- HyperPod now auto-configures partition-level network topology for Slurm clusters, letting each partition use tree or block topology based on its instance types.
- This topology-aware scheduling improves distributed training performance by optimizing GPU‑to‑GPU communication and NCCL collectives.
- Enabled by default on Slurm 25.11+; requires no manual setup and adapts automatically during scaling or node replacement.