Starting with a trained PyTorch model, it builds optimized C++ binaries for running inference (not training) on NVidia and AMD GPUs. Various optimizations mentioned a lot, so presumably models run faster than with just running them via regular PyTorch.