Use torch.repeat instead of expand on key & value in Triton MQA to prevent NaNs with certain h_dims #442
Conversation
This file contains 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Outdated suggestions cannot be applied.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.
For h_dim=8, we see NaNs due to a
.expandof the key value tensors, which can be resolved with.repeat. While h_dim=8 is an edge case, we are not sure if there are other cases of h_dims for which this might be problematic, or if there might be silent failures.Note that this does come at a performance hit, see MFU for 7b, so it may be desirable to revert this change in the future.
(blue curve:
.expand()green curve.repeat()red curve.expand().clone())