"""
The ``mlflow.pyfunc.model`` module defines logic for saving and loading custom "python_function"
models with a user-defined ``PythonModel`` subclass.
"""
import bz2
import gzip
import inspect
import logging
import lzma
import os
import shutil
from abc import ABCMeta, abstractmethod
from pathlib import Path
from typing import Any, Generator, Optional, Union
import cloudpickle
import pandas as pd
import yaml
import mlflow.pyfunc
import mlflow.utils
from mlflow.entities.span import SpanType
from mlflow.environment_variables import MLFLOW_LOG_MODEL_COMPRESSION
from mlflow.exceptions import MlflowException
from mlflow.models import Model
from mlflow.models.model import MLMODEL_FILE_NAME, MODEL_CODE_PATH
from mlflow.models.rag_signatures import ChatCompletionRequest, SplitChatMessagesRequest
from mlflow.models.signature import (
_extract_type_hints,
_is_context_in_predict_function_signature,
_TypeHints,
)
from mlflow.models.utils import _load_model_code_path
from mlflow.protos.databricks_pb2 import INVALID_PARAMETER_VALUE
from mlflow.pyfunc.utils import pyfunc
from mlflow.pyfunc.utils.data_validation import (
_check_func_signature,
_get_func_info_if_type_hint_supported,
_wrap_predict_with_pyfunc,
wrap_non_list_predict_pydantic,
)
from mlflow.pyfunc.utils.input_converter import _hydrate_dataclass
from mlflow.tracking.artifact_utils import _download_artifact_from_uri
from mlflow.types.agent import (
ChatAgentChunk,
ChatAgentMessage,
ChatAgentRequest,
ChatAgentResponse,
ChatContext,
)
from mlflow.types.llm import (
ChatCompletionChunk,
ChatCompletionResponse,
ChatMessage,
ChatParams,
)
from mlflow.types.utils import _is_list_dict_str, _is_list_str
from mlflow.utils.annotations import deprecated, experimental
from mlflow.utils.databricks_utils import (
_get_databricks_serverless_env_vars,
is_in_databricks_serverless_runtime,
)
from mlflow.utils.environment import (
_CONDA_ENV_FILE_NAME,
_CONSTRAINTS_FILE_NAME,
_PYTHON_ENV_FILE_NAME,
_REQUIREMENTS_FILE_NAME,
_mlflow_conda_env,
_process_conda_env,
_process_pip_requirements,
_PythonEnv,
)
from mlflow.utils.file_utils import TempDir, get_total_file_size, write_to
from mlflow.utils.model_utils import _get_flavor_configuration, _validate_infer_and_copy_code_paths
from mlflow.utils.pydantic_utils import IS_PYDANTIC_V2_OR_NEWER
from mlflow.utils.requirements_utils import _get_pinned_requirement
CONFIG_KEY_ARTIFACTS = "artifacts"
CONFIG_KEY_ARTIFACT_RELATIVE_PATH = "path"
CONFIG_KEY_ARTIFACT_URI = "uri"
CONFIG_KEY_PYTHON_MODEL = "python_model"
CONFIG_KEY_CLOUDPICKLE_VERSION = "cloudpickle_version"
CONFIG_KEY_COMPRESSION = "python_model_compression"
_SAVED_PYTHON_MODEL_SUBPATH = "python_model.pkl"
_DEFAULT_CHAT_MODEL_METADATA_TASK = "agent/v1/chat"
_DEFAULT_CHAT_AGENT_METADATA_TASK = "agent/v2/chat"
_COMPRESSION_INFO = {
"lzma": {"ext": ".xz", "open": lzma.open},
"bzip2": {"ext": ".bz2", "open": bz2.open},
"gzip": {"ext": ".gz", "open": gzip.open},
}
_DEFAULT_RESPONSES_AGENT_METADATA_TASK = "agent/v1/responses"
_logger = logging.getLogger(__name__)
[docs]def get_default_pip_requirements():
"""
Returns:
A list of default pip requirements for MLflow Models produced by this flavor. Calls to
:func:`save_model()` and :func:`log_model()` produce a pip environment that, at minimum,
contains these requirements.
"""
return [_get_pinned_requirement("cloudpickle")]
[docs]def get_default_conda_env():
"""
Returns:
The default Conda environment for MLflow Models produced by calls to
:func:`save_model() <mlflow.pyfunc.save_model>`
and :func:`log_model() <mlflow.pyfunc.log_model>` when a user-defined subclass of
:class:`PythonModel` is provided.
"""
return _mlflow_conda_env(additional_pip_deps=get_default_pip_requirements())
def _log_warning_if_params_not_in_predict_signature(logger, params):
if params:
logger.warning(
"The underlying model does not support passing additional parameters to the predict"
f" function. `params` {params} will be ignored."
)
[docs]class PythonModel:
"""
Represents a generic Python model that evaluates inputs and produces API-compatible outputs.
By subclassing :class:`~PythonModel`, users can create customized MLflow models with the
"python_function" ("pyfunc") flavor, leveraging custom inference logic and artifact
dependencies.
