zest-transfer 0.4.0

P2P acceleration for ML model distribution

zest — P2P Acceleration for ML Model Distribution

zest accelerates ML model downloads by adding a peer-to-peer layer on top of HuggingFace's Xet storage. Models download from nearby peers first, falling back to HuggingFace CDN — never slower than vanilla hf_xet.

Install

pip install zest-transfer

Authentication

zest needs a HuggingFace token to download models. Set it up once:

# option 1: environment variable
export HF_TOKEN=hf_xxxxxxxxxxxxxxxxxxxxx

# option 2: huggingface-cli (token saved to ~/.cache/huggingface/token)
pip install huggingface_hub
huggingface-cli login

Get your token at huggingface.co/settings/tokens.

Quick Start

CLI

# Pull a model (uses P2P when peers available, CDN fallback)
zest pull meta-llama/Llama-3.1-8B

# Files land in standard HF cache — transformers.from_pretrained() just works
python -c "from transformers import AutoModel; AutoModel.from_pretrained('meta-llama/Llama-3.1-8B')"

Python API

import zest

# One-line activation — monkey-patches huggingface_hub
zest.enable()

# Or pull directly
path = zest.pull("meta-llama/Llama-3.1-8B")

Environment Variable

# Auto-enable on import
ZEST=1 python train.py

How It Works

HuggingFace's Xet protocol breaks files into content-addressed ~64KB chunks grouped into xorbs. zest adds a BitTorrent-compatible peer swarm so these immutable xorbs can be served by anyone who already downloaded them.

For each xorb needed:
  1. Check local cache
  2. Ask peers (BitTorrent protocol)
  3. Fall back to CDN (presigned S3 URLs)

Every download makes the network faster for the next person.

P2P Testing

# Server A: pull a model and seed it
zest pull gpt2
zest serve

# Server B: pull from Server A
zest pull gpt2 --peer <server-a-ip>:6881

Links

• GitHub
• Xet Protocol