Matthias Fey

Likely no, since this will result in a lot of issues regarding compilation - if we require `torch` in `setup.py`, it will install PyTorch with CUDA 10.2 which might not...

Well, that‘s because transformers is Python only - it does not need to compile CUDA kernels.

Are you using an older version of `torch-cluster`? In the current version, we no longer make use of `scipy` for this, see https://github.com/rusty1s/pytorch_cluster/blob/master/torch_cluster/radius.py.

Yes, our GPU implementation is non-deterministic while the CPU one is deterministic (it will pick the closest points). Do you think this is a problem? I am happy to add...

Sounds like a great idea!

Seems like I missed this issue, really sorry. Batch-wise computation is done through the `batch` vectors, which assigns each point to a single example.

Thanks for reporting. I need to look into this. Any help or reference is highly appreciated as I'm pretty unfamiliar with CUDA graphs.

For CPU, we are currently using `scipy.cluster.vq`, while you can also make use of `scipy.spatial.KDTree` to maintain the information about the kd-tree. Adding kd-tree CUDA support is something that I...

Also pinging @mrjel here who promised me some time ago to look into this ;)

Thanks for the pointer. Pinging @mrjel here as well as he always shared interest in adding this feature to `torch-cluster` :)