mlflow.pyfunc

The python_function model flavor serves as a default model interface for MLflow Python models. Any MLflow Python model is expected to be loadable as a python_function model.

In addition, the mlflow.pyfunc module defines a generic filesystem format for Python models and provides utilities for saving to and loading from this format. The format is self contained in the sense that it includes all necessary information for anyone to load it and use it. Dependencies are either stored directly with the model or referenced via a Conda environment.

The mlflow.pyfunc module also defines utilities for creating custom pyfunc models using frameworks and inference logic that may not be natively included in MLflow. See Models From Code for Custom Models.

Inference API

Python function models are loaded as an instance of PyFuncModel, which is an MLflow wrapper around the model implementation and model metadata (MLmodel file). You can score the model by calling the predict() method, which has the following signature:

predict(
  model_input: [pandas.DataFrame, numpy.ndarray, scipy.sparse.(csc_matrix | csr_matrix),
  List[Any], Dict[str, Any], pyspark.sql.DataFrame]
) -> [numpy.ndarray | pandas.(Series | DataFrame) | List | Dict | pyspark.sql.DataFrame]

All PyFunc models will support pandas.DataFrame as input and PyFunc deep learning models will also support tensor inputs in the form of Dict[str, numpy.ndarray] (named tensors) and numpy.ndarrays (unnamed tensors).

Here are some examples of supported inference types, assuming we have the correct model object loaded.

Input Type

Example

pandas.DataFrame

 
import pandas as pd

x_new = pd.DataFrame(dict(x1=[1, 2, 3], x2=[4, 5, 6]))
model.predict(x_new)

numpy.ndarray

 
import numpy as np

x_new = np.array([[1, 4][2, 5], [3, 6]])
model.predict(x_new)

scipy.sparse.csc_matrix or scipy.sparse.csr_matrix

 
import scipy

x_new = scipy.sparse.csc_matrix([[1, 2, 3], [4, 5, 6]])
model.predict(x_new)

x_new = scipy.sparse.csr_matrix([[1, 2, 3], [4, 5, 6]])
model.predict(x_new)

python List

 
x_new = [[1, 4], [2, 5], [3, 6]]
model.predict(x_new)

python Dict

 
x_new = dict(x1=[1, 2, 3], x2=[4, 5, 6])
model.predict(x_new)

pyspark.sql.DataFrame

 
from pyspark.sql import SparkSession

spark = SparkSession.builder.getOrCreate()

data = [(1, 4), (2, 5), (3, 6)]  # List of tuples
x_new = spark.createDataFrame(data, ["x1", "x2"])  # Specify column name
model.predict(x_new)

Filesystem format

The Pyfunc format is defined as a directory structure containing all required data, code, and configuration:

./dst-path/
    ./MLmodel: configuration
    <code>: code packaged with the model (specified in the MLmodel file)
    <data>: data packaged with the model (specified in the MLmodel file)
    <env>: Conda environment definition (specified in the MLmodel file)

The directory structure may contain additional contents that can be referenced by the MLmodel configuration.

MLModel configuration

A Python model contains an MLmodel file in python_function format in its root with the following parameters:

  • loader_module [required]:

    Python module that can load the model. Expected as module identifier e.g. mlflow.sklearn, it will be imported using importlib.import_module. The imported module must contain a function with the following signature:

    _load_pyfunc(path: string) -> <pyfunc model implementation>

    The path argument is specified by the data parameter and may refer to a file or directory. The model implementation is expected to be an object with a predict method with the following signature:

    predict(
  model_input: [pandas.DataFrame, numpy.ndarray,
  scipy.sparse.(csc_matrix | csr_matrix), List[Any], Dict[str, Any]],
  pyspark.sql.DataFrame
) -> [numpy.ndarray | pandas.(Series | DataFrame) | List | Dict | pyspark.sql.DataFrame]
  • code [optional]:

    Relative path to a directory containing the code packaged with this model. All files and directories inside this directory are added to the Python path prior to importing the model loader.

  • data [optional]:

    Relative path to a file or directory containing model data. The path is passed to the model loader.

  • env [optional]:

    Relative path to an exported Conda environment. If present this environment should be activated prior to running the model.

  • Optionally, any additional parameters necessary for interpreting the serialized model in pyfunc format.

Example

tree example/sklearn_iris/mlruns/run1/outputs/linear-lr

├── MLmodel
├── code
│   ├── sklearn_iris.py
│
├── data
│   └── model.pkl
└── mlflow_env.yml

cat example/sklearn_iris/mlruns/run1/outputs/linear-lr/MLmodel

python_function:
  code: code
  data: data/model.pkl
  loader_module: mlflow.sklearn
  env: mlflow_env.yml
  main: sklearn_iris

Models From Code for Custom Models

Tip

MLflow 2.12.2 introduced the feature “models from code”, which greatly simplifies the process of serializing and deploying custom models through the use of script serialization. It is strongly recommended to migrate custom model implementations to this new paradigm to avoid the limitations and complexity of serializing with cloudpickle. You can learn more about models from code within the Models From Code Guide.

The section below illustrates the process of using the legacy serializer for custom Pyfunc models. Models from code will provide a far simpler experience for logging of your models.

Creating custom Pyfunc models

MLflow’s persistence modules provide convenience functions for creating models with the pyfunc flavor in a variety of machine learning frameworks (scikit-learn, Keras, Pytorch, and more); however, they do not cover every use case. For example, you may want to create an MLflow model with the pyfunc flavor using a framework that MLflow does not natively support. Alternatively, you may want to build an MLflow model that executes custom logic when evaluating queries, such as preprocessing and postprocessing routines. Therefore, mlflow.pyfunc provides utilities for creating pyfunc models from arbitrary code and model data.

The save_model() and log_model() methods are designed to support multiple workflows for creating custom pyfunc models that incorporate custom inference logic and artifacts that the logic may require.

An artifact is a file or directory, such as a serialized model or a CSV. For example, a serialized TensorFlow graph is an artifact. An MLflow model directory is also an artifact.

Workflows

save_model() and log_model() support the following workflows:

  1. Programmatically defining a new MLflow model, including its attributes and artifacts.

    Given a set of artifact URIs, save_model() and log_model() can automatically download artifacts from their URIs and create an MLflow model directory.

    In this case, you must define a Python class which inherits from PythonModel, defining predict() and, optionally, load_context(). An instance of this class is specified via the python_model parameter; it is automatically serialized and deserialized as a Python class, including all of its attributes.

  2. Interpreting pre-existing data as an MLflow model.

    If you already have a directory containing model data, save_model() and log_model() can import the data as an MLflow model. The data_path parameter specifies the local filesystem path to the directory containing model data.

    In this case, you must provide a Python module, called a loader module. The loader module defines a _load_pyfunc() method that performs the following tasks:

    • Load data from the specified data_path. For example, this process may include deserializing pickled Python objects or models or parsing CSV files.

    • Construct and return a pyfunc-compatible model wrapper. As in the first use case, this wrapper must define a predict() method that is used to evaluate queries. predict() must adhere to the Inference API.

    The loader_module parameter specifies the name of your loader module.

    For an example loader module implementation, refer to the loader module implementation in mlflow.sklearn.

Which workflow is right for my use case?

We consider the first workflow to be more user-friendly and generally recommend it for the following reasons:

  • It automatically resolves and collects specified model artifacts.

  • It automatically serializes and deserializes the python_model instance and all of its attributes, reducing the amount of user logic that is required to load the model

  • You can create Models using logic that is defined in the __main__ scope. This allows custom models to be constructed in interactive environments, such as notebooks and the Python REPL.

You may prefer the second, lower-level workflow for the following reasons:

  • Inference logic is always persisted as code, rather than a Python object. This makes logic easier to inspect and modify later.

  • If you have already collected all of your model data in a single location, the second workflow allows it to be saved in MLflow format directly, without enumerating constituent artifacts.