"""
__metaclass__ = ABCMeta
[docs] def load_context(self, context):
"""
Loads artifacts from the specified :class:`~PythonModelContext` that can be used by
:func:`~PythonModel.predict` when evaluating inputs. When loading an MLflow model with
:func:`~load_model`, this method is called as soon as the :class:`~PythonModel` is
constructed.
The same :class:`~PythonModelContext` will also be available during calls to
:func:`~PythonModel.predict`, but it may be more efficient to override this method
and load artifacts from the context at model load time.
Args:
context: A :class:`~PythonModelContext` instance containing artifacts that the model
can use to perform inference.
"""
@deprecated("predict_type_hints", "2.20.0")
def _get_type_hints(self):
return self.predict_type_hints
@property
def predict_type_hints(self) -> _TypeHints:
"""
Internal method to get type hints from the predict function signature.
"""
if hasattr(self, "_predict_type_hints"):
return self._predict_type_hints
if _is_context_in_predict_function_signature(func=self.predict):
self._predict_type_hints = _extract_type_hints(self.predict, input_arg_index=1)
else:
self._predict_type_hints = _extract_type_hints(self.predict, input_arg_index=0)
return self._predict_type_hints
def __init_subclass__(cls, **kwargs) -> None:
super().__init_subclass__(**kwargs)
# automatically wrap the predict method with pyfunc to ensure data validation
# NB: skip wrapping for built-in classes defined in MLflow e.g. ChatModel
if not cls.__module__.startswith("mlflow."):
# TODO: ChatModel uses dataclass type hints which are not supported now, hence
# we need to skip type hint based validation for user-defined subclasses
# of ChatModel. Once we either (1) support dataclass type hints or (2) migrate
# ChatModel to pydantic, we can remove this exclusion.
# NB: issubclass(cls, ChatModel) does not work so we use a hacky attribute check
if getattr(cls, "_skip_type_hint_validation", False):
return
predict_attr = cls.__dict__.get("predict")
if predict_attr is not None and callable(predict_attr):
func_info = _get_func_info_if_type_hint_supported(predict_attr)
setattr(cls, "predict", _wrap_predict_with_pyfunc(predict_attr, func_info))
predict_stream_attr = cls.__dict__.get("predict_stream")
if predict_stream_attr is not None and callable(predict_stream_attr):
_check_func_signature(predict_stream_attr, "predict_stream")
else:
cls.predict._is_pyfunc = True
[docs] @abstractmethod
def predict(self, context, model_input, params: Optional[dict[str, Any]] = None):
"""
Evaluates a pyfunc-compatible input and produces a pyfunc-compatible output.
For more information about the pyfunc input/output API, see the :ref:`pyfunc-inference-api`.
Args:
context: A :class:`~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.
"""
[docs] def predict_stream(self, context, model_input, params: Optional[dict[str, Any]] = None):
"""
Evaluates a pyfunc-compatible input and produces an iterator of output.
For more information about the pyfunc input API, see the :ref:`pyfunc-inference-api`.
Args:
context: A :class:`~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.
"""
raise NotImplementedError()
class _FunctionPythonModel(PythonModel):
"""
When a user specifies a ``python_model`` argument that is a function, we wrap the function
in an instance of this class.
"""
def __init__(self, func, signature=None):
self.signature = signature
# only wrap `func` if @pyfunc is not already applied
if not getattr(func, "_is_pyfunc", False):
self.func = pyfunc(func)
else:
self.func = func
@property
def predict_type_hints(self):
if hasattr(self, "_predict_type_hints"):
return self._predict_type_hints
self._predict_type_hints = _extract_type_hints(self.func, input_arg_index=0)
return self._predict_type_hints
def predict(
self,
model_input,
params: Optional[dict[str, Any]] = None,
):
"""
Args:
model_input: A pyfunc-compatible input for the model to evaluate.
params: Additional parameters to pass to the model for inference.
Returns:
Model predictions.
"""
# callable only supports one input argument for now
return self.func(model_input)
[docs]class PythonModelContext:
"""
A collection of artifacts that a :class:`~PythonModel` can use when performing inference.
:class:`~PythonModelContext` objects are created *implicitly* by the
:func:`save_model() <mlflow.pyfunc.save_model>` and
:func:`log_model() <mlflow.pyfunc.log_model>` persistence methods, using the contents specified
by the ``artifacts`` parameter of these methods.
"""
def __init__(self, artifacts, model_config):
"""
Args:
artifacts: A dictionary of ``<name, artifact_path>`` entries, where ``artifact_path``
is an absolute filesystem path to a given artifact.
model_config: The model configuration to make available to the model at
loading time.
"""
self._artifacts = artifacts
self._model_config = model_config
@property
def artifacts(self):
"""
A dictionary containing ``<name, artifact_path>`` entries, where ``artifact_path`` is an
absolute filesystem path to the artifact.
"""
return self._artifacts
@property
def model_config(self):
"""
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.
"""
return self._model_config
[docs]@deprecated("ResponsesAgent", "3.0.0")
class ChatModel(PythonModel, metaclass=ABCMeta):
"""
.. tip::
Since MLflow 3.0.0, we recommend using
:py:class:`ResponsesAgent <mlflow.pyfunc.ResponsesAgent>`
instead of :py:class:`ChatModel <mlflow.pyfunc.ChatModel>` unless you need strict
compatibility with the OpenAI ChatCompletion API.
