With the current way CUDA version is detected, it is not possible to detect the version from build system and only fall back to a specific version if detection fails (i.e. on CI, or when a user wants to work on the code without having to install CUDA).

I'm not 100% sure what the best way to handle this is, but there should be some kind of way to gracefully fall back after failing to detect the version through NVCC. There's no good way to handle this downstream, so this would need to be implemented in cudarc itself.