---

name: large-scale-graphs description: "Use DIG's large-scale graph stack for GraphFMOB/GraphFMIB-style workflows, METIS partitioning, subgraph loaders, feature momentum, and async graph-memory helpers." disable-model-invocation: true metadata: disco-role: operating license: GPL 3.0

Large-Scale Graphs

Use this sub-skill for DIG workflows that scale graph learning to large node-classification graphs or the GraphFMOB/GraphFMIB code paths.

Include

• dig.lsgraph.dataset: get_data, SubgraphLoader, EvalSubgraphLoader.
• dig.lsgraph.method.FM: FeatureMomentum.
• dig.lsgraph.method.GraphFMOB: AsyncIOPool, metis, permute, compute_micro_f1, gen_masks, dropout, and GraphFMOB loaders when dig_ext is available.
• examples/lsgraph/GraphFMOB style workflows and the GraphFMIB example path.
• OGB/Reddit/Flickr/Yelp/SBM-style large-graph loading patterns.

Exclude

• Molecular, 3D, SSL, explainability, GOOD, augmentation, or fairness workflows unless they are used only as supporting examples.

Start Here

• Read references/api-reference.md for the loader and memory helpers.
• Read references/workflows.md for the partition/load/train/inference pattern.
• Read references/troubleshooting.md when dig_ext or sparse extensions are missing.
• Run scripts/lsgraph_feature_momentum_smoke.py for a safe, extension-aware smoke check.

Core Workflows

• Partitioned training: use METIS partitioning, SubgraphLoader, and EvalSubgraphLoader to batch a large graph into local neighborhoods.
• Feature memory: use FeatureMomentum and AsyncIOPool to keep node embeddings or async transfer buffers aligned with the partitioned loader.
• Model evaluation: use compute_micro_f1 on the loaded data and masks after training.

Important Limitation

dig.lsgraph.dataset and the async helper code depend on a compiled dig_ext extension in the source tree. If that extension is not present, document the limitation and restrict yourself to extension-independent API guidance. FeatureMomentum can be imported directly from dig.lsgraph.method.FM, but some CPU-only PyTorch builds still fail its pinned-memory allocation and should be reported as a backend limitation rather than treated as a dataset failure.