Quick Start Series

Welcome to the NVIDIA FLARE Quick Start Series! This guide provides a set of hello-world examples to help you quickly learn how to build federated learning programs using NVIDIA FLARE.

Make sure you have completed the Installation steps before proceeding.

Run Modes

FLARE supports three modes for different stages of your workflow:

Simulator (NVIDIA FLARE FL Simulator) – Runs jobs on a single system for fast testing and algorithm development.
POC (Proof Of Concept (POC) Command) – Simulates deployment on one host with separate processes for clients and server.
Production (Provisioning and startup package distribution) – Distributed deployment using startup kits from provisioning.

Start with the Simulator for development, then validate with POC before going to Production.

Convert Your ML Code to Federated

Converting existing training code to federated learning requires just 3 changes:

Step 1: Add FLARE imports to your training script

import nvflare.client as flare

Step 2: Initialize FLARE and wrap your training loop

flare.init()

while flare.is_running():
    input_model = flare.receive()           # receive global model
    model.load_state_dict(input_model.params)

    # ... your existing training code here ...

    output_model = flare.FLModel(
        params=model.cpu().state_dict(),
        metrics={"accuracy": accuracy},
    )
    flare.send(output_model)                # send updated model back

Step 3: Create a job recipe to define the FL workflow

from nvflare.app_opt.pt.recipes import FedAvgRecipe

recipe = FedAvgRecipe(
    name="my-fedavg-job",
    min_clients=2,
    num_rounds=5,
    train_script="train.py",
)
recipe.execute()

That’s it. Your training logic stays the same – FLARE handles the communication, aggregation, and orchestration. For the full Client API reference, see Client API. For pre-built recipes, see Available Recipes.

Hello-world Examples

The following hello-world examples demonstrate different federated learning algorithms and workflows. Each example includes instructions and code to help you get started.

Hello PyTorch - Federated averaging with PyTorch models and training loops. Hello PyTorch
Hello Lightning - Example using PyTorch Lightning for streamlined model training. Hello Pytorch Lightning
Hello Differential Privacy - Federated learning with differential privacy using PyTorch and Opacus for privacy-preserving training.
Hello TensorFlow - Federated averaging using TensorFlow models.
Hello JAX - Federated averaging using JAX, Flax, and Optax on MNIST.
Hello Logistic Regression - Federated logistic regression example using scikit-learn.
Hello Cyclic - Cyclic federated learning workflow example.
Hello Tabular Statistics - Federated statistics computation example.
Hello Flower - Running Flower apps in FLARE.
Hello XGBoost - Federated XGBoost example demonstrating gradient boosting for tabular data in a federated setting.

Let’s start with Hello PyTorch: Hello PyTorch