Repository navigation
Conversation
Signed-off-by: Micah Williamson <micah.williamson@amd.com>
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
With
--mm-encoder-tp-mode data, ranks that get no images all-gather placeholder embeddings whose dtype comes fromnext(vision_model.parameters()). Since #64111, MiniMax-M3 keeps its patch embedding in fp32 while the tower outputs bf16. Ranks with images then gather bf16 and ranks without gather fp32, so the all-gather mismatches and startup hangs.This uses the vision model's
dtypeattribute when it has one, and falls back to the first parameter's dtype otherwise. Other models are unchanged: theirdtypeeither matches the first parameter or isn't defined.Test plan
amd/MiniMax-M3-MXFP4with-tp 4 --mm-encoder-tp-mode datanow starts and serves image requests. Before the fix, it hung after the encoder profile run.--mm-encoder-tp-mode data.