It would be fantastic if TensorRT support is added to TorchSharp directly (at least on the CUDA variants). TensorRT performance (due to kernel optimizations) beats regular PyTorch inference performance by magnitudes. It would be great to see this implemented.

https://github.com/NVIDIA-AI-IOT/torch2trt