Installing xFormers and Triton

[Rogala]

I don't know if Triton gives any advantages when it is installed, but when I check the value True when xFormers is installed, it warms my heart. xFormers speeds up generation, not significantly on my PC, but you may have different results.

The whole setup comes down to the fact that you need to update PyTorch to 2.7.0.

If you have the portable version before, you need to open the console (cmd) in the python_embedded folder and update torch and install xformers.

First, enter the nvidia-smi command to determine which version of cuda is installed on your PC (we are talking about Windows, under Linux I think there will be no problems). We need according to the description 12.6 or 12.8

If you have a cuda version not >=12.8, just change the number at the end of the cu128 link to cu126

Console commands in order:

1. Update the pip
2. See what version of torch is on our system
3. Update it
4. Install triton-windows
5. Install xformers

.\python.exe -m pip install --upgrade pip
.\python.exe -m pip list
.\python.exe -m pip install --upgrade torch==2.7.0 torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu128
.\python.exe -m pip install -U triton-windows
.\python.exe -m pip install -U xformers==0.0.30 --index-url https://download.pytorch.org/whl/cu128

This is the check line of the installation scripts, it will show you all the options that are available when working with xFormers on your system
.\python.exe -m xformers.info

If you have downloaded a version of Fooocus and are working with a virtual environment (.venv) to install Phyton libraries, the commands and actions look like this, open the console (cmd) in the folder where the virtual environment is located.

.\venv\Scripts\activate.bat
pip install --upgrade pip
pip list
pip install --upgrade torch==2.7.0 torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu128
pip install -U triton-windows
pip install -U xformers==0.0.30 --index-url https://download.pytorch.org/whl/cu128
python -m xformers.info

I think this will help in general systems where 60 steps take more than 1 minute, for others it will speed up no more than 1-2 seconds