Re: plan of deep learning team for next stable release?
Hi Steffen,
On Tue, Dec 01, 2020 at 07:54:39PM +0100, Steffen Möller wrote:
> Dear non-free autobuilders,
>
> On debian-ai we are wondering about how to tag packages that need
> Debian's cuda packages installed to build, i.e. for their runtime tests.
> That said, is possibly the mere addition of a cuda package as build
> dependency sufficient to point to a particular subset of buildds?
The problem is not to prepare nvidia-cuda-toolkit on the buildd for
*building*
https://salsa.debian.org/science-team/caffe-contrib/-/blob/master/debian/control#L54
but having nvidia-driver and a real nvidia card while *testing*.
> The number of packages in need of a CUDA-savvy buildd I expect to be
> less than 20 in 2020 from debian-ai and also for Debian Med I only have
> about two packages in mind that would make immediate use of CUDA.
>
> Since packages needing a non-free autobuild need to register anyway,
> maybe there is not too much overhead for a support of CUDA on the
> non-free buildd? I would be prepared to help with a machine if that is a
> blocker.
Nvidia team has already addressed the autobuild issue. In my fuzzy
memory only two steps are required to build cuda version of software
(in the contrib section) on our buildd: 1) b-d on cuda and set
XS-Auto-Build: yes ; 2) register the package in the whitelist by
mailing wanna-build. See devref for detail.
Reply to: