Parameterized Stripe Attention for Efficient Video Generation
PSA, a parameterized stripe attention that formalizes the observed stripe regularity, unifying diverse attention patterns for efficient mask generation, enables a single hardware-efficient CUDA kernel to process all sparse patterns, achieving FlashAttention-3-level Model FLOPs Utilization.
Xing-Yu Jia, Bao-Le Ai, Ang Wang et al.
· 0 citations