How to Fix xformers and Torch Errors When Launching Stable Diffusion

A red wall of text mentioning xformers the moment you launch Stable Diffusion looks alarming, but the cause is almost always mundane: xformers has to match your installed PyTorch and CUDA versions exactly, and when it does not, the import fails before the interface even loads. Nearly every launch-time ImportError involving xformers is a version mismatch, and the fix is to put the matching pieces back together rather than to hunt for an exotic bug.

Quick Answer

xformers and torch launch errors come from a version mismatch. xformers is compiled for one specific PyTorch and CUDA combination, and if your installed torch differs, the import breaks. Fix it by reinstalling a wheel that matches your torch and CUDA, or by letting the WebUI rebuild xformers against your current PyTorch. A typical message reads "xFormers was built for PyTorch X but you have Y."

Why the versions have to line up

xformers is a low-level acceleration library that hooks deep into PyTorch's internals, so it is built against an exact PyTorch and CUDA version. When you install or update one without the other, the build xformers expects and the torch you actually have no longer agree, and Python refuses to load the mismatched module. The error usually spells this out plainly, naming the PyTorch version xformers was built for and the one you have.

This is why the trouble so often appears right after an update. A WebUI that auto-updates its libraries can pull a newer torch while leaving an older xformers in place, breaking the pairing. The cure is to realign them, not to remove xformers and lose the speed it provides.

Fix one: install a matching xformers wheel

The direct fix is to install an xformers build made for your exact torch and CUDA. First check what you have. In your environment, list the installed versions to see your torch and CUDA build, then install the xformers wheel that names the same PyTorch and CUDA combination.

The reliable way is to install torch and xformers together from the same source so they cannot drift apart, specifying the CUDA build that matches your card. Installing them in one command, pointed at the correct CUDA index, guarantees the pairing rather than leaving it to chance. After installing, restart the WebUI and the launch error should be gone.

Fix two: let the WebUI rebuild xformers

If pinning a wheel feels fiddly, many WebUIs can build xformers themselves against whatever PyTorch you currently have, which sidesteps the matching problem entirely. Launching with the appropriate xformers flag tells the WebUI to fetch or compile a compatible build for your installed torch.

The trade-off is that a from-source build takes longer the first time and needs the right build tools present, but once it succeeds the versions are guaranteed to agree because it built against your actual torch. This is the easier route for anyone who does not want to track wheel filenames, and it is often what the WebUI does automatically when xformers is missing.

Stop it breaking again

Once it works, the most common way to re-break it is an auto-update silently bumping torch. Pinning the xformers package version so the WebUI stops trying to update it on every launch keeps the pairing stable. Setting the xformers package as a fixed version in your launch configuration freezes it, so an update no longer pulls a mismatched torch out from under it. Older GPUs and constrained setups feel these dependency tangles more sharply, and a card with current driver and CUDA support reduces the friction, which you can weigh across the best-selling GPUs or the dedicated AI PC range.

Frequently Asked Questions

How do I know the error is a version mismatch?

The message itself usually names the PyTorch version xformers was built for and the one you have installed. Any ImportError on launch that mentions xformers and a PyTorch version is the classic mismatch signature.

Should I install a wheel or let the WebUI rebuild xformers?

Either works. Installing a matching wheel is fast if you know your exact torch and CUDA. Letting the WebUI build xformers against your current torch is easier because it guarantees a match, at the cost of a slower first build.

Why did this start after I updated?

Updates can bump torch while leaving an older xformers in place, breaking the pairing they need. That is why the error so often appears right after a WebUI auto-update or a manual torch upgrade.

Can I just remove xformers instead of fixing it?

You can disable it and the WebUI will still run, but you lose the memory and speed benefits it provides. Realigning the versions is usually a few minutes of work and worth keeping rather than dropping the acceleration.

How do I stop the error from returning?

Pin the xformers package version in your launch configuration so the WebUI stops auto-updating it. With xformers frozen, a future update cannot quietly pull a mismatched torch and break the pairing again.