Intel / article

Scaling Federated Learning Across Docker, Kubernetes, and Slurm with NVIDIA FLARE

Source: developer.nvidia.com
Author: Elizabeth Goodman
Date: Published Sep 15, 2026

Context

NVIDIA FLARE lets organizations collaborate on model training while using different local computing systems. Persistent processes coordinate the federation; separate workers start when a job needs resources and exit when it finishes. Each site translates the job's resource requirements into its own Docker containers, Kubernetes pods or Slurm allocations. The deployment article explains how this avoids requiring every participant to adopt the same infrastructure.

The earlier workflow coverage explains how sites train on local data and exchange model updates. This deployment layer handles where that work runs. Site operators still control approved resources and enforce local workload policies.

Key takeaways · AI-distilled

NVIDIA FLARE splits federated learning into two layers: persistent processes that just coordinate, and short-lived job workers launched on demand, so coordination never occupies a GPU and each site runs workers on whatever it already has (Docker, Kubernetes, or Slurm).
Jobs describe resource needs abstractly (GPUs, CPU units, memory) via a portable resource_spec, and each site's launcher translates that into native settings: Docker gets nano_cpus and mem_limit, Kubernetes gets pod requests/limits, Slurm gets --cpus-per-task and --mem flags.
FLARE's "studies" give logical multi-tenant boundaries inside one shared federation: each study maps to its own datasets, secrets, approved images, and scheduling policy per site, so a hospital can run a GPU-heavy pathology study and a CPU-only outcomes study without mixing data.
Because dataset paths and secret values live in site-owned config rather than the submitted job, a data scientist can trigger a study's job without ever seeing, or being able to redirect, another site's local file paths or credentials.

Read the source developer.nvidia.com