Serving

Models logged with the autogluon flavor include a python_function flavor, so all standard MLflow deployment targets work.

Local REST server

mlflow models serve -m "models:/my-model/1" -p 5001

Score with the standard /invocations payloads:

curl -s http://localhost:5001/invocations \
  -H "Content-Type: application/json" \
  -d '{
        "dataframe_split": {
          "columns": ["num_a", "num_b", "cat_a"],
          "data": [[0.12, -1.3, "x"], [0.5, 0.7, "z"]]
        }
      }'

{"predictions": [1, 0]}

Class probabilities over REST

Pass the predict_method inference param in the payload:

curl -s http://localhost:5001/invocations \
  -H "Content-Type: application/json" \
  -d '{
        "dataframe_split": {
          "columns": ["num_a", "num_b", "cat_a"],
          "data": [[0.12, -1.3, "x"]]
        },
        "params": {"predict_method": "predict_proba"}
      }'

{"predictions": [{"0": 0.31, "1": 0.69}]}

Batch scoring from the CLI

mlflow models predict \
  -m runs:/<run_id>/model \
  -i input.json \
  -o predictions.json \
  --content-type json \
  --env-manager local

Environment reproduction

The logged model pins autogluon.tabular to the training version in requirements.txt. With the default --env-manager virtualenv, MLflow rebuilds that environment before serving; --env-manager local reuses the current one.

Tip

The test suite exercises this exact path: tests/test_serving.py runs mlflow models predict against a saved model, including the predict_proba param route, on every CI build.