CUDA 11.2 + Driver 460.xx

[Andreas Beckmann]

On 1/23/21 2:46 PM, Graham Inggs wrote:
> I've tested upgrading an Ubuntu 21.04 (hirsute) am64 system, with all
> of the above packages installed, to nvidia-cuda-toolkit 11.2.0-1 built
> in my PPA and didn't find any issues.

Good, so the version number magic is working ;-)

I also tried to rebuild the rdepends in sid amd64 (and some arm64) and
only had failures in the already failing and rc-buggy eztrace-contrib
and pyhst2.

> On Mon, 18 Jan 2021 at 19:31, Andreas Beckmann <anbe at debian.org> wrote:
>> 11.2 is on its way to NEW/experimental, but it might be a bit late for
>> still getting this into bullseye, or do you have a good reason why we
>> definitively should have 11.2 instead of 11.1.1 in bullseye?
> 
> Let's turn the question around.  Seeing this isn't a transition, is
> there any good reason not to get 11.2 into bullseye?

I already saw that there was no transition tracker created and thought
the same ;-)

We only need need NEW processing for nvidia-graphics-drivers-tesla-460
and nvidia-settings-tesla-460 (and a binNMU for the latter one, while
I'll have to reupload the driver with ppc64el source included).
Can you try to ask ftp-master for quick processing?

And then we can probably kick tesla-450 out of bullseye.

> I'm currently waiting for 11.1.1-4 to migrate in Ubuntu.  The latest
> packages took several retries on each architecture before they built.
> They failed without logs and I suspect the time taken to download the
> source packages is becoming an issue.  It seems to take more retries
> every release.  Once 11.1.1-4 has migrated, I will consider a sync of
> 11.2.0-1 from experimental.

It has some useless alternative dependencies on 450-driver (>= 460), but
that's only cosmetic, will be fixed in -2 to sid.


There is this "Forward-Compatible Upgrade Path" where you can use
nvidia.ko from an older driver and libcuda.so from a newer driver in
order to support newer cuda toolkits without upgrading the kernel
module. We don't support this by the strict dependencies in our
packages, but I'm not sure whether someone would be interested in that.

https://docs.nvidia.com/deploy/cuda-compatibility/index.html

Andreas