Can't access to cuda on containers using usernamespace mapping

Hello there I’m trying to improve our project containerisation by using user namespaces. The problem is one of the container using cuda and it used to work just fine with runtime: nvidia but when I add the “userns-remap”:“username” to /etc/docker/daemon.json and restart docker or reboot the device it fails to find cuda inside the container. I have tried mounting /dev/nvidia* resources too but to no avail. If anyone knows how to achieve this guidance would be much appreciated. Thank you.

Device is jetson orin nano with jetpack 5.1.2. The said container uses dustynv/l4t-pytorch:r35.4.1 as base image and was working fine until usernamespace mapping.


When you add the namespace mapping, do you also launch the container with nvidia runtime?
More, fail to find cuda indicating no cuda files/libraries or GPU not working?