Grove: New Tool Enables Seamless Distributed ML Training Over Apple's AirDrop Protocol
Key Takeaways
- ▸Grove enables zero-configuration distributed ML training across Apple devices using AirDrop's underlying AWDL protocol, eliminating the need to manually specify machine addresses
- ▸The tool works on restricted networks like university WiFi that block peer-to-peer communication, addressing a significant limitation of existing distributed training frameworks
- ▸Uses a hybrid Python+Swift architecture where Python handles ML workloads while a compiled Swift binary manages wireless peer-to-peer connections through a Unix socket IPC channel
Summary
Grove is a new open-source distributed machine learning training framework that leverages Apple's AirDrop protocol to enable zero-configuration model training across multiple MacBooks without requiring network setup or connectivity. The tool uses AWDL (Apple Wireless Direct Link), Apple's proprietary peer-to-peer wireless protocol, to allow devices to discover and communicate with each other automatically, even on restricted networks like university eduroam that employ client isolation. This solves a critical pain point in distributed training where users typically must manually configure machine addresses and network connectivity beforehand.
The system uses an innovative hybrid architecture combining Python for machine learning operations with a lightweight Swift sidecar that manages AWDL connections through Apple's Network.framework. Discovery occurs over AWDL (which works anywhere), while the system intelligently upgrades to faster direct WiFi connections when available. Grove joins existing distributed training tools but takes a fundamentally different approach by prioritizing automatic discovery and working on isolated networks, making it particularly suitable for collaborative training scenarios in academic settings and ad-hoc environments.
- Implements intelligent network fallback: discovery always uses AWDL, but upgrades to faster direct WiFi when available, with transparent fallback if direct connections aren't supported
Editorial Opinion
Grove represents a clever exploitation of existing Apple infrastructure to solve a genuine distributed training friction point. By building on AirDrop's proven reliability and leveraging AWDL's network-agnostic discovery, the creator has found an elegant niche where neither traditional networking tools nor specialized frameworks like EXO Labs excel. The Python-first interface with a hidden Swift backend shows thoughtful API design. If it gains adoption in academic and research communities, this could become the de facto standard for collaborative ML training on Apple hardware.
