microsoft/onnxruntime

Supports inference across diverse ML frameworks (PyTorch, TensorFlow, scikit-learn, XGBoost, LightGBM) through the ONNX standard, applying graph optimizations and hardware acceleration (CPUs, GPUs, NPUs) for optimal performance. Training acceleration targets PyTorch transformer models on multi-GPU setups with minimal code changes. Operates as a portable runtime layer abstracting hardware and framework differences across Windows, Linux, and macOS.

19,534 stars and 76,261,123 monthly downloads. Used by 150 other packages. Actively maintained with 160 commits in the last 30 days. Available on PyPI and npm.