mlflow/mlflow · error · RuntimeError

Unknown parameter search model type {type(param_search_model

Error message

Unknown parameter search model type {type(param_search_model)}.

What it means

MLflow's PySpark autologging extracts validation metrics from parameter-search results, supporting CrossValidatorModel and TrainValidationSplitModel. If the fitted model is neither of these types, MLflow cannot find validationMetrics and raises this RuntimeError.

Source

Thrown at mlflow/pyspark/ml/__init__.py:513

    - For TrainValidationSplitModel, the result dict contains metrics for each param map.

    `best_index` is the best index of trials.
    """
    from pyspark.ml.tuning import CrossValidatorModel, TrainValidationSplitModel

    metrics_dict = {}

    metric_key = param_search_estimator.getEvaluator().getMetricName()
    if isinstance(param_search_model, CrossValidatorModel):
        avg_metrics = param_search_model.avgMetrics
        metrics_dict["avg_" + metric_key] = avg_metrics
        if hasattr(param_search_model, "stdMetrics"):
            metrics_dict["std_" + metric_key] = param_search_model.stdMetrics
    elif isinstance(param_search_model, TrainValidationSplitModel):
        avg_metrics = param_search_model.validationMetrics
        metrics_dict[metric_key] = avg_metrics
    else:
        raise RuntimeError(f"Unknown parameter search model type {type(param_search_model)}.")

    if param_search_estimator.getEvaluator().isLargerBetter():
        best_index = np.argmax(avg_metrics)
    else:
        best_index = np.argmin(avg_metrics)

    return metrics_dict, best_index


def _log_estimator_params(param_map):
    # Chunk model parameters to avoid hitting the log_batch API limit
    for chunk in _chunk_dict(param_map, chunk_size=MAX_PARAMS_TAGS_PER_BATCH):
        truncated = _truncate_dict(chunk, MAX_ENTITY_KEY_LENGTH, MAX_PARAM_VAL_LENGTH)
        mlflow.log_params(truncated)


class _AutologgingMetricsManager:
    """

View on GitHub (pinned to 6a27f2decc)

Solutions

1. Log the CrossValidatorModel/TrainValidationSplitModel itself, not its .bestModel.
2. Confirm the tuning estimator is pyspark.ml.tuning.CrossValidator or TrainValidationSplit.
3. Unwrap any custom tuner to the underlying Spark model before autolog logging.
4. Upgrade MLflow if a newer Spark tuner type is the cause.

Example fix

# before
best = cv_model.bestModel
mlflow.pyspark.ml.log_posttraining_metadata(...)  # expects tuner model, got best estimator
# after
mlflow.pyspark.ml.log_posttraining_metadata(...)  # pass cv_model (the tuner model)

from pyspark.ml.tuning import CrossValidatorModel, TrainValidationSplitModel
if not isinstance(tuner_model, (CrossValidatorModel, TrainValidationSplitModel)):
    raise TypeError(f"Expected tuner model, got {type(tuner_model)}")

from pyspark.ml.tuning import CrossValidatorModel, TrainValidationSplitModel
def is_param_search_model(m) -> bool:
    return isinstance(m, (CrossValidatorModel, TrainValidationSplitModel))

try:
    log_tuning_metadata(model)
except RuntimeError as e:
    if "Unknown parameter search model type" in str(e):
        log_tuning_metadata(model._original_tuner_model)

When it happens

Trigger: Passing a fitted model to _log_posttraining_metadata / _create_child_runs_for_parameter_search that is not a CrossValidatorModel or TrainValidationSplitModel, e.g. a custom tuner, a wrapped model, or a tune step whose result was unwrapped to the best model.

Common situations: Autologging custom hyperparameter-tuning wrappers; calling .bestModel and logging the result as if it were the tuner model; Spark version changes introducing new tuner classes not yet supported.

Related errors

• INVALID_PARAMETER_VALUE
• Pipeline stages should be iterable, but found object {origin
• Tuning params should not include params not owned by the tun
• Failed to set up OpenTelemetry instrumentation for Agno V2.
• Unknown content type: {content_type['type']}. Please make su

AI-assisted analysis of mlflow/mlflow@6a27f2decc (2026-08-29). Data as JSON: /api/errors/7583941916ebafcc. Report an issue: GitHub.