Bug#1107726: unblock: pytorch{,-cuda}/2.6.0+dfsg-8 (pre-approval)
On 2025-06-13 16:28:38 -0400, M. Zhou wrote:
> On Fri, 2025-06-13 at 19:20 +0200, Santiago Vila wrote:
> > >
> > > We need CUDA >= 12.4 to fix the src:pytorch-cuda FTBFS #1105066,
> >
> > Hi. Not a release manager, but am I right to think that if we
> > do nothing and just wait for CUDA 12.4 to enter testing, the bug
> > would be solved as well?
>
> You are right. If pytorch-cuda were in main section, that would
> be true. But it is contrib with non-free dependencies.
>
> > Is there a reason what it really needs a rebuild?
> > (For example, being statically linked)
>
> We can ask the release team to schedule binNMU, and then the buildd
> will refuse to build them because CUDA is from non-free section.
> The build process will also pull nvidia-cudnn from internet. Even
> if we toggle XS-Autobuild: Yes and allow the CUDA dependency for
> buildd, the no-internet restriction during sbuild will fail the
> installation of another key dependency nvidia-cudnn[1].
>
> The only way to update the package is manually rebuild the binary
> packages for every single supported architecture, and upload the
> source and binary packages.
No, you can also upload maintainer-built rebuilds. If there is no source
change at all and only a rebuild, this would be preferred. But please do
not forget to file an unblock request for the rebuilds as they also will
not migrate automatically.
Cheers
--
Sebastian Ramacher
Reply to: