As AI workloads scale to thousands of GPUs, organizations need scheduling capabilities that Kubernetes doesn't provide out of the box: gang scheduling for multi-node jobs, fair-share policies across teams, and topology-aware placement for optimal GPU communication. Slurm has solved these problems for decades in HPC, but traditionally requires dedicated bare-metal infrastructure.

In this talk, we show how to run Slurm natively on Kubernetes using Slinky, an open-source operator. This approach lets you keep existing Slurm workflows (job scripts, accounting, fair-share policies) while gaining Kubernetes benefits: declarative configuration, rolling updates, Helm deployments, and ecosystem integration. We operate this at scale with 8,000+ GPUs for distributed AI training.