Function-based Model vs Class-based Model

When creating custom PyFunc models, you can choose between two different interfaces: a function-based model and a class-based model. In short, a function-based model is simply a python function that does not take additional params. The class-based model, on the other hand, is subclass of PythonModel that supports several required and optional methods. If your use case is simple and fits within a single predict function, a function-based approach is recommended. If you need more power, such as custom serialization, custom data processing, or to override additional methods, you should use the class-based implementation.

Before looking at code examples, it’s important to note that both methods are serialized via cloudpickle. cloudpickle can serialize Python functions, lambda functions, and locally defined classes and functions inside other functions. This makes cloudpickle especially useful for parallel and distributed computing where code objects need to be sent over network to execute on remote workers, which is a common deployment paradigm for MLflow.

That said, cloudpickle has some limitations.

  • Environment Dependency: cloudpickle does not capture the full execution environment, so in MLflow we must pass pip_requirements, extra_pip_requirements, or an input_example, the latter of which is used to infer environment dependencies. For more, refer to the model dependency docs.

  • Object Support: cloudpickle does not serialize objects outside of the Python data model. Some relevant examples include raw files and database connections. If your program depends on these, be sure to log ways to reference these objects along with your model.

Function-based Model

If you’re looking to serialize a simple python function without additional dependent methods, you can simply log a predict method via the keyword argument python_model.

Note

Function-based model only supports a function with a single input argument. If you would like to pass more arguments or additional inference parameters, please use the class-based model below.

import mlflow
import pandas as pd


# Define a simple function to log
def predict(model_input):
    return model_input.apply(lambda x: x * 2)


# Save the function as a model
with mlflow.start_run():
    mlflow.pyfunc.log_model("model", python_model=predict, pip_requirements=["pandas"])
    run_id = mlflow.active_run().info.run_id

# Load the model from the tracking server and perform inference
model = mlflow.pyfunc.load_model(f"runs:/{run_id}/model")
x_new = pd.Series([1, 2, 3])

prediction = model.predict(x_new)
print(prediction)

Class-based Model

If you’re looking to serialize a more complex object, for instance a class that handles preprocessing, complex prediction logic, or custom serialization, you should subclass the PythonModel class. MLflow has tutorials on building custom PyFunc models, as shown here, so instead of duplicating that information, in this example we’ll recreate the above functionality to highlight the differences. Note that this PythonModel implementation is overly complex and we would recommend using the functional-based Model instead for this simple case.

import mlflow
import pandas as pd


class MyModel(mlflow.pyfunc.PythonModel):
    def predict(self, context, model_input, params=None):
        return [x * 2 for x in model_input]


# Save the function as a model
with mlflow.start_run():
    mlflow.pyfunc.log_model(
        "model", python_model=MyModel(), pip_requirements=["pandas"]
    )
    run_id = mlflow.active_run().info.run_id

# Load the model from the tracking server and perform inference
model = mlflow.pyfunc.load_model(f"runs:/{run_id}/model")
x_new = pd.Series([1, 2, 3])

print(f"Prediction:
    {model.predict(x_new)}")

The primary difference between the this implementation and the function-based implementation above is that the predict method is wrapped with a class, has the self parameter, and has the params parameter that defaults to None. Note that function-based models don’t support additional params.

In summary, use the function-based Model when you have a simple function to serialize. If you need more power, use the class-based model.

class mlflow.pyfunc.EnvType[source]

Bases: object

CONDA = 'conda'
VIRTUALENV = 'virtualenv'
class mlflow.pyfunc.PyFuncModel(model_meta: mlflow.models.model.Model, model_impl: Any, predict_fn: str = 'predict', predict_stream_fn: Optional[str] = None)[source]

Bases: object

MLflow ‘python function’ model.

Wrapper around model implementation and metadata. This class is not meant to be constructed directly. Instead, instances of this class are constructed and returned from load_model().

model_impl can be any Python object that implements the Pyfunc interface, and is returned by invoking the model’s loader_module.

model_meta contains model metadata loaded from the MLmodel file.

get_raw_model()[source]

Note

Experimental: This function may change or be removed in a future release without warning.

Get the underlying raw model if the model wrapper implemented get_raw_model function.

property input_example

The input example provided when the model was saved.

property loader_module

Model’s flavor configuration

Note

Experimental: This property may change or be removed in a future release without warning.

property metadata

Model metadata.

property model_config

Model’s flavor configuration

Note

Experimental: This property may change or be removed in a future release without warning.

predict(data: Union[pandas.core.frame.DataFrame, pandas.core.series.Series, numpy.ndarray, csc_matrix, csr_matrix, List[Any], Dict[str, Any], datetime.datetime, bool, bytes, float, int, str, pyspark.sql.dataframe.DataFrame], params: Optional[dict] = None) → Union[pandas.core.frame.DataFrame, pandas.core.series.Series, numpy.ndarray, list, str, pyspark.sql.dataframe.DataFrame][source]
predict_stream(data: Union[dict, bool, bytes, float, int, str], params: Optional[dict] = None) → Iterator[Union[dict, str]][source]
unwrap_python_model()[source]

Note

Experimental: This function may change or be removed in a future release without warning.

Unwrap the underlying Python model object.

This method is useful for accessing custom model functions, while still being able to leverage the MLflow designed workflow through the predict() method.

Returns

The underlying wrapped model object

Example 
import mlflow


# define a custom model
class MyModel(mlflow.pyfunc.PythonModel):
    def predict(self, context, model_input, params=None):
        return self.my_custom_function(model_input, params)

    def my_custom_function(self, model_input, params=None):
        # do something with the model input
        return 0


some_input = 1
# save the model
with mlflow.start_run():
    model_info = mlflow.pyfunc.log_model(artifact_path="model", python_model=MyModel())

# load the model
loaded_model = mlflow.pyfunc.load_model(model_uri=model_info.model_uri)
print(type(loaded_model))  # <class 'mlflow.pyfunc.model.PyFuncModel'>
unwrapped_model = loaded_model.unwrap_python_model()
print(type(unwrapped_model))  # <class '__main__.MyModel'>

# does not work, only predict() is exposed
# print(loaded_model.my_custom_function(some_input))
print(unwrapped_model.my_custom_function(some_input))  # works
print(loaded_model.predict(some_input))  # works

# works, but None is needed for context arg
print(unwrapped_model.predict(None, some_input))
mlflow.pyfunc.add_to_model(model, loader_module, data=None, code=None, conda_env=None, python_env=None, model_config=None, model_code_path=None, **kwargs)[source]

Add a pyfunc spec to the model configuration.

Defines pyfunc configuration schema. Caller can use this to create a valid pyfunc model flavor out of an existing directory structure. For example, other model flavors can use this to specify how to use their output as a pyfunc.

Note

All paths are relative to the exported model root directory.

Parameters

  • model – Existing model.

  • loader_module – The module to be used to load the model.

  • data – Path to the model data.

  • code – Path to the code dependencies.

  • conda_env – Conda environment.

  • python_env – Python environment.

  • model_config –

    The model configuration to apply to the model. This configuration is available during model loading.

    Note

    Experimental: This parameter may change or be removed in a future release without warning.

  • model_code_path – Path to the model code.

  • kwargs – Additional key-value pairs to include in the pyfunc flavor specification. Values must be YAML-serializable.

Returns

Updated model configuration.

mlflow.pyfunc.build_model_env(model_uri, save_path)[source]

Prebuild model python environment and generate an archive file saved to provided save_path.

Typical usages:

  • Pre-build a model’s environment in Databricks Runtime and then download the prebuilt python environment archive file. This pre-built environment archive can then be used in mlflow.pyfunc.spark_udf for remote inference execution when using Databricks Connect to remotely connect to a Databricks environment for code execution.

