GPU Scheduling & Priority Queue
A priority-ordered scheduler for shared accelerator capacity that selects higher-priority waiting work next while allowing the active request to finish.
Shared accelerators create a simple but consequential queueing problem: interactive work can otherwise sit behind background jobs that are already waiting.
Attributed requests enter a priority-ordered waiting queue for an eligible execution target. Health-aware dispatch chooses where work can run, bounded active capacity limits simultaneous execution, and queue metadata accompanies the accepted request.
Priority changes the order of waiting work only. The scheduler lets the active request finish, then selects the next eligible request; unavailable execution or backend failure is surfaced instead of being described as an invisible pause-and-resume cycle.