An error occurred while loading data using torch_geometric.datasets.Planetoid

[LinkJam0]

When using torch_geometry.datasets.Planetoid to load data on a Mac, an error message Could not run 'pyg::index_sort' with arguments from the 'CPU' backend. The error message states that pyg::index_sort 'is only available for these backups: [CUDA, MPS, Meta, BackendSelect, Python, FuncTorchDynamicLayerBackMode, Functionalize, Named, Conjugate, Negative, ZeroTens: Or, ADInplaceOrView, AutogradOther, AutogradCPU, AutogradCUDA, AutogradXLA, AutogradMPS, AutogradXPU, AutogradHPU, AutogradLazy, AutogradMeta, Tracer, AutocastCPU, AutocastCUDA, FuncTorchBatched, BatchedNestedTensor, FuncTorchVmapMode, Batched, VmapMode, FuncTorchGradWrapper, Python TLSSnap, FuncTorchDynamicLayerFrontMode, PreDisp [attach, Python Dispatcher]

After debugging the source code, I found that there was an issue with the if check in the function definition of index_sort(). This is because when loading the dataset, CPU mode is used by default, and the value of is_cpu should be true. In this case, torch.sort() should be used instead of torch.ops.pyg.index_sort(). This error will not appear on devices using CUDA, but will cause errors in dataset loading on devices using CPU by default.

[rusty1s]

I don't think that's correct. This seems to be more of a problem with a failed pyg-lib installation (and not correctly treating this downstream in PyG). index_sort in pyg-lib only has a CPU implementation, and so we only call it in case the tensor sits on CPU.