---

name: "configuration-and-extension" description: "Guides MMGeneration config authoring, dataset layout, registry extension, and runtime customization workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Configuration and Extension

Use this sub-skill when you need to read, edit, or extend MMGeneration configs rather than run a full training or evaluation job.

Typical triggers

• "How do I change this config?"
• "How do I register a custom loss or hook?"
• "What dataset layout does this config expect?"
• "How do I print the resolved config?"
• "How do I add a new generator or optimizer?"

Include here

• Config inheritance, _base_, _delete_, and custom_imports
• --cfg-options overrides and config printing
• Dataset layout and pipeline keys
• Registry patterns for models, modules, losses, optimizers, and hooks
• Runtime customizations: LR schedules, momentum schedules, checkpoint/log/eval hooks
• Safe config-only validation before training or evaluation

Exclude here

• Actually launching training jobs -> training-and-distribution
• Running metrics or sampling jobs -> evaluation-and-metrics or inference-and-sampling
• Latent editing and TorchServe packaging -> applications-and-deployment

Read these files first

• references/configuration.md
• references/troubleshooting.md
• ../../references/api-reference.md
• ../../references/data-formats.md
• ../../references/cli-reference.md

Bundled helper

• scripts/print_config.py — print the fully resolved config and verify --cfg-options overrides.

What good guidance looks like

A future agent should be able to:

1. Inspect a config file and see the final merged result.
2. Decide whether a dataset or pipeline shape matches the intended workflow.
3. Add a custom module or hook with the correct registry/import pattern.
4. Understand when custom_imports, _delete_, or data_info is required.
5. Distinguish config-only validation from a real training or evaluation run.

Common failure modes

• A custom class is written but never imported into the registry namespace.
• A nested override does not match the original config structure.
• Paired/unpaired dataset keys do not match the pipeline expectations.
• A loss expects a different output key than the model actually emits.
• A runtime hook or optimizer is described in config but not imported or registered.

For symptom-by-symptom recovery, read references/troubleshooting.md.