Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Accommodate FSDP full-precision param_dtype training with PyTorch < 2.0 #18278

Conversation

speediedan
Copy link
Contributor

@speediedan speediedan commented Aug 10, 2023

Fixes #18277

When FSDP training with full precision param_dtypes (16-mixed, bf16-mixed and 32-true configurations) and PyTorch < 2.0, FSDP training will encounter this assertion error.

This is because FSDP uses the noneness of param_dtype as a proxy for the _uses_param_mixed_precision property and FSDPPrecisionPlugin currently sets the default param_dtype to torch.float32 when training in full precision.

This PR:

  1. Sets MixedPrecision param_dtype to None when FSDP training with full precision param_dtypes and PyTorch < 2.0
  2. Updates relevant Lightning and Fabric 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

… with full precision ``param_dtype`` s and PyTorch < 2.0, condition ``lightning_module_state_dict`` FSDP state dict import locations on PyTorch version, emit ``UserWarnings`` rather than attempting to load/save optim state dicts with FSDP and PyTorch < 2.0
@github-actions github-actions bot added fabric lightning.fabric.Fabric pl Generic label for PyTorch Lightning package labels Aug 10, 2023
…ategy_load_optimizer_states`, improve `test_configure_model` to accommodate SGD issues with FSDP and PyTorch `2.0`
@speediedan speediedan marked this pull request as ready for review August 10, 2023 23:14
@speediedan
Copy link
Contributor Author

I believe the currently failing check (Test Fabric / fabric-cpu (ubuntu-20.04, lightning, 3.8, 1.11): test_collectives_distributed[2]) is not related to this PR.

@speediedan speediedan changed the title Accommodate FSDP full-precision param_dtype training and issue optim state dict loading/saving warnings with PyTorch < 2.0 Accommodate FSDP full-precision param_dtype training with PyTorch < 2.0 Aug 13, 2023
@awaelchli awaelchli added strategy: fsdp Fully Sharded Data Parallel community This PR is from the community labels Aug 13, 2023
@awaelchli awaelchli added this to the 2.0.x milestone Aug 13, 2023
@carmocca carmocca merged commit c081b48 into Lightning-AI:master Aug 14, 2023
103 checks passed
@mergify mergify bot added the ready PRs ready to be merged label Aug 14, 2023
Borda pushed a commit that referenced this pull request Aug 28, 2023
lantiga pushed a commit that referenced this pull request Aug 30, 2023
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
community This PR is from the community fabric lightning.fabric.Fabric pl Generic label for PyTorch Lightning package ready PRs ready to be merged strategy: fsdp Fully Sharded Data Parallel
Projects
None yet
Development

Successfully merging this pull request may close these issues.

FSDP full-precision param_dtype training with PyTorch < 2.0 triggers FSDP assertion error
3 participants