Environment variables¶

The pipeline keeps personal paths and credentials out of committed YAML. Set runtime locations in the shell that invokes the orchestrator; generated jobs receive the MLflow variables they need.

Operator-set variables¶

Variable	Used by	Default/requirement
`MLFLOW_TRACKING_URI`	`lan-sbatch`, MLflow clients, `lan-publish`	`lan-sbatch` defaults to `sqlite:///mlflow.db`; set explicitly for publication and cluster work
`MLFLOW_ARTIFACT_LOCATION`	Experiment creation and publication records	Optional MLflow-managed location; use an absolute shared path on a cluster
`INSPECT_ONNX`	marimo network inspector	Required; path to the candidate ONNX
`INSPECT_MODEL`	marimo network inspector	`ddm`

Hugging Face authentication is handled by huggingface_hub. Authenticate on the operator machine using its supported credential mechanism; never put a token in pipeline YAML, a generated script, or documentation.

Documentation convention¶

The guides use LAN_PIPELINE_ROOT as a shell convenience:

export LAN_PIPELINE_ROOT="/oscar/data/frankmj/$USER/proj_hssm_pipeline"

No Python code reads this name. The shell expands it in --output-path, --training-data-folder, and MLflow variables before the CLI runs. You may use another name or pass literal paths.

Variables injected into jobs¶

When MLflow initialization succeeds and the command is not --script-only, the generated script exports:

Variable	Value
`MLFLOW_TRACKING_URI`	The configured URI; a relative SQLite URI is converted to an absolute path
`MLFLOW_ARTIFACT_LOCATION`	The configured artifact root, when set
`MLFLOW_EXPERIMENT_NAME`	`{model}-data-generation` or `{model}-training`

Generation also receives a run-name argument containing $SLURM_ARRAY_TASK_ID; a fanned-out submission adds the lane index so names do not collide between arrays. Command arguments permit expansion only for their allowlisted Slurm identifiers and shell-quote all other config-derived values. The modules list is a deliberate exception: each entry is emitted verbatim as module load <entry>. Treat module names as administrator-controlled cluster configuration, never as untrusted or user-submitted text.

Path rules¶

The submit host and compute nodes must resolve the same MLflow tracking and artifact locations.
Prefer an absolute SQLite URI for shared work. sqlite:////path/to/file.db has four slashes because the final slash begins an absolute POSIX path.
Keep Slurm output roots free of whitespace; scheduler directives cannot quote those paths safely.
Keep large data, artifacts, and tracking stores off small home volumes.
The generated script changes into the checkout that produced it, then adds $HOME/.local/bin to PATH to find a standard uv installation.

Precedence¶

MLflow-aware downstream commands generally resolve an explicit CLI argument before an environment variable and then a package default. The orchestrator's scheduling resources follow a separate YAML/CLI precedence documented in the configuration reference.