SkyPilot has updated its scheduler for Nvidia Grace Blackwell clusters

The SkyPilot team has introduced a new job scheduler specifically designed to optimize the training of large language models on Nvidia NVL72 rack-scale systems with the Grace Blackwell architecture.

The new resource allocation algorithm keeps tensor parallelism and mixture-of-experts (MoE) parallelism within the high-speed NVLink interface. As a result, the developers managed to achieve a 43% increase in throughput compared to traditional InfiniBand connections.

Furthermore, the system addresses hardware fragmentation through topology-aware job placement. According to the creators, this approach allows for 27% more efficient GPU utilization on the same hardware racks.

The update is aimed at large data centers and research laboratories that use arrays of Blackwell chips to scale heavy workloads without additional capital expenditure on hardware.