Skip to content

Conversation

@Cyrilvallez
Copy link
Member

#34858 modified Llama, but Aria relying on Llama got merged in the meantime. This correctly reapplies modular to Aria (order of RotaryEmbedding got changed because now the Attention does not need it anymore, thus it got closer to the class actually needing it)

@ydshieh
Copy link
Collaborator

ydshieh commented Dec 11, 2024

We probably also need to check how check_modular_conversion works to avoid it passes on PR but fail on main. But that could be in another PR and by me or anyone wants to work on it.

@ydshieh
Copy link
Collaborator

ydshieh commented Dec 11, 2024

BTW, it's always a good idea to trigger a slow CI :-)

@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.

@Cyrilvallez
Copy link
Member Author

Cyrilvallez commented Dec 11, 2024

Humm, it seems very tricky to always avoid race conditions, it would probably require some check at merge time 🤔 I thought about this becoming an issue as we scale modular, but this is already an issue with the "copied from" (except if there is additional CI magic for copied from?). Here's the scenario:

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.

Thanks

@ydshieh
Copy link
Collaborator

ydshieh commented Dec 11, 2024

@Cyrilvallez Thanks for explaining. Yes, this happens from time to time. There is Update branch button, but click it all the time is also frustrating. We will try to think of solutions.

@Cyrilvallez Cyrilvallez merged commit 33c12e4 into main Dec 11, 2024
12 checks passed
@Cyrilvallez Cyrilvallez deleted the fix-regression branch December 11, 2024 13:24
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