Automated installers for SageAttention on Windows for ComfyUI portable.
SageAttention accelerates AI video generation significantly. If you're using models like WAN in ComfyUI, SageAttention can make your renders up to twice as fast — but installing it on Windows has always been a pain. This installer handles everything automatically so you can get the speed boost without the headache.
Automatically detects your Python, PyTorch, and CUDA versions, selects the correct wheel, installs triton-windows, and creates a ready-to-use launcher.
| File | Description |
|---|---|
install-sage-only.bat |
Installs SageAttention + Triton and creates a launcher |
install-comfyui-setup.bat |
Same as above, plus optionally installs ComfyUI-Manager and network access (--listen) |
Install these before running the installer:
- Visual C++ Redistributable
- Git (only needed for ComfyUI-Manager in
install-comfyui-setup.bat)
- Windows 10/11
- ComfyUI portable (with embedded Python)
- NVIDIA GPU with CUDA 12.x or 13.x
- PyTorch 2.4+
| GPU | Minimum PyTorch | Minimum CUDA |
|---|---|---|
| RTX 50xx (Blackwell) | 2.7 | 12.8 |
| RTX 40xx (Ada) | 2.4 | 12.4 |
| RTX 30xx (Ampere) | 2.4 | 12.4 |
GTX 16xx / RTX 20xx (Turing) support was dropped in triton-windows 3.3+ (PyTorch 2.7+). Use PyTorch 2.6 or older.
- Download the latest version or clone this repository
- Double-click
install-sage-only.batorinstall-comfyui-setup.bat - Enter the path to your ComfyUI portable folder (e.g.
D:\ComfyUI_windows_portable) - The installer auto-detects your environment and installs everything
- Use the generated
run_nvidia_gpu_sage.batin your ComfyUI folder to launch with SageAttention enabled
- c10.dll error — Install the Visual C++ Redistributable
- ComfyUI won't start — Update your NVIDIA drivers to the latest version
- Wheel auto-detection fails — The installer falls back to manual wheel selection
- All-black / dark video on RTX 50xx — SageAttention's Blackwell (sm_120) fp8 kernel overflows fp16 on models like WAN (fp8 weights run fp16 math), producing NaNs → black frames. This is most commonly seen when working with 2:1 resolutions (e.g. 1280x640). The installer offers an optional fp16 black-frame fix (drops a small node into
custom_nodes\sage_fp16_fix) that routes fp16 attention through the stable bf16 Triton backend. It activates automatically; setSAGE_FP16_FIX=0to disable, orSAGE_FP16_FIX_GUARD=1for an extra fp32 fallback safety net.
See the THIRD_PARTY folder for licenses of bundled components:
- SageAttention — Apache 2.0
- triton-windows — MIT
- CPython — PSF License v2 (include/libs)