Support pulling models from OCI registries (Docker Hub, etc.)

[ericcurtin]

It would be not so tricky to add OCI pulling code in Swift, to be able to pull models from here:

https://hub.docker.com/u/ai

or any OCI registry private or public. I can push the PR.

[erusev]

I'm not familiar with OCI registries.

Could you share a brief explanation of what it is and how it could be useful in LlamaBarn?

[ericcurtin]

Docker Hub is an example of an OCI registry but there are many others quay.io, ghcr, etc. It's common to have private OCI registries in house in data centres, rather than always pulling things from the public internet. It allows one to push and pull models to/from many places using existing infrastructure, democratizes transport of models I guess, so that huggingface isn't the only source for model pushing and pulling.

[erusev]

Thanks for the proposal! A few questions to understand the use case:

1. What problem does this solve? Environment without Hugging Face access, or something else?
2. Would you pull LlamaBarn's existing catalog models from OCI registries, or arbitrary models?
3. LlamaBarn's catalog includes compatibility checks, footprint calculations, etc. For arbitrary models, how would this metadata be provided?

Understanding the scope will help evaluate whether this fits LlamaBarn's curated, compatibility-checked approach.

[ericcurtin]

Thanks for the proposal! A few questions to understand the use case:

1. What problem does this solve? Environment without Hugging Face access, or something else?

Yes there are many OCI-registries, they are everywhere, like git repos, easy to set up in your private networks. Models are huge we are talking about trillion parameter models these days. Do we always want to pull over public internet? Is that secure?

2. Would you pull LlamaBarn's existing catalog models from OCI registries, or arbitrary models?

Open-minded. I would say do both.

3. LlamaBarn's catalog includes compatibility checks, footprint calculations, etc. For arbitrary models, how would this metadata be provided?

This is metadata:

https://www.docker.com/blog/oci-artifacts-for-ai-model-packaging/

Understanding the scope will help evaluate whether this fits LlamaBarn's curated, compatibility-checked approach.

[erusev]

Thanks for the detailed answers! OCI support sounds promising for secure, private model distribution. I'm converting this to a Discussion. Would like to see community interest before prioritizing development.