NVIDIA FLARE 2.8 and 2.9 introduce a two-tier architecture designed to support federated learning across diverse infrastructures. The system separates long-running federation services from job-specific workers, allowing organizations to maintain control over their local data, GPU resources, and security policies.
The platform now supports Docker, Kubernetes, and Slurm as execution environments. Each site uses a local launcher to translate task requirements—such as CPU, memory, and GPU needs—into native configurations for their specific environment, whether it is a container, a pod, or a batch allocation.
The architecture utilizes a parent-worker model where parent processes coordinate the federation while job workers are created on-demand. This approach ensures that GPU resources are only occupied when a task is actively running, optimizing infrastructure utilization across different sites.
Logical multitenancy is managed through 'study' configurations. Each study defines participating sites and maps tasks to local resources, such as datasets and secrets, without exposing sensitive data. This allows multiple research projects to run concurrently on the same infrastructure while maintaining strict isolation.
