Build cuda without cuda?

[AntoinePrv]

Hello,

I have build my project, and resolve its dependencies when building a Singularity image (think Docker without sudo rights for executing).
I need sudo rights to build the image, so I do it on my machine, without GPU, and run the image on a cluster with GPU (there is a special flag to expose the GPU driver).

When I build my image, I take the PyTorch version with GPU, even though I don't have any when I do the install. This does not give me any problem and I can run GPU code when I deploy.

Now, resolving pytorch_scatter (et. al.), this resolves to False and the Cuda extension is not built. I get at run-time:

/usr/lib/python3.7/site-packages/torch_scatter/utils/ext.py:5: in <module>
    import scatter_cuda
E   ModuleNotFoundError: No module named 'scatter_cuda'

Is it possible to change the test torch.cuda.is_available() in setup.py to something less strong, i.e. ensuring that the extension can be built but not that there is a full Cuda run-time?
I'm not even sure that's possible...

Thanks

[rusty1s]

Hi,
interesting question. I looked up the source code for the cpp_extension file and we could try to replace the call to torch.cuda.is_available() with something like _find_cuda_home(). Could you do me a favour and try if this fits your needs?

[AntoinePrv]

Sure, will let you know as soon as I have something.