Skip to content

Conversation

@tomaarsen
Copy link
Member

What does this PR do?

  1. Add integration tests with the now-released checkpoint + 2 checkpoints under hf-internal-testing (thanks @xenova !)
  2. Reuse GemmaRotaryEmbedding via modular functionality

Note that I based the expected values from the integration tests on the main implementation, not the RotaryEmbedding refactored implementation. In short: the test passes indicate that the refactor preserves the performance.

This is the refactor that I promised earlier offline.

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?

@ArthurZucker

cc @warner-benjamin

  • Tom Aarsen

@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
Contributor

@warner-benjamin warner-benjamin left a comment

Choose a reason for hiding this comment

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

LGTM

Copy link
Collaborator

@ArthurZucker ArthurZucker left a comment

Choose a reason for hiding this comment

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

Super nice to add the tests

class ModernBertRotaryEmbedding(GemmaRotaryEmbedding):
def __init__(self, config: ModernBertConfig, dim: int, base: float, device: Optional[torch.device] = None):
super().__init__(self, config=config, device=device)
self.rope_kwargs = {"dim": dim, "base": base}
Copy link
Collaborator

Choose a reason for hiding this comment

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

I think a PR #35589 might have removed the kwargs! but happy to add them back if need or just not use the gemma one!

Copy link
Member Author

@tomaarsen tomaarsen Jan 9, 2025

Choose a reason for hiding this comment

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

I'll reintroduce the rope_kwargs in LlamaRotaryEmbedding.

Copy link
Member Author

Choose a reason for hiding this comment

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

Done

Copy link
Collaborator

@ArthurZucker ArthurZucker left a comment

Choose a reason for hiding this comment

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

Actually it makes more sense to juste inherit and overwrite the call to inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, **self.rope_kwargs) rather than reverting the commit!

Copy link
Collaborator

@ArthurZucker ArthurZucker left a comment

Choose a reason for hiding this comment

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

Good to go! Thanks

@tomaarsen tomaarsen merged commit 6b73ee8 into huggingface:main Jan 10, 2025
16 checks passed
ArthurZucker pushed a commit that referenced this pull request Jan 10, 2025
#35459)

* Introduce 5 integration tests for the 4 model classes + torch export

* ModernBert: reuse GemmaRotaryEmbedding via modular

* Revert #35589, keep rope_kwargs; rely on them in modular_modernbert

* Revert "Revert #35589, keep rope_kwargs; rely on them in modular_modernbert"

This reverts commit 11b44b9.

* Don't set rope_kwargs; override 'self.rope_init_fn' call instead
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.

4 participants