[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

NVIDIA drivers version

Hi everyone!

I've just subscribed to this mailing list (looong time since the last time
I was subscribed to any Debian ml) to ask about NVIDIA version policies
and Debian and Ubuntu packages. Sorry if this is not the best place to
ask about such a thing :-?

Ok, I'll run the risk! 

The point is: I'm using CUDA and some NVIDIA deep learning stuff on top
of it, all of it in a docker instance. The docker image is based on 
CoCalc [1][2], project built using Ubuntu 18.04 as base distro. The host
machine is a Debian Buster/Sid/Experimental system.

So far, I managed to keep the same NVIDIA Driver on both the host and
the docker image, something apparently necessary to have all the stuff
running properly in the nvidia-docker envinronment. Specifically, the
version I'm using is 410.104.

This version (410.104) could be easily installed in both the host
(Debian) and the guest (Ubuntu) using deb packages in the official
repos. And NVIDIA also provides all the stuff I need for this version in
convenient deb packages (CUDA 10.0, NCCL, CUDNN, nvinfer, etc.) here:

So far, so good :-)

Now, I want to update part of the software stack and also the NVIDIA
driver, to use the version that will be the definitive in Buster
(418.74, I guess). Unfortunatelly, the 418.x drivers in the NVIDIA repos
I need for some of the stuff (the mentioned [3] and [4]) are not 418.74,
neither any of the versions I can found packaged in Debian/Ubuntu.
Specifically, in those repos I can get stuff for 418.39, 418.40 and
418.67 :-?

Isn't that weird? :-? Why are they using those versions and
Debian/Ubuntu are packaging others? (after reading some stuff in Debian I
assume the versions being packaged are a kind of long-support versions,

Can you think in an alternative that allows me to stick with the deb
packages setup (yeah, I know I can solve the problem if I avoid the deb
packages and use other installation methods, but I will like to keep it
that way if possible).


[1] https://github.com/sagemathinc/cocalc
[2] https://github.com/emilioj/CuCalc
[3] https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64
[4] http://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64

Reply to: