minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[diffusion] feat: support cube sparse attention for minimax h3 (#34893)

Browse Source 
Co-authored-by: zhenaozhenfu <zhenaozhenfu@minimaxi.com>
Co-authored-by: Reynor <reynor@minimaxi.com>
This commit is contained in:
Mick
2026-09-02 15:21:33 +08:00
committed by GitHub
co-authored by zhenaozhenfu Reynor
parent 9175590aa0
commit 4b329482e8
18 changed files with 2586 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/docs/sglang-diffusion/compatibility_matrix.mdx
+6
View File
@@ -910,3 +910,9 @@ The entries below simply reflect configurations that have been manually validate
### Sliding Tile Attention
- Currently, only Hopper GPUs (H100s) are supported.
### Cube Sparse Attention
- Available only for MiniMax-H3 (`--component-attention-backends transformer=cube_sparse_attn`). It sparsifies only the packed sequence's 3D visual streams; text, audio, standalone reference images, and the text-only token refiner remain dense.
- Requires `--attention-backend-config` with both `local_cube_size` and `topk_ratio_list`. `topk_ratio_list` must have one entry per denoise step, each in `(0, 1]`. See the [MiniMax-H3 cookbook](/cookbook/diffusion/MiniMax/MiniMax-H3) for a worked example.
- Runs on pure PyTorch plus FlexAttention, so it has no third-party kernel dependency.