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

[fix bug] logits's shape different from label's shape in preprocess_logits_for_metrics #31447

Merged
merged 5 commits into from
Jul 3, 2024

Conversation

wiserxin
Copy link
Contributor

@wiserxin wiserxin commented Jun 17, 2024

What does this PR do?

After update my tranformers 4.39.3 --> 4.41.2, training LLM with deepspeed DDP method, something wrong when it goes to my preprocess_logits_for_metrics function: logits's shape is different from label's shape.

Then i discover that in latest trainer code, before passing to preprocess_logits_for_metrics, logits are padded by function pad_across_processes but labels are not.

It seems like a bug unexpectedly introduced .

Before submitting

  • This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
  • Did you read the contributor guideline,
    Pull Request section?
  • Was this discussed/approved via a Github issue or the forum? Please add a link
    to it if that's the case.
  • Did you make sure to update the documentation with your changes? Here are the
    documentation guidelines, and
    here are tips on formatting docstrings.
  • Did you write any new necessary tests?

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.

@amyeroberts
Copy link
Collaborator

cc @muellerzr @SunMarc

@SunMarc
Copy link
Member

SunMarc commented Jun 17, 2024

Hi @wiserxin, this bug was indeed introduced by this PR #30267. The padding of the labels happens after preprocess_logits_for_metrics, hence the issue that you are experiencing. @qubvel, is it safe to put it back where it was ?

@qubvel
Copy link
Member

qubvel commented Jun 17, 2024

Oh, that's true. @SunMarc yes, should be safe. Alternatively, we can just move the "if" block with labels padding before the "if" block with logits.

@SunMarc
Copy link
Member

SunMarc commented Jun 17, 2024

Alternatively, we can just move the "if" block with labels padding before the "if" block with logits

Yeah, I think that's would be better for readability + it was the previous behavior ! Could you make the change @wiserxin ?

@wiserxin
Copy link
Contributor Author

Alternatively, we can just move the "if" block with labels padding before the "if" block with logits

Yeah, I think that's would be better for readability + it was the previous behavior ! Could you make the change @wiserxin ?

sure

labels can't use  `gather` before pass to `preprocess_logits_for_metrics`, so must split into 2 if-block
Copy link
Member

@SunMarc SunMarc left a comment

Choose a reason for hiding this comment

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

LGTM ! Thanks for the PR @wiserxin !

@SunMarc SunMarc requested a review from amyeroberts June 18, 2024 11:56
@qubvel
Copy link
Member

qubvel commented Jun 18, 2024

@SunMarc @wiserxin let's maybe add a comment about the order of this "if" block to avoid the same bug in the future. :)

@HuggingFaceDocBuilderDev

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.

Copy link
Collaborator

@amyeroberts amyeroberts left a comment

Choose a reason for hiding this comment

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

Thanks for fixing!

+1 on @qubvel's comment

@wiserxin
Copy link
Contributor Author

@SunMarc @wiserxin let's maybe add a comment about the order of this "if" block to avoid the same bug in the future. :)

good idea

Copy link
Contributor

@muellerzr muellerzr left a comment

Choose a reason for hiding this comment

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

Thanks a bunch!

@muellerzr muellerzr merged commit 534cbf8 into huggingface:main Jul 3, 2024
21 checks passed
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.

6 participants