Compact minibatches while preserving global node identity

Assign each required global node a local index in a batch table, then rewrite every sampled edge endpoint through that map. Keep the inverse mapping for feature fetches, labels and output interpretation. PyTorch Geometric 2.7.0 NeighborLoader exposes n_id for global node identities and distinguishes the seed-node batch size from the full sampled context. Code that assumes every local node has a training target can accidentally supervise context nodes or compare predictions with labels in the wrong order.

A global identifier includes its namespace and, for heterogeneous graphs, its node type. User 17 and item 17 need not identify the same entity. Graph partitioning can also relabel nodes, so a partition-local ID is not automatically a dataset-global ID. Test identity round trips and edge endpoints using deliberately nonconsecutive IDs. Sequential toy IDs can allow a mistaken array-position assumption to survive until production data introduces larger offsets or a different node ordering.

A source shared by two destinations still contributes to both edges. A repeated replacement draw can require repeated contribution or an equivalent weight even if its feature row is stored once. During backpropagation, gradients from every use must accumulate into the appropriate source representation or embedding row. Collapsing edge records merely because one endpoint repeats changes the graph operator. If parallel edges have distinct relation attributes or times, preserve those distinctions as well as their endpoint identities.

Representations at different layers are different tensors even when their node IDs agree. Shared input storage does not imply the same hidden vector can replace both h at layer one and h at layer two. A minibatch can use bipartite computation blocks with separate source and destination index sets, so remapping must follow each block’s contract. Similarly, two uses of a node under different temporal cutoffs or feature versions cannot safely share a computed state unless the model and version rules prove that state is the same.

After compaction, verify that output positions map back to the intended seeds before computing loss or serving responses. Preserve edge sampling probabilities and any normalization data through reordering. Measure occurrences, unique feature rows, sampled edges and computation-block sizes separately; one reduced count does not establish a smaller peak allocation. Use a fixed sampled trace to compare compacted and unshared implementations numerically. Small floating-point differences from reduction order can be acceptable under a declared tolerance, while missing contributions are a semantic failure.

The cache lab demonstrates sharing only at an abstract address level. Reset and set both C=4 and s=4; the first sixteen requests contain four unique addresses, yielding four misses and twelve hits. Step adds a seventeenth request and predicts one further hit. The stream uses N=min(4096,16+floor(2t)) and begins empty. A hit avoids another abstract fetch but does not erase that request’s logical use. This LRU instrument has no graph edges, gradients, global-to-local maps or layer-specific tensors.