A subclass of :class:`~PythonModel` that makes it more convenient to implement models
that are compatible with popular LLM chat APIs. By subclassing :class:`~ChatModel`,
users can create MLflow models with a ``predict()`` method that is more convenient
for chat tasks than the generic :class:`~PythonModel` API. ChatModels automatically
define input/output signatures and an input example, so manually specifying these values
when calling :func:`mlflow.pyfunc.save_model() <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.
.. list-table::
:header-rows: 1
:widths: 20 40 40
* -
- 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.
"""
_skip_type_hint_validation = True
[docs] @abstractmethod
def predict(
self, context, messages: list[ChatMessage], params: ChatParams
) -> ChatCompletionResponse:
"""
Evaluates a chat input and produces a chat output.
Args:
context: A :class:`~PythonModelContext` instance containing artifacts that the model
can use to perform inference.
messages (List[:py:class:`ChatMessage <mlflow.types.llm.ChatMessage>`]):
A list of :py:class:`ChatMessage <mlflow.types.llm.ChatMessage>`
objects representing chat history.
params (:py:class:`ChatParams <mlflow.types.llm.ChatParams>`):
A :py:class:`ChatParams <mlflow.types.llm.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 :py:class:`ChatCompletionResponse <mlflow.types.llm.ChatCompletionResponse>`
object containing the model's response(s), as well as other metadata.
"""
[docs] def predict_stream(
self, context, messages: list[ChatMessage], params: ChatParams
) -> Generator[ChatCompletionChunk, None, None]:
"""
Evaluates a chat input and produces a chat output.
Override this function to implement a real stream prediction.
Args:
context: A :class:`~PythonModelContext` instance containing artifacts that the model
can use to perform inference.
messages (List[:py:class:`ChatMessage <mlflow.types.llm.ChatMessage>`]):
A list of :py:class:`ChatMessage <mlflow.types.llm.ChatMessage>`
objects representing chat history.
params (:py:class:`ChatParams <mlflow.types.llm.ChatParams>`):
A :py:class:`ChatParams <mlflow.types.llm.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 :py:class:`ChatCompletionChunk <mlflow.types.llm.ChatCompletionChunk>`
object containing the model's response(s), as well as other metadata.
"""
raise NotImplementedError(
"Streaming implementation not provided. Please override the "
"`predict_stream` method on your model to generate streaming "
"predictions"
)
[docs]class ChatAgent(PythonModel, metaclass=ABCMeta):
"""
.. tip::
Since MLflow 3.0.0, we recommend using
:py:class:`ResponsesAgent <mlflow.pyfunc.ResponsesAgent>`
instead of :py:class:`ChatAgent <mlflow.pyfunc.ChatAgent>`.
**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 :py:class:`ChatModel <mlflow.pyfunc.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 :py:class:`ChatAgentRequest <mlflow.types.agent.ChatAgentRequest>` schema is similar to,
but not strictly compatible with the OpenAI ChatCompletion schema. ChatAgent adds additional
functionality and diverges from OpenAI
:py:class:`ChatCompletionRequest <mlflow.types.llm.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 :py:class:`ChatAgentResponse <mlflow.types.agent.ChatAgentResponse>` schema diverges from
:py:class:`ChatCompletionResponse <mlflow.types.llm.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 :py:class:`ChatAgentResponse <mlflow.types.agent.ChatAgentResponse>`
detailing a tool call:
.. code-block:: python
{
"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
:py:func:`ChatAgent.predict_stream <mlflow.pyfunc.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.
.. code-block:: python
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
:py:class:`ChatAgentRequest <mlflow.types.agent.ChatAgentRequest>` schema for ease of testing.
.. code-block:: python
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 :py:class:`ChatAgentState <mlflow.langchain.chat_agent_langgraph.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 <https://mlflow.org/docs/latest/model/models-from-code.html>`_ logging
approach.
.. code-block:: python
with mlflow.start_run():
logged_agent_info = mlflow.pyfunc.log_model(
name="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
:py:class:`ChatAgentRequest <mlflow.types.agent.ChatAgentRequest>` schema. Under the hood, it
will be converted into the python objects expected by your ``predict`` and ``predict_stream``
methods.
