How to install xFormers ? xformers==0.0.30 requires torch==2.7.0pip

[Shippdenn]

Hi There,
So I'm using Stability Matrix to run ForgeUI, and I wanted to install xFormers to boost performance and use ControlNet. I tried installing via pip, but hit this error:

xformers==0.0.30 requires torch==2.7.0pip
And updating Torch breaks compatibility with the other

= -----> Loop

Python 3.10
Torch 2.3.1+cu121

the Cuda I manually installed :1.24

Then I tried the "[Forge with CUDA 12.4 + Pytorch 2.4]<- Fastest, but MSVC may be broken, xformers may not work" package, but it seems it doesn't have xFormers either.

Does anyone have any workaround to get this working ?

Thanks in advance!

Edit : Overall, is all of this worth it? If I manage to get it working, will it be stable, or am I likely to run into crashes and issues? Would it be better to just stick with [CUDA 12.1 + PyTorch 2.3.1] instead? Also, regarding Stability Matrix — is there a way to import my manual installation into it?

[lvlontheinside]

Use the method for webui.

Edit forge's webui-user.bat to include 2 new lines BEFORE "set COMMANDLINE_ARGS" like below. Change the xformer and torch information to suit your needs, I left mine in for reference. DO NOT delete them afterwards or forge will/may downgrade your torch/xformers.

set XFORMERS_PACKAGE=xformers==0.0.31.dev1032
set TORCH_COMMAND=pip install torch==2.7.0 torchvision --extra-index-url https://download.pytorch.org/whl/cu128

Now add the ONE TIME reinstall augments for torch (--reinstall-torch) and xformers (--reinstall-xformers). Remember to delete them after the upgrade but not the two things mentioned above. When doing this myself, I reinstall torch alone then closed webui afterwards to then do xformers by itself. I doubt there is an issue doing them together but just an FYI.

set COMMANDLINE_ARGS=<whatever crap you use here, like --autolaunch --update-all-extensions --cuda-malloc --force-enable-xformers --xformers> --reinstall-torch --reinstall-xformers

After you deleted those lines, it should look something along these lines, depending on a webui install, but you get the idea...

set VENV_DIR=%A1111_HOME%/venv
set XFORMERS_PACKAGE=xformers==0.0.31.dev1032
set TORCH_COMMAND=pip install torch==2.7.0 torchvision --extra-index-url https://download.pytorch.org/whl/cu128
set COMMANDLINE_ARGS=--autolaunch --update-all-extensions --force-enable-xformers --xformers --cuda-malloc --no-half-vae

Note: I've done this all on webui to make it work with the 50xx series GPU. Forge will run fine using that venv but I did not use forge to reinstall torch/xformers. But it will try to reinstall torch 2.4 if I leave out torch_command. So if the reinstall args do not work, do it manually while using the two set lines to ensure it doesn't downgrade.

[davehornik]

Will try this asap I got the time to mess again with this, was trying like 10 times with different approaches - ultimately gave up. And will report back with outcome, using RTX5060Ti from MSI on one version behind (latest nvidia studio driver is unstable due to DX12 errors in games)

[Shippdenn]

Use the method for webui.

Edit forge's webui-user.bat to include 2 new lines BEFORE "set COMMANDLINE_ARGS" like below. Change the xformer and torch information to suit your needs, I left mine in for reference. DO NOT delete them afterwards or forge will/may downgrade your torch/xformers.

set XFORMERS_PACKAGE=xformers==0.0.31.dev1032 set TORCH_COMMAND=pip install torch==2.7.0 torchvision --extra-index-url https://download.pytorch.org/whl/cu128

Now add the ONE TIME reinstall augments for torch (--reinstall-torch) and xformers (--reinstall-xformers). Remember to delete them after the upgrade but not the two things mentioned above. When doing this myself, I reinstall torch alone then closed webui afterwards to then do xformers by itself. I doubt there is an issue doing them together but just an FYI.

set COMMANDLINE_ARGS=<whatever crap you use here, like --autolaunch --update-all-extensions --cuda-malloc --force-enable-xformers --xformers> --reinstall-torch --reinstall-xformers

After you deleted those lines, it should look something along these lines, depending on a webui install, but you get the idea...

set VENV_DIR=%A1111_HOME%/venv set XFORMERS_PACKAGE=xformers==0.0.31.dev1032 set TORCH_COMMAND=pip install torch==2.7.0 torchvision --extra-index-url https://download.pytorch.org/whl/cu128 set COMMANDLINE_ARGS=--autolaunch --update-all-extensions --force-enable-xformers --xformers --cuda-malloc --no-half-vae

Note: I've done this all on webui to make it work with the 50xx series GPU. Forge will run fine using that venv but I did not use forge to reinstall torch/xformers. But it will try to reinstall torch 2.4 if I leave out torch_command. So if the reinstall args do not work, do it manually while using the two set lines to ensure it doesn't downgrade.

Thanks for taking the time to write this, I appreciate it.

Adding to your solution: Fortunately, I was able to get it installed after maybe two days. The key was to install xformers alongside torch so that pip itself could automatically search for compatible versions:

pip install torch==2.4.1+cu124 torchaudio==2.4.1+cu124 torchvision==0.19.1+cu124 --index-url https://download.pytorch.org/whl/cu124 xformers

I found this tip in a Reddit post, and it worked.

However, when trying to launch, you might encounter an error where some packages like gradio, fastapi, starlette, uvicorn, etc., are missing. So install them. And then, you'll likely run into another error stating that it couldn't find the requested version, something like this:

The conflict is caused by:
The user requested gradio-client==0.10.1
gradio 4.40.0 depends on gradio-client==1.2.0

To fix this you could try to:

1. loosen the range of package versions you've specified
2. remove package versions to allow pip to attempt to solve the dependency conflict

That’s when I found another solution Here

Just change pydantic to pydantic==2.8.2 in the requirements_versions file inside the package folder. (pydantic==2.0.0 doesn't work, just 2.8.2)

Done:

pytorch version: 2.4.1+cu124
WARNING:xformers:A matching Triton is not available, some optimizations will not be enabled
Traceback (most recent call last):
File "C:\Users\XXXX\AppData\Roaming\StabilityMatrix\Packages\stable-diffusion-webui-forge\venv\lib\site-packages\xformers_init_.py", line 57, in _is_triton_available
import triton # noqa
ModuleNotFoundError: No module named 'triton'
xformers version: 0.0.28.post1

As you can see, there's a warning about Triton, but It's not Important. because :
First: it’s a Linux-only thing. not available on Windows.
Second: based on what I’ve read, it doesn’t noticeably improve performance anyway. That said, feel free to look into it — there are some ways to install Triton on Windows.

(Optional) Since I installed Forge via Stability Matrix, I first uninstalled torch and related packages from the virtual environment. I'm not sure if that's necessary, so I’ll call it optional — but for a clean setup, you can go into the venv and uninstall torch, torchaudio, and xformers before doing the steps above.