ONNX Runtime

ONNX Runtime is Microsoft's cross-platform accelerator for machine learning inference and training. It runs models in the ONNX interchange format exported from deep learning frameworks like PyTorch and TensorFlow or from classical libraries such as scikit-learn, LightGBM, and XGBoost, applying graph optimizations and dispatching to hardware-specific execution providers. Backends cover CPU, NVIDIA GPUs through CUDA and TensorRT, DirectML on Windows, and mobile targets, so the same model can move between server, desktop, and edge deployments without rewriting application code. On the training side it accelerates transformer training on multi-node NVIDIA GPU clusters with minimal changes to existing PyTorch scripts. Released under the MIT license, it is widely embedded in production software by teams that need consistent, fast model execution across heterogeneous hardware without tying the application to a single training framework.