Build both cpu and gpu binaries

[kkleidal]

This way, a docker image containing the GPU package will still run when on a CPU-only machine, as might be the case during testing. This is the same behavior as the pytorch package.

[rusty1s]

Thanks, I will try to merge by the end of the year.

[kkleidal]

Hi @rusty1s . Just checking in on this. Let me know if there is anything I can do to help 😄

[rusty1s]

I will try to merge this soon. Sorry for the delay. Finally have time to look into it.

[rusty1s]

Looks like there is indeed no elegant way to allow this behavior without building separate libraries. What a bummer. One solution would be to separate CPU and CUDA functionality in C++ code, but this results in a lot of duplicated code, and we would need to call the correct method in Python dependent on whether we are operating on CUDA or CPU tensors.
Interestingly, torchvision does not allow CUDA wheels on CPU-only machines either.

Nevertheless, I pushed some experimental changes to torch-spline-conv to see how those affect build times in our CI. In case they are there reasonable, I will merge this in.

[rusty1s]

I already updated torch-spline-conv and torch-cluster with the proposed changes. This one is next :) Wheels for torch-scatter will be released soon. Thank you very much!

[kkleidal]

Thank you! 😃

[rusty1s]

Done! Please confirm that it is working as expected.

[kkleidal]

It's working as expected! Thank you! It feels good to get rid of the private fork and use the official release 😄