Federated learning without shipping anyone's data anywhere

What it does

NVIDIA FLARE is a domain-agnostic Python SDK for federated learning and the broader idea it calls federated computing: multiple independently controlled sites run approved tasks on their local data and exchange only permitted results — model updates, statistics, evaluations, intersections. A server coordinates workflows and aggregates, but raw datasets never leave their source. It supports PyTorch, TensorFlow, and XGBoost, plus federated analytics and custom multi-site computation.

The interesting bit

The federated-learning part is built on a general federated-computing core — orchestration, task execution, communication, security, lifecycle — so training is one application of the runtime rather than the whole thing. That means the same machinery serves statistics jobs, private set intersection, or split learning and swarm-style decentralized workflows where clients, not a server, call the shots.

Key highlights

• Data stays at its source; site policies decide what results may leave
• Works with PyTorch, TensorFlow, and XGBoost, including LLM fine-tuning (Hugging Face LoRA, NeMo)
• Privacy extras: differential privacy, homomorphic encryption, PSI, confidential computing
• Runs a two-client simulation on CPU with synthetic data — no dataset download needed to try it
• Extensive example catalog, tutorials, and a self-paced curriculum; Apache 2.0

Caveats

• The nvflare examples CLI requires 2.10.0 or later, which the README notes isn’t on PyPI yet — you’d install from main, which is explicitly the unstable development branch
• Example setup is version-sensitive: each example’s README owns its own dependencies, and older requirements.txt files can clobber your installed NVFlare version

Verdict

Worth a look if you’re in healthcare, finance, or any domain where data can’t cross organizational lines but collaboration must. If your data already lives in one place, this is machinery you don’t need.