awswarm
Run one model too big for any single GPU across a pool of small ones.
The Aither World is an operating system for agents — a Linux you can hand to one, the runtimes it works in, and the tools it works with. awswarm is one of its 71 bricks — each installs on its own, runs offline, and needs no account. 5 more are in development and 15 are planned.
The problem
A frontier MoE checkpoint (Kimi K3, ~1.56TB) fits no consumer card, and "rent a bigger box" stops being an option past a certain size. Splitting a model finely enough to run on heterogeneous, unreliable consumer GPUs -- some too small to hold even one full layer -- is a placement and scheduling problem with no shipped general answer. The money-losing failure mode is retrying a fleet acquisition that was never going to complete, because nothing scored the odds before spending on it.
that a model either fits your GPU or it doesn't run at all
a placement plan and an acquisition-probability estimate before you spend on a run
Adopt it
pip install awswarmFeed it one layer's shape and a pool of heterogeneous GPU specs and get back a sub-layer placement plan, plus the probability that plan actually assembles.
Pairs with
Pairing is composition, never dependency — awswarm installs and runs on its own.
The rest of the Aither World
Every brick below is live, drawn from that repository's own published manifest.