Skip to content

cli_reference.md

Latest commit

 

History

History
148 lines (109 loc) · 4.02 KB

cli_reference.md

File metadata and controls

148 lines (109 loc) · 4.02 KB
Raw

CLI Reference

So what: the repo now has a real CLI surface for core TEVO runs, TrainRecipe export/render, and autoresearch transfer workflows. This page is the fast map of the commands you are most likely to need.

evo-loop CLI

The package installs an evo-loop CLI with these top-level command families:

evo-loop --help
  • Core search:
    • run
    • resume-state
    • frontier
    • export-seed
  • TrainRecipe bridge:
    • train-recipe-export
    • train-recipe-render
  • Transfer workflows:
    • autoresearch-at-home-handoff
    • cuda-transfer-prepare
    • cuda-transfer-benchmark
    • cuda-transfer-report
    • mlx-transfer-prepare
    • mlx-transfer-benchmark
    • mlx-transfer-audit
    • mlx-transfer-continuation-summary
    • mlx-transfer-report
  • Maintenance:
    • prune-checkpoints
    • cleanup-run
    • convert-checkpoints
    • cache
    • version

Core TEVO Run

Run an evolutionary search directly:

evo-loop run configs/live_smoke.yaml \
  --device cpu \
  --generations 3 \
  --steps 40 \
  --eval-batches 2 \
  --seed 0

Inspecting Results

evo-loop frontier runs/<run>/frontier.json

evo-loop export-seed runs/<run>/frontier.json \
  <candidate_id> \
  configs/seed_winner.yaml

Use the reporting scripts when you want more context than the CLI summary:

python scripts/report_motifs.py runs/<run>/frontier.json \
  --lineage runs/<run>/frontier_lineage.json --top 15

TrainRecipe Bridge

evo-loop train-recipe-export runs/<run>/frontier.json \
  --candidate-id <candidate_id> \
  --out artifacts/train_recipes/<candidate_id>.yaml

evo-loop train-recipe-render artifacts/train_recipes/<candidate_id>.yaml \
  --backend autoresearch_cuda \
  --train-py /path/to/autoresearch/train.py

evo-loop train-recipe-render artifacts/train_recipes/<candidate_id>.yaml \
  --backend autoresearch_at_home_cuda \
  --train-py /path/to/autoresearch-at-home/train.py

See train_recipe_bridge.md for the compatibility rules and projection behavior.

autoresearch@home Handoff

Use this when you already have one TEVO candidate you want to contribute into the collaborative autoresearch@home loop:

evo-loop autoresearch-at-home-handoff \
  --frontier runs/<run>/frontier.json \
  --candidate-id <candidate_id> \
  --run-root runs/at_home_handoff

Useful options:

  • --autoresearch-repo /path/to/autoresearch-at-home to stage against an existing local checkout
  • --autoresearch-repo-url <git_url> to override the default repo URL
  • --autoresearch-ref <git_ref> to pin a specific branch or commit

The command writes a staged repo/ workspace with the chosen candidate already installed as train.py. See autoresearch_at_home_handoff.md for the workflow and collaboration boundary.

CUDA Transfer Workflow

evo-loop cuda-transfer-prepare \
  --run-root runs/cuda_transfer_demo \
  --config configs/exp_train_recipe_bridge_owt_10m_v1.yaml \
  --modal-gpu A10G \
  --autoresearch-flavor upstream

evo-loop cuda-transfer-benchmark runs/cuda_transfer_demo \
  --repeat 3 \
  --timeout-minutes 10

evo-loop cuda-transfer-report runs/cuda_transfer_demo

See cuda_transfer_demo.md and motif_transfer_demo.md for the recommended public-facing CUDA path.

MLX Transfer Workflow

evo-loop mlx-transfer-prepare /path/to/autoresearch-mlx \
  --run-root runs/mlx_transfer_demo \
  --config configs/exp_train_recipe_bridge_owt_10m_v1.yaml

evo-loop mlx-transfer-benchmark runs/mlx_transfer_demo

evo-loop mlx-transfer-report runs/mlx_transfer_demo

See mlx_transfer_demo.md for the full flow.

Disk Hygiene

Runs can accumulate checkpoints quickly. Useful cleanup tools:

evo-loop cleanup-run runs/<run_dir>/frontier.manifest.json --apply

evo-loop prune-checkpoints runs/<run_dir>/checkpoints \
  --state-path runs/<run_dir>/frontier.state.json

evo-loop convert-checkpoints runs/<run_dir>/checkpoints --dtype fp16 --apply