.. code-block:: python
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
:py:class:`ChatAgentRequest <mlflow.types.agent.ChatAgentRequest>` and
:py:class:`ChatAgentResponse <mlflow.types.agent.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
- Providing 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
:py:class:`ChatAgentRequest <mlflow.types.agent.ChatAgentRequest>` schema
- .. code-block:: python
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
:py:class:`List[ChatMessage] <mlflow.types.llm.ChatMessage>` and
:py:class:`ChatParams <mlflow.types.llm.ChatParams>` and outputs a
:py:class:`ChatCompletionResponse <mlflow.types.llm.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 :py:class:`ChatAgentResponse <mlflow.types.agent.ChatAgentResponse>` instead of a
:py:class:`ChatCompletionResponse <mlflow.types.llm.ChatCompletionResponse>`
For example, we can convert the ChatModel from the
`Chat Model Intro <https://mlflow.org/docs/latest/llms/chat-model-intro/index.html#building-your-first-chatmodel>`_
to a ChatAgent:
.. code-block:: python
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,
)
.. code-block:: python
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:
- LangGraph in the
:py:class:`ChatAgentState <mlflow.langchain.chat_agent_langgraph.ChatAgentState>` docstring
"""
_skip_type_hint_validation = True
def __init_subclass__(cls, **kwargs) -> None:
super().__init_subclass__(**kwargs)
for attr_name in ("predict", "predict_stream"):
attr = cls.__dict__.get(attr_name)
if callable(attr):
setattr(
cls,
attr_name,
wrap_non_list_predict_pydantic(
attr,
ChatAgentRequest,
"Invalid dictionary input for a ChatAgent. Expected a dictionary with the "
"ChatAgentRequest schema.",
unpack=True,
),
)
def _convert_messages_to_dict(self, messages: list[ChatAgentMessage]):
return [m.model_dump_compat(exclude_none=True) for m in messages]
# nb: We use `messages` instead of `model_input` so that the trace generated by default is
# compatible with mlflow evaluate. We also want `custom_inputs` to be a top level key for
# ease of use.
[docs] @abstractmethod
def predict(
self,
messages: list[ChatAgentMessage],
context: Optional[ChatContext] = None,
custom_inputs: Optional[dict[str, Any]] = None,
) -> ChatAgentResponse:
"""
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 :py:class:`ChatAgentRequest <mlflow.types.agent.ChatAgentRequest>` schema for ease
of testing.
.. code-block:: python
chat_agent = ChatAgent()
chat_agent.predict(
{
"messages": [{"role": "user", "content": "What is 10 + 10?"}],
"context": {"conversation_id": "123", "user_id": "456"},
}
)
Args:
messages (List[:py:class:`ChatAgentMessage <mlflow.types.agent.ChatAgentMessage>`]):
A list of :py:class:`ChatAgentMessage <mlflow.types.agent.ChatAgentMessage>`
objects representing the chat history.
context (:py:class:`ChatContext <mlflow.types.agent.ChatContext>`):
A :py:class:`ChatContext <mlflow.types.agent.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 :py:class:`ChatAgentResponse <mlflow.types.agent.ChatAgentResponse>` object containing
the model's response, as well as other metadata.
"""
# nb: We use `messages` instead of `model_input` so that the trace generated by default is
# compatible with mlflow evaluate. We also want `custom_inputs` to be a top level key for
# ease of use.
[docs] def predict_stream(
self,
messages: list[ChatAgentMessage],
context: Optional[ChatContext] = None,
custom_inputs: Optional[dict[str, Any]] = None,
) -> Generator[ChatAgentChunk, None, None]:
"""
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
:py:class:`ChatAgentRequest <mlflow.types.agent.ChatAgentRequest>`
schema for ease of testing.
.. code-block:: python
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
:py:class:`ChatAgentChunk <mlflow.types.agent.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
:py:class:`ChatAgentChunk <mlflow.types.agent.ChatAgentChunk>` should be aggregated by
the consuming client. See the example below.
.. code-block:: python
{"delta": {"role": "assistant", "content": "Born", "id": "123"}}
{"delta": {"role": "assistant", "content": " in", "id": "123"}}
{"delta": {"role": "assistant", "content": " data", "id": "123"}}
Args:
messages (List[:py:class:`ChatAgentMessage <mlflow.types.agent.ChatAgentMessage>`]):
A list of :py:class:`ChatAgentMessage <mlflow.types.agent.ChatAgentMessage>`
objects representing the chat history.
context (:py:class:`ChatContext <mlflow.types.agent.ChatContext>`):
A :py:class:`ChatContext <mlflow.types.agent.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 :py:class:`ChatAgentChunk <mlflow.types.agent.ChatAgentChunk>`
objects containing the model's response(s), as well as other metadata.
"""
raise NotImplementedError(
"Streaming implementation not provided. Please override the "
"`predict_stream` method on your model to generate streaming predictions"
)
def _check_compression_supported(compression):
if compression in _COMPRESSION_INFO:
return True
if compression:
supported = ", ".join(sorted(_COMPRESSION_INFO))
mlflow.pyfunc._logger.warning(
f"Unrecognized compression method '{compression}'"
f"Please select one of: {supported}. Falling back to uncompressed storage/loading."
)
return False
def _maybe_compress_cloudpickle_dump(python_model, path, compression):
file_open = _COMPRESSION_INFO.get(compression, {}).get("open", open)
with file_open(path, "wb") as out:
cloudpickle.dump(python_model, out)
def _maybe_decompress_cloudpickle_load(path, compression):
_check_compression_supported(compression)
file_open = _COMPRESSION_INFO.get(compression, {}).get("open", open)
with file_open(path, "rb") as f:
return cloudpickle.load(f)
if IS_PYDANTIC_V2_OR_NEWER:
from mlflow.types.responses import (
ResponsesAgentRequest,
ResponsesAgentResponse,
ResponsesAgentStreamEvent,
)
[docs] @experimental(version="3.0.0")
class ResponsesAgent(PythonModel, metaclass=ABCMeta):
"""
A base class for creating ResponsesAgent models. It can be used as a wrapper around any
agent framework to create an agent model that can be deployed to MLflow. Has a few helper
methods to help create output items that can be a part of a ResponsesAgentResponse or
ResponsesAgentStreamEvent.
