SageMaker HyperPod supports partition-level topology for Slurm clusters

SageMaker HyperPod now supports network topology configuration at the partition level for Slurm orchestrated clusters. This allows different topologies, such as tree and block, to coexist within a single cluster, chosen based on the instance types in each partition. The system automatically maintains this configuration during cluster scaling events.

By aligning job placement with instance interconnect characteristics, this feature accelerates GPU-to-GPU communication, enhances NCCL collective operation efficiency, and boosts overall training throughput. Topology-aware scheduling is enabled by default for supported instance types.

HyperPod automatically assigns block topology to partitions with instance types like ml.p6e.gb200, and tree topology to those with hierarchical-interconnect types like ml.p5.48xlarge. Partitions without specific topology information remain fully schedulable. The feature requires Slurm 25.11 or later and is available in all AWS Regions where SageMaker HyperPod is supported.