[Modular Diffusers] Repo ID to custom path mapping

[Sector14]

It would be useful if there was a way to alias a huggingface repo such that all references to it for models or custom blocks are short-cut and replaced before usage by a local path.

For example

"pretrained_model_name_or_path": "diffusers-internal-dev/canny-filtering"

Assumes you are using the huggingface cache and have not already downloaded the model / custom blocks in the past to a --local-dir. Whilst it's possible to replace the repo name above with an absolute path to the local download, this would become quite tedious to do if you have downloaded a lot of custom blocks as these can reference each other.

It also makes sharing a custom block harder since if I have made edits to reference my local blocks by path I would have to un-edit everything to refer to my actual shared repo for it to be useful to others.

Whilst I could use the huggingface cache for downloads, I do like to keep a local backup of models/workflows and it can be easier to share across machines using a single local-dir based directory structure.

If the huggingface library provided a way to specify a config file of repo / path mappings, stored in a well-known path in HF_HOME with path override available via an ENV such as LOCAL_REPO_MAPPING=/some/path/local_repo_mapping.json, this could enable any reference to "diffusers-internal-dev/canny-filtering" to be replaced at runtime before a download is needed with the mapped absolute path "/some/local/dir/canny-filtering"

repo_mapping: {
  "diffusers-internal-dev/canny-filtering": "/some/local/path/canny-filtering",
  ...
}

Perhaps my usage is a niche case and everyone is happy with the hf cache handling this. I do use it too in some cases, but I really value the flexibility to also have a local-dir downloaded version (either via hf download --local-dir ... or copying a given snapshot out of the hf cache folder) as a backup without having to modify a lot of repo paths to absolute and back again if I wish to share with the community.

Would something like this be feasible? Hopefully there would be a single point where a hook could be placed to short circuit checking the cache and performing a download to instead use the mapped dir.

Any strong reasons against this, or perhaps a better way to support local-dir downloads alongside hf cache ?