---

name: cli-config description: "Routes DI-engine CLI, config, and launch-parameter workflows." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0 disable-model-invocation: true

CLI and config

Use this sub-skill for the package entry points, config compilation, and the lightweight launch wrappers that sit in front of the training/runtime code.

Owns

• ding and ditask command-line usage
• ding.config helpers: read_config, read_config_directly, read_config_with_system, save_config, save_config_py, compile_config, and compile_config_parallel
• config file shape: main_config, create_config, and the optional system_config used by distributed launch paths
• predefined config lookup through ding.entry.predefined_config
• generic launch wrappers such as ding -m serial, ding -m parallel, ding -m dist, and ding -m eval

Does not own

• The detailed serial/offline/eval loops themselves; those live in serial-pipelines
• Task/middleware examples and multi-process routing internals; those live in framework-runtime
• Env-wrapper behavior and environment shape debugging; those live in env-integration

Read this first when the user asks

• how to choose a ding mode or ditask flag
• why a config file fails to load or compile
• how create_config and main_config are supposed to pair up
• how to run the repo-provided serial or parallel wrappers from a prepared environment
• how to query the registry or choose a predefined CartPole/Pendulum config

Workflow

1. Decide whether the request is about the CLI surface or the config surface.
2. Use references/cli-reference.md for commands, flags, and launch modes.
3. Use references/config-reference.md for config object structure and compile/save behavior.
4. Use scripts/cli_smoke.py when you want a bundled check for CLI help output plus config compilation in the installed environment.
5. If the request is really about training logic or env behavior, route to the neighboring sub-skill instead of growing this one.

Common decision points

• Prefer ding -m serial for the legacy single-process pipeline family.
• Prefer ding -m parallel when the workflow is using Parallel-based message routing and multiple workers on one machine.
• Prefer ding -m dist only when the config contains distributed system fields and the user explicitly wants coordinator/collector/learner launch control.
• Prefer ditask when the user is working with the task router or the modern multi-process framework entry model.

Helpful bundle links

• references/cli-reference.md for CLI flags, modes, and launch semantics.
• references/config-reference.md for config object shape and compile rules.
• references/troubleshooting.md for common parse, mode, and registry errors.
• scripts/cli_smoke.py for a small check that the installed package still parses the CLI and compiles a representative config.