I’m writing commercial software using CUDA for acceleration. I want it to run as broadly as possible, of course. I have a number of potential customers running Debian; I see that there’s no official NV CUDA package for any Debian release. I don’t expect my customers to start building drivers from source, as has been recommended a few places on this forum. Are there any other options for delivering commercial CUDA software on Linux?
One possibility I’m thinking about is the customers should only need to install the latest NVIDIA graphics driver (but not the CUDA toolkit), and I could package libcudart.so with my software, and install it in the appropriate place. Would that stand any chance of working?
I’m not familiar with debian/apt and friends, but if I wanted to this on Gentoo, I’d pack up the stuff with my own ebuild install (called apt-package or some such I guess in deb format), and have the package depend on the nvidia-driver and cuda-toolkit. I assume debians system will allow you to do the same.