CLI Reference¶

OpenWAM provides six console commands for configuration, training, and inference.

Stable Commands¶

Command	Purpose
openwam-train	Train from an experiment YAML
openwam-eval	Offline evaluation from experiment or eval YAML
openwam-inspect-config	Load and print typed config
openwam-validate-config	Static YAML validation without model imports
openwam-sanity	Check loading, training, evaluation, and rollout-style inference
openwam-sim-rollout	Run a registered simulator adapter in closed loop

All six commands work from an installed package. Use --help for their options. The uv run examples below assume a source checkout; after a pip installation with the relevant extras, invoke the commands directly without that prefix.

For complete train, full-state resume, offline evaluation, and benchmark rollout examples, use Training and Inference. It is the complete usage guide.

Command Examples¶

Installed extensions load before experiment construction. Repeat --extension module[:hook] when a config uses out-of-tree dataset, policy, or decoder registrations:

uv run --extra train openwam-train \
  --extension acme_open_wam \
  --cfg /path/to/acme_joint.yaml

uv run openwam-validate-config configs/examples/public_tiny_synthetic_contract.yaml

openwam-validate-config checks authored experiment and evaluation YAML. Do not use it to lint checkpoint-generated resolved_config.yaml files: those artifacts serialize typed defaults, including fields intentionally omitted from an authored method config. Load such files through the runtime checkpoint path, which applies checkpoint compatibility when required. For a current-schema artifact, openwam-inspect-config can display the typed configuration without treating it as authored YAML.

uv run --extra train openwam-train \
  --cfg configs/examples/public_tiny_synthetic_contract.yaml \
  --save-root runs/public-tiny \
  --disable-wandb

uv run --extra eval openwam-eval \
  --cfg configs/evals/public_tiny_synthetic_contract.yaml \
  --checkpoint runs/public-tiny/checkpoints/checkpoint_step_1/model_state.pt \
  --device cpu \
  --max-batches 1 \
  --output-json outputs/eval.json \
  --provenance-mode full

These two commands form the data-free train/eval example. Benchmark configs require the local artifacts and resources documented in Training and Inference.

uv run --extra sim openwam-sim-rollout \
  --cfg configs/examples/calvin_npz_raw7_sanity.yaml \
  --benchmark calvin \
  --max-steps 10 \
  --zero-policy

Checkpoint loading is strict by default. --allow-partial-checkpoint permits missing or unexpected model keys only for an intentional migration diagnostic; results produced under that opt-in should not be reported as normal evals.

Third-party simulator adapters use --extension, an application-owned --benchmark identifier, and repeatable --sim-option KEY=VALUE values. See the simulator adapter cookbook.

Source-Checkout Utilities¶

Model conversion, dataset generation, research diagnostics, and specialized artifact visualization can require a source checkout. In particular, the Wan conversion scripts and scripts/generate_video_only_rollout.py are not installed console commands and are not included in release distributions. Their guides label them as source-checkout integrations. They may compose package APIs, but they are not stable package entrypoints.