SageAttention Missing

What you see in the logs

When ComfyUI starts, the following message appears in the terminal:

Warning: Could not load sageattention: No module named 'sageattention'

What this is

SageAttention is an attention quantization acceleration library developed at Tsinghua University (ICLR 2025). It applies INT8/FP8 quantization to attention computation, achieving 2-5x faster inference than FlashAttention with no loss in generation quality.

Some ComfyUI custom nodes (such as TeaCache, CogVideoX Wrapper, etc.) try to load it. When it is missing, ComfyUI falls back to its default attention implementation and everything works normally.

Severity

Low -- this is a warning, not an error. ComfyUI runs fine without it.

Is it worth installing

Bottom line: Installing SageAttention on Windows is non-trivial, and the benefit is limited for most users.

The actual speed-up depends on your GPU and workflow:

• RTX 4000/5000 series: SageAttention 2's CUDA kernels are specifically optimized for Ada/Blackwell architectures (sm89/sm120) and can deliver a 30-40% speed-up -- the biggest gain.
• RTX 3000 series: Falls back to a Triton kernel path (equivalent to SageAttention 1), yielding roughly 10-15% acceleration.
• RTX 2000 series (Turing): Also uses the Triton fallback path with a similar but smaller improvement. Older versions had compatibility issues; use a recent wheel.
• GTX 16 series and older: Not recommended -- minimal benefit and poor compatibility.

If you mainly run Stable Diffusion image generation and do not use heavy workflows like video generation, ignoring this warning is the safest choice.

How to fix it

Option 1: Ignore it (recommended)

No action needed. This warning does not affect ComfyUI's functionality or output quality.

Option 2: Install SageAttention

On Windows, pip install sageattention tries to compile CUDA kernels from source and usually fails (it requires the Visual Studio C++ build tools). Use a community-provided prebuilt wheel instead.

Prerequisites:

• NVIDIA GPU (GTX 16 series or newer)
• PyTorch >= 2.6 with CUDA (not the CPU-only build)
• Python >= 3.9
• triton-windows installed (SageAttention depends on Triton)

Step 1: Check your environment

Run the following in the Wonderful Launcher terminal:

python -c "import torch; print(f'PyTorch {torch.__version__}, CUDA {torch.version.cuda}')"

Note the PyTorch and CUDA version numbers -- you will need them to pick the right wheel.

Step 2: Install Triton

SageAttention depends on Triton. If you have not installed it yet:

pip install triton-windows

Step 3: Install the prebuilt wheel

Go to the woct0rdho/SageAttention Releases page, download the .whl file that matches your PyTorch and CUDA versions, then install it:

pip install sageattention-2.2.0+cu128torch2.8.0.post2-cp39-abi3-win_amd64.whl

The filename above is just an example. Pick the wheel that matches your setup:

• cu128 / cu126 etc. corresponds to your CUDA version
• torch2.8.0 / torch2.6.0 etc. corresponds to your PyTorch version (minor version differences are usually fine)
• cp39-abi3 means compatible with Python 3.9 and above

Step 4: Verify the installation

python -c "from sageattention import sageattn; print('SageAttention OK')"

Step 5: Restart ComfyUI

After restarting, the warning should be gone.

Common installation failures

Error message	Cause and fix
No module named 'triton'	Triton is not installed. Run pip install triton-windows
error: Microsoft Visual C++ 14.0 or greater is required	You ran pip install sageattention which tries to compile from source. Use a prebuilt wheel instead
RuntimeError: CUDA 12.8 or higher is required	The wheel's CUDA version is higher than your environment. Pick a matching wheel or upgrade PyTorch
AttributeError involving triton.jit	Corrupted Triton cache. Delete the C:\Users\<your-username>\.triton\ folder and try again
Install succeeds but ComfyUI still reports No module named 'sageattention'	Installed into the wrong Python environment. Make sure you run pip from ComfyUI's Python

Option 3: Use a ComfyUI custom node for automatic installation

Community nodes like ComfyUI_EasySageAttention can auto-detect your environment and install the matching wheel. If you prefer not to do things manually, give these a try. Note that these are third-party projects and are outside our support scope.

Contact us

If you encounter errors not listed in the table above during installation, or if ComfyUI behaves abnormally after installing SageAttention, reach out to support via the "Contact Us" button in the app.