See https://www.mlflow.org/docs/latest/llms/responses-agent-intro/ for more details.
"""
_skip_type_hint_validation = True
[docs] @staticmethod
def responses_agent_output_reducer(
chunks: list[Union[ResponsesAgentStreamEvent, dict[str, Any]]],
):
output_items = []
for chunk in chunks:
# Handle both dict and pydantic object formats
if isinstance(chunk, dict):
chunk_type = chunk.get("type")
if chunk_type == "response.output_item.done":
output_items.append(chunk.get("item"))
else:
# Pydantic object (ResponsesAgentStreamEvent)
if hasattr(chunk, "type") and chunk.type == "response.output_item.done":
output_items.append(chunk.item)
return ResponsesAgentResponse(output=output_items).model_dump(exclude_none=True)
def __init_subclass__(cls, **kwargs) -> None:
super().__init_subclass__(**kwargs)
for attr_name in ("predict", "predict_stream"):
attr = cls.__dict__.get(attr_name)
if callable(attr):
# Only apply trace decorator if it is not already traced with mlflow.trace
if getattr(attr, "__mlflow_traced__", False):
mlflow.pyfunc._logger.warning(
f"You have manually traced {attr_name} with @mlflow.trace, but this is "
"unnecessary with ResponsesAgent subclasses. You can remove the "
"@mlflow.trace decorator and it will be automatically traced."
)
traced_attr = attr
else:
# Apply trace decorator first
if attr_name == "predict_stream":
traced_attr = mlflow.trace(
span_type=SpanType.AGENT,
output_reducer=cls.responses_agent_output_reducer,
)(attr)
else:
traced_attr = mlflow.trace(span_type=SpanType.AGENT)(attr)
# Then wrap with pydantic wrapper
wrapped_attr = wrap_non_list_predict_pydantic(
traced_attr,
ResponsesAgentRequest,
"Invalid dictionary input for a ResponsesAgent. "
"Expected a dictionary with the ResponsesRequest schema.",
)
setattr(cls, attr_name, wrapped_attr)
[docs] @abstractmethod
def predict(self, request: ResponsesAgentRequest) -> ResponsesAgentResponse:
"""
Given a ResponsesAgentRequest, returns a ResponsesAgentResponse.
You can see example implementations at
https://www.mlflow.org/docs/latest/llms/responses-agent-intro#simple-chat-example and
https://www.mlflow.org/docs/latest/llms/responses-agent-intro#tool-calling-example.
"""
[docs] def predict_stream(
self, request: ResponsesAgentRequest
) -> Generator[ResponsesAgentStreamEvent, None, None]:
"""
Given a ResponsesAgentRequest, returns a generator of ResponsesAgentStreamEvent objects.
See more details at
https://www.mlflow.org/docs/latest/llms/responses-agent-intro#streaming-agent-output.
You can see example implementations at
https://www.mlflow.org/docs/latest/llms/responses-agent-intro#simple-chat-example and
https://www.mlflow.org/docs/latest/llms/responses-agent-intro#tool-calling-example.
"""
raise NotImplementedError(
"Streaming implementation not provided. Please override the "
"`predict_stream` method on your model to generate streaming predictions"
)
[docs] def create_text_delta(self, delta: str, item_id: str) -> dict[str, Any]:
"""Helper method to create a dictionary conforming to the text delta schema for
streaming.
Read more at https://www.mlflow.org/docs/latest/llms/responses-agent-intro/#streaming-agent-output.
"""
return {
"type": "response.output_text.delta",
"item_id": item_id,
"delta": delta,
}
[docs] def create_text_output_item(self, text: str, id: str) -> dict[str, Any]:
"""Helper method to create a dictionary conforming to the text output item schema.
Read more at https://www.mlflow.org/docs/latest/llms/responses-agent-intro/#creating-agent-output.
Args:
text (str): The text to be outputted.
id (str): The id of the output item.
"""
return {
"id": id,
"content": [
{
"text": text,
"type": "output_text",
}
],
"role": "assistant",
"type": "message",
}
[docs] def create_function_call_item(
self, id: str, call_id: str, name: str, arguments: str
) -> dict[str, Any]:
"""Helper method to create a dictionary conforming to the function call item schema.
Read more at https://www.mlflow.org/docs/latest/llms/responses-agent-intro/#creating-agent-output.
Args:
id (str): The id of the output item.
call_id (str): The id of the function call.
name (str): The name of the function to be called.
arguments (str): The arguments to be passed to the function.
"""
return {
"type": "function_call",
"id": id,
"call_id": call_id,
"name": name,
"arguments": arguments,
}
[docs] def create_function_call_output_item(self, call_id: str, output: str) -> dict[str, Any]:
"""Helper method to create a dictionary conforming to the function call output item
schema.
