procedures recursively to sub-dictionaries (instead of replacing them).
• Strategy._convert_stateful_objects_in_state: This method again has to recursively walk into sub-dictionaries and turn stateful objects into their state-dicts.
• Strategy.load_checkpoint and anything overriding it: Again, we need to recursively walk into the sub-dictionaries and possibly load the state-dicts of stateful objects.
• Strategy.save_checkpoint: Perhaps depending on the strategy there is also some relevant code here (not the case for the abstract strategy class).

Any strategy that overrides load_checkpoint will also have to be updated.
I am not sure whether these are all the changes required to make this work, but preliminary tests from my side look okay. If someone more familiar with the codebase could give some feedback on whether this covers everything, I'd be grateful.

TODO:

• Update Fabric.load
• Update Strategy._convert_stateful_objects_in_state and Strategy.load_checkpoint
• Possibly update the load_checkpoint and save_checkpoint methods of strategies overriding the Strategy methods
• Add tests

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--20210.org.readthedocs.build/en/20210/

[lantiga]

hey @Markus28, makes sense to me. Can you start adding a few tests to the PR?