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

Noisy loss fix #293

Merged
merged 4 commits into from
Nov 9, 2023
Merged

Noisy loss fix #293

merged 4 commits into from
Nov 9, 2023

Conversation

bocchris-aws
Copy link
Contributor

This reverts the changes of the sampler from commit 757886b as it leads to loss curves that diverge from GPU baseline. Closes #253.

…ng and noisy loss with TP>1.""

This reverts commit 757886b.
@bocchris-aws
Copy link
Contributor Author

bocchris-aws commented Nov 6, 2023

Loss curves align with GPU for t5-small and TP=8. Additional tests for larger models are on its way.
Gray: Diverging loss curve from current main branch.
Blue: GPU baseline
Magenta: Loss curve with this fix.

Screenshot 2023-11-06 at 16 13 04

@HuggingFaceDocBuilderDev

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

@bocchris-aws bocchris-aws changed the title Rolled back changes in sampler as it leads to diverging and noisy loss with TP>1 Noisy loss fix Nov 6, 2023
@dacorvo
Copy link
Collaborator

dacorvo commented Nov 9, 2023

You need to reformat two of the files you modified. You can apply style changes automatically by typing make style in your local directory.

Copy link
Member

@michaelbenayoun michaelbenayoun left a comment

Choose a reason for hiding this comment

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

LGTM!

The code quality / style check does not pass, can you run the following command please?

make style

@michaelbenayoun michaelbenayoun merged commit 6ed2eca into main Nov 9, 2023
7 of 13 checks passed
@michaelbenayoun michaelbenayoun deleted the fix_noisy_loss branch November 9, 2023 17:30
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

Loss curve mismatch with tensor parallelism
4 participants