onnx-tensorrt

Why we included this project

Teams deploying ONNX models on NVIDIA GPUs use this parser when latency matters. It takes an ONNX graph and compiles it into a TensorRT engine, so a model trained and validated in PyTorch or another framework runs on TensorRT's optimized kernels instead of a generic interpreter. It is available as a C++ library, a Python backend, and an executable, so you can slot it into a serving pipeline at whatever level you work. Dynamic shapes and full-dimension graphs are supported, which helps if your input batch or resolution changes at runtime. Plan for it being a compiler rather than a drop-in runtime: you build the engine ahead of time and execute it later, and not every ONNX operator has a native mapping, a gap the included operator support matrix spells out.