load_nvidia_uvm module

A script to load the nvidia uvm device

Some (older) Linux Operating systems do not load /dev/nvidia-uvm on boot to runlevel 3 (headless). This results in the nvidia-uvm module not being loaded.

Unfortunately, a simple modprobe does not fix the issue, but a CUDA call on the host (not in a container) will.

This scripts attempts to locate a libcudart.so library and calls the cudaGetDeviceCount function, which loads the /dev/nvidia-uvm driver. If it cannot locate the cuda runtime, you can give it the location as an argument.

The CUDA Runtime is required on the host.

load_nvidia_uvm.find_cudart(search_dirs)[source]

load_nvidia_uvm.main(args=[])[source]