How to properly use webui with conda environment?

[Infinitay]

Preface: Lately I've been trying to update all my modules to their latest versions that are compatible with Pytorch 2.1 and CUDA 12.1. After updating xformers I realized that all this time despite running webui-user.bat on the terminal within my SD conda environment, it was using old modules. I noticed that it was still using venv instead of conda.

I know I could use venv instead of conda and just delete my existing venv folder in order to update; however, it's downloading all my modules twice. Once within my conda environment and once more within the venv folder. After what seemed like 30 minutes of digging around, I found a comment saying to modify webuser-ui.bat and change the respective line to set VENV_DIR=- (note the dash). This worked and successfully used my conda environment, but then I started to have issues with my extensions for some reason.

Is there a proper way to install and use SD Web UI with conda?

[Infinitay]

I finally managed to set up my environment. I will list all the steps I did in case anyone else needs it, and so that someone could review my steps and let me know if anything is wrong.

1. Create new conda environment

   conda create --name stablediffusion python=3.10
   conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia
   pip install ninja
   pip install -v -U git+https://github.com/facebookresearch/xformers.git@main#egg=xformers
   pip install bitsandbytes --prefer-binary --extra-index-url=https://jllllll.github.io/bitsandbytes-windows-webui
   pip uninstall opencv-python-headless
   pip install opencv-python-headless

   • I have to build xformers myself because I am using CUDA 12.1 and Pytorch 2.1.0. They don't seem to have builds for those.
   • Since I am on windows and bitsandbytes still doesn't officially support it, I have to use the precompiled binary from @jllllll
   • The last two commands are because it was installing opencv-python-headless==4.5.5.64. I suppose that was because I had an outdated extension. You may not need it. You can replace the last two commands with one command to just upgrade the package if you know pip.

2. Modify webui-user.bat to not use venv, modify args, and allow for left-click conda start

   call conda activate stablediffusion
   
   @echo off
   
   set PYTHON=
   set GIT=
   set VENV_DIR=-
   set COMMANDLINE_ARGS=--xformers --api --update-check
   
   set SAFETENSORS_FAST_GPU=1
   
   call webui.bat
   

3. Run webui-user.bat

---

In some cases you may have to run some of the commands again from above. Since I was trying to upgrade everything in my existing webui directory, I deleted my venv folder. I started a new conda environment. I ran the commands up to bitsandbytes, but when I attempted to launch webui, it was failing because it kept installing an older version due to the old extensions. I didn't know if there was a command to force-update extensions without having to launch the web-ui and I didn't bother uninstall and reinstalling them. So I manually updated the outdated modules. I had to remove the older bitsandbyte it install and reran the command from above. Afterwards, I had issues with a cv2 module. I saw that the other two opencv modules were up to date, but opencv-python-headless was being forced to a lower version. I manually uninstalled and updated it. At that point everything worked fine and now I have updated my extensions and everything seems to be working.

[AyoKeito]

Currently xformers do not compile under Windows. But you don't need them for modern GPUs anyway:

conda create --name stablediffusion python=3.10
conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia
python -m pip install bitsandbytes --prefer-binary --extra-index-url=https://jllllll.github.io/bitsandbytes-windows-webui
pip install -r requirements.txt

Then, in webui-user.bat or in a separate file:

call conda activate stablediffusion
@echo off

set PYTHON=
set GIT=
set VENV_DIR=-
set COMMANDLINE_ARGS=--no-half-vae --opt-sdp-attention
set SAFETENSORS_FAST_GPU=1

call webui.bat

In Settings -> Optimizations:
Cross attention optimization set to sdp-no-mem - scaled dot product without memory efficient attention or sdp - scaled dot product

If you're still sure that you want xformers, use guide above, but instead of pip install -v -U git+https://github.com/facebookresearch/xformers.git@main#egg=xformers run pip3 install -U xformers --index-url https://download.pytorch.org/whl/cu121

[dsent]

This might be common knowledge already, but for anyone like me, finding this discussion as a n00b: bitsandbytes support Windows now. We can just pip install bitsandbytes and be done with it.

Another thing is that the pip install -r requirements.txt should be executed from the stable-diffusion-webui folder.

Also, if your Anaconda/Miniconda install is the default (i.e. without adding any python/env to PATH), then instead of call conda activate, you should use the full path to your activate.bat. Finally, I found it convenient to move my launcher batch file outside of the stable-diffusion-webui folder, so the A1111 installation pulled from git is separated from the scripts I add (I keep it just directly outside of the stable-diffusion-webui folder, in the parent directory). So my final webui-conda.bat looks like this:

call C:\Users\user\miniconda3\Scripts\activate.bat stablediffusion
@echo off
cd stable-diffusion-webui

set PYTHON=
set GIT=
set VENV_DIR=-
set COMMANDLINE_ARGS=--no-half-vae --opt-sdp-attention
set SAFETENSORS_FAST_GPU=1

call webui.bat

[thiagojramos]

I do things a bit differently (though the outcome is the same). I set the .exe for Python from the Conda environment and in set VENV_DIR= I define it with a -. For set GIT, I don’t set anything because it picks up the Git I have installed globally from the Windows path. Apart from PyTorch, the packages built into the Miniconda3 environment, and Dlib (for which I use the link from the Conda Forge file because a regular command forces the installation of CUDA 11.8), I install everything else via pip install. And yes, I install them "manually." Actually, I install all the packages, including the requirements of the extensions in a single command line (I bulk collect the packages defined in the requirements.txt, and when available, from the install.py of the extensions). Here's my webui-user.bat:

@echo off
git pull
set PYTHON="C:\ProgramData\miniconda3\envs\automatic\python.exe"
set GIT=
set VENV_DIR=-
set COMMANDLINE_ARGS= --xformers --skip-python-version-check --skip-install --update-all-extensions --disable-nan-check --disable-safe-unpickle --enable-insecure-extension-access --localizations-dir "C:\\AI\\Localization\\WebUI" --loglevel ERROR --log-startup
set ACCELERATE=True
set SADTALKER_CHECKPOINTS="C:\\AI\\Models\\SadTalker\\checkpoints"
set NUMEXPR_MAX_THREADS=9
set DREAMBOOTH_SKIP_INSTALL=True
set TF_ENABLE_ONEDNN_OPTS=0
set REQS_FILE=.\\extensions\\sd_dreambooth_extension\\requirements.txt


call webui.bat

[joshgura]

"call conda" sounds cool

but what's the Linux equivalent? Just would like to write "conda activate" somewhere in my webui-user.sh - writing 'conda acitvate ...' inside of a bash script is next to impossible because conda is meant for the terminal, not a script.

[matthu017]

call conda activate is specific to windows
Try adding the following lines to your .sh script as mentioned by:
https://stackoverflow.com/questions/34534513/calling-conda-source-activate-from-bash-script

eval "$(conda shell.bash hook)"
conda activate stablediffusion

[ruixing76]

@joshgura You can simply change the python_cmd to your conda python location. It works for me.

[Nislaco]

Thank you!

[viking1304]

Check #16214 (comment)
I do not want to repost what I already wrote there