Details
Description
Loading /dev/nvidia-uvm (and installing a script to make sure it loads on reboot) is not technically part of the official Nvidia driver installation process. The rationale being that CUDA applications typically load this device on-demand if they need it. Unfortunately, it can't load it if mesos hasn't made it available to the container running the CUDA application though.
We should add support to have the mesos agent auto-load this device when running the GPU isolator.