This repository contains an experimental torrent-style download for model weights in Hugging Face.
The core implementation is main_hub.py
, which downloads files from Hugging Face and converts them into seedable torrents.
pip install hf-torrent
Before launching downloads, make sure you have aria2c installed and
Then you can start using HF-Torrent with the following command:
hf-torrent <your target repo>
Or if you have your preferred torrent clients, you can pull the torrent file only
hf-torrent --get-torrent <your target repo>
To install, run the following command:
pip install -r requirements.txt
It is easy to contribute to seeding! First, change --seed-ratio=0
(from 10 to 0) when launching the aria2c
RPC. Setting 0
means the daemon will seed infinitely.
Next, change the upload bandwidth limit to a value that suits you, for example: --max-upload-limit=20M
(4M by default).
After making these changes, launching seeding is as simple as downloading models. You can start seeding by running the following script and help distribute your favorite models!
python download.py <your target repo>
To create a popular torrent, run the following command:
python generate_popular_torrents.py -r efederici/retnet-120m
This script will (1) download model files from Hugging Face, (2) iterate through the folder, (3) generate torrents for each file, and then put them under hf-torrent-store/{repo}
.
We are looking for a wanted model list. Please open an issue if you have any favorite models to share.