Maybe the simple service should use CUDA images as base too?

[veritas9872]

A lot of libraries need to compile stuff during runtime. Using an Ubuntu base image is needlessly complicated.

[veritas9872]

But then, how would simple be any different from train?

[veritas9872]

After taking a look at the train image, I think I should just get rid of the fetch-torch option and force that image to always be a version that builds PyTorch. I will make the simple service the default to make life easier.
For this, the simple service needs to be able to take in more base images.

[veritas9872]

Also, I should probably rename the train service to compile or something and make it non-default.

[veritas9872]

Maybe custom?

[veritas9872]

This would also involve totally rewriting the README, which is a lot of work...

[veritas9872]

Also, the simple service can get rid of the frozen conda requirements, I think. I don't use it and no one has asked about it after I removed the frozen environment file.

[veritas9872]

Also, the README is getting a bit embarrassing after all these years. Maybe make it more sober? Although it might be better staying rather colorful. It should most definitely be shorter, though.