The “Failed to initialize NVML: Driver/library version mismatch?” error generally means the CUDA Driver is still running an older release that is incompatible with the CUDA toolkit version currently in use. Rebooting the compute nodes will generally resolve this issue.
If you do not wish to reboot the compute node, you will need to remove the existing Nvidia kernel module and load the new module.
On the compute node:
Remove the existing Nvidia kernel module:
# modprobe -r nvidia nvidia_uvm
Reload the systemd units:
Build and load the new kernel module:
systemctl restart cuda-driver
If the old Nvidia Kernel module is still loading, you may need to delete the module from the software image and node. You can check this with:
node001$ find /lib/modules | grep nvidia
headnode1$ find /cm/images/default-image/lib/modules | grep nvidia