Wednesday 28 December 2022

Stability AI Stable Diffusion ¶

Repo on Github: https://github.com/Stability-A…
Clone repo then cd into it
Installed Anaconda first yay -S anaconda
Activate anaconda env source /opt/anaconda/bin/activate root
Install pytorch conda install pytorch==1.12.1 torchvision==0.13.1 -c pytorch
Install required packages pip install transformers==4.19.2 diffusers invisible-watermark
Got this error

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
daal4py 2021.6.0 requires daal==2021.4.0, which is not installed.
numba 0.55.1 requires numpy<1.22,>=1.18, but you have numpy 1.24.1 which is incompatible.

Installed daal manually pip install daal==2021.4.0
run pip install -e .
Download https://huggingface.co/stabili… I didn’t know about HuggingFace
Install Cuda sudo pacman -S cuda
to install xformers

export CUDA_HOME=/opt/cuda
sudo conda install -c nvidia/label/cuda-11.4.0 cuda-nvcc
sudo conda install -c conda-forge gcc
sudo conda install -c conda-forge gxx_linux-64==9.5.0

I added sudo because without sudo I got this error

EnvironmentNotWritableError: The current user does not have write permissions to the target environment.
  environment location: /opt/anaconda
  uid: 1000
  gid: 1000

Previous commands took too long to finish (couple hours). I’m not sure where the slowness is coming from. I noticed it’s maxing out 1 logical CPU out of 16 ☹️
Clone xformers https://github.com/facebookres… and cd into it
Setup the rest of requirements

git submodule update --init --recursive
pip install -r requirements.txt
sudo pip install -e .

this last line shows an error: ModuleNotFoundError: No module named 'torch' so I ran sudo conda install pytorch then sudo pip install -e .
I got some errors of missing packages so

pip install omegaconf
pip install torchvision
pip install pytorch_lightning
pip install pytorch-lightning
pip install open_clip_torch

I got another error that prompted me to run this sudo python setup.py build develop
also found error about missing triton so pip install triton==2.0.0.dev20221120
created a dir output and ran

python scripts/txt2img.py --prompt "a professional photograph of an astronaut riding a horse" --ckpt v2-1_768-ema-pruned.ckpt --config configs/stable-diffusion/v2-inference-v.yaml --H 768 --W 768  --outdir output

I got an error

making attention of type 'vanilla-xformers' with 512 in_channels
building MemoryEfficientAttnBlock with 512 in_channels...
Killed

tried reducing number of steps and other params but I got the same error

python scripts/txt2img.py --prompt "a professional photograph of an astronaut riding a horse" --ckpt v2-1_768-ema-pruned.ckpt --config configs/stable-diffusion/v2-inference-v.yaml --ddim_eta 0.0 --n_samples 3 --n_iter 3 --scale 5.0  --steps 100 --H 192 --W 192 --outdir output

I guess at this point I should give up on this approach and try another setup like this one https://github.com/basujindal/…

Docker compose fork for CompVis Stable Diffusion ¶

cloned this repo and cd into it gh repo clone basujindal/stable-diffusion
moved the model to another directory mv ../stablediffusion/v2-1_768-ema-pruned.ckpt ../sd-data/model.ckpt
Install nvidia-container-toolking yay -S nvidia-container-toolkit
uncomment no-cgroups = false in /etc/nvidia-container-runtime/config.toml
and restart docker sudo systemctl restart docker
I had to update my machine at this point as it couldn’t install bmake so that’s another 4gb or downloads and 14gb installation side. 🥲
build the containers docker-compose build
and run the containers docker-composer run
I got an error, it seems the model for Stable Diffusion from Stability won’t work on the CompVis version. so I downloaded the latest version from their model https://huggingface.co/CompVis
then run docker-composer run again.
I got this interface