We are happy to announce the availability of MLflow 1.14.0!

In addition to bug and documentation fixes, MLflow 1.14.0 includes the following features and improvements:

• MLflow's model inference APIs (mlflow.pyfunc.predict), built-in model serving tools (mlflow models serve), and model signatures now support tensor inputs. In particular, MLflow now provides built-in support for scoring PyTorch, TensorFlow, Keras, ONNX, and Gluon models with tensor inputs. For more information, see https://mlflow.org/docs/latest/models.html#deploy-mlflow-models (#3808, #3894, #4084, #4068 @wentinghu; #4041 @tomasatdatabricks, #4099, @arjundc-db)
• Add new mlflow.shap.log_explainer, mlflow.shap.load_explainer APIs for logging and loading shap.Explainer instances (#3989, @vivekchettiar)
• The MLflow Python client is now available with a reduced dependency set via the mlflow-skinny PyPI package (#4049, @eedeleon)
• Add new RequestHeaderProvider plugin interface for passing custom request headers with REST API requests made by the MLflow Python client (#4042, @jimmyxu-db)
• mlflow.keras.log_model now saves models in the TensorFlow SavedModel format by default instead of the older Keras H5 format (#4043, @harupy)
• mlflow_log_model now supports logging MLeap models in R (#3819, @yitao-li)
• Add mlflow.pytorch.log_state_dict, mlflow.pytorch.load_state_dict for logging and loading PyTorch state dicts (#3705, @shrinath-suresh)
• mlflow gc can now garbage-collect artifacts stored in S3 (#3958, @sklingel)

For a comprehensive list of changes, see the release change log, and check out the latest documentation on mlflow.org.