luml-mlflow

MLflow plugin that redirects MLflow tracking to a LUML orbit. Training code keeps calling the standard MLflow API, such as mlflow.start_run, mlflow.log_param, mlflow.log_metric and mlflow.sklearn.log_model. The plugin records each run in a local store and, when the run finishes, uploads it to the orbit as a LUML collection. No LUML SDK calls are needed in the training script.

The plugin registers a tracking store, an artifact repository and a placeholder model-registry store for the luml:// URI scheme. MLflow discovers them automatically.

Requires Python 3.12 or later and MLflow 3.1 or later.

Installation

pip install luml-mlflow

Uploading runs needs a LUML API key, created in the LUML web interface under account settings. The plugin reads it from LUML_API_KEY, then ~/.luml.json, then the system keyring where Flow stores it. Logging to the local store works without a key.

export LUML_API_KEY="luml_your_api_key"

Usage

Set the tracking URI to an organization and an orbit, by name or id. It is the only LUML-specific line:

import mlflow
import mlflow.sklearn
from sklearn.ensemble import RandomForestClassifier

mlflow.set_tracking_uri("luml://my-org/my-orbit")
mlflow.set_experiment("churn_prediction")

with mlflow.start_run(run_name="random_forest_v1"):
    mlflow.log_param("n_estimators", 100)
    model = RandomForestClassifier(n_estimators=100).fit(X, y)
    mlflow.log_metric("accuracy", model.score(X, y))
    mlflow.sklearn.log_model(model, name="churn_model")

When the run finishes, it is uploaded to a collection named after the experiment. A logged model becomes a LUML model in the .luml format, other artifacts become attachments, and parameters, metrics and tags go into an experiment snapshot. The upload is best-effort: a failure is recorded on the run and never crashes the training script.

luml://local keeps runs on disk without uploading them. Runs are staged in the same SQLite store that Flow reads (~/.luml/experiments, set by LUML_BACKEND_STORE_URI), so they show up in lumlflow ui as well.

Configuration

• LUML_MLFLOW_AUTOSYNC=0 turns off the automatic upload.
• LUML_MLFLOW_UPLOAD_MODE=separate uploads each model and the experiment snapshot as separate artifacts. The default, auto, embeds the snapshot in the model when a run logged exactly one model.

Command line

luml-mlflow uploads runs on demand, for example runs logged with autosync off or runs whose upload failed:

luml-mlflow sync --run <run_id>
luml-mlflow sync --experiment churn_prediction
luml-mlflow sync --all
luml-mlflow status --run <run_id>

Already uploaded runs are skipped unless --force is passed.

Documentation

Logging MLflow runs to LUML