Ai2 replaces GPU priorities with research budgets and time sharing

Ai2 says it has replaced its priority-based GPU scheduler with a system that allocates GPU time through research budgets set by managers. The institute manages thousands of GPUs for about 150 internal researchers, with outstanding requests for 2-3x more GPUs than are available. The scheduler tracks each group's use against its allocation over a default 7-day window, giving precedence to groups that have used less of their budgets. Jobs receive protection from interruption for a declared minimum runtime; afterward, the scheduler can interrupt and requeue resumable workloads. Spare capacity remains available without budget charges, but jobs using it can be interrupted by any budget-funded request.