onnxruntime

Why we included this project

ONNX Runtime is the engine you reach for when a trained model has to run reliably in production. It executes ONNX-format models exported from PyTorch, TensorFlow/Keras, or classical ML libraries like scikit-learn, LightGBM, and XGBoost, so teams can keep training in whatever framework they prefer and still deploy through one runtime. Graph optimizations, quantization, and hardware-specific execution providers cut latency and cost across a wide range of hardware, from edge CPUs to NVIDIA GPUs. The same runtime runs on Windows, Linux, and macOS with bindings for Python, C++, C#, Java, and JavaScript, which helps when one model needs to serve in several environments. There's also a training mode that speeds up transformer training on multi-node NVIDIA GPUs with a one-line change to existing PyTorch scripts, a useful option if you want more throughput without rewriting your training code.