https://huggingface.co/docs/transformers/installation#offline-mode

It would be nice to expose APIs to download models explicitly, and otherwise run huggingface in an offline/firewall compatible way only reading models from a local cache. Huggingface uses caching automatically, but there are still calls to their APIs to check things like the latest model version the first time it is used in a process. If they are unreachable for some reason, the cache will not be accessible, and new processes will not be able to execute models.