FSDP grad accum fix #34645
Merged
FSDP grad accum fix #34645
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@ArthurZucker @muellerzr for review |
muellerzr
approved these changes
Nov 7, 2024
|
is it possible to also trigger the slow tests for this PR too to make sure the new test passes? |
SunMarc
approved these changes
Nov 15, 2024
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
ydshieh
approved these changes
Nov 15, 2024
ArthurZucker
approved these changes
Nov 15, 2024
2015aroras
pushed a commit
to 2015aroras/transformers
that referenced
this pull request
Nov 15, 2024
* add gradient accumulation steps tests for fsdp * invert no_sync context to fix training for fsdp
ArthurZucker
pushed a commit
that referenced
this pull request
Nov 18, 2024
* add gradient accumulation steps tests for fsdp * invert no_sync context to fix training for fsdp
4 tasks
4 tasks
BernardZach
pushed a commit
to BernardZach/transformers
that referenced
this pull request
Dec 5, 2024
* add gradient accumulation steps tests for fsdp * invert no_sync context to fix training for fsdp
BernardZach
pushed a commit
to innovationcore/transformers
that referenced
this pull request
Dec 6, 2024
* add gradient accumulation steps tests for fsdp * invert no_sync context to fix training for fsdp
4 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
8c62a92#diff-ed55888e6665791fe92cc8fc0c499da54f4ace6738551cd9a2591881cda076deL2473-R2480 seems to have broken FSDP training as we need to wrap the train step in no_sync on every step except the last step in the gradient accumulation batch. Without this, during the optimizer.step(), I get the error:
this wasn't caught by the fsdp tests since gradient accumulation wasn't part of those tests, so I also added tests for that as well.
see https://muellerzr.github.io/blog/gradient_accumulation.html#what-is-the-right-way-then
Fixes # (issue)
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.