Bug#1094440: transition: xnnpack and onednn for PyTorch 2.6

[Shengqi Chen <harry@debian.org>]

Hi,

> 2025年2月2日 11:00，Emilio Pozuelo Monfort <pochu@debian.org> 写道：
> 
> Yes, I noticed that xnnpack in experimental depends on pytorch/experimental.

I think this should be the reversed way (pytorch depends on xnnpack)?

> But pytorch is causing autopkgtest regressions on a couple of rdeps:

Thanks & noticed.

* pytorch-cuda is just a non-free variant which would be uploaded
  after pytorch.
* pytorch-vision itself has autopkgtest issues and is not in testing
  now.

> Also, have you test rebuilt rdeps against the new packages?

@lumin has run rate on them:

For onednn, xnnpack:

* onnxruntime: FTBFS solved by new 1.20.1+dfsg-1~exp1

For pytorch{,-cuda}:

* pytorch-{cluster,scatter,vision,ignite}, baler: builds without issue
* pytorch-sparse: failed, in testing, @lumin is looking into it
* skorch, pytorch-{audio, geometric}, python-array-api-compat: failed, but not in testing

Thanks,
Shengqi Chen