Amazon SageMaker HyperPod now supports partition-level topology for Slurm orchestrated clusters

Amazon SageMaker HyperPod now supports partition-level network topology configuration for Slurm orchestrated clusters, allowing different topologies optimized for each instance type.

• Single cluster can run tree topology in one partition and block topology in another
• UltraServer instances use block topology; hierarchical-interconnect instances use tree topology
• Improves GPU-to-GPU communication speed and NCCL collective operation efficiency
• Topology automatically maintained during scale-up, scale-down, and node replacement
• Enabled by default on Slurm 25.11 or later with no configuration required
• Available in all AWS Regions supporting Amazon SageMaker HyperPod

Topology-aware scheduling optimizes distributed training performance by aligning job placement with each instance type's interconnect characteristics.