Pirate Face Hub

Plug and play torrent downloads for Hugging Face models. Add one package. Change nothing else.

uv run --with pirateface python your_script.py

or install it once and forget about it:

pip install pirateface

That's the whole setup. No code changes, no config, no new API. Your script still calls transformers or huggingface_hub like it always did, and the files now come from Pirate Face seeders over BitTorrent:

from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("org/model")  # unchanged, now from the swarm

The same goes for anything that downloads through huggingface_hub: diffusers, sentence-transformers, hf_hub_download, snapshot_download and the hf download CLI.

Every file is checked against Hugging Face's own fingerprint (SHA-256 for large files, git SHA-1 for the rest) before it goes into the normal Hugging Face cache. If Pirate Face doesn't have the model, a peer is missing, or a check fails, that file comes from Hugging Face as usual. Uninstall it and your code runs exactly as before.

There's also a small CLI:

pf download org/model [--revision REV]

How it hooks in

The package installs pirateface.pth, which Python runs at startup. It does nothing until something imports huggingface_hub, then fetches each file's bytes from the swarm. Hugging Face still answers which commit and fingerprint to expect, and every file from the swarm is checked against it. If Pirate Face has no torrent, or no peer sends the file list within 30 s, the file comes from Hugging Face as usual. Cache layout, symlinks and locks stay with huggingface_hub.

Settings

Flags are on for any value except empty or 0.

Needs Python 3.12 or 3.13 (libtorrent has no wheels for 3.14 yet). uvx picks the newest Python it has, so pass the version:

uvx -p 3.13 --with pirateface --from huggingface_hub hf download org/model

Development

uv run --group dev ruff check