Note

The build_model_env API is intended to only work when executed within Databricks runtime, serving the purpose of capturing the required execution environment that is needed for remote code execution when using DBConnect. The environment archive is designed to be used when performing remote execution using mlflow.pyfunc.spark_udf in Databricks runtime or Databricks Connect client and has no other purpose. The prebuilt env archive file cannot be used across different Databricks runtime versions or different platform machines. As such, if you connect to a different cluster that is running a different runtime version on Databricks, you will need to execute this API in a notebook and retrieve the generated archive to your local machine. Each environment snapshot is unique to the the model, the runtime version of your remote Databricks cluster, and the specification of the udf execution environment. When using the prebuilt env in mlflow.pyfunc.spark_udf, MLflow will verify whether the spark UDF sandbox environment matches the prebuilt env requirements and will raise Exceptions if there are compatibility issues. If these occur, simply re-run this API in the cluster that you are attempting to attach to.

Example 
from mlflow.pyfunc import build_model_env

# Create a python environment archive file at the path `prebuilt_env_uri`
prebuilt_env_uri = build_model_env(f"runs:/{run_id}/model", "/path/to/save_directory")
Parameters

  • model_uri – URI to the model that is used to build the python environment.

  • save_path – The directory path that is used to save the prebuilt model environment archive file path. The path can be either local directory path or mounted DBFS path such as ‘/dbfs/…’ or mounted UC volume path such as ‘/Volumes/…’.

Returns

Return the path of an archive file containing the python environment data.

mlflow.pyfunc.get_model_dependencies(model_uri, format='pip')[source]

Downloads the model dependencies and returns the path to requirements.txt or conda.yaml file.

Warning

This API downloads all the model artifacts to the local filesystem. This may take a long time for large models. To avoid this overhead, use mlflow.artifacts.download_artifacts("<model_uri>/requirements.txt") or mlflow.artifacts.download_artifacts("<model_uri>/conda.yaml") instead.

Parameters

  • model_uri – The uri of the model to get dependencies from.

  • format – The format of the returned dependency file. If the "pip" format is specified, the path to a pip requirements.txt file is returned. If the "conda" format is specified, the path to a "conda.yaml" file is returned . If the "pip" format is specified but the model was not saved with a requirements.txt file, the pip section of the model’s conda.yaml file is extracted instead, and any additional conda dependencies are ignored. Default value is "pip".

Returns

The local filesystem path to either a pip requirements.txt file (if format="pip") or a conda.yaml file (if format="conda") specifying the model’s dependencies.

mlflow.pyfunc.load_model(model_uri: str, suppress_warnings: bool = False, dst_path: Optional[str] = None, model_config: Optional[Union[str, pathlib.Path, dict]] = None) → mlflow.pyfunc.PyFuncModel[source]

Load a model stored in Python function format.

Parameters

  • model_uri –

    The location, in URI format, of the MLflow model. For example:

    • /Users/me/path/to/local/model

    • relative/path/to/local/model

    • s3://my_bucket/path/to/model

    • runs:/<mlflow_run_id>/run-relative/path/to/model

    • models:/<model_name>/<model_version>

    • models:/<model_name>/<stage>

    • mlflow-artifacts:/path/to/model

    For more information about supported URI schemes, see Referencing Artifacts.

  • suppress_warnings – If True, non-fatal warning messages associated with the model loading process will be suppressed. If False, these warning messages will be emitted.

  • dst_path – The local filesystem path to which to download the model artifact. This directory must already exist. If unspecified, a local output path will be created.

  • model_config –

    The model configuration to apply to the model. The configuration will be available as the model_config property of the context parameter in PythonModel.load_context() and PythonModel.predict(). The configuration can be passed as a file path, or a dict with string keys.

    Note

    Experimental: This parameter may change or be removed in a future release without warning.

mlflow.pyfunc.load_pyfunc(model_uri, suppress_warnings=False)[source]

Warning

mlflow.pyfunc.load_pyfunc is deprecated since 1.0. This method will be removed in a future release. Use mlflow.pyfunc.load_model instead.

Load a model stored in Python function format.

Parameters

  • model_uri –

    The location, in URI format, of the MLflow model. For example:

    • /Users/me/path/to/local/model

    • relative/path/to/local/model

    • s3://my_bucket/path/to/model

    • runs:/<mlflow_run_id>/run-relative/path/to/model

    • models:/<model_name>/<model_version>

    • models:/<model_name>/<stage>

    • mlflow-artifacts:/path/to/model

    For more information about supported URI schemes, see Referencing Artifacts.

  • suppress_warnings – If True, non-fatal warning messages associated with the model loading process will be suppressed. If False, these warning messages will be emitted.

mlflow.pyfunc.log_model(artifact_path, loader_module=None, data_path=None, code_path=None, code_paths=None, infer_code_paths=False, conda_env=None, python_model=None, artifacts=None, registered_model_name=None, signature: mlflow.models.signature.ModelSignature = None, input_example: Union[pandas.core.frame.DataFrame, numpy.ndarray, dict, list, csr_matrix, csc_matrix, str, bytes, tuple] = None, await_registration_for=300, pip_requirements=None, extra_pip_requirements=None, metadata=None, model_config=None, example_no_conversion=None, streamable=None, resources: Optional[Union[str, list]] = None, auth_policy: Optional[mlflow.models.auth_policy.AuthPolicy] = None, prompts: Optional[list] = None)[source]

Log a Pyfunc model with custom inference logic and optional data dependencies as an MLflow artifact for the current run.

For information about the workflows that this method supports, see Workflows for creating custom pyfunc models and Which workflow is right for my use case?. You cannot specify the parameters for the second workflow: loader_module, data_path and the parameters for the first workflow: python_model, artifacts together.

Parameters

  • artifact_path – The run-relative artifact path to which to log the Python model.

  • loader_module –

    The name of the Python module that is used to load the model from data_path. This module must define a method with the prototype _load_pyfunc(data_path). If not None, this module and its dependencies must be included in one of the following locations:

    • The MLflow library.

    • Package(s) listed in the model’s Conda environment, specified by the conda_env parameter.

    • One or more of the files specified by the code_path parameter.

  • data_path – Path to a file or directory containing model data.

  • code_path – Deprecated The legacy argument for defining dependent code. This argument is replaced by code_paths and will be removed in a future version of MLflow.

  • code_paths –

    A list of local filesystem paths to Python file dependencies (or directories containing file dependencies). These files are prepended to the system path when the model is loaded. Files declared as dependencies for a given model should have relative imports declared from a common root path if multiple files are defined with import dependencies between them to avoid import errors when loading the model.

    You can leave code_paths argument unset but set infer_code_paths to True to let MLflow infer the model code paths. See infer_code_paths argument doc for details.

    For a detailed explanation of code_paths functionality, recommended usage patterns and limitations, see the code_paths usage guide.

  • infer_code_paths –

    If set to True, MLflow automatically infers model code paths. The inferred

    code path files only include necessary python module files. Only python code files under current working directory are automatically inferable. Default value is False.

    Warning

    Please ensure that the custom python module code does not contain sensitive data such as credential token strings, otherwise they might be included in the automatic inferred code path files and be logged to MLflow artifact repository.

    If your custom python module depends on non-python files (e.g. a JSON file) with a relative path to the module code file path, the non-python files can’t be automatically inferred as the code path file. To address this issue, you should put all used non-python files outside your custom code directory.

    If a python code file is loaded as the python __main__ module, then this code file can’t be inferred as the code path file. If your model depends on classes / functions defined in __main__ module, you should use cloudpickle to dump your model instance in order to pickle classes / functions in __main__.

    Note

    Experimental: This parameter may change or be removed in a future release without warning.

  • conda_env –

    Either a dictionary representation of a Conda environment or the path to a conda environment yaml file. If provided, this describes the environment this model should be run in. At a minimum, it should specify the dependencies contained in get_default_conda_env(). If None, a conda environment with pip requirements inferred by mlflow.models.infer_pip_requirements() is added to the model. If the requirement inference fails, it falls back to using get_default_pip_requirements(). pip requirements from conda_env are written to a pip requirements.txt file and the full conda environment is written to conda.yaml. The following is an example dictionary representation of a conda environment:

    {
    "name": "mlflow-env",
    "channels": ["conda-forge"],
    "dependencies": [
        "python=3.8.15",
        {
            "pip": [
                "scikit-learn==x.y.z"
            ],
        },
    ],
}

  • python_model –

    An instance of a subclass of PythonModel or a callable object with a single argument (see the examples below). The passed-in object is serialized using the CloudPickle library. The python_model can also be a file path to the PythonModel which defines the model from code artifact rather than serializing the model object. Any dependencies of the class should be included in one of the following locations:

    • The MLflow library.

