gbtb/nix-stable-diffusion
1
1
Fork 0
mirror of https://github.com/gbtb/nix-stable-diffusion.git synced 2024-11-05 04:55:20 +03:00
Code Issues Packages Projects Releases Wiki Activity
Flake for running SD on NixOS
60 Commits 5 Branches 8 Tags 189 KiB
Nix 94.6%
Shell 5.4%
vNext
Go to file
gbtb 3507cf8c42 fixed safetensors cargoDeps hash
2023-03-25 22:34:18 +10:00
packages fixed safetensors cargoDeps hash 2023-03-25 22:34:18 +10:00
.gitmodules give up on making paths for webui consistent with InvokeAI, lets ship it as-is 2023-03-18 21:32:31 +10:00
flake.lock readme update, made packages runnable 2023-03-18 22:50:38 +10:00
flake.nix final bits 2023-03-19 11:08:36 +10:00
python-relax-deps-hook.sh provided all deps, made fix for pythonRelaxDepsHook,invokeai.amd is runnable now 2023-03-05 19:58:32 +10:00
README.md mistakes were fixed 2023-03-19 11:12:43 +10:00

README.md

Table of contents

nix-stable-diffusion

Flake for running SD on NixOS

What's done

  • Nix flake capable of running InvokeAI's and stable-diffusion-webui flavors of SD without need to reach for pip or conda (including AMD ROCM support)
  • ...???
  • PROFIT

How to use it?

InvokeAI

  1. Clone repo
  2. Run nix run .#invokeai.{default,amd} -- --web --root_dir "folder for configs and models", wait for package to build
    1. .#invokeai.default builds package with default torch-bin that has CUDA-support by default
    2. .#invokeai.amd builds package which overrides torch packages with ROCM-enabled bin versions
  3. Weights download
    1. Built-in CLI way. Upon first launch InvokeAI will check its default config dir (~/invokeai) and suggest you to run build-in TUI startup configuration script that help you to download default models or supply existing ones to InvokeAI. Follow the instructions and finish configuration. Note: you can also pass option --root_dir to pick another location for configs/models installation. More fine-grained directory setup options also available - run nix run .#invokeai.amd -- --help for more info.
    2. Build-in GUI way. Recent version of InvokeAI added GUI for model managing. See upstream docs on that matter.
  4. CLI arguments for invokeai itself can be supplied after -- part of the nix run command
  5. If you need to run additional scripts (like invokeai-merge, invokeai-ti), then you can run nix build .#invokeai.amd and call those scripts manually like that: ./result/bin/invokeai-ti.

stable-diffusion-webui aka 111AUTOMATIC111 fork

  1. Clone repo
  2. Run nix run .#webui.{default,amd} -- --data-dir "runtime folder for webui stuff" --ckpt-dir "folder with pre-downloaded main SD models", wait for packages to build
    1. .#webui.default builds package with default torch-bin that has CUDA-support by default
    2. .#webui.amd builds package which overrides torch packages with ROCM-enabled bin versions
  3. Webui is not a proper python package by itself, so I had to make a multi-layered wrapper script which sets required env and args. bin/flake-launch is a top-level wrapper, which sets default args and is running by default. bin/launch.py is a thin wrapper around original launch.py which only sets PYTHONPATH with required packages. Both wrappers pass additional args further down the pipeline. To list all available args you may run nix run .#webui.amd -- --help.

Hardware quirks

AMD

If you get an error "hipErrorNoBinaryForGpu: Unable to find code object for all current devices!", then probably your GPU is not fully supported by ROCM (only several gpus are by default) and you have to set env variable to trick ROCM into running - export HSA_OVERRIDE_GFX_VERSION=10.3.0

Nvidia

  • Please note, that I don't have an nvidia gpu and therefore I can't test that CUDA functionality actually work. If something is broken in that department, please open an issue, or even better - submit a PR with a proposed fix.
  • xformers for CUDA hasn't been tested. Python package added to the flake, but it's missing triton compiler. It might partially work, so please test it and report back :)

What's (probably) needed to be done

  • Most popular missing packages definitions should be submitted to Nixpkgs
  • Try to make webui to use same paths and filenames for weights, as InvokeAI (through patching/args/symlinks)
  • Should create a PR to pynixify with "skip-errors mode" so that no ugly patches would be necessary
  • Increase reproducibility by replacing models, downloaded in runtime, to proper flake inputs

Current versions

  • InvokeAI 2.3.1.post2
  • stable-diffusion-webui 12.03.2023

Meta

Contributions

Contributions are welcome. I have no intention to keep up with development pace of these apps, especially the Automatic's fork :) . However, I will ocasionally update at least InvokeAI's flake. Considering versioning, I will try to follow semver with respect to submodules as well, which means major version bump for submodule = major version bump for this flake.

Acknowledgements

Many many thanks to https://github.com/cript0nauta/pynixify which generated all the boilerplate for missing python packages.
Also thanks to https://github.com/colemickens/stable-diffusion-flake and https://github.com/skogsbrus/stable-diffusion-nix-flake for inspiration and some useful code snippets.

Similar projects

You may want to check out Nixified-AI. It aims to support broader range (e.g. text models) of AI models in NixOS.