Re: Backporting torch to Ubuntu 20.04
On Sun, 2023-09-10 at 18:39 +0200, Christian Kastner wrote:
> On 2023-09-10 16:28, M. Zhou wrote:
> > On Sat, 2023-09-09 at 22:45 -0700, Dima Kogan wrote:
> > > Are there known
> > > backports already available somewhere?
> >
> > No as far as I know. I do not have plan to backport to debian
> > stable
> > neither.
>
> I intend to backport the ROCm stack to Debian stable eventually (once
> we're 5.6 or 5.7 across the board), and I'd be happy to also cover
> pytorch and any necessary dependencies.
>
> I've got a bit of tooling set up to automate the process, and of
> course
> intend to use our CI to cover this.
Well, I have a debian stable machine with nvidia gpu which can serve
testing the backported package. If people request that, I can slowly
backport the dependency tree to stable, since it is not expected
to pose extra patching work for me.
The current 1.13.1 in stable is not that old though. 2.0.1 still has
some minor issues to fix in order to reach a backportable state.
I uploaded pytorch-cuda 2.0.1 to NEW with an minor issue of loading
the nvfuser library, because I thought it may take a long time to
pass the NEW queue -- but ftp-masters accepted it instantly!
Backporting the torch dependencies are easy. But I'll really need
some time to think of dh_pytorch before pushing multiple variants
into stable backports.
Reply to: