---

name: dataloading-graphbolt description: "Build stochastic minibatch, DGL dataloading, and GraphBolt pipelines for DGL training." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

DGL Dataloading and GraphBolt

Use this sub-skill when a task needs stochastic minibatch training, neighbor sampling, block/MFG handoff, GraphBolt datapipes, ItemSet/ItemSampler, OnDiskDataset task sets, feature fetching, or GPU/UVA dataloader choices.

Start Here

• Read references/api-reference.md for classic dgl.dataloading and dgl.graphbolt constructors, outputs, and object contracts.
• Read references/workflows.md for node/link neighbor sampling, DataLoader vs GraphBolt decisions, feature prefetch/fetch, and GPU/UVA setup.
• Read references/data-formats.md when preparing or validating GraphBolt metadata.yaml and OnDiskDataset task splits.
• Read references/troubleshooting.md for device mismatches, CUDA worker rules, GraphBolt native-library issues, feature-key errors, and empty seed batches.
• Run python scripts/graphbolt_pipeline_sanity.py --help for available CPU-safe checks.

Route Elsewhere

• Full GNN layer implementation, block-compatible model forward() methods, losses, and training loops belong in ../message-passing-training/.
• Raw CSV validation, DGLDataset, CSVDataset, graph serialization, and initial OnDiskDataset file preparation belong in ../datasets-and-io/.
• Graph construction, heterograph canonical edge types, dgl.to_block() details, batching, and graph transforms belong in ../graph-apis/.
• Cluster launch, graph partitioning, distributed sampling/training, and DDP process orchestration belong in ../distributed-tools/.

Safety Defaults

• Prefer tiny CPU synthetic graphs and local metadata validation for smoke checks; do not download datasets by default.
• Use GraphBolt for pipelined CPU/pinned-memory large-graph workflows, but use classic dgl.dataloading.DataLoader for GPU neighbor sampling and UVA sampling.
• Keep scripts and examples independent from external repositories; pass user data paths explicitly.