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

[FSDP] Adding Native FSDP Strategy #12447

Merged
merged 13 commits into from
May 5, 2022

Conversation

sisilmehta2000
Copy link
Contributor

@sisilmehta2000 sisilmehta2000 commented Mar 24, 2022

What does this PR do?

Pytorch 1.11.0 makes torch.distributed.fsdp available for public use (PR #72084) https://github.com/pytorch/pytorch/releases/tag/v1.11.0

Currently there is a strategies/fully_sharded.py strategy that uses the fairscale implementation of fully sharded data parallel technique. With the introduction of fsdp in torch.distributed we want to create a new strategy called strategies/fully_sharded2.py

Reasons to add a new strategy:

  • Torch.distributed.fsdp doesn’t currently implement all the features of the fairscale version
  • We don't want the torch distributed API to feel restricted based on current integrations with fairscale. By keeping them separate, we'll adapt to whatever recommendation the torch.distributed team provides
  • It's cheap/easy for us to add a new strategy class
  • We want to have people start using the new torch.distributed.fsdp version without introducing a breaking change to those using the fairscale version

Fixes #12334

Does your PR introduce any breaking changes? If yes, please list them.

None

Before submitting

  • Was this discussed/approved 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)
  • [n] 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 internal minor changes/refactorings)

PR review

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

  • 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

Did you have fun?

Make sure you had fun coding 🙃

@sisilmehta2000
Copy link
Contributor Author

So some of the tests like conda(1.0, 1.8) are failing because it can't find torch.distributed.fsdp. That only exists in torch version 1.11.

@sisilmehta2000 sisilmehta2000 force-pushed the strategy/fsdp_native branch 3 times, most recently from a064b70 to 7903625 Compare March 27, 2022 01:30
@carmocca carmocca added strategy: fsdp Fully Sharded Data Parallel and removed strategy: fairscale fsdp (removed) Fully Sharded Data Parallel labels Mar 28, 2022
@mergify mergify bot added the has conflicts label Apr 5, 2022
@sisilmehta2000 sisilmehta2000 force-pushed the strategy/fsdp_native branch 2 times, most recently from ef4c67e to 118a0ea Compare April 6, 2022 20:26
@mergify mergify bot added has conflicts and removed ready PRs ready to be merged labels May 4, 2022
@mergify mergify bot added ready PRs ready to be merged and removed has conflicts ready PRs ready to be merged labels May 4, 2022
@SeanNaren SeanNaren enabled auto-merge (squash) May 5, 2022 10:40
@SeanNaren SeanNaren disabled auto-merge May 5, 2022 11:44
@SeanNaren SeanNaren enabled auto-merge (squash) May 5, 2022 11:47
@SeanNaren SeanNaren self-requested a review May 5, 2022 12:03
Copy link
Member

@Borda Borda left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

goooo 🎉

@SeanNaren SeanNaren merged commit d337374 into Lightning-AI:master May 5, 2022
@SeanNaren SeanNaren mentioned this pull request May 5, 2022
12 tasks
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
feature Is an improvement or enhancement 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.

Use PyTorch's Fully Sharded Data Parallel API with 1.11+