On 2015-08-21 12:28, Luca Boccassi wrote: > On 21 August 2015 at 11:15, François Legendre <f.legen...@u-pec.fr> wrote: >> I thought IMHO that the maintainers of nvidia-cuda-toolkit should set >> a dependency : nvidia-cuda-toolkit depends on nvidia-driver as nouveau >> drivers are not able to run a CUDA application.
Thinking about headless systems, it's not nvidia-driver that is needed, but just libcuda1 + the kernel module + blacklist nouveau + load the nvidia module. > I think it's good that there is a clear separation between the "build" > stack and the "runtime" stack, as it is often the case. Ack. > The toolkit dependency chain already has a "suggests" on the > nvidia-driver package (there might even be a "recommends" along the > way, can't remember), nvidia-cuda-toolkit Depends: nvidia-cuda-dev nvidia-cuda-dev Depends: libcuda1 libcuda1 Recommends: nvidia-kernel-dkms (= $VER) | nvidia-kernel-$VER So in a default installation which has --install-recommends enabled the nvidia kernel module should be there. The only bug I see is that such an installation without nvidia-driver (or just libgl1-nvidia-glx or xserver-xorg-video-nvidia) currently does not activate the nvidia alternative, but that is easily fixed in glx-alternative-nvidia. Andreas
