Skip to content

[Mllama] Fix workaround compile#44850

Closed
vasqu wants to merge 2 commits intohuggingface:mainfrom
vasqu:fix-mllama
Closed

[Mllama] Fix workaround compile#44850
vasqu wants to merge 2 commits intohuggingface:mainfrom
vasqu:fix-mllama

Conversation

@vasqu
Copy link
Copy Markdown
Contributor

@vasqu vasqu commented Mar 19, 2026

See #44458

This is a deep issue tbh - the cross attentions are reshaped into a different shape than the text input leading to a mismatch between batch sizes. This only gets noticed during compile as it is more strict about the concrete shapes and indices. Tested locally that it works.

@github-actions
Copy link
Copy Markdown
Contributor

[For maintainers] Suggested jobs to run (before merge)

run-slow: mllama

@vasqu
Copy link
Copy Markdown
Contributor Author

vasqu commented Mar 19, 2026

Does not work, only for the initial steps...

@vasqu vasqu closed this Mar 19, 2026
@HuggingFaceDocBuilderDev
Copy link
Copy Markdown

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.

@vasqu vasqu reopened this Mar 19, 2026
@vasqu
Copy link
Copy Markdown
Contributor Author

vasqu commented Mar 19, 2026

@themavik please stop spamming reviews (of no real value), it's not helpful and disturbs workflows.

This is a warning but if I see it another time I will block you.

@vasqu
Copy link
Copy Markdown
Contributor Author

vasqu commented Mar 26, 2026

Closing, not planned on our side - seems like a genuine torch bug

@vasqu vasqu closed this Mar 26, 2026
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.

2 participants