Conda packaging: env failure

[dominiquesydow]

Conda packaging for 2023.05.2 fails due to

The reported errors are:
- Encountered problems while solving:
   - nothing provides __cuda needed by pytorch-1.13.0-cuda112py310he33e0d6_200

---

conda_build.exceptions.DependencyNeedsBuildingError: Unsatisfiable dependencies for platform linux-64: {MatchSpec("pytorch==1.13.0=cuda112py310he33e0d6_200"), MatchSpec("__cuda")}

Full logs: https://dev.azure.com/conda-forge/feedstock-builds/_build/results?buildId=717588&view=logs&jobId=656edd35-690f-5c53-9ba3-09c10d0bea97&j=656edd35-690f-5c53-9ba3-09c10d0bea97&t=e5c8ab1d-8ff9-5cae-b332-e15ae582ed2d

Ideas

• Currently, we install pytorch from the pytorch channel. conda-forge states: "As a general rule: all dependencies have to be packaged by conda-forge as well. This is necessary to assure ABI compatibility for all our packages." --- can we install the package from conda-forge channel instead? https://anaconda.org/conda-forge/pytorch

[dominiquesydow]

I am trying pytorch from the conda-forge channel for our teachopencadd conda package here: https://dev.azure.com/conda-forge/feedstock-builds/_build/results?buildId=717595&view=logs&j=656edd35-690f-5c53-9ba3-09c10d0bea97

-> This seems to work for conda-forge, their CI is passing.

[dominiquesydow]

Pinging @AndreaVolkamer and @mbackenkoehler, could you please check with the DL team if the switch of channels still works fine for the DL talktorials?

[AndreaVolkamer]

@PaulaKramer @joschka-gross @Old-Shatterhand can you comment on this?

[Old-Shatterhand]

That's an interesting error.

Intuitively, I would have said, it doesn't matter. To be on the safe side, adjust the CI/CD pipeline on GitHub and test it.

But then I checked the actual conda-packages on their respective channels: conda-forge and PyTorch. They are not just copies of each other (as it is the case for RDKit, afaik). The conda-forge version has no support for win-64 and is one minor version behind the PyTorch channel. Also, the conda-forge builds are way smaller than the PyTorch builds. So, I'd say, the packages are not identical (as it might be the case for RDKit), but I doubt that they differ substantially and will brake the notebooks.

TL;DR: I'd update the CI/CD pipeline to conda-forge packages of PyTorch and rerun it. It will not work for Windows (as there is just no win-64 build of PyTorch on conda-forge), but I think on general Linux/MaxOS, this should not change the behavior of the notebooks.

[AndreaVolkamer]

adressed in https://github.com/volkamerlab/teachopencadd/pull/394

[AndreaVolkamer]

We will continue to work on this, but will merge the post-release fixes for now.