See if the host is configured properly: check if /run/opengl-driver/lib/libcuda.so really exists. If it doesn’t, check hardware.opengl.enable. You’ve already discussed the driver earlier in the thread
When you run a container, you still need to explicitly expose the gpus. E.g. if you use docker run, you 'll need to pass something like --gpus all
As was mentioned, do describe what you run. E.g. you could upload the logs that feature the commands you’re running and the output (the exact errors) they produce