Views
No views yet
.whl for woct0rdho/sageattention-autotune, targeted specifically at NVIDIA's Blackwell architecture (sm_120).abi3 compatible)1# ComfyUI Portable example
2.\python_embeded\python.exe -m pip uninstall sageattention
3
4# Standard venv example
5pip uninstall sageattention
61# ComfyUI Portable example
2.\python_embeded\python.exe -m pip install sageattention-2.2.0-cp310-abi3-win_amd64.whl
3
4# Standard venv example
5pip install sageattention-2.2.0-cp310-abi3-win_amd64.whl
6| Resolution | SageAttention 2 (s/it) | PyTorch Attention (s/it) |
|---|---|---|
| 0.5MP | 1.15 | 1.09 |
| 1.0MP | 2.10 | 2.00 |
| 1.5MP | 2.85 | 3.23 |
| 2.0MP | 3.80 | 4.81 |
| 2.5MP | 3.98 | 6.55 |
| 3.0MP | 4.51 | 8.49 |
| 3.5MP | 5.55 | 11.00 |
| 4.0MP | 6.55 | 14.08 |
sageattn_qk_int8_pv_fp16_cuda or auto.--use-sage-attention launch argument, as it may force the unstable Triton backend on Blackwell.sageattention-autotune repo with TORCH_CUDA_ARCH_LIST=12.0 to natively target sm_120. The -DTORCH_STABLE_ONLY flag was removed to ensure compatibility with PyTorch nightly builds.