    • Package(s) listed in the model’s Conda environment, specified by the conda_env parameter.

    • One or more of the files specified by the code_path parameter.

    Note: If the class is imported from another module, as opposed to being defined in the __main__ scope, the defining module should also be included in one of the listed locations.

    Examples

    Class model

    from typing import List
import mlflow


class MyModel(mlflow.pyfunc.PythonModel):
    def predict(self, context, model_input: List[str], params=None) -> List[str]:
        return [i.upper() for i in model_input]


with mlflow.start_run():
    model_info = mlflow.pyfunc.log_model(
        artifact_path="model",
        python_model=MyModel(),
    )


loaded_model = mlflow.pyfunc.load_model(model_uri=model_info.model_uri)
print(loaded_model.predict(["a", "b", "c"]))  # -> ["A", "B", "C"]

    Functional model

    Note

    Experimental: Functional model support is experimental and may change or be removed in a future release without warning.

    from typing import List
import mlflow


def predict(model_input: List[str]) -> List[str]:
    return [i.upper() for i in model_input]


with mlflow.start_run():
    model_info = mlflow.pyfunc.log_model(
        artifact_path="model", python_model=predict, input_example=["a"]
    )


loaded_model = mlflow.pyfunc.load_model(model_uri=model_info.model_uri)
print(loaded_model.predict(["a", "b", "c"]))  # -> ["A", "B", "C"]

    Model from code

    Note

    Experimental: Model from code model support is experimental and may change or be removed in a future release without warning.

    # code.py
from typing import List
import mlflow


class MyModel(mlflow.pyfunc.PythonModel):
    def predict(self, context, model_input: List[str], params=None) -> List[str]:
        return [i.upper() for i in model_input]


mlflow.models.set_model(MyModel())

# log_model.py
import mlflow

with mlflow.start_run():
    model_info = mlflow.pyfunc.log_model(
        artifact_path="model",
        python_model="code.py",
    )

    If the predict method or function has type annotations, MLflow automatically constructs a model signature based on the type annotations (unless the signature argument is explicitly specified), and converts the input value to the specified type before passing it to the function. Currently, the following type annotations are supported:

    • List[str]

    • List[Dict[str, str]]

  • artifacts –

    A dictionary containing <name, artifact_uri> entries. Remote artifact URIs are resolved to absolute filesystem paths, producing a dictionary of <name, absolute_path> entries. python_model can reference these resolved entries as the artifacts property of the context parameter in PythonModel.load_context() and PythonModel.predict(). For example, consider the following artifacts dictionary:

    {"my_file": "s3://my-bucket/path/to/my/file"}

    In this case, the "my_file" artifact is downloaded from S3. The python_model can then refer to "my_file" as an absolute filesystem path via context.artifacts["my_file"].

    If None, no artifacts are added to the model.

  • registered_model_name – This argument may change or be removed in a future release without warning. If given, create a model version under registered_model_name, also creating a registered model if one with the given name does not exist.

  • signature –

    ModelSignature describes model input and output Schema. The model signature can be inferred from datasets with valid model input (e.g. the training dataset with target column omitted) and valid model output (e.g. model predictions generated on the training dataset), for example:

    from mlflow.models import infer_signature

train = df.drop_column("target_label")
predictions = ...  # compute model predictions
signature = infer_signature(train, predictions)

  • input_example – one or several instances of valid model input. The input example is used as a hint of what data to feed the model. It will be converted to a Pandas DataFrame and then serialized to json using the Pandas split-oriented format, or a numpy array where the example will be serialized to json by converting it to a list. Bytes are base64-encoded. When the signature parameter is None, the input example is used to infer a model signature.

  • await_registration_for – Number of seconds to wait for the model version to finish being created and is in READY status. By default, the function waits for five minutes. Specify 0 or None to skip waiting.

  • pip_requirements – Either an iterable of pip requirement strings (e.g. ["scikit-learn", "-r requirements.txt", "-c constraints.txt"]) or the string path to a pip requirements file on the local filesystem (e.g. "requirements.txt"). If provided, this describes the environment this model should be run in. If None, a default list of requirements is inferred by mlflow.models.infer_pip_requirements() from the current software environment. If the requirement inference fails, it falls back to using get_default_pip_requirements(). Both requirements and constraints are automatically parsed and written to requirements.txt and constraints.txt files, respectively, and stored as part of the model. Requirements are also written to the pip section of the model’s conda environment (conda.yaml) file.

  • extra_pip_requirements –

    Either an iterable of pip requirement strings (e.g. ["pandas", "-r requirements.txt", "-c constraints.txt"]) or the string path to a pip requirements file on the local filesystem (e.g. "requirements.txt"). If provided, this describes additional pip requirements that are appended to a default set of pip requirements generated automatically based on the user’s current software environment. Both requirements and constraints are automatically parsed and written to requirements.txt and constraints.txt files, respectively, and stored as part of the model. Requirements are also written to the pip section of the model’s conda environment (conda.yaml) file.

    Warning

    The following arguments can’t be specified at the same time:

    • conda_env

    • pip_requirements

    • extra_pip_requirements

    This example demonstrates how to specify pip requirements using pip_requirements and extra_pip_requirements.

  • metadata – Custom metadata dictionary passed to the model and stored in the MLmodel file.

  • model_config –

    The model configuration to apply to the model. The configuration will be available as the model_config property of the context parameter in PythonModel.load_context() and PythonModel.predict(). The configuration can be passed as a file path, or a dict with string keys.

    Note

    Experimental: This parameter may change or be removed in a future release without warning.

  • example_no_conversion – This parameter is deprecated and will be removed in a future release. It’s no longer used and can be safely removed. Input examples are not converted anymore.

  • streamable – A boolean value indicating if the model supports streaming prediction, If None, MLflow will try to inspect if the model supports streaming by checking if predict_stream method exists. Default None.

  • resources –

    A list of model resources or a resources.yaml file containing a list of

    resources required to serve the model.

    Note

    Experimental: This parameter may change or be removed in a future release without warning.

  • auth_policy –

    Specifies the authentication policy for the model, which includes two key components.

    Note that only one of auth_policy or resources should be defined.

    • System Auth Policy: A list of resources required to serve this model.

    • User Auth Policy: A minimal list of scopes that the user should have access to

      ,in order to invoke this model.

    Note

    Experimental: This parameter may change or be removed in a future release without warning.

  • prompts –

    A list of prompt URIs registered in the MLflow Prompt Registry, to be associated with the model. Each prompt URI should be in the form prompt:/<name>/<version>. The prompts should be registered in the MLflow Prompt Registry before being associated with the model.

    This will create a mutual link between the model and the prompt. The associated prompts can be seen in the model’s metadata stored in the MLmodel file. From the Prompt Registry UI, you can navigate to the model as well.

    import mlflow

prompt_template = "Hi, {name}! How are you doing today?"

# Register a prompt in the MLflow Prompt Registry
mlflow.prompts.register_prompt("my_prompt", prompt_template, description="A simple prompt")

# Log a model with the registered prompt
with mlflow.start_run():
    model_info = mlflow.pyfunc.log_model(
        MyModel(),
        artifact_path="model",
        prompts=["prompt:/my_prompt/1"]
    )

print(model_info.prompts)
# Output: ['prompt:/my_prompt/1']

# Load the prompt
prompt = mlflow.load_prompt(model_info.prompts[0])

Returns

A ModelInfo instance that contains the metadata of the logged model.

mlflow.pyfunc.save_model(path, loader_module=None, data_path=None, code_path=None, code_paths=None, infer_code_paths=False, conda_env=None, mlflow_model=None, python_model=None, artifacts=None, signature: mlflow.models.signature.ModelSignature = None, input_example: Union[pandas.core.frame.DataFrame, numpy.ndarray, dict, list, csr_matrix, csc_matrix, str, bytes, tuple] = None, pip_requirements=None, extra_pip_requirements=None, metadata=None, model_config=None, example_no_conversion=None, streamable=None, resources: Optional[Union[str, list]] = None, auth_policy: Optional[mlflow.models.auth_policy.AuthPolicy] = None, **kwargs)[source]

Save a Pyfunc model with custom inference logic and optional data dependencies to a path on the local filesystem.

