Refactor bias initialization and enhance bias computation in FlashDMAttnFunc#191

Merged

LoserCheems merged 2 commits intomainfrom

Oct 12, 2025

Commits on Oct 12, 2025

Refactors mha_bwd to use torch::zeros for bias initialization and removes unnecessary zeroing of dbias_expanded
algo-home
committed
Enhances FlashDMAttnFunc to track original sequence length bias and adjusts dbias computation based on its value
algo-home
committed