I trained models across MacBooks using Apple's AirDrop protocol. grove is a distributed training library for Apple Silicon. Devices discover each other over AWDL, a direct radio link. If there's a shared WiFi network it upgrades to that for speed, otherwise everything goes over the direct link. No router,... no cloud, no setup. grove start -n 4 grove joinshow more

AWDL is the protocol behind AirDrop. I tested this on eduroam where all peer-to-peer traffic is blocked and it just works. The devices talk directly over radio, bypassing the network entirely. Works offline too.

Each device trains independently using DiLoCo, then syncs pseudo-gradients. SparseLoCo compression brings that down to 1-3% of the payload for slower links. Async overlap hides communication behind compute.

Still early (v0.1) and would love to hear feedback, built on MLX. Wrote about the architecture and technical decisions here: GitHub: pip install grove-ml

i’m using this to mine crypto on everyone’s macbooks at work when they’re having lunch

incredibly cool work

Wow, this is awesome.. AWDL was the bane of my existence 4 years ago 😅

Very interesting work, isn’t it a bit slow for training though? We have been using thunderbolt for distributed training over Macs -

Cool work! 🔥

thanks for the compute🫡

@alexocheema maybe for exo....

Magical! I have to try it. Thanks to @RobiCeraolo for the heads up, I lost this one yesterday. 🙌🏻

This is brilliant - using AWDL for device discovery solves the network configuration nightmare that kills most distributed training setups. The automatic WiFi upgrade when available is exactly the kind of seamless handoff that makes tools actually usable in practice.

@bleuonbase so cool

this is super cool

🔥

Quite impressive, great work!

Insane

Yo @evilsocket

this is awesome! what's the bandwidth? gb/sec?

This is insane for having no setup. Having just a minimal setup, thunderbolt over awdl and wifi will get an order of magnitude decrement in latency Great work!

Insane!

LFG

Super cool!!

airdrop as a distributed training fabric. the model knows no bounds — only WiFi passwords 😎

Nice

Can we use that to train an AI model that will finally solve that pesky “Searching for Device” bug that sometimes pops in and out, and plagues millions of Apple users worldwide

Dope

new infiniband just dropped

@Rafa_Schwinger this is super cool

this is nuts, another great example of SparseLoCo

damn this is super cool

can you do the reverse, run llms on distributed computing

@zmanian yo this is pretty sick

This is going to end well 😜

followed. good work!

@grok what protocol does airdrop use and what is the max data throughput. Please compare it to WiFi and Bluetooth if it is not Bluetooth.

insane