Grove: Distributed ML Training Across MacBooks via AirDrop's AWDL Protocol — Zero Network Configuration

Swarnim Jain / Hacker News·low signal

Grove is a Python library enabling distributed ML model training across multiple MacBooks using AWDL (Apple Wireless Direct Link) — the same protocol powering AirDrop — requiring zero network configuration. Run 'grove start train.py -n 4' on one machine and 'grove join' on others. Implements DiLoCo (independent training for 500 steps before syncing pseudo-gradients) and SparseLoCo (1-3% gradient compression for ~32x payload reduction). Uses a compiled Swift sidecar via Unix domain sockets since AWDL is locked behind Apple's Network.framework. 32 points on HN. Creative approach to democratizing ML training using consumer Apple hardware.