Clear output of Torch SDPA for masked pieces #360
Merged
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
danieldk
force-pushed
the
bugfix/sdpa-attention-nan
branch
from
be5af05 to
814d04d
Compare
Since Torch 2.1, the Torch memory-efficient SDPA GPU kernel returns NaN for pieces that are completely masked out. This leads to NaN propagation in the next attention layer, because masked pieces get an attention of zero, but zero times NaN is still NaN. In this we fix this by setting masked tokens to zero to clear out any NaNs. We currently rely on the query dimension of the mask to be singular, but in the future we should probably redesign the `AttentionMask` class to account for the differences between attention masks and causal masks.
danieldk
force-pushed
the
bugfix/sdpa-attention-nan
branch
from
5031bc8 to
31631f8
Compare
danieldk
force-pushed
the
bugfix/sdpa-attention-nan
branch
from
121d69b to
fc783bb
Compare
1 task
shadeMe
reviewed
Feb 8, 2024
Co-authored-by: Madeesh Kannan <[email protected]>
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Outdated suggestions cannot be applied.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.
Description
Since Torch 2.1, the Torch memory-efficient SDPA GPU kernel returns NaN for pieces that are completely masked out. This leads to NaN propagation in the next attention layer, because masked pieces get an attention of zero, but zero times NaN is still NaN.
In this we fix this by setting masked tokens to zero to clear out any NaNs.
We currently rely on the query dimension of the mask to be singular, but in the future we should probably redesign the
AttentionMaskclass to account for the differences between attention masks and causal masks.Checklist