Docker support (Tested on RTX3080) + ignoring reproducable files / cache

[peterwilli]

Intro

I added support for Docker, and made sure cache and build files do not hit the repository, making it easier to work for others on this.

The Docker image downloads Python from source and uses optimized compilation making sure for max possible speed. It replicates the Conda environment and gives easier access for others as well as deploying this to servers!

I added a clause in the README explaining how to install the software on Docker.

Also offers a fix for #49 (this issue won't appear on Docker)

Tested on:

• Tuxedo Stellaris 15 (AMD Ryzen 9 5900HX - 32GB RAM - RTX 3080 16GB)

"A large blue whale on a freight ship, vector art"

[srelbo]

Just curious, why not put these in a requirements file so its easier to read?

[peterwilli]

We can do that, except RUN python3 -m pip install pip==20.3 because the goal was to stay as close to the Conda environment as possible to make sure we won't run into issues that they wouldn't have otherwise when they chose to install with Conda.

[srelbo]

Thank you @peterwilli this is very useful.

[srelbo]

^ This will probably not work on older GPUs. Perhaps we should state it in the documentation somewhere?

[peterwilli]

I'm thinking of using build args for this. However, I don't know the exact configuration for each GPU. I used an RTX3080, I think that we should just expose the choice of Torch versions as build arg, and then let people choose themselves. We can set a default version that works for most GPU's.

But I don't think GPU with less than 12GB VRAM works anyway, so it narrows down the old GPU's supported, depending on how old we're talking about!

[peterwilli]

I wonder, is there a better way to get the right Torch version supporting the GPU of the machine running the image builder I'm not aware of? Currently, all my Dockerfiles are configured for my GPU, but it makes it less portable.

[peterwilli]

Thank you @peterwilli this is very useful.

Thank you, it's useful for me too. I'll work with you to resolve the comments.

[peterwilli]

Hey @srelbo! I uploaded a new commit, sorry for the long wait. The Dockerfile now uses Conda rather than trying hard to replicate the same environment. The only exception is #49 which can safely be deleted after that is fixed.

In addition, we use the pre-made nvcr.io images as base. These are officially made by Nvidia, so they should be trustworthy and widely supported. I had good experience with them before in my work at LAION medical (https://github.com/LAION-AI/medical) so I think this will be the best we can get with Docker...

The README is slightly adjusted for the new command structure (you have to call python now, something I did because now, for debugging reasons or otherwise, one can bash into the Docker container should they want to).

[srelbo]

That's awesome! Thank you so much @peterwilli ! We will merge your commit to our fork of this branch. But I think it should get it into mainline too, so everyone can benefit from your work.

@rromb can you please review and merge?

Also @oguzelibol FYI

[Utopiah]

FWIW just tried it, trying to follow the instructions as well as I could. The image was built but running it, using the suggested parameters I get ModuleNotFoundError: No module named 'kornia'.

PS: if that helps somehow Ubuntu 22.04 with nvidia-docker2 showing nvidia-smi working.