For information about the workflows that this method supports, please see “workflows for creating custom pyfunc models” and “which workflow is right for my use case?”. Note that the parameters for the second workflow: loader_module, data_path and the parameters for the first workflow: python_model, artifacts, cannot be specified together.

Parameters

  • path – The path to which to save the Python model.

  • loader_module –

    The name of the Python module that is used to load the model from data_path. This module must define a method with the prototype _load_pyfunc(data_path). If not None, this module and its dependencies must be included in one of the following locations:

    • The MLflow library.

    • Package(s) listed in the model’s Conda environment, specified by the conda_env parameter.

    • One or more of the files specified by the code_path parameter.

  • data_path – Path to a file or directory containing model data.

  • code_path – Deprecated The legacy argument for defining dependent code. This argument is replaced by code_paths and will be removed in a future version of MLflow.

  • code_paths –

    A list of local filesystem paths to Python file dependencies (or directories containing file dependencies). These files are prepended to the system path when the model is loaded. Files declared as dependencies for a given model should have relative imports declared from a common root path if multiple files are defined with import dependencies between them to avoid import errors when loading the model.

    You can leave code_paths argument unset but set infer_code_paths to True to let MLflow infer the model code paths. See infer_code_paths argument doc for details.

    For a detailed explanation of code_paths functionality, recommended usage patterns and limitations, see the code_paths usage guide.

  • infer_code_paths –

    If set to True, MLflow automatically infers model code paths. The inferred

    code path files only include necessary python module files. Only python code files under current working directory are automatically inferable. Default value is False.

    Warning

    Please ensure that the custom python module code does not contain sensitive data such as credential token strings, otherwise they might be included in the automatic inferred code path files and be logged to MLflow artifact repository.

    If your custom python module depends on non-python files (e.g. a JSON file) with a relative path to the module code file path, the non-python files can’t be automatically inferred as the code path file. To address this issue, you should put all used non-python files outside your custom code directory.

    If a python code file is loaded as the python __main__ module, then this code file can’t be inferred as the code path file. If your model depends on classes / functions defined in __main__ module, you should use cloudpickle to dump your model instance in order to pickle classes / functions in __main__.

    Note

    Experimental: This parameter may change or be removed in a future release without warning.

  • conda_env –

    Either a dictionary representation of a Conda environment or the path to a conda environment yaml file. If provided, this describes the environment this model should be run in. At a minimum, it should specify the dependencies contained in get_default_conda_env(). If None, a conda environment with pip requirements inferred by mlflow.models.infer_pip_requirements() is added to the model. If the requirement inference fails, it falls back to using get_default_pip_requirements(). pip requirements from conda_env are written to a pip requirements.txt file and the full conda environment is written to conda.yaml. The following is an example dictionary representation of a conda environment:

    {
    "name": "mlflow-env",
    "channels": ["conda-forge"],
    "dependencies": [
        "python=3.8.15",
        {
            "pip": [
                "scikit-learn==x.y.z"
            ],
        },
    ],
}

  • mlflow_model – mlflow.models.Model configuration to which to add the python_function flavor.

  • python_model –

    An instance of a subclass of PythonModel or a callable object with a single argument (see the examples below). The passed-in object is serialized using the CloudPickle library. The python_model can also be a file path to the PythonModel which defines the model from code artifact rather than serializing the model object. Any dependencies of the class should be included in one of the following locations:

    • The MLflow library.

    • Package(s) listed in the model’s Conda environment, specified by the conda_env parameter.

    • One or more of the files specified by the code_path parameter.

    Note: If the class is imported from another module, as opposed to being defined in the __main__ scope, the defining module should also be included in one of the listed locations.

    Examples

    Class model

    from typing import List, Dict
import mlflow


class MyModel(mlflow.pyfunc.PythonModel):
    def predict(self, context, model_input: List[str], params=None) -> List[str]:
        return [i.upper() for i in model_input]


mlflow.pyfunc.save_model("model", python_model=MyModel(), input_example=["a"])
model = mlflow.pyfunc.load_model("model")
print(model.predict(["a", "b", "c"]))  # -> ["A", "B", "C"]

    Functional model

    Note

    Experimental: Functional model support is experimental and may change or be removed in a future release without warning.

    from typing import List
import mlflow


def predict(model_input: List[str]) -> List[str]:
    return [i.upper() for i in model_input]


mlflow.pyfunc.save_model("model", python_model=predict, input_example=["a"])
model = mlflow.pyfunc.load_model("model")
print(model.predict(["a", "b", "c"]))  # -> ["A", "B", "C"]

    Model from code

    Note

    Experimental: Model from code model support is experimental and may change or be removed in a future release without warning.

    # code.py
from typing import List
import mlflow


class MyModel(mlflow.pyfunc.PythonModel):
    def predict(self, context, model_input: List[str], params=None) -> List[str]:
        return [i.upper() for i in model_input]


mlflow.models.set_model(MyModel())

# log_model.py
import mlflow

with mlflow.start_run():
    model_info = mlflow.pyfunc.log_model(
        artifact_path="model",
        python_model="code.py",
    )

    If the predict method or function has type annotations, MLflow automatically constructs a model signature based on the type annotations (unless the signature argument is explicitly specified), and converts the input value to the specified type before passing it to the function. Currently, the following type annotations are supported:

    • List[str]

    • List[Dict[str, str]]

  • artifacts –

    A dictionary containing <name, artifact_uri> entries. Remote artifact URIs are resolved to absolute filesystem paths, producing a dictionary of <name, absolute_path> entries. python_model can reference these resolved entries as the artifacts property of the context parameter in PythonModel.load_context() and PythonModel.predict(). For example, consider the following artifacts dictionary:

    {"my_file": "s3://my-bucket/path/to/my/file"}

    In this case, the "my_file" artifact is downloaded from S3. The python_model can then refer to "my_file" as an absolute filesystem path via context.artifacts["my_file"].

    If None, no artifacts are added to the model.

  • signature –

    ModelSignature describes model input and output Schema. The model signature can be inferred from datasets with valid model input (e.g. the training dataset with target column omitted) and valid model output (e.g. model predictions generated on the training dataset), for example:

    from mlflow.models import infer_signature

train = df.drop_column("target_label")
predictions = ...  # compute model predictions
signature = infer_signature(train, predictions)

  • input_example – one or several instances of valid model input. The input example is used as a hint of what data to feed the model. It will be converted to a Pandas DataFrame and then serialized to json using the Pandas split-oriented format, or a numpy array where the example will be serialized to json by converting it to a list. Bytes are base64-encoded. When the signature parameter is None, the input example is used to infer a model signature.

  • pip_requirements – Either an iterable of pip requirement strings (e.g. ["scikit-learn", "-r requirements.txt", "-c constraints.txt"]) or the string path to a pip requirements file on the local filesystem (e.g. "requirements.txt"). If provided, this describes the environment this model should be run in. If None, a default list of requirements is inferred by mlflow.models.infer_pip_requirements() from the current software environment. If the requirement inference fails, it falls back to using get_default_pip_requirements(). Both requirements and constraints are automatically parsed and written to requirements.txt and constraints.txt files, respectively, and stored as part of the model. Requirements are also written to the pip section of the model’s conda environment (conda.yaml) file.

  • extra_pip_requirements –

    Either an iterable of pip requirement strings (e.g. ["pandas", "-r requirements.txt", "-c constraints.txt"]) or the string path to a pip requirements file on the local filesystem (e.g. "requirements.txt"). If provided, this describes additional pip requirements that are appended to a default set of pip requirements generated automatically based on the user’s current software environment. Both requirements and constraints are automatically parsed and written to requirements.txt and constraints.txt files, respectively, and stored as part of the model. Requirements are also written to the pip section of the model’s conda environment (conda.yaml) file.

