CUDA: fix overflow in MMA kernel without stream-k #17939
Merged
+6
−6
Conversation
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
github-actions
bot
added
Nvidia GPU
ggerganov
approved these changes
Dec 12, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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Outdated suggestions cannot be applied.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.
Fixes #17931 .
On Ampere and older a stream-k decomposition is only used if the efficiency of tiling is bad. However, the code that is being run is still the exact same, there is only a change in the launch configuration so that each CUDA block works on a single tile. Since the number of tiles is proportional to the physical batch size, raising it can cause a numerical overflow in an intermediate result. This PR simply makes it so that 64 bit integers are used for that part rather than 32 bit integers.