Matthias Fey

Can you try to install via ``` git clone https://github.com/rusty1s/pytorch_scatter.git python setup.py install ```

We would have to bring https://github.com/rusty1s/pytorch_cluster/pull/208 also to `torch-scatter`.

This is only relevant for the C++ API. Since you are using Python, `pip install` is sufficient.

This is currently expected, since the custom ops by `torch-scatter` are not supported in `torch.compile`. There exists two options: * Disallow the use of `compile` for certain ops * Fallback...

For this, we added [`utils.scatter`](https://pytorch-geometric.readthedocs.io/en/latest/modules/utils.html#torch_geometric.utils.scatter), which will pick up the best computation path depending on your input arguments. Also works with `torch.compile`.

Yes, if you want `torch.compile` support, then this is the recommended way.

That's strange. We increased the mac OS version we use to build the wheels which might explain this. I am very confused the universal builds work with ARM at all.

How did you install PyTorch? Are you using the `conda-forge` package?

No, that sounds good then. Is this ARM or Intel mac?

Got it. What happens if you simply install via ``` pip uninstall torch-scatter pip install torch-scatter ```