    Warning

    The following arguments can’t be specified at the same time:

    • conda_env

    • pip_requirements

    • extra_pip_requirements

    This example demonstrates how to specify pip requirements using pip_requirements and extra_pip_requirements.

  • metadata – Custom metadata dictionary passed to the model and stored in the MLmodel file.

  • model_config –

    The model configuration to apply to the model. The configuration will be available as the model_config property of the context parameter in PythonModel.load_context() and PythonModel.predict(). The configuration can be passed as a file path, or a dict with string keys.

    Note

    Experimental: This parameter may change or be removed in a future release without warning.

  • example_no_conversion – This parameter is deprecated and will be removed in a future release. It’s no longer used and can be safely removed. Input examples are not converted anymore.

  • streamable – A boolean value indicating if the model supports streaming prediction, If None, MLflow will try to inspect if the model supports streaming by checking if predict_stream method exists. Default None.

  • resources –

    A list of model resources or a resources.yaml file containing a list of

    resources required to serve the model.

    Note

    Experimental: This parameter may change or be removed in a future release without warning.

  • auth_policy –

    Specifies the authentication policy for the model, which includes two key components.

    Note that only one of auth_policy or resources should be defined.

    • System Auth Policy: A list of resources required to serve this model.

    • User Auth Policy: A minimal list of scopes that the user should have access to

      ,in order to invoke this model.

    Note

    Experimental: This parameter may change or be removed in a future release without warning.

  • kwargs – Extra keyword arguments.

mlflow.pyfunc.spark_udf(spark, model_uri, result_type=None, env_manager=None, params: Optional[dict] = None, extra_env: Optional[dict] = None, prebuilt_env_uri: Optional[str] = None, model_config: Optional[Union[str, pathlib.Path, dict]] = None)[source]

A Spark UDF that can be used to invoke the Python function formatted model.

Parameters passed to the UDF are forwarded to the model as a DataFrame where the column names are ordinals (0, 1, …). On some versions of Spark (3.0 and above), it is also possible to wrap the input in a struct. In that case, the data will be passed as a DataFrame with column names given by the struct definition (e.g. when invoked as my_udf(struct(‘x’, ‘y’)), the model will get the data as a pandas DataFrame with 2 columns ‘x’ and ‘y’).

If a model contains a signature with tensor spec inputs, you will need to pass a column of array type as a corresponding UDF argument. The column values of which must be one dimensional arrays. The UDF will reshape the column values to the required shape with ‘C’ order (i.e. read / write the elements using C-like index order) and cast the values as the required tensor spec type.

If a model contains a signature, the UDF can be called without specifying column name arguments. In this case, the UDF will be called with column names from signature, so the evaluation dataframe’s column names must match the model signature’s column names.

The predictions are filtered to contain only the columns that can be represented as the result_type. If the result_type is string or array of strings, all predictions are converted to string. If the result type is not an array type, the left most column with matching type is returned.

Note

Inputs of type pyspark.sql.types.DateType are not supported on earlier versions of Spark (2.4 and below).

Note

When using Databricks Connect to connect to a remote Databricks cluster, the Databricks cluster must use runtime version >= 16, and when ‘spark_udf’ param ‘env_manager’ is set as ‘virtualenv’, the ‘prebuilt_env_uri’ param is required to be specified.

Note

Please be aware that when operating in Databricks Serverless, spark tasks run within the confines of the Databricks Serverless UDF sandbox. This environment has a total capacity limit of 1GB, combining both available memory and local disk capacity. Furthermore, there are no GPU devices available in this setup. Therefore, any deep-learning models that contain large weights or require a GPU are not suitable for deployment on Databricks Serverless.

Example 
from pyspark.sql.functions import struct

predict = mlflow.pyfunc.spark_udf(spark, "/my/local/model")
df.withColumn("prediction", predict(struct("name", "age"))).show()
Parameters

  • spark – A SparkSession object.

  • model_uri –

    The location, in URI format, of the MLflow model with the mlflow.pyfunc flavor. For example:

    • /Users/me/path/to/local/model

    • relative/path/to/local/model

    • s3://my_bucket/path/to/model

    • runs:/<mlflow_run_id>/run-relative/path/to/model

    • models:/<model_name>/<model_version>

    • models:/<model_name>/<stage>

    • mlflow-artifacts:/path/to/model

    For more information about supported URI schemes, see Referencing Artifacts.

  • result_type –

    the return type of the user-defined function. The value can be either a pyspark.sql.types.DataType object or a DDL-formatted type string. Only a primitive type, an array pyspark.sql.types.ArrayType of primitive type, or a struct type containing fields of above 2 kinds of types are allowed. If unspecified, it tries to infer result type from model signature output schema, if model output schema is not available, it fallbacks to use double type.

    The following classes of result type are supported:

    • ”int” or pyspark.sql.types.IntegerType: The leftmost integer that can fit in an int32 or an exception if there is none.

    • ”long” or pyspark.sql.types.LongType: The leftmost long integer that can fit in an int64 or an exception if there is none.

    • ArrayType(IntegerType|LongType): All integer columns that can fit into the requested size.

    • ”float” or pyspark.sql.types.FloatType: The leftmost numeric result cast to float32 or an exception if there is none.

    • ”double” or pyspark.sql.types.DoubleType: The leftmost numeric result cast to double or an exception if there is none.

    • ArrayType(FloatType|DoubleType): All numeric columns cast to the requested type or an exception if there are no numeric columns.

    • ”string” or pyspark.sql.types.StringType: The leftmost column converted to string.

    • ”boolean” or “bool” or pyspark.sql.types.BooleanType: The leftmost column converted to bool or an exception if there is none.

    • ArrayType(StringType): All columns converted to string.

    • ”field1 FIELD1_TYPE, field2 FIELD2_TYPE, …”: A struct type containing multiple fields separated by comma, each field type must be one of types listed above.

  • env_manager –

    The environment manager to use in order to create the python environment for model inference. Note that environment is only restored in the context of the PySpark UDF; the software environment outside of the UDF is unaffected. If prebuilt_env_uri parameter is not set, the default value is local, and the following values are supported:

    • virtualenv: Use virtualenv to restore the python environment that was used to train the model.

    • conda: (Recommended) Use Conda to restore the software environment that was used to train the model.

    • local: Use the current Python environment for model inference, which may differ from the environment used to train the model and may lead to errors or invalid predictions.

    If the prebuilt_env_uri parameter is set, env_manager parameter should not be set.

  • params – Additional parameters to pass to the model for inference.

  • extra_env – Extra environment variables to pass to the UDF executors. For overrides that need to propagate to the Spark workers (i.e., overriding the scoring server timeout via MLFLOW_SCORING_SERVER_REQUEST_TIMEOUT).

  • prebuilt_env_uri –

    The path of the prebuilt env archive file created by mlflow.pyfunc.build_model_env API. This parameter can only be used in Databricks Serverless notebook REPL, Databricks Shared cluster notebook REPL, and Databricks Connect client environment. The path can be either local file path or DBFS path such as ‘dbfs:/Volumes/…’, in this case, MLflow automatically downloads it to local temporary directory, “MLFLOW_MODEL_ENV_DOWNLOADING_TEMP_DIR” environmental variable can be set to specify the temporary directory to use.

    If this parameter is set, env_manger parameter must not be set.

  • model_config – The model configuration to set when loading the model. See ‘model_config’ argument in mlflow.pyfunc.load_model API for details.

