Optimized Docker image for cloud and local use, with external model cache

[ebr]

What this does

• Initially intended/tested for cloud deployments, further adapted for local use
• Can be used locally on CUDA-capable systems. The only dependencies are docker with the nvidia runtime, and make for convenience. (e.g. not even a local python is required). Fully tested on Linux only (Ubuntu 22.04/CUDA 11.7). Will most likely work on Windows/WSL2.
• Significant optimizations to Docker build:
  • No git clone happens inside docker build - only context is used
  • Attempts to utilize docker layer cache as efficiently as possible by splitting out local package installation
  • reduces image size by at least 2GB compared to the current Docker x86_64 image, might be possible to optimize further.
• Supports the latest unified caching mechanism specified by $INVOKE_ROOT or --root options
• Includes a Makefile for easy building/running (was quick to throw together, but it can be easily rewritten as bash or docker-compose, as this is currently not used in the project)
• adds Github actions for automated image building and pushing, ready for use with e.g. an existing INVOKE_ROOT cache on a cloud runner. No special privileges or secrets are required for image build/push.

Testing/usage:

• cd docker-build
• make load-models (Huggingface token will be prompted for if not already set via env var)
• make run
• access the web UI on http://localhost:9090
• examine the ~/invokeai directory which will be populated with the expected INVOKE_ROOT contents

This PR does not touch any existing local docker setup to avoid surprises to existing users, except un-ignoring some paths from .dockerignore at build time. Should have no practical effect on existing users. But it can easily and optimally replace the existing docker setup. Future work can replace Makefile with docker-compose, if desirable.

[mauwii]

Does the CUDA Image work for ATI as well?

[ebr]

Do you mean AMD? i have no way to test, but I doubt it will, since AMD doesn't use CUDA...

[mauwii]

Yes, ment AMD, sorry, old man sitting here xD

[ebr]

No worries, that's how I also knew what you meant 👴 ;)

[damian0815]

ohh. any hints for making the cache rehydration work with low-tech cheap remote GPU hosts like https://runpod.io (which doesn't expose any system config other than a docker image path)?

[ebr]

I'm not familiar with runpod, but you should be able to run this very image with preload_models.py insead of invoke.py (take a look at the Makefile), as long as you can mount some directory or volume into the container. Basically anywhere you can run a container you should be able to do it. but maybe i'm misunderstanding your question

[damian0815]

actually yeah that should work - thanks.

[ebr]

@mauwii I made significant updates to the PR, rebased on latest in development, etc. Could you please take a look so it doesn't go stale/outdated. Thx

[ebr]

will reopen

[mauwii]

@mauwii I made significant updates to the PR, rebased on latest in development, etc. Could you please take a look so it doesn't go stale/outdated. Thx

Sorry, I was sick the last days and did not open my MacBook at all 🤧

[ebr]

no problem @mauwii, hope you're feeling better. I opened a new PR that I'm currently testing on Runpod.

[mauwii]

I know, I already added a review which instead of lstein is no in my decission :-o

And don't feel ofenden plz, because I updated my image now as well 🙈

[ebr]

@mauwii I don't feel offended, no worries. It does look like some of the ideas from this PR inspired the changes in your Dockerfile, but 🙈 😄 . I'm glad it helped.

[mauwii]

I started this way before your PR came in, then stopped it since the Models where far away from being centralized (would need to look in my fork if I have a PR about this, but I think you won't think I am lieing anyway ;P )