NVIDIA FLARE lets each training site keep its own job scheduler

NVIDIA FLARE separates always-on coordination from temporary training workers, letting participants in a federated-learning project use Docker, Kubernetes, or Slurm at different sites. Jobs request compute resources while local launchers translate those requests into containers, pods, or batch allocations. Coordination can stay available without holding training GPUs between jobs, and each organization keeps its own operating policies.

NVIDIA Technical Blog · Sep 15, 2026

Why it matters

A useful DevOps design for collaborating across organizations without forcing every team onto the same infrastructure.

More context

FLARE 2.8 supports Docker and Kubernetes; 2.9 adds Slurm. Operators map each study to approved images, datasets, secrets and scheduling rules. Evaluate the deployment documentation against your existing cluster controls before trying a small multi-site job.

Keep in mind

Study boundaries provide logical separation. Separate administrators, public-key infrastructure or failure boundaries may require separate deployments. 1min has not tested workload isolation.

Original read