license: apache-2.0 tags:
- sageattention
- blackwell
- sm_120
- rtx-5070-ti
- windows
- comfyui
SageAttention (Autotune) - Windows Wheel for Blackwell (sm_120)
Precompiled Windows .whl for woct0rdho/sageattention-autotune, targeted specifically at NVIDIA's Blackwell architecture (sm_120).
Supported GPUs
- RTX 5070
- RTX 5070 Ti
- RTX 5080
- RTX 5090
- RTX 5070
Requirements
- OS: Windows 10/11
- GPU Architecture: sm_120 (Blackwell)
- CUDA: 13.0+
- PyTorch: 2.14.0.dev+cu130 (Nightly required for sm_120 support)
- Python: 3.10+ (Wheel is
abi3compatible)
Installation
Before installation, remove the previous version of SageAttention, if it exists:
# ComfyUI Portable example
.\python_embeded\python.exe -m pip uninstall sageattention
# Standard venv example
pip uninstall sageattention
Download the wheel and install it using your ComfyUI portable Python or venv:
# ComfyUI Portable example
.\python_embeded\python.exe -m pip install sageattention-2.2.0-cp310-abi3-win_amd64.whl
# Standard venv example
pip install sageattention-2.2.0-cp310-abi3-win_amd64.whl
Benchmarks (Ideogram 4, RTX 5070 Ti)
| 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 |
While standard PyTorch Attention is slightly faster at very low resolutions, SageAttention scales drastically better, cutting 4MP generation time by over 50%.
Usage in ComfyUI
- Install ComfyUI-KJNodes.
- Add the "Patch Sage Attention" node to your workflow.
- Set the backend to
sageattn_qk_int8_pv_fp16_cudaorauto.
Do NOT use the --use-sage-attention launch argument, as it may force the unstable Triton backend on Blackwell.
Build Details
Compiled from the 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.