Fix: use cu_chunk instead of cu_seqlens for flash_attn during corner case where top-k equals 1 #26
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Outdated suggestions cannot be applied.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.
This PR fixes an inconsistency in MoBA. When top-k > 1, the self attn is calculated using
cu_seqlens_q=self_attn_cu_seqlen
, whereself_attn_cu_seqlen = cu_chunk
. Meanwhile when top-k = 1, the self attn is calculated usingcu_seqlens_q=cu_seqlens
, which behaves the same as standard attention. Because the current block is always chosen, the corner case should usecu_chunk
instead ofcu_seqlens
.