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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
In our vector and NAX attention we keep the scale factor in fp32. But in the non-NAX fused attention it gets downcast to bf16 which is made worse by the fact that it is multiplied by another scale as well.
It seems to have a real impact on model quality in some cases: ml-explore/mlx-lm#868 (comment)
In terms of performance I think the regression is acceptable as it's less than 1% (at most 0.5%) for all the cases I tried
And in terms of accuracy the difference between the fused attention in bf16 and fp32 with a scaling factor is noticeably lower:
The maximum absolute difference goes down by a factor of 6. For some random inputs with a typical scaling factor based on the head dimension:
Pre: 0.00585938
Post: 0.000976562