Comment by conradry

You may find this library I wrote a couple years ago interesting: https://github.com/conradry/prtm. Curious about why you chose to make separate images for each model instead of copy-pasting source code into a big monorepo (similar to Huggingface transformers).