I’d like to hear a resolution to this problem, too, because I have a similar problem in Linux. My production machines have GPUs in them, but the machine building the code does not. Since the driver installer refuses to run (and the driver package includes libcuda.so) I can’t build on my build system.
I worked around it by copying libcuda.so from another machine. This is not an ideal situation. Can libcuda please be included in the toolkit as well as the driver package?