Returns

Spark UDF that applies the model’s predict method to the data and returns a type specified by result_type, which by default is a double.

mlflow.pyfunc.update_signature_for_type_hint_from_example(input_example: Any, signature: mlflow.models.signature.ModelSignature)[source]
mlflow.pyfunc.get_default_pip_requirements()[source]
Returns

A list of default pip requirements for MLflow Models produced by this flavor. Calls to save_model() and log_model() produce a pip environment that, at minimum, contains these requirements.

mlflow.pyfunc.get_default_conda_env()[source]
Returns

The default Conda environment for MLflow Models produced by calls to save_model() and log_model() when a user-defined subclass of PythonModel is provided.

class mlflow.pyfunc.PythonModelContext[source]

A collection of artifacts that a PythonModel can use when performing inference. PythonModelContext objects are created implicitly by the save_model() and log_model() persistence methods, using the contents specified by the artifacts parameter of these methods.

property artifacts

A dictionary containing <name, artifact_path> entries, where artifact_path is an absolute filesystem path to the artifact.

property model_config

A dictionary containing <config, value> entries, where config is the name of the model configuration keys and value is the value of the given configuration.

Note

Experimental: This property may change or be removed in a future release without warning.

class mlflow.pyfunc.PythonModel[source]

Represents a generic Python model that evaluates inputs and produces API-compatible outputs. By subclassing PythonModel, users can create customized MLflow models with the “python_function” (“pyfunc”) flavor, leveraging custom inference logic and artifact dependencies.

load_context(context)[source]

Loads artifacts from the specified PythonModelContext that can be used by predict() when evaluating inputs. When loading an MLflow model with load_model(), this method is called as soon as the PythonModel is constructed.

The same PythonModelContext will also be available during calls to predict(), but it may be more efficient to override this method and load artifacts from the context at model load time.

Parameters

context – A PythonModelContext instance containing artifacts that the model can use to perform inference.

abstract predict(context, model_input, params: Optional[dict] = None)[source]

Evaluates a pyfunc-compatible input and produces a pyfunc-compatible output. For more information about the pyfunc input/output API, see the Inference API.

Parameters

  • context – A PythonModelContext instance containing artifacts that the model can use to perform inference.

  • model_input – A pyfunc-compatible input for the model to evaluate.

  • params – Additional parameters to pass to the model for inference.

Tip

Since MLflow 2.20.0, context parameter can be removed from predict function signature if it’s not used. def predict(self, model_input, params=None) is valid.

predict_stream(context, model_input, params: Optional[dict] = None)[source]

Evaluates a pyfunc-compatible input and produces an iterator of output. For more information about the pyfunc input API, see the Inference API.

Parameters

  • context – A PythonModelContext instance containing artifacts that the model can use to perform inference.

  • model_input – A pyfunc-compatible input for the model to evaluate.

  • params – Additional parameters to pass to the model for inference.

Tip

Since MLflow 2.20.0, context parameter can be removed from predict_stream function signature if it’s not used. def predict_stream(self, model_input, params=None) is valid.

property predict_type_hints

Internal method to get type hints from the predict function signature.

class mlflow.pyfunc.ChatModel[source]

Note

Experimental: This class may change or be removed in a future release without warning.

Tip

Since MLflow 2.20.2, we recommend using ChatAgent instead of ChatModel unless you need strict compatibility with the OpenAI ChatCompletion API.

A subclass of PythonModel that makes it more convenient to implement models that are compatible with popular LLM chat APIs. By subclassing ChatModel, users can create MLflow models with a predict() method that is more convenient for chat tasks than the generic PythonModel API. ChatModels automatically define input/output signatures and an input example, so manually specifying these values when calling mlflow.pyfunc.save_model() is not necessary.

See the documentation of the predict() method below for details on that parameters and outputs that are expected by the ChatModel API.

ChatModel

PythonModel

When to use

Use when you want to develop and deploy a conversational model with standard chat schema compatible with OpenAI spec.

Use when you want full control over the model’s interface or customize every aspect of your model’s behavior.

Interface

Fixed to OpenAI’s chat schema.

Full control over the model’s input and output schema.

Setup
Quick. Works out of the box for conversational applications, with pre-defined

model signature and input example.

Custom. You need to define model signature or input example yourself.

Complexity

Low. Standardized interface simplified model deployment and integration.
High. Deploying and integrating the custom PythonModel may not be straightforward.

E.g., The model needs to handle Pandas DataFrames as MLflow converts input data to DataFrames before passing it to PythonModel.
abstract predict(context, messages: list, params: mlflow.types.llm.ChatParams) → mlflow.types.llm.ChatCompletionResponse[source]

Evaluates a chat input and produces a chat output.

Parameters

  • context – A PythonModelContext instance containing artifacts that the model can use to perform inference.

  • messages (List[ChatMessage]) – A list of ChatMessage objects representing chat history.

  • params (ChatParams) – A ChatParams object containing various parameters used to modify model behavior during inference.

Tip

Since MLflow 2.20.0, context parameter can be removed from predict function signature if it’s not used. def predict(self, messages: list[ChatMessage], params: ChatParams) is valid.

Returns

A ChatCompletionResponse object containing the model’s response(s), as well as other metadata.

predict_stream(context, messages: list, params: mlflow.types.llm.ChatParams) → Generator[mlflow.types.llm.ChatCompletionChunk, None, None][source]

Evaluates a chat input and produces a chat output. Override this function to implement a real stream prediction.

Parameters

  • context – A PythonModelContext instance containing artifacts that the model can use to perform inference.

  • messages (List[ChatMessage]) – A list of ChatMessage objects representing chat history.

  • params (ChatParams) – A ChatParams object containing various parameters used to modify model behavior during inference.

Tip

Since MLflow 2.20.0, context parameter can be removed from predict_stream function signature if it’s not used. def predict_stream(self, messages: list[ChatMessage], params: ChatParams) is valid.

Returns

A generator over ChatCompletionChunk object containing the model’s response(s), as well as other metadata.

class mlflow.pyfunc.ChatAgent[source]

Note

Experimental: This class may change or be removed in a future release without warning.

What is the ChatAgent Interface?

The ChatAgent interface is a chat schema specification that has been designed for authoring conversational agents. ChatAgent allows your agent to do the following:

  • Return multiple messages

  • Return intermediate steps for tool calling agents

  • Confirm tool calls

  • Support multi-agent scenarios

ChatAgent should always be used when authoring an agent. We also recommend using ChatAgent instead of ChatModel even for use cases like simple chat models (e.g. prompt-engineered LLMs), to give you the flexibility to support more agentic functionality in the future.

The ChatAgentRequest schema is similar to, but not strictly compatible with the OpenAI ChatCompletion schema. ChatAgent adds additional functionality and diverges from OpenAI ChatCompletionRequest in the following ways:

  • Adds an optional attachments attribute to every input/output message for tools and internal agent calls so they can return additional outputs such as visualizations and progress indicators

  • Adds a context attribute with a conversation_id and user_id attributes to enable modifying the behavior of the agent depending on the user querying the agent

  • Adds the custom_inputs attribute, an arbitrary dict[str, Any] to pass in any additional information to modify the agent’s behavior

The ChatAgentResponse schema diverges from ChatCompletionResponse schema in the following ways:

  • Adds the custom_outputs key, an arbitrary dict[str, Any] to return any additional information

  • Allows multiple messages in the output, to improve the display and evaluation of internal tool calls and inter-agent communication that led to the final answer.

Here’s an example of a ChatAgentResponse detailing a tool call:

{
    "messages": [
        {
            "role": "assistant",
            "content": "",
            "id": "run-04b46401-c569-4a4a-933e-62e38d8f9647-0",
            "tool_calls": [
                {
                    "id": "call_15ca4fcc-ffa1-419a-8748-3bea34b9c043",
                    "type": "function",
                    "function": {
                        "name": "generate_random_ints",
                        "arguments": '{"min": 1, "max": 100, "size": 5}',
                    },
                }
            ],
        },
        {
            "role": "tool",
            "content": '{"content": "Generated array of 2 random ints in [1, 100]."',
            "name": "generate_random_ints",
            "id": "call_15ca4fcc-ffa1-419a-8748-3bea34b9c043",
            "tool_call_id": "call_15ca4fcc-ffa1-419a-8748-3bea34b9c043",
        },
        {
            "role": "assistant",
            "content": "The new set of generated random numbers are: 93, 51, 12, 7, and 25",
            "name": "llm",
            "id": "run-70c7c738-739f-4ecd-ad18-0ae232df24e8-0",
        },
    ],
    "custom_outputs": {"random_nums": [93, 51, 12, 7, 25]},
}

Streaming Agent Output with ChatAgent

Please read the docstring of ChatAgent.predict_stream for more details on how to stream the output of your agent.

