ByzPy

ByzPy is a unified Python framework for Byzantine-robust distributed machine learning. It provides a powerful actor runtime, declarative computation graphs, and batteries-included aggregators and attack simulators.

Installation

Install from PyPI:

pip install byzpy              # CPU baseline
pip install "byzpy[gpu]"       # add CUDA/UCX extras
pip install "byzpy[dev]"       # add development tools (pytest, coverage)

Quick Start

from byzpy.aggregators.coordinate_wise.median import CoordinateWiseMedian
import torch

# Create aggregator
aggregator = CoordinateWiseMedian(chunk_size=4096)

# Aggregate gradients
gradients = [torch.randn(1000) for _ in range(10)]
result = aggregator.aggregate(gradients)

Features

• Byzantine-robust aggregators – Krum, MDA, trimmed mean, geometric median, and more
• Unified actor runtime – threads, processes, GPUs, and TCP/UCX remotes share a single API
• Declarative computation graphs – build heterogeneous pipelines with deterministic scheduling
• Complete examples – parameter-server and peer-to-peer demos included

CLI Tools

byzpy version                # installed version
byzpy doctor                 # environment diagnostics
byzpy list aggregators       # discover built-ins

Documentation

📖 Full documentation: https://byzpy.github.io/byzpy/

Links:

• Homepage: https://github.com/Byzpy/byzpy
• Documentation: https://byzpy.github.io/byzpy/
• Issues: https://github.com/Byzpy/byzpy/issues