SageAttention Wheel Won't Install in ComfyUI? Fix It

Q: What’s the difference between cp39-abi3 and cp310-abi3?

Both use Python’s stable ABI, meaning the wheel works across multiple Python versions without needing an exact match. cp39-abi3 supports Python 3.9 and newer; cp310-abi3 supports Python 3.10 and newer. If your Python is 3.10 or newer, either tag works for you.

⚡ Quick Answer

A SageAttention wheel (a pre-built package file ending in .whl) fails to install in ComfyUI almost always because it doesn't match your CUDA version, your PyTorch version, or your Python version. The wheel's filename tells you exactly which versions it needs — check your own three versions first, then pick the wheel whose filename matches all three.

If you've already installed Triton and tried to install SageAttention, but pip either refuses the file outright or installs it and then ComfyUI can't find it, the problem is almost never SageAttention itself. It's a mismatch between the wheel file you downloaded and the CUDA, PyTorch, or Python setup already on your computer. This guide shows you how to check your own versions, how to read a SageAttention wheel filename so you know exactly what it needs, and how to fix the three most common pip errors that show up when the versions don't line up.

This guide assumes Triton is already installed and working. If you haven't installed Triton yet, or you're setting up SageAttention from scratch, follow the full SageAttention install guide first, then come back here if the wheel step is the part giving you trouble.

What You Need Before Matching a SageAttention Wheel

Triton already installed. Triton is a tool that lets your computer compile GPU code — SageAttention needs it to run. If import triton fails, or you haven't installed it yet, finish that step first.
A terminal (command prompt) open inside your ComfyUI_windows_portable folder. This is the folder that contains python_embeded and ComfyUI as subfolders — your ComfyUI portable installation root. Every command in this guide is typed here.

Warning: Every command below starts with .\python_embeded\python.exe. That's not decoration — it's the whole reason wheel installs go wrong for a lot of people. If you type just python or pipinstead, Windows may run a completely different, unrelated copy of Python that isn't the one ComfyUI actually uses, and the install will silently go to the wrong place.

How Do You Check Your CUDA, PyTorch, and Python Versions in ComfyUI?

A SageAttention wheel is built for one specific combination of Python, CUDA, and PyTorch. Before you can pick the right one, you need to know your own three numbers. Run these three commands in order and write down what each one tells you.

Check Your Embedded Python Version

.\python_embeded\python.exe --version

You should see something like Python 3.12.7 or Python 3.13.11 printed on screen. Write this number down — it decides which "ABI tag" your wheel needs (explained in the next section).

Terminal output of the embedded Python version command in ComfyUI portable — Terminal output of the Python version check.

Check Your CUDA Version

CUDA is the software NVIDIA GPUs use to run heavy calculations like AI image generation. Your GPU driver has one CUDA version installed, and PyTorch was built against a specific CUDA version too — the wheel needs to match the one PyTorch is using, not necessarily your driver's newest supported version.

.\python_embeded\python.exe -c "import torch; print(torch.version.cuda)"

This prints a number like 12.8 or 13.0. If it prints None instead, PyTorch isn't installed yet, or it's installed as a CPU-only build — install PyTorch with CUDA support before continuing (the main install guide's PyTorch step covers this).

Terminal output showing the CUDA version reported by PyTorch in ComfyUI — Terminal output of the CUDA version check.

Check Your PyTorch Version

.\python_embeded\python.exe -m pip show torch

Look for the line starting with Version: — for example Version: 2.10.0.dev20260815+cu130. That full string, including everything after the +, matters. If this command prints nothing useful, run this instead:

.\python_embeded\python.exe -m pip freeze

and look for a line starting with torch==.

Terminal output of pip show torch displaying the installed PyTorch version — Terminal output of the PyTorch version check.

How Do You Read a SageAttention Wheel Filename?

Once you have your three version numbers, you need to be able to read a SageAttention wheel filename and know whether it matches you. Here's a real example, broken into its parts:

sageattention-2.2.0+cu130torch2.10.0andhigher.post6-cp310-abi3-win_amd64.whl

sageattention-2.2.0 — the SageAttention version itself.
cu130 — the CUDA version this wheel was built for. This must match the CUDA number you checked above (12.8 → look for cu128, 13.0 → look for cu130).
torch2.10.0andhigher — the minimum PyTorch version this wheel works with. "andhigher" means it supports that version and later ones, but only within the same CUDA tag — a cu130 wheel still needs a cu130 build of PyTorch, even if the PyTorch version number itself is higher.
post6 — a release/build number. Higher usually means newer and more bug fixes, not a different requirement.
cp310-abi3 — the Python compatibility tag. cp310-abi3 means it works with Python 3.10 and any newer version, because it uses the "stable ABI" (application binary interface — the part of Python that doesn't change between minor versions). An older wheel might instead say cp39-abi3, meaning Python 3.9 and newer.
win_amd64 — confirms it's built for 64-bit Windows. If you're not on 64-bit Windows, this wheel isn't for you at all.

Tip: If your Python version is 3.10 or newer, a cp39-abi3 or cp310-abi3wheel will both work — ABI3 wheels are backward-compatible with any Python version at or above the number in the tag. You don't need an exact Python match, just one that's the same or newer.

Which SageAttention Wheel Do You Actually Need?

Match your CUDA version and PyTorch version to the wheel's cuXXX and torchX.X.Xandhigher tags. The two rows below are real filename patterns from recent SageAttention releases, shown so you can see the pattern — always confirm against the current release list before downloading, since new wheels are published often and older ones get replaced.

