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

[diffusion] Accelerate SANA-Video linear attention in quality=high (#35728)

Browse Source
This commit is contained in:
Xiaoyu Zhang
2026-08-21 18:05:43 +08:00
committed by GitHub
parent a5c52a9358
commit 39d4d65a51
9 changed files with 163 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/docs/sglang-diffusion/fused_kernels.mdx
+2
View File
@@ -57,6 +57,7 @@ These fusion families mount under `quality="high"`:
| LTX-2 RMSNorm + modulate | `rms_norm(x) * (1 + scale) + shift` in one launch |
| Gate RMSNorm (BF16-native) | `RMSNorm + tanh + mul + add` in one pass |
| HunyuanVideo strided QK RMSNorm | Per-head QK RMSNorm over the packed QKV layout |
| SANA-Video BF16-input linear attention | Keeps the first linear-attention GEMM's inputs in BF16 with FP32 accumulation/output; the second GEMM remains FP32 |
## Kernel inventory
@@ -146,6 +147,7 @@ Kernels are written against a specific eager chain in a specific model, so cover
| LTX-2.5 decoder | paired 3D RoPE with shared axis-table cache |
| HunyuanVideo | QKV+RoPE pack, strided QK RMSNorm, linear+GELU |
| Sana | LN+modulate, GLUMB bias+SiLU / bias+GLU, residual-gate add |
| SANA-Video | Packed QKV/KV; BF16-input linear attention at `quality=high` |
| Sana-WM | bidirectional gated delta-net, fused QK inverse-RMS |
| Wan | temb table slices; VAE cat+pad and DupUp3D add, `channels_last_3d` RMSNorm+SiLU |
| Cosmos3 / Krea2 / MiniMax-H3 | QK-norm + RoPE (Krea2 also CuTe-DSL norm+scale/shift; MiniMax-H3 also indexed modulation) |