I replace flash_attn_v2 with sparse_sageattention in wan2.1-t2v-1.3B. However, the time cost increases.

[a5704607]

[jt-zhang]

Hi, For the fastest state-of-the-art sparse attention API, use the APIs in SpargeAttn.

[a5704607]

已收到，请放心。

[jt-zhang]

Hi, there’s an update I’d like to share. The code in SpargeAttn has been updated, and I recommend using the spas_sage2_attn_meansim_topk_cuda API.