Skip to content

Latest commit

 

History

History
35 lines (25 loc) · 2.18 KB

File metadata and controls

35 lines (25 loc) · 2.18 KB

CLI implementation (rengu)

Entry points

Script Module
rengu rengu_flow.cli:main
rengu-flow (deprecated) rengu_flow.cli:main
rengu-flow-ui (deprecated) rengu_flow_ui.cli:main — separate entry point; prefer rengu ui …

Repo-root rengu runs uv sync if needed, then exec .venv/bin/rengu (requires uv on PATH).

Layout

Module Role
rengu_flow/cli/main.py Dispatcher, Linux guard, legacy rengu --config
rengu_flow/cli/init_cmd.py init = local TOML + uv sync
rengu_flow/cli/update_cmd.py update = uv sync only
rengu_flow/cli/train_cmd.py train, validate, cache, dump-dataset
rengu_flow/cli/ui_cmd.py ui start / serve / dev / build / reset-db
rengu_flow/cli/train_launcher.py DeepSpeed argv + [training.env] merge
rengu_flow/config/local_config.py Parse rengu.local.toml, apply UI env
rengu_flow/install_profiles.py Profile names → uv sync --extra
rengu_flow/main.py parse_args / run_prepared — still used by DeepSpeed --module rengu_flow.main

Training launcher

rengu_flow_ui/jobs.py imports base_train_command and training_subprocess_env from train_launcher.py (wrapping the former in its own build_train_command) so UI jobs and rengu train share defaults from rengu.local.toml.

DeepSpeed subprocesses use --module rengu_flow.main (DeepSpeed's launcher flag for a module target — not -m, which it rejects) so the CLI does not need to be on PATH inside workers.

Extending install profiles

Add an entry to PROFILE_EXTRAS and labels in rengu_flow/install_profiles.py. Maintenance UI reads DEP_PROFILES built from the same helpers.