fix: handle 3D input tensors in flashAttention#45
Merged
MankyDanky merged 1 commit intostagingfrom Apr 17, 2026
Merged
Conversation
flashAttention was hardcoded to assume 4D [B, H, T, D] input, but models commonly pass 3D [B*H, T, D] after merging batch and head dims. With 3D input, qShape[3] was undefined, causing NaN to propagate through the entire forward pass. Now reads T and D from the last two dimensions regardless of rank.
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
flashAttention was hardcoded to assume 4D [B, H, T, D] input, but models commonly pass 3D [B*H, T, D] after merging batch and head dims. With 3D input, qShape[3] was undefined, causing NaN to propagate through the entire forward pass.
Now reads T and D from the last two dimensions regardless of rank.