Per this answer,
on JetPack 5 [CUDA/cuDNN/TensorRT/etc] are installed inside the containers themselves.
Previously on L4T r32.7.3, I was installing the packages cuda-toolkit-10-2 libcudnn8 nvidia-container-csv-cuda nvidia-container-csv-cudnn nvidia-l4t-cuda to my custom rootfs.
Now on r35.1, if the only programs that use CUDA and cuDNN are executed inside containers, do I need to install anything to the host (Jetson device)? My interpretation of the above quote is that the answer is “no”, but I don’t want to miss something like a kernel module that’s needed on the container host.