Authoring a ChatAgent

Authoring an agent using the ChatAgent interface is a framework-agnostic way to create a model with a standardized interface that is loggable with the MLflow pyfunc flavor, can be reused across clients, and is ready for serving workloads.

To write your own agent, subclass ChatAgent, implementing the predict and optionally predict_stream methods to define the non-streaming and streaming behavior of your agent. You can use any agent authoring framework - the only hard requirement is to implement the predict interface.

def predict(
    self,
    messages: list[ChatAgentMessage],
    context: Optional[ChatContext] = None,
    custom_inputs: Optional[dict[str, Any]] = None,
) -> ChatAgentResponse: ...

In addition to calling predict and predict_stream methods with an input matching their type hints, you can also pass a single input dict that matches the ChatAgentRequest schema for ease of testing.

chat_agent = MyChatAgent()
chat_agent.predict(
    {
        "messages": [{"role": "user", "content": "What is 10 + 10?"}],
        "context": {"conversation_id": "123", "user_id": "456"},
    }
)

See an example implementation of predict and predict_stream for a LangGraph agent in the ChatAgentState docstring.

Logging the ChatAgent

Since the landscape of LLM frameworks is constantly evolving and not every flavor can be natively supported by MLflow, we recommend the Models-from-Code logging approach.

with mlflow.start_run():
    logged_agent_info = mlflow.pyfunc.log_model(
        artifact_path="agent",
        python_model=os.path.join(os.getcwd(), "agent"),
        # Add serving endpoints, tools, and vector search indexes here
        resources=[],
    )

After logging the model, you can query the model with a single dictionary with the ChatAgentRequest schema. Under the hood, it will be converted into the python objects expected by your predict and predict_stream methods.

loaded_model = mlflow.pyfunc.load_model(tmp_path)
loaded_model.predict(
    {
        "messages": [{"role": "user", "content": "What is 10 + 10?"}],
        "context": {"conversation_id": "123", "user_id": "456"},
    }
)

To make logging ChatAgent models as easy as possible, MLflow has built in the following features:

  • Automatic Model Signature Inference

    • You do not need to set a signature when logging a ChatAgent

    • An input and output signature will be automatically set that adheres to the ChatAgentRequest and ChatAgentResponse schemas

  • Metadata

    • {“task”: “agent/v2/chat”} will be automatically appended to any metadata that you may pass in when logging the model

  • Input Example

    • Providng an input example is optional, mlflow.types.agent.CHAT_AGENT_INPUT_EXAMPLE will be provided by default

    • If you do provide an input example, ensure it’s a dict with the ChatAgentRequest schema

    • input_example = {
    "messages": [{"role": "user", "content": "What is MLflow?"}],
    "context": {"conversation_id": "123", "user_id": "456"},
}

Migrating from ChatModel to ChatAgent

To convert an existing ChatModel that takes in List[ChatMessage] and ChatParams and outputs a ChatCompletionResponse, do the following:

  • Subclass ChatAgent instead of ChatModel

  • Move any functionality from your ChatModel’s load_context implementation into the __init__ method of your new ChatAgent.

  • Use .model_dump_compat() instead of .to_dict() when converting your model’s inputs to dictionaries. Ex. [msg.model_dump_compat() for msg in messages] instead of [msg.to_dict() for msg in messages]

  • Return a ChatAgentResponse instead of a ChatCompletionResponse

For example, we can convert the ChatModel from the Chat Model Intro to a ChatAgent:

class SimpleOllamaModel(ChatModel):
    def __init__(self):
        self.model_name = "llama3.2:1b"
        self.client = None

    def load_context(self, context):
        self.client = ollama.Client()

    def predict(
        self, context, messages: list[ChatMessage], params: ChatParams = None
    ) -> ChatCompletionResponse:
        ollama_messages = [msg.to_dict() for msg in messages]
        response = self.client.chat(model=self.model_name, messages=ollama_messages)
        return ChatCompletionResponse(
            choices=[{"index": 0, "message": response["message"]}],
            model=self.model_name,
        )

class SimpleOllamaModel(ChatAgent):
    def __init__(self):
        self.model_name = "llama3.2:1b"
        self.client = None
        self.client = ollama.Client()

    def predict(
        self,
        messages: list[ChatAgentMessage],
        context: Optional[ChatContext] = None,
        custom_inputs: Optional[dict[str, Any]] = None,
    ) -> ChatAgentResponse:
        ollama_messages = self._convert_messages_to_dict(messages)
        response = self.client.chat(model=self.model_name, messages=ollama_messages)
        return ChatAgentResponse(**{"messages": [response["message"]]})

ChatAgent Connectors

MLflow provides convenience APIs for wrapping agents written in popular authoring frameworks with ChatAgent. See examples for:

abstract predict(messages: list, context: Optional[mlflow.types.agent.ChatContext] = None, custom_inputs: Optional[dict] = None) → mlflow.types.agent.ChatAgentResponse[source]

Given a ChatAgent input, returns a ChatAgent output. In addition to calling predict with an input matching the type hints, you can also pass a single input dict that matches the ChatAgentRequest schema for ease of testing.

chat_agent = ChatAgent()
chat_agent.predict(
    {
        "messages": [{"role": "user", "content": "What is 10 + 10?"}],
        "context": {"conversation_id": "123", "user_id": "456"},
    }
)
Parameters

  • messages (List[ChatAgentMessage]) – A list of ChatAgentMessage objects representing the chat history.

  • context (ChatContext) – A ChatContext object containing conversation_id and user_id. Optional Defaults to None.

  • custom_inputs (Dict[str, Any]) – An optional param to provide arbitrary additional inputs to the model. The dictionary values must be JSON-serializable. Optional Defaults to None.

Returns

A ChatAgentResponse object containing the model’s response, as well as other metadata.

predict_stream(messages: list, context: Optional[mlflow.types.agent.ChatContext] = None, custom_inputs: Optional[dict] = None) → Generator[mlflow.types.agent.ChatAgentChunk, None, None][source]

Given a ChatAgent input, returns a generator containing streaming ChatAgent output chunks. In addition to calling predict_stream with an input matching the type hints, you can also pass a single input dict that matches the ChatAgentRequest schema for ease of testing.

chat_agent = ChatAgent()
for event in chat_agent.predict_stream(
    {
        "messages": [{"role": "user", "content": "What is 10 + 10?"}],
        "context": {"conversation_id": "123", "user_id": "456"},
    }
):
    print(event)

To support streaming the output of your agent, override this method in your subclass of ChatAgent. When implementing predict_stream, keep in mind the following requirements:

  • Ensure your implementation adheres to the predict_stream type signature. For example, streamed messages must be of the type ChatAgentChunk, where each chunk contains partial output from a single response message.

  • At most one chunk in a particular response can contain the custom_outputs key.

  • Chunks containing partial content of a single response message must have the same id. The content field of the message and usage stats of the ChatAgentChunk should be aggregated by the consuming client. See the example below.

{"delta": {"role": "assistant", "content": "Born", "id": "123"}}
{"delta": {"role": "assistant", "content": " in", "id": "123"}}
{"delta": {"role": "assistant", "content": " data", "id": "123"}}
Parameters

  • messages (List[ChatAgentMessage]) – A list of ChatAgentMessage objects representing the chat history.

  • context (ChatContext) – A ChatContext object containing conversation_id and user_id. Optional Defaults to None.

  • custom_inputs (Dict[str, Any]) – An optional param to provide arbitrary additional inputs to the model. The dictionary values must be JSON-serializable. Optional Defaults to None.

Returns

A generator over ChatAgentChunk objects containing the model’s response(s), as well as other metadata.