Matthias Fey

Can you clarify what you wan to do exactly?

`torch-sparse` can handle more than 2 dimensions (but only the first two can be sparse). Would that be sufficient in your case?

These should be automatically linked by PyTorch via the `CPUExtension`, see [here](https://github.com/rusty1s/pytorch_sparse/blob/master/setup.py#L117). Wondering why this is not the case for you.

There exists a few edge cases where METIS sadly silently just crashes :( In your case, you need to make sure that your `SparseTensor` is correctly defined, and holds a...

I think too small graphs is one edge case. The other one is that METIS expects to operate on undirected graphs.

Broadcasting is indeed not supported. For addition, we currently just assume the output size is `max(size_x, size_y)`.

I can reproduce this, but this is not an issue on the Python level and it looks like this is coming from PyTorch CSR@CSR matmul.

Currently, PyG's SchNet expects an input feature vector of shape `[num_atoms]`, while it looks that your input is two-dimensional.

Which version of `pylibcugraphops` did you have installed?

Yes, PyG `DataLoader` is just a wrapper around `torch.utils.data.DataLoader` with a custom `collate_fn`. As such, this is the only argument that cannot be overridden. Let me clarify this in the...