[feat] VSA-256 fastpath on Blackwell via FA4 CuTe block-sparse attention (#1354)

2026-05-26 12:58:29 -07:00
parent 6a610e2bc9
commit be548a78ea
11 changed files with 1211 additions and 44 deletions
@@ -3,6 +3,7 @@ from .version import __version__
 from fastvideo_kernel.ops import (
    sliding_tile_attention,
    video_sparse_attn,
+    video_sparse_attn_bshd,
 )

 from fastvideo_kernel.block_sparse_attn import (
@@ -27,6 +28,7 @@ from fastvideo_kernel.turbodiffusion_ops import (
 __all__ = [
    "sliding_tile_attention",
    "video_sparse_attn",
+    "video_sparse_attn_bshd",
    "block_sparse_attn",
    "block_sparse_attn_from_indices",
    "moba_attn_varlen",