Run the CI-backed local workflow¶

The quickest meaningful check is local_test_run.sh. GitHub Actions runs this same script after the unit tests, so the tutorial stays tied to an executable integration path rather than a separate set of hand-maintained commands.

What you will run¶

The quick workflow:

installs and invokes ssm-simulators to generate a small DDM dataset;
retrieves the data-generation experiment ID from MLflow;
invokes LANfactory's JAX trainer with that experiment ID;
writes local MLflow metadata and artifacts for inspection.

It takes a few minutes on a typical development machine. The first dependency sync takes longer because ssm-simulators includes compiled extensions.

Install the tracked environment¶

From the repository root:

uv sync --locked

On macOS, install the native prerequisites first:

brew install gsl libomp

On Debian or Ubuntu, install GSL before syncing:

sudo apt-get update
sudo apt-get install libgsl-dev

Run the workflow¶

bash local_test_run.sh

The tutorial output is disposable

The script removes and recreates local_test_data/ in the current working directory. Run it from the repository root and do not place your own files in that directory. It also creates mlflow.db and mlflow_artifacts/.

A successful run ends with === Test Complete ===. Inspect these outputs:

Output	What it demonstrates
`local_test_data/`	Generated training data and the trained network artifacts
`mlflow.db`	Generation and training metadata share one tracking store
`mlflow_artifacts/`	Run artifacts are separate from tracking metadata

Inspect the lineage in MLflow¶

Start the UI against the same tracking database:

uv run mlflow ui --backend-store-uri sqlite:///mlflow.db

Open the MLflow UI at localhost port 5000. Find the {model}-data-generation experiment and the {model}-training experiment. The training run should carry the data_generation_experiment_id that points back to the generation experiment.

Stop the UI with Ctrl+C when you are done.

Exercise the production entry point without Slurm¶

CI also renders both job types without submitting them. Do the same for data generation:

uv run python sbatch_scripts/gen_sbatch.py generate \
  --config-path configs/quick_test/data_generation.yaml \
  --output-path /tmp/lan-pipeline-generate \
  --n-jobs-in-array 2 \
  --script-only

The command prints one JSON object. Its sbatch_script field names the generated file under /tmp/lan-pipeline-generate/runs/. Read that file before moving to a real cluster; --script-only does not submit a job or create an MLflow experiment.

Check the repository gates¶

uv run pytest tests/ -q
uv run ruff check .

The unit tests cover resource resolution, script generation, the validation gate, and publish staging. Together with local_test_run.sh, these are the same core checks used by the repository's CI workflow.

Next¶

Continue with the operator guides to configure personal cluster lanes and shared MLflow storage. Keep the quick-test configs for rehearsal; copy an example config and review its scale before scheduling production work.