Similar experience here. NVidia is pretending to become more open but their software engineers really cannot handle their hardware. I finally bought a non-NVidia device which happened to become my daily driver. As soon as I go back to ML, my plan is to buy a powerful desktop computer and install stable a stable version of cuda etc there, and then execute everything via ssh from my laptop. Might even use my Linode to forward the port so I can work remotely from anywhere.