Velocity-FL¶
The uv of Federated Learning: Rust speed, Python ergonomics.
Rust core Python API Up to 138× faster FedAvg
What is Velocity-FL?
Federated learning where the round's hot path, aggregation and attack simulation, runs in Rust, and everything you touch stays Python: Hugging Face, PEFT and PyTorch.
Aggregation in compiled code
4.0 ms vs 545 ms
FedAvg at 1M params42.2 ms vs 5.82 s
FedAvg at 10M params~138×
faster than the pure-Python fallback
Aggregation only, not end-to-end training, on the latest idle-box snapshot. Methodology and every tier
One round, end to end
- Configure A server, a strategy and a round count, in Python or TOML.
- Collect Client updates cross into Rust, zero-copy where possible.
- Aggregate Nine strategies, all in Rust, from FedAvg to Bulyan.
- Attack and record Byzantine attacks run in the core; Prefect records every round.
Explore the docs
Getting Started
Install and run your first round.
CLI Reference
Every velocity command.
Architecture
The Rust crate, PyO3 bindings and Python layer.
Configuration
Every server, strategy and attack field.
Strategies
The nine aggregators, and which to pick.
Attacks
Poisoning, Sybils, noise, label flipping.
API Reference
VelocityServer, Strategy, ClientUpdate.
Benchmarks
Rust against Python, tier by tier.
Rust + PyO3maturin + uvPrefectTyperPydanticHugging FacePEFTPyTorch