Hummingbird compiles trained ML models into tensor computations that can dramatically accelerate inference without further re-engineering. Specifically, trained “traditional” ML models from packages such as sklearn and xgboost can be converted to PyTorch, TorchScript, and ONNX, thereby benefiting from any native hardware acceleration and/or current/future optimizations to these libraries. In addition, Hummingbird provides a single unified inference API modeled after sklearn’s, so sklearn models do not require any inference code changes.