mlflow/mlflow · error · MlflowException

INVALID_PARAMETER_VALUE

INVALID_PARAMETER_VALUE

Error message

Dependencies schema should be set only once to the callback.

What it means

MLflow's DSPy callback stores a single dependencies schema used to enrich traces. Calling set_dependencies_schema() more than once on the same callback instance would silently overwrite tracing context, so MLflow raises INVALID_PARAMETER_VALUE on the second call.

Source

Thrown at mlflow/dspy/callback.py:76

        # call_id: (LiveSpan, OTel token)
        self._call_id_to_span: dict[str, SpanWithToken] = {}
        self._call_id_to_module: dict[str, Any] = {}

        ###### state management for optimization process ######
        # The current callback logic assumes there is no optimization running in parallel.
        # The state management may not work when multiple optimizations are running in parallel.
        # optimizer_stack_level is used to determine if the callback is called within compile
        # we cannot use boolean flag because the callback can be nested
        self.optimizer_stack_level = 0
        # call_id: (key, step)
        self._call_id_to_metric_key: dict[str, tuple[str, int]] = {}
        self._evaluation_counter = defaultdict(int)
        self._disabled_eval_call_ids = set()
        self._eval_runs_started: set[str] = set()

    def set_dependencies_schema(self, dependencies_schema: dict[str, Any]):
        if self._dependencies_schema:
            raise MlflowException(
                "Dependencies schema should be set only once to the callback.",
                error_code=MlflowException.INVALID_PARAMETER_VALUE,
            )
        self._dependencies_schema = dependencies_schema

    @skip_if_trace_disabled
    def on_module_start(self, call_id: str, instance: Any, inputs: dict[str, Any]):
        span_type = self._get_span_type_for_module(instance)
        attributes = self._get_span_attribute_for_module(instance)

        # The __call__ method of dspy.Module has a signature of (self, *args, **kwargs),
        # while all built-in modules only accepts keyword arguments. To avoid recording
        # empty "args" key in the inputs, we remove it if it's empty.
        if "args" in inputs and not inputs["args"]:
            inputs.pop("args")

        self._start_span(
            call_id,

View on GitHub (pinned to 6a27f2decc)

Solutions

1. Call set_dependencies_schema only once per callback/tracer lifecycle (e.g. at app startup)
2. Check whether the schema is already set before calling, or use a module-level 'initialized' flag
3. Create a fresh callback instance if a new schema truly must be installed
4. Remove duplicate invocations in per-request code paths

Example fix

// before
def handler(req):
    mlflow.dspy.set_dependencies_schema(schema)  # called every request
// after
init_schema():
    mlflow.dspy.set_dependencies_schema(schema)
def handler(req):
    pass  # schema already installed at startup

_schema_installed = False

def install_schema(schema):
    global _schema_installed
    if _schema_installed:
        return
    mlflow.dspy.set_dependencies_schema(schema)
    _schema_installed = True

try:
    callback.set_dependencies_schema(schema)
except MlflowException as e:
    if 'only once' in str(e):
        logger.debug('schema already set; skipping')
    else:
        raise

When it happens

Trigger: Invoking mlflow.dspy.set_dependencies_schema (via _set_dependency_schema_to_tracer) twice within one trace/session — e.g. calling it in both an initialization routine and a per-request handler with the same tracer/callback.

Common situations: A web app calls it at module import and again per request; a test suite reuses a tracer fixture across tests; framework autotuning re-invokes setup code on hot reload.

Related errors

• An MLflow Tracking URI is required, please provide the track
• An MLflow experiment ID is required, please provide the expe
• The MLflow Tracing client is not configured. Please call ini
• INVALID_PARAMETER_VALUE
• No deployments target has been set. Please either set the ML

AI-assisted analysis of mlflow/mlflow@6a27f2decc (2026-08-29). Data as JSON: /api/errors/a4290263cc7caf48. Report an issue: GitHub.