mlflow/mlflow · error · MlflowException

Unable to load model metadata. Ensure the source path of the

Error message

Unable to load model metadata. Ensure the source path of the model being registered points to a valid MLflow model directory (see https://mlflow.org/docs/latest/models.html#storage-format) containing a model signature (https://mlflow.org/docs/latest/models.html#model-signature) specifying both input and output type specifications.

What it means

Before creating or validating a model version in Unity Catalog, MLflow loads the MLmodel metadata from the model's source directory and re-raises any failure as this MlflowException. The source must be a valid MLflow model directory whose MLmodel file is loadable and carries a signature with both input and output specs. The original exception is chained ('from e').

Source

Thrown at mlflow/store/_unity_catalog/registry/rest_store.py:244

def _raise_unsupported_method(method, message=None):
    messages = [
        f"Method '{method}' is unsupported for models in the Unity Catalog.",
    ]
    if message is not None:
        messages.append(message)
    raise MlflowException(" ".join(messages))


def _load_model(local_model_dir):
    # Import Model here instead of in the top level, to avoid circular import; the
    # mlflow.models.model module imports from MLflow tracking, which triggers an import of
    # this file during store registry initialization
    from mlflow.models.model import Model

    try:
        return Model.load(local_model_dir)
    except Exception as e:
        raise MlflowException(
            "Unable to load model metadata. Ensure the source path of the model "
            "being registered points to a valid MLflow model directory "
            "(see https://mlflow.org/docs/latest/models.html#storage-format) containing a "
            "model signature (https://mlflow.org/docs/latest/models.html#model-signature) "
            "specifying both input and output type specifications."
        ) from e


def get_feature_dependencies(model_dir):
    """
    Gets the features which a model depends on. This functionality is only implemented on
    Databricks. In OSS mlflow, the dependencies are always empty ("").
    """
    model = _load_model(model_dir)
    if (
        model.flavors.get("python_function", {}).get("loader_module")
        == mlflow.models.model._DATABRICKS_FS_LOADER_MODULE
    ):

View on GitHub (pinned to 6a27f2decc)

Solutions

1. Log the model with a flavor API (e.g. mlflow.sklearn.log_model) including infer_signature(input, output) so MLmodel and signature exist
2. Verify the source path points to a directory containing a valid MLmodel file (inspect it locally after downloading artifacts)
3. If registering from run artifacts, pass the correct runs:/<run_id>/<artifact_path> source rather than a custom path

Example fix

// before
with mlflow.start_run():
    mlflow.log_artifact("model.pkl", "model")
    client.create_model_version("m", "runs:/<run>/model")
// after
with mlflow.start_run():
    mlflow.sklearn.log_model(sk_model, "model", signature=infer_signature(X, preds))
    client.create_model_version("m", f"runs:/{run.info.run_id}/model")

import os
from mlflow.models import Model

def validate_model_source(model_dir):
    mlmodel = os.path.join(model_dir, "MLmodel")
    assert os.path.isfile(mlmodel), f"No MLmodel at {model_dir}"
    m = Model.load(model_dir)
    assert m.signature and m.signature.inputs and m.signature.outputs, "Model must have input and output signature"

try:
    client.create_model_version(name, source)
except MlflowException as e:
    if "Unable to load model metadata" in str(e):
        raise RuntimeError(f"Re-log the model with a flavor API and infer_signature; bad source: {source}") from e
    raise

When it happens

Trigger: Registering a model (or querying feature/model-version dependencies) in UC where the artifact source is not a valid MLflow model directory, the MLmodel file is missing/corrupt, or the signature is absent/incomplete.

Common situations: Registering raw artifacts (e.g. a bare pickle or checkpoint dir) instead of an mlflow.*.log_model output; manually edited or truncated MLmodel files; logging models without an explicit or inferred signature.

Related errors

• UC Model Versions gathered through search_model_versions do
• UC Model Versions gathered through search_model_versions do
• UC Registered Models gathered through search_registered_mode
• UC Registered Models gathered through search_registered_mode
• Argument '{arg_name}' is unsupported for models in the Unity

AI-assisted analysis of mlflow/mlflow@6a27f2decc (2026-08-29). Data as JSON: /api/errors/ddcece9fcb2b4e2d. Report an issue: GitHub.