Read more at https://www.mlflow.org/docs/latest/llms/responses-agent-intro/#creating-agent-output.
Args:
call_id (str): The id of the function call.
output (str): The output of the function call.
"""
return {
"type": "function_call_output",
"call_id": call_id,
"output": output,
}
def _save_model_with_class_artifacts_params(
path,
python_model,
signature=None,
artifacts=None,
conda_env=None,
code_paths=None,
mlflow_model=None,
pip_requirements=None,
extra_pip_requirements=None,
model_config=None,
streamable=None,
model_code_path=None,
infer_code_paths=False,
):
"""
Args:
path: The path to which to save the Python model.
python_model: An instance of a subclass of :class:`~PythonModel`. ``python_model``
defines how the model loads artifacts and how it performs inference.
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, (e.g. {"file": "absolute_path"}).
``python_model`` can reference these resolved entries as the ``artifacts`` property
of the ``context`` attribute. If ``<artifact_name, 'hf:/repo_id'>``(e.g.
{"bert-tiny-model": "hf:/prajjwal1/bert-tiny"}) is provided, then the model can be
fetched from huggingface hub using repo_id `prajjwal1/bert-tiny` directly. If ``None``,
no artifacts are added to the model.
conda_env: Either a dictionary representation of a Conda environment or the path to a Conda
environment yaml file. If provided, this decsribes the environment this model should be
run in. At minimum, it should specify the dependencies contained in
:func:`get_default_conda_env()`. If ``None``, the default
:func:`get_default_conda_env()` environment is added to the model.
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 before the
model is loaded.
mlflow_model: The model to which to add the ``mlflow.pyfunc`` flavor.
model_config: The model configuration for the flavor. Model configuration is available
during model loading time.
.. Note:: Experimental: This parameter may change or be removed in a future release
without warning.
model_code_path: The path to the code that is being logged as a PyFunc model. Can be used
to load python_model when python_model is None.
.. Note:: Experimental: This parameter may change or be removed in a future release
without warning.
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.
"""
if mlflow_model is None:
mlflow_model = Model()
custom_model_config_kwargs = {
CONFIG_KEY_CLOUDPICKLE_VERSION: cloudpickle.__version__,
}
if callable(python_model):
python_model = _FunctionPythonModel(func=python_model, signature=signature)
saved_python_model_subpath = _SAVED_PYTHON_MODEL_SUBPATH
compression = MLFLOW_LOG_MODEL_COMPRESSION.get()
if compression:
if _check_compression_supported(compression):
custom_model_config_kwargs[CONFIG_KEY_COMPRESSION] = compression
saved_python_model_subpath += _COMPRESSION_INFO[compression]["ext"]
else:
compression = None
# If model_code_path is defined, we load the model into python_model, but we don't want to
# pickle/save the python_model since the module won't be able to be imported.
if not model_code_path:
try:
_maybe_compress_cloudpickle_dump(
python_model, os.path.join(path, saved_python_model_subpath), compression
)
except Exception as e:
raise MlflowException(
"Failed to serialize Python model. Please save the model into a python file "
"and use code-based logging method instead. See"
"https://mlflow.org/docs/latest/models.html#models-from-code for more information."
) from e
custom_model_config_kwargs[CONFIG_KEY_PYTHON_MODEL] = saved_python_model_subpath
if artifacts:
saved_artifacts_config = {}
with TempDir() as tmp_artifacts_dir:
saved_artifacts_dir_subpath = "artifacts"
hf_prefix = "hf:/"
for artifact_name, artifact_uri in artifacts.items():
if artifact_uri.startswith(hf_prefix):
try:
from huggingface_hub import snapshot_download
except ImportError as e:
raise MlflowException(
"Failed to import huggingface_hub. Please install huggingface_hub "
f"to log the model with artifact_uri {artifact_uri}. Error: {e}"
)
repo_id = artifact_uri[len(hf_prefix) :]
try:
snapshot_location = snapshot_download(
repo_id=repo_id,
local_dir=os.path.join(
path, saved_artifacts_dir_subpath, artifact_name
),
local_dir_use_symlinks=False,
)
except Exception as e:
raise MlflowException.invalid_parameter_value(
"Failed to download snapshot from Hugging Face Hub with artifact_uri: "
f"{artifact_uri}. Error: {e}"
)
saved_artifact_subpath = (
Path(snapshot_location).relative_to(Path(os.path.realpath(path))).as_posix()
)
else:
tmp_artifact_path = _download_artifact_from_uri(
artifact_uri=artifact_uri, output_path=tmp_artifacts_dir.path()
)
relative_path = (
Path(tmp_artifact_path)
.relative_to(Path(tmp_artifacts_dir.path()))
.as_posix()
)
saved_artifact_subpath = os.path.join(
saved_artifacts_dir_subpath, relative_path
)
saved_artifacts_config[artifact_name] = {
CONFIG_KEY_ARTIFACT_RELATIVE_PATH: saved_artifact_subpath,
CONFIG_KEY_ARTIFACT_URI: artifact_uri,
}
shutil.move(tmp_artifacts_dir.path(), os.path.join(path, saved_artifacts_dir_subpath))
custom_model_config_kwargs[CONFIG_KEY_ARTIFACTS] = saved_artifacts_config
if streamable is None:
streamable = python_model.__class__.predict_stream != PythonModel.predict_stream
if model_code_path:
loader_module = mlflow.pyfunc.loaders.code_model.__name__
elif python_model:
loader_module = _get_pyfunc_loader_module(python_model)
