xformers performance boost about 20% - installing for Windows (how I did it)

[cmp-nct]

Given that not one of the tutorials worked and I'm new to Python I wrote the steps together.
My 3090 improved from average 13 iterations to 16 with xformers enabled

1) I use Anaconda to launch webui despite it having an own environment, I'm not sure if it's necessary but I used it anyway.
2) https://visualstudio.microsoft.com/downloads/?q=build+tools#build-tools-for-visual-studio-2022 (Install build tools)
3) https://developer.nvidia.com/cuda-11.3.0-download-archive (install development, runtime and VS integration in custom install)
   # Use network install, if it complains about Visual Studio not being found, ignore it.
4) Anaconda -> cd X:\...webui\venv\Scripts -> activate # now we are in the right python environment 
5) cd webui\repositories
6) git clone https://github.com/facebookresearch/xformers.git; cd xformers; git submodule update --init --recursive
Optional) set TORCH_CUDA_ARCH_LIST=8.6 # (for RTX 3 series 8.6 is right), this tells xformers to code only for this arch. I believe default is to generate code for all available archs
7) pip install -r requirements.txt
8) python setup.py build; python setup.py develop; python setup.py install # takes 10-30 minutes and looks messy
9) webui.bat : set COMMANDLINE_ARGS=--xformers
From here on it should work, on 3090 I see about 20% performance improvements

[TheCompAce]

### Some notes:
Before Step 8
Make sure your CDA_HOME path variable is set to your CUDA install location (i.e. "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7")
Also ensure to the path for cl.exe (i.e. "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.33.31629\bin\Hostx64\x64") was added to the PATH variable.

Instead of Step 9
I edited the webui.bat file and added --xformers to launch.py line
:launch %PYTHON% launch.py %* pause exit /b

to
:launch %PYTHON% launch.py --xformers %* pause exit /b

After that it speed up a little, I am sure it is my PC, but it did speed up about 10%.
Mainly did this to try to speed up training, but the amount it saved was only 10%.
Creating images was about 15% (It was slower with HiRes and Face Fix, but that If I understand correctly run separate and this won't help with that.)