Domain-aware placement
The placement groups are built to be aware of NVLink domains, meaning they take into account the high-speed interconnect topology when assigning GPUs. This allows the scheduler to place GPUs that need to communicate frequently on the same NVLink domain, reducing latency and improving throughput. The approach is particularly relevant for systems like the GB300 NVL72, which is designed for demanding AI training and inference tasks.
Multi-node GPU scheduling is a complex problem. When a model spans multiple servers, the way GPUs are grouped
