Runs ONNX-format ML models with cross-platform, hardware-accelerated inference and optional training support — using execution providers (CUDA/TensorRT, OpenVINO, CoreML, etc.), graph optimizations, and runtime plumbing for cloud and edge deployment.
