Max Berrendorf

Hey. The issue is that `ERMLPInteraction` has internal parameters whose size depends on the embedding dimension. Thus, the interaction's `embedding_dim` must coincide with the entity representations' dimension. The short-term fix...

Somewhere here, I guess? https://github.com/mberr/torch-max-mem/blob/fb9d91e5e5e9d566fe83410798fea8521098afb8/.github/workflows/cruft.yml#L35-L40

> I was thinking I could check the element sizes with footprint = x_batch.nelement() * x_batch.element_size(), but this doesn't seem to match up to the error, so I guess I'm...

Hm, looks like the segfault still occurs. This discussion might be related: https://discuss.pytorch.org/t/segmentation-fault-with-pytorch-2-3/203381

Re-enabling tests on `macos-latest` does not seem to work 🤔 https://github.com/mberr/torch-max-mem/actions/runs/14681583536/job/41204893694?pr=19

Hey 👋 Instead of providing the `edge_index`, you can also directly pass the adjacency matrix. The latter is not further normalized, but is checked by `torch_ppr.utils.validate_adjacency`. For creating such matrices...