5p00ky$ cat ~/notes/coordinating-local-inference-without-a-heavyweight-scheduler.md
Coordinating local inference without a heavyweight scheduler
Inferlock uses conservative exclusive leases so cooperating inference engines can share one GPU without pretending to be a cluster manager.
Local inference stacks often have several engines that can individually consume most of a GPU. A full scheduler is excessive, but hoping they never overlap is not coordination.
Inferlock exposes a deliberately small lease service. Cooperating clients acquire exclusive GPU access, renew while working and release when finished.
The narrow scope is the point. It coordinates known engines without claiming to police every process on the machine.