---

name: distributed-tools description: "Plan safe DGL distributed graph partitioning, launch commands, DistGraph usage, and partition metadata validation." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

DGL Distributed Tools

Use this sub-skill when a task needs DGL distributed partitioning, DistGraph runtime setup, partition metadata validation, ip_config.txt, launch-command construction, or GraphBolt distributed partitioning notes.

Start Here

• Read references/api-reference.md for dgl.distributed.initialize, DistGraph, distributed dataloaders, node_split/edge_split, partition concepts, and launcher argument semantics.
• Read references/workflows.md for partition preflight, safe launch-command construction, trainer script changes, GraphBolt partition handoffs, and skip conditions.
• Read references/data-formats.md for ip_config.txt, partition config JSON essentials, part-* entries, and canonical edge-type string rules.
• Read references/troubleshooting.md when distributed setup fails due to shared workspace, SSH assumptions, relative paths, process-count mismatches, metadata errors, heterograph edge-type ambiguity, or cluster limits.
• Run python scripts/check_partition_config.py --part-config PATH to inspect partition metadata without loading partition graphs or mutating files.
• Run python scripts/dgl_distributed_command_builder.py --help to build a dry, printable tools/launch.py-style command; the script never SSHes or launches jobs.

Route Elsewhere

• Single-machine sampling, dgl.dataloading.DataLoader, and GraphBolt ItemSampler pipelines belong in dataloading-graphbolt.
• Model layers, GraphConv/SAGEConv/GATConv, losses, and non-distributed training loops belong in message-passing-training.
• Raw CSV dataset layout, DGLDataset, graph serialization, and dataset cache design belong in datasets-and-io.
• Homogeneous/heterogeneous graph construction and canonical edge-type basics belong in graph-apis.

Safety Notes

• Do not run launchers, SSH, rsync, dispatch data, or start cluster processes unless the user explicitly confirms the target cluster, shared workspace, and failure/cleanup plan.
• Prefer workspace-relative --part_config and --ip_config values for launch commands; validate them before constructing commands.
• Stop and ask for cluster details when network ports, passwordless SSH, shared storage, GPU/CUDA availability, or multi-host process counts are unknown.