NameError: name 'xformers' is not defined

[papersplease]

Describe the bug While trying to generate an image with --force-enable-xformers on GTX970, I'm getting the error: NameError: name 'xformers' is not defined

Full traceback is here: https://gist.githubusercontent.com/papersplease/d60b889881b27e6ebc5484707ad0fd9e/raw/f6cb1d789bc82e85b1926da582c0568c5ad68d90/gistfile1.txt

To Reproduce

1. Build and pip install the xformers library (I had to pass my compute capability as TORCH_CUDA_ARCH_LIST=5.2 for it to work)
2. Launch webui with --force-enable-xformers (other parameters are irrelevant, the same error happens regardless)
3. Enter the prompt, press generate
4. Get the aforementioned error.

Desktop:

• OS: Windows 10
• Commit revision 050a6a798cec90ae2f881c2ddd3f0221e69907dc
• Nvidia GTX970, driver version 511.65

[C43H66N12O12S2]

Can you please test removing the try block and just importing without checks in modules/sd_hijack_optimizations.py? Remove the cmd_opts check block entirely and input this:

import xformers.ops
import functorch
xformers._is_functorch_available = True
shared.xformers_available = True

This will likely fail as well but it'll output a more useful error.

[papersplease]

Remove the cmd_opts check block entirely and input this:

import xformers.ops
import functorch
xformers._is_functorch_available = True
shared.xformers_available = True

It throws the ModuleNotFoundError: No module named 'xformers' in import xformers.ops. I take it I botched the installation somehow...

[C43H66N12O12S2]

Are you using venv? If yes, you need to install xformers inside venv.

[papersplease]

Yes I'm using venv and installed xformers inside it; the library compiled as intended, and the binary is physically at the intended path inside venv. I'll try to clean it up and recompile.

[papersplease]

Turns out the leftovers from the previous failed compilation somehow prevented the normal functioning, it's working now.

[Mikian01]

Turns out the leftovers from the previous failed compilation somehow prevented the normal functioning, it's working now.

can you elaborate on how to "clean it up and recompile" plz? i have the exact same error :c

[XiaoMouz]

Turns out the leftovers from the previous failed compilation somehow prevented the normal functioning, it's working now.

i have this problem too,how to "clean it up and recompile"?

[Mikian01]

Turns out the leftovers from the previous failed compilation somehow prevented the normal functioning, it's working now.

i have this problem too,how to "clean it up and recompile"?

i solved it, my issue was that the program didnt built the xformers at all, and the reason was bcuz i somehow installed automatic111 with python3.8, i noticed this cuz i erased the venv folder to re-do the whole xformers installation process, which then gave me this error: The detected CUDA version (11.8) mismatches the version that was used to compile PyTorch (11.3). Please make sure to use the same CUDA versions.

so after swaping my default python from 3.8 to 3.10 and reinstalling the whole automatic111webui, i was finally able to install the xformers with no issues

[Zueuk]

Launching Web UI with arguments: --force-enable-xformers
Cannot import xformers
Traceback (most recent call last):
  File "Z:\stable-diffusion-webui\modules\sd_hijack_optimizations.py", line 18, in <module>
    import xformers.ops
ModuleNotFoundError: No module named 'xformers'

I am getting the same error. It compiles, produces a xformers-0.0.14.dev0-cp310-cp310-win_amd64.whl, which installs using pip without any problems. How do I even check what's wrong here?

[0xdevalias]

Rather than manually installing each missing dependency one by one, this is the 'proper' and more sustainable way to fix any missing dependency issues:

You can modify your run_webui_mac.sh to add a line that tries to install the required dependencies with pip install -r requirements.txt each time it starts, after it pulls the latest code. This would probably make a sensible default, and is fairly quick to run, so it might be worth someone telling the original author to include it in their setup script to avoid this sort of error for end users in future:

Here is what mine looks like:

#!/usr/bin/env bash -l

pyenv local anaconda3-2022.05

# This should not be needed since it's configured during installation, but might as well have it here.
conda env config vars set PYTORCH_ENABLE_MPS_FALLBACK=1

# Activate conda environment
conda activate web-ui

# Pull the latest changes from the repo
git pull --rebase

+ # Update the dependencies if needed
+ pip install -r requirements.txt

# Run the web ui
python webui.py --deepdanbooru --precision full --no-half --use-cpu Interrogate GFPGAN CodeFormer $@

# Deactivate conda environment
conda deactivate

Originally posted by @0xdevalias in https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/4109#issuecomment-1304747007

---

For the others who said that pip install -r requirements.txt didn't work for them, or finding that despite pip installing the individual requirements they still don't seem to 'be there', it might be an issue with your conda environment, and which pip is being used. Sometimes the version you're calling doesn't actually install the packages to the correct place, and so they can't be found later.

I have a new theory for you, based on this StackOverflow:

• https://stackoverflow.com/questions/41060382/using-pip-to-install-packages-to-anaconda-environment

What do you see when you activate your conda environment, then run which -a pip?

If it's only something like:

/opt/conda/bin/pip

And not something like:

/opt/conda/envs/web-ui/bin/pip
/opt/conda/bin/pip

Then you can likely fix it by either doing a conda install pip then using pip as normal, or using python -m pip install FOO instead of using pip directly.

Originally posted by @0xdevalias in https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/4109#issuecomment-1308357941

I did a conda install pip and then used python -m, it looked like it wasn't going to work but then I changed my run script a tiny bit,

Curious, what made you think it wasn't going to work? And what was the specific change to the run script that made it work for you? Was it using requirements_versions.txt rather than requirements.txt?

python -m pip install -r requirements_versions.txt

Originally posted by @0xdevalias in https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/4109#issuecomment-1309356932

---

Hopefully the above helps people who are still running into this issue 🖤

Originally posted by @0xdevalias in https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/4061#issuecomment-1314398808

[LukeLIN-web]

Meet same problem