// the find
pytorch/TensorRT
PyTorch/TorchScript/FX compiler for NVIDIA GPUs using TensorRT
Torch-TensorRT is NVIDIA's compiler for PyTorch models. It lowers the supported parts of a model graph to TensorRT engines, either through torch.compile(backend='tensorrt') or ahead of time with torch_tensorrt.compile using the dynamo frontend. It is for teams already serving PyTorch models on NVIDIA GPUs who want faster inference without rewriting the model.
The dynamo path compiles from torch.export graphs and saves an ExportedProgram that loads back in plain PyTorch, so the artifact does not require the compiler at load time. Unsupported ops are partitioned out and run in PyTorch instead of failing the whole compile, which is what makes it usable on real models. The TorchScript output gives a route to libtorch deployment with no Python dependency. The repo has Linux x86-64, Linux SBSA and Windows build and test workflows, and it documents a six-month migration window for deprecations, which is more than most ML tooling offers.
The README pins its verified stack to a PyTorch nightly, CUDA 13.4 and TensorRT 11.3, and says other combinations are not guaranteed to pass tests, so a production team on stable releases will spend time finding a working combination. Serialized TensorRT engines are generally tied to the GPU architecture and TensorRT version they were built with, so the saved artifact is less portable than the original model. Windows is dynamo-only and Jetson is source-build only, so those platforms get less of the workflow. The 5x speedup in the header is a vendor figure with no workload attached, and the README lists a tool for resolving graph breaks as coming soon, which suggests graph breaks are still a pain point.