Point HF_ENDPOINT at Hangar

Hugging Face libraries read the Hub's address from HF_ENDPOINT, and a Token from HF_TOKEN. Set both and they use Hangar:

export HF_ENDPOINT=https://hangar.example.com
export HF_TOKEN=hgr_...
export DISABLE_SAFETENSORS_CONVERSION=1

hangar env prints these for the Instance you signed in to:

eval "$(hangar env)"

DISABLE_SAFETENSORS_CONVERSION=1 matters for models that only have pickle weights (pytorch_model.bin). Without it, transformers may send the Repo ID, and a Token you passed to it, to a conversion service on huggingface.co. Hangar also refuses the request transformers makes before that call, so the leak cannot happen, but setting the variable avoids the attempt.

Now every Hub call goes to Hangar. The Repo must be in the Arsenal, so pull it first, as in Pull your first model; a Repo that is not there answers as if it did not exist, unless an admin turned on Pull-through:

from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("hf-internal-testing/tiny-random-bert")
tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-bert")

Files land in the usual cache, ~/.cache/huggingface/hub, which the hangar CLI shares.

In Python only

import os
os.environ["HF_ENDPOINT"] = "https://hangar.example.com"   # before importing huggingface_hub

or pass endpoint= to HfApi, hf_hub_download and snapshot_download.

Datasets

datasets also calls datasets-server.<host>. Point that name at Hangar too, which serves the Dataset Viewer API there; see Datasets.

More: Python libraries.