else:
raise MlflowException(
"Either `python_model` or `model_code_path` must be provided to save the model.",
error_code=INVALID_PARAMETER_VALUE,
)
mlflow.pyfunc.add_to_model(
model=mlflow_model,
loader_module=loader_module,
code=None,
conda_env=_CONDA_ENV_FILE_NAME,
python_env=_PYTHON_ENV_FILE_NAME,
model_config=model_config,
streamable=streamable,
model_code_path=model_code_path,
**custom_model_config_kwargs,
)
if size := get_total_file_size(path):
mlflow_model.model_size_bytes = size
# `mlflow_model.save` must be called before _validate_infer_and_copy_code_paths as it
# internally infers dependency, and MLmodel file is required to successfully load the model
mlflow_model.save(os.path.join(path, MLMODEL_FILE_NAME))
saved_code_subpath = _validate_infer_and_copy_code_paths(
code_paths,
path,
infer_code_paths,
mlflow.pyfunc.FLAVOR_NAME,
)
mlflow_model.flavors[mlflow.pyfunc.FLAVOR_NAME][mlflow.pyfunc.CODE] = saved_code_subpath
# `mlflow_model.code` is updated, re-generate `MLmodel` file.
mlflow_model.save(os.path.join(path, MLMODEL_FILE_NAME))
if conda_env is None:
if pip_requirements is None:
default_reqs = get_default_pip_requirements()
extra_env_vars = (
_get_databricks_serverless_env_vars()
if is_in_databricks_serverless_runtime()
else None
)
# To ensure `_load_pyfunc` can successfully load the model during the dependency
# inference, `mlflow_model.save` must be called beforehand to save an MLmodel file.
inferred_reqs = mlflow.models.infer_pip_requirements(
path,
mlflow.pyfunc.FLAVOR_NAME,
fallback=default_reqs,
extra_env_vars=extra_env_vars,
)
default_reqs = sorted(set(inferred_reqs).union(default_reqs))
else:
default_reqs = None
conda_env, pip_requirements, pip_constraints = _process_pip_requirements(
default_reqs,
pip_requirements,
extra_pip_requirements,
)
else:
conda_env, pip_requirements, pip_constraints = _process_conda_env(conda_env)
with open(os.path.join(path, _CONDA_ENV_FILE_NAME), "w") as f:
yaml.safe_dump(conda_env, stream=f, default_flow_style=False)
# Save `constraints.txt` if necessary
if pip_constraints:
write_to(os.path.join(path, _CONSTRAINTS_FILE_NAME), "\n".join(pip_constraints))
# Save `requirements.txt`
write_to(os.path.join(path, _REQUIREMENTS_FILE_NAME), "\n".join(pip_requirements))
_PythonEnv.current().to_yaml(os.path.join(path, _PYTHON_ENV_FILE_NAME))
def _load_context_model_and_signature(
model_path: str, model_config: Optional[dict[str, Any]] = None
):
pyfunc_config = _get_flavor_configuration(
model_path=model_path, flavor_name=mlflow.pyfunc.FLAVOR_NAME
)
signature = mlflow.models.Model.load(model_path).signature
if MODEL_CODE_PATH in pyfunc_config:
conf_model_code_path = pyfunc_config.get(MODEL_CODE_PATH)
model_code_path = os.path.join(model_path, os.path.basename(conf_model_code_path))
python_model = _load_model_code_path(model_code_path, model_config)
if callable(python_model):
python_model = _FunctionPythonModel(python_model, signature=signature)
else:
python_model_cloudpickle_version = pyfunc_config.get(CONFIG_KEY_CLOUDPICKLE_VERSION, None)
if python_model_cloudpickle_version is None:
mlflow.pyfunc._logger.warning(
"The version of CloudPickle used to save the model could not be found in the "
"MLmodel configuration"
)
elif python_model_cloudpickle_version != cloudpickle.__version__:
# CloudPickle does not have a well-defined cross-version compatibility policy. Micro
# version releases have been known to cause incompatibilities. Therefore, we match on
# the full library version
mlflow.pyfunc._logger.warning(
"The version of CloudPickle that was used to save the model, `CloudPickle %s`, "
"differs from the version of CloudPickle that is currently running, `CloudPickle "
"%s`, and may be incompatible",
python_model_cloudpickle_version,
cloudpickle.__version__,
)
python_model_compression = pyfunc_config.get(CONFIG_KEY_COMPRESSION, None)
python_model_subpath = pyfunc_config.get(CONFIG_KEY_PYTHON_MODEL, None)
if python_model_subpath is None:
raise MlflowException("Python model path was not specified in the model configuration")
python_model = _maybe_decompress_cloudpickle_load(
os.path.join(model_path, python_model_subpath), python_model_compression
)
artifacts = {}
for saved_artifact_name, saved_artifact_info in pyfunc_config.get(
CONFIG_KEY_ARTIFACTS, {}
).items():
artifacts[saved_artifact_name] = os.path.join(
model_path, saved_artifact_info[CONFIG_KEY_ARTIFACT_RELATIVE_PATH]
)
context = PythonModelContext(artifacts=artifacts, model_config=model_config)
python_model.load_context(context=context)
return context, python_model, signature
def _load_pyfunc(model_path: str, model_config: Optional[dict[str, Any]] = None):
context, python_model, signature = _load_context_model_and_signature(model_path, model_config)
return _PythonModelPyfuncWrapper(
python_model=python_model,
context=context,
signature=signature,
)
def _get_first_string_column(pdf):
iter_string_columns = (col for col, val in pdf.iloc[0].items() if isinstance(val, str))
return next(iter_string_columns, None)
class _PythonModelPyfuncWrapper:
"""
Wrapper class that creates a predict function such that
predict(model_input: pd.DataFrame) -> model's output as pd.DataFrame (pandas DataFrame)
"""
def __init__(self, python_model: PythonModel, context, signature):
"""
Args:
python_model: An instance of a subclass of :class:`~PythonModel`.
