flashattention-

Definition

The versioned form of FlashAttention — FlashAttention-2, FlashAttention-3 — where the number marks a successive implementation. The original works by tiling the attention computation and never materialising the full score matrix in high-bandwidth memory, cutting both time and memory; later versions improve work partitioning across warps and exploit newer GPU features. Crucially it is exact, computing the same attention as the naive version, unlike sparse or low-rank approximations that trade accuracy for speed.