If you have...	Look for a filename containing...
CUDA 12.8, PyTorch 2.9.0 or newer	`cu128torch2.9.0andhigher`
CUDA 13.0, PyTorch 2.10.0 or newer	`cu130torch2.10.0andhigher`

Go to the SageAttention releases page on GitHub: https://github.com/woct0rdho/SageAttention/releases. Scroll to the latest release and find the .whl file whose CUDA tag and PyTorch tag match your two numbers from earlier, and whose ABI tag (cp39-abi3 or cp310-abi3) is at or below your Python version.

SageAttention GitHub releases page showing a list of downloadable .whl files — The current wheel list on the SageAttention GitHub releases page.

Warning: SageAttention's newer "SageAttention2++" kernels — the ones that give the biggest speed boost — only work on RTX 40xx and RTX 50xx GPUs with CUDA 12.8 or higher and PyTorch 2.7 or higher. Any wheel will still install and run on older GPUs, but without that extra speedup.

Tip: As of ComfyUI version 0.32.0 (August 2026), ComfyUI has its own built-in fast attention option called comfy kitchen, turned on with a different startup flag (--use-ck-attention) and needing no wheel install at all. If you're on a recent ComfyUI version, it's worth checking whether you need SageAttention specifically or whether comfy kitchen already covers what you need — that comparison is a separate topic, but it's worth knowing the option exists before you spend time matching wheel versions.

How to Install the Matching Wheel

Download the .whl file you matched above and move it into your ComfyUI_windows_portable folder — the same folder where your terminal is open.
Install it with pip, using the real filename (not a placeholder) — see the example below.
Check for the success message. You should see a line near the bottom that says something like Successfully installed sageattention-2.2.0+.... If you see that, the wheel installed correctly.

.\python_embeded\python.exe -m pip install sageattention-2.2.0+cu130torch2.10.0andhigher.post6-cp310-abi3-win_amd64.whl

Replace the filename above with whatever you actually downloaded.

Tip: Type .\python_embeded\python.exe -m pip install sageattention- and then press the Tabkey — Windows will auto-complete the rest of the filename for you, so you don't have to type the whole thing by hand.

Terminal showing the Successfully installed sageattention confirmation line — A successful install ends with this confirmation line.

Installing the wheel is not the same as turning SageAttention on. ComfyUI still needs to be told to use it at startup with the --use-sage-attention flag — that step, plus how to confirm it's actually running, is covered in the full SageAttention install guide.

Why Is Pip Rejecting Your SageAttention Wheel?

Running into something not covered below? Our general ComfyUI troubleshooting guide covers errors that show up across every workflow, not just this one.

"No matching distribution found for sageattention==X.X.X"

What causes it: This shows up when you try to install SageAttention by name — for example pip install sageattention==2.2.0 — instead of installing a downloaded .whl file directly. SageAttention's Windows wheels aren't published to PyPI (the normal online package index pip searches by default), so pip has nowhere to find them by name.

How to fix it: Download the exact .whl file from the GitHub releases page yourself, and install it by its local filename, as shown in the install steps above — not by typing a package name and version.

"... is not a supported wheel on this platform"

What causes it: The wheel's ABI tag (cp39-abi3, cp310-abi3) or the win_amd64 part doesn't match your embedded Python version or your operating system.

Re-run the Python version check from earlier in this guide.
Go back to the wheel filename breakdown and confirm the cpXX-abi3 number is equal to or lower than your Python version.
Download a different wheel from the releases page if the one you chose doesn't match, and reinstall.

"No module named 'sageattention'" (after pip said the install succeeded)

What causes it: The wheel installed into the wrong copy of Python on your computer — usually your system-wide Python instead of ComfyUI's own python_embeded copy. ComfyUI can only see packages installed into python_embeded.

Open a terminal inside ComfyUI_windows_portable again (not any other folder).
Reinstall using the full path, exactly as shown earlier: .\python_embeded\python.exe -m pip install <your-file>.whl.
Confirm it worked by running .\python_embeded\python.exe -c "import sageattention" — if nothing prints, it worked. An error means it's still not installed in the right place.

Frequently Asked Questions

It means the wheel supports that PyTorch version and any newer one, as long as the CUDA tag still matches. For example, torch2.10.0andhigher on a cu130 wheel works with PyTorch 2.10.0 or later, but only PyTorch builds that also use CUDA 13.0.

Yes, as long as the CUDA tag in the filename still matches your PyTorch’s CUDA build. The version number in "andhigher" is a minimum, not an exact match — but the CUDA tag (cu128, cu130) has to match exactly.

This almost always means the wheel was installed into the wrong Python. Reinstall using the full .\python_embeded\python.exe -m pip install command from inside the ComfyUI_windows_portable folder, not a shortened pip install command.

Yes, and the requirements are stricter in the other direction. RTX 20xx (Turing) support was dropped in newer Triton and SageAttention builds, so you may need an older PyTorch version (2.6 or earlier) and a matching older SageAttention wheel rather than the newest release.

Both use Python’s stable ABI, which means the wheel works across multiple Python versions without needing an exact match. cp39-abi3 supports Python 3.9 and newer; cp310-abi3 supports Python 3.10 and newer. If your Python is 3.10 or newer, either tag works for you.

Not necessarily. ComfyUI versions from August 2026 onward include a built-in fast attention option called comfy kitchen that needs no separate wheel install. If you’re only trying to speed up generation and you’re on a recent ComfyUI version, it may be simpler to try --use-ck-attention first before going through wheel-matching at all.

What to Do Next

Turn SageAttention on and confirm it's running.

A matched, installed wheel doesn't do anything until ComfyUI is told to use it. The full install guide covers the launch flag and how to check the console for confirmation.

Full SageAttention Install Guide →General Troubleshooting Guide →