A1111+SD install issues, wants unavailable Python 3.11.7 for xformers, not 3.11.11

[playertoo]

tl;dr - I'm looking source a trustworthy debian flavored Python 3.11.7 & modules as I don't want to want to compile from source, entering that dep-hell. Then I've to figure out how to get the specific PyTorch installed (as mentioned later below.) I'll have to do the same for fedora 41 next.

---

Although Automatic1111 works for me with the default installed v1-5-pruned-emaonly.safetensors and other =<4 GB file size checkpoints. When I try to load ponyDiffusionV6XL_v6StartWithThisOne.safetensors (>6 GB), it throws an OOM and crashes -- stating it needs 20 GB VRAM.

Before I can troubleshoot that OOM situation, I need to tackle this error being an earlier listed error for A1111's PyTorch dependency. I suspect with this failure, like it says, memory management is not possible -- which constrains me to smaller and pruned checkpoints.

Looking for an assist to overcome the following error:

WARNING:xformers:WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
    PyTorch 2.1.2+cu121 with CUDA 1201 (you have 2.6.0+cu124)
    Python  3.11.7 (you have 3.11.11)
  Please reinstall xformers (see https://github.com/facebookresearch/xformers#installing-xformers)
  Memory-efficient attention, SwiGLU, sparse and more won't be available.
  Set XFORMERS_MORE_DETAILS=1 for more details

So, I'm looking source a trustworthy debian flavored Python 3.11.7 & modules as I don't want to want to compile from source, entering that dep-hell. Then I've to figure out how to get the specific PyTorch installed (as mentioned later below.)

FYI, Others have had A1111 working on a GTX1660 6GB without issue with startup flags. So I don't think I'm necessarily hardware constrained -- things will just take longer to render.

Here's what I'm having work with and done so far on my Low Profile system (that use-case constrains me.)

• Low Profile RTX 3050 6GB - (finally an RTX I can fit with the case cover on and bus powered.)
• AMD Ryzen 5 2600, 32GB RAM (A1111 not being CPU based, the LP workstation gets the job done. )
• Fresh Mint 22.1 XFCE install (Ubuntu 24.04 based) to a new, separate SSD just for this AI hobby/learning (my single NVMe is for my main OS).
  • nvidia-driver-550 (the recommend highest version release available in Mint)
  • kernel 6.11.0-19 (switch from 6.8.0 LTS)

Here's the build out and what I've done like so many HOWTOs have said:

# Ubuntu 24.04 (the base for Mint 22.1)
# home to older pre-made, community trusted Python releases:
# https://askubuntu.com/questions/1398568/installing-python-who-is-deadsnakes-and-why-should-i-trust-them
sudo add-apt-repository ppa:deadsnakes/ppa
sudo apt update
sudo apt install python3.11 python3.11-venv

mkdir ~/aiapps/
cd ~/aiapps/
python3.11 -m venv .cp311
source .cp311/bin/activate
# I prefer my venv outside the app so I can re-use it considering the size of PyTorch and all.
# so long as it's activated is all that matters

git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
cd stable-diffusion-webui/
pip install -r requirements.txt

~/aiapps/stable-diffusion-webui/webui.sh
# profit - default and =<4 GB checkpoints work fine

I also confirmed CUDA is available:

(.cp311) me@pc:~/aiapps/stable-diffusion-webui$ python3 -c "import torch; print(torch.version.cuda)"
12.4
(.cp311) me@pc:~/aiapps/stable-diffusion-webui$ python3 -c "import torch; print(torch.cuda.is_available())"
True
(.cp311) me@pc:~/aiapps/stable-diffusion-webui$ python3 -c "import torch; print(torch.cuda.get_arch_list())"
['sm_50', 'sm_60', 'sm_70', 'sm_75', 'sm_80', 'sm_86', 'sm_90']
(.cp311) me@pc:~/aiapps/stable-diffusion-webui$ python3 -c "import torch; print(torch.__version__)"
2.6.0+cu124

With only ponyDiffusionV6XL_v6StartWithThisOne.safetensors installed, I tried the following:

source .cp311/bin/activate
export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512
export COMMANDLINE_ARGS=--xformers --no-half --opt-sdp-attention --medvram
~/aiapps/stable-diffusion-webui/webui.sh

and was crashing with the OOM error. I then noticed the earlier message (having quickly scrolled off the screen). And double checked with the default v1-5-pruned-emaonly.safetensors -- yup still there even though I could use A1111 with =< 4 GB file sized checkpoints.

WARNING:xformers:WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
    PyTorch 2.1.2+cu121 with CUDA 1201 (you have 2.6.0+cu124)
    Python  3.11.7 (you have 3.11.11)
  Please reinstall xformers (see https://github.com/facebookresearch/xformers#installing-xformers)
  Memory-efficient attention, SwiGLU, sparse and more won't be available.
  Set XFORMERS_MORE_DETAILS=1 for more details

I then tried to re-install the transformers:

source .cp311/bin/activate
export COMMANDLINE_ARGS=--reinstall-xformers
~/aiapps/stable-diffusion-webui/webui.sh

Nope. Didn't re-install. Then I dove down deeper rabbit holes. For example,

• The deadsnakes repo only offers python 3.11.11, not a 3.11.7 release. Supposedly A1111 works with python 3.10 but considering the deadsnakes release has the same date stamp as their 3.11, I'm not confident I won't have the same situation with their python 3.10.
• If I had to manually install PyTorch cu121 in the venv, that should be possible using pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121 given it still exists. I just don't know yet how to force it to install torch-2.1.2+cpu.cxx11.abi-cp311-cp311-linux_x86_64.whl considering torch-2.6.0+cpu.cxx11.abi-cp311-cp311-linux_x86_64.whl is likely what would be installed being the latest from that repo.
• Looking into Containers, those have their warts too. And, some Dockerfiles are using the ppa:deadsnakes/ppa as well putting right back where I'm at now. The container adding another layer of hurt (pun intended.)

If there's someone who could provide some guidance, I would be very thankful. I really would like to have A1111 working. Having the prompt and settings captured in the PNG image text blocks is extremely helpful. IIRC, Comfy has a similar feature.

I find the WebUI a little more intuitive vs Comfy's "flowcharting" (even though it looks to be more powerful & flexible.) And, A1111 is what my cohorts are using on our project so I'm kind of locked into getting A1111 operational.

Thanks in advance for the assist. Cheers.

FYI, I'd rather avoid using Conda, miniconda might be tolerable. I'll also need to figure out how to get it working for ROCm on my Framework 13 running the Fedora 41 Silverblue. I suspect I'll have a similar challenge there finding an older, available Python 3.11.7 and modules.

[missionfloyd]

You could try --opt-sdp-attention instead of xformers.
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Optimizations

• If I had to manually install PyTorch cu121 in the venv, that should be possible using pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121 given it still exists. I just don't know yet how to force it to install torch-2.1.2+cpu.cxx11.abi-cp311-cp311-linux_x86_64.whl considering torch-2.6.0+cpu.cxx11.abi-cp311-cp311-linux_x86_64.whl is likely what would be installed being the latest from that repo.

Specific versions of packages can be installed like so:

pip install torch==2.1.2+cpu --extra-index-url https://download.pytorch.org/whl/cu121

But that's the CPU version of torch, which is very slow. Use torch==2.1.2+cu121 for nvidia cards.