Fixed: "FutureWarning: mlflow.models.utils.validate_serving_input is deprecated since 3.13.0"
I checked a logged model's serving input with mlflow.models.validate_serving_input, and MLflow answered with the predictions and a FutureWarning: ``mlflow.models.utils.validate_serving_input`` is deprecated since 3.13.0. This method will be removed in a future release. Use ``mlflow.models.predict`` instead. The warning names its replacement, and mlflow.models.predict does a stronger check: it predicts in a fresh environment built from the model's own requirements.
What I Was Doing
Before deploying a model, I wanted to make sure it accepts a request in the format MLflow's REST API uses. On MLflow 3.18.0, I turned three rows of input into a serving request with convert_input_example_to_serving_input and passed it to validate_serving_input, for a linear regression trained on scikit-learn's diabetes data:
import sys
import mlflow
from mlflow.models import convert_input_example_to_serving_input, validate_serving_input
from sklearn.datasets import load_diabetes
mlflow.set_tracking_uri("http://127.0.0.1:8080")
X, _ = load_diabetes(return_X_y=True, as_frame=True)
serving_input = convert_input_example_to_serving_input(X.head(3))
print(validate_serving_input(sys.argv[1], serving_input))
uv run python validate_old.py models:/m-4bc61ef4ed934f89b2764c65df636f69
What I Expected
The three predictions, with no warnings.
What I Got
The predictions, after a warning:
/Users/you/source/mlflow-demo/validate_old.py:11: FutureWarning: ``mlflow.models.utils.validate_serving_input`` is deprecated since 3.13.0. This method will be removed in a future release. Use ``mlflow.models.predict`` instead.
print(validate_serving_input(sys.argv[1], serving_input))
[206.11667725 68.07103297 176.88279035]

What I Checked
MLflow's reference for mlflow.models carries the same warning on validate_serving_input, and adds: "This API is deprecated. Use mlflow.models.predict() instead, which validates the input example by running prediction in an isolated environment (e.g. with env_manager="uv") that closely mirrors the serving environment."
MLflow's guide to validating models before deployment uses mlflow.models.predict and doesn't mention validate_serving_input.
Why It Happens
MLflow 3.13.0 deprecated validate_serving_input in favor of mlflow.models.predict. The old function still works in 3.18.0, which is why it printed the predictions, but every call warns, and a future release removes it.
The Fix
Call mlflow.models.predict instead. It takes the input as it is, so the conversion to a serving request goes too, and it predicts in a new environment built by uv:
import sys
import mlflow
from sklearn.datasets import load_diabetes
mlflow.set_tracking_uri("http://127.0.0.1:8080")
X, _ = load_diabetes(return_X_y=True, as_frame=True)
mlflow.models.predict(model_uri=sys.argv[1], input_data=X.head(3), env_manager="uv")
After the Fix
uv run python fresh_env.py models:/m-4bc61ef4ed934f89b2764c65df636f69
2026/11/21 14:31:30 INFO mlflow.models.flavor_backend_registry: Selected backend for flavor 'python_function'
2026/11/21 14:31:31 INFO mlflow.utils.virtualenv: Creating a new environment in /var/folders/…/T/tmpqupj33l7/envs/virtualenv_envs/mlflow-7ce281718f4d3c0008005aff97ca470a54f7fb7a with python version 3.15.2 using uv
...
{"predictions": [206.11667724510568, 68.07103297306874, 176.882790351053]}

No warning, and the same three predictions, this time from an environment that holds only the model's recorded requirements. The ... stands for the environment's setup and the list of installed packages. How to Test an MLflow Model in a Fresh Environment walks through that output, and why the first call takes a minute.
Common Mistakes
| Mistake | What happens | Do this instead |
|---|---|---|
Silencing the FutureWarning | The code breaks when a release removes the function | Move to mlflow.models.predict now |
env_manager="local" to make it faster | The test uses your environment, not the model's | env_manager="uv" |
Troubleshooting
| Symptom | Likely cause | Fix |
|---|---|---|
mlflow.models.predict waits a minute after installing | Python compiles the new environment's packages on first import | Expected for a fresh environment |
FAQ
Can I still pass a serving request to predict?
Yes: input_path= takes a JSON file, such as the model's own serving_input_example.json, with content_type="json".
Why is the new check better?
validate_serving_input predicted in your current environment; mlflow.models.predict with env_manager="uv" predicts in one built from the model's requirements, so it also catches a missing package.
Conclusion
validate_serving_input still returned predictions in MLflow 3.18.0, but warned that it's deprecated since 3.13.0. mlflow.models.predict(..., env_manager="uv") returned the same predictions without the warning, and checked the model's environment too.
Additional Resources
- mlflow.models —
predict, and the deprecation note onvalidate_serving_input. - MLflow Models: Validate Models before Deployment — MLflow's guide to
mlflow.models.predict.