context: A :class:`~PythonModelContext` instance containing artifacts that
``python_model`` may use when performing inference.
signature: :class:`~ModelSignature` instance describing model input and output.
"""
self.python_model = python_model
self.context = context
self.signature = signature
def _convert_input(self, model_input):
hints = self.python_model.predict_type_hints
# we still need this for backwards compatibility
if isinstance(model_input, pd.DataFrame):
if _is_list_str(hints.input):
first_string_column = _get_first_string_column(model_input)
if first_string_column is None:
raise MlflowException.invalid_parameter_value(
"Expected model input to contain at least one string column"
)
return model_input[first_string_column].tolist()
elif _is_list_dict_str(hints.input):
if (
len(self.signature.inputs) == 1
and next(iter(self.signature.inputs)).name is None
):
if first_string_column := _get_first_string_column(model_input):
return model_input[[first_string_column]].to_dict(orient="records")
if len(model_input.columns) == 1:
return model_input.to_dict("list")[0]
return model_input.to_dict(orient="records")
elif isinstance(hints.input, type) and (
issubclass(hints.input, ChatCompletionRequest)
or issubclass(hints.input, SplitChatMessagesRequest)
):
# If the type hint is a RAG dataclass, we hydrate it
# If there are multiple rows, we should throw
if len(model_input) > 1:
raise MlflowException(
"Expected a single input for dataclass type hint, but got multiple rows"
)
# Since single input is expected, we take the first row
return _hydrate_dataclass(hints.input, model_input.iloc[0])
return model_input
def predict(self, model_input, params: Optional[dict[str, Any]] = None):
"""
Args:
model_input: Model input data as one of dict, str, bool, bytes, float, int, str type.
params: Additional parameters to pass to the model for inference.
Returns:
Model predictions as an iterator of chunks. The chunks in the iterator must be type of
dict or string. Chunk dict fields are determined by the model implementation.
"""
parameters = inspect.signature(self.python_model.predict).parameters
kwargs = {}
if "params" in parameters:
kwargs["params"] = params
else:
_log_warning_if_params_not_in_predict_signature(_logger, params)
if _is_context_in_predict_function_signature(parameters=parameters):
return self.python_model.predict(
self.context, self._convert_input(model_input), **kwargs
)
else:
return self.python_model.predict(self._convert_input(model_input), **kwargs)
def predict_stream(self, model_input, params: Optional[dict[str, Any]] = None):
"""
Args:
model_input: LLM Model single input.
params: Additional parameters to pass to the model for inference.
Returns:
Streaming predictions.
"""
parameters = inspect.signature(self.python_model.predict_stream).parameters
kwargs = {}
if "params" in parameters:
kwargs["params"] = params
else:
_log_warning_if_params_not_in_predict_signature(_logger, params)
if _is_context_in_predict_function_signature(parameters=parameters):
return self.python_model.predict_stream(
self.context, self._convert_input(model_input), **kwargs
)
else:
return self.python_model.predict_stream(self._convert_input(model_input), **kwargs)
def _get_pyfunc_loader_module(python_model):
if isinstance(python_model, ChatModel):
return mlflow.pyfunc.loaders.chat_model.__name__
elif isinstance(python_model, ChatAgent):
return mlflow.pyfunc.loaders.chat_agent.__name__
elif IS_PYDANTIC_V2_OR_NEWER and isinstance(python_model, ResponsesAgent):
return mlflow.pyfunc.loaders.responses_agent.__name__
return __name__
class ModelFromDeploymentEndpoint(PythonModel):
"""
A PythonModel wrapper for invoking an MLflow Deployments endpoint.
This class is particularly used for running evaluation against an MLflow Deployments endpoint.
"""
def __init__(self, endpoint, params):
self.endpoint = endpoint
self.params = params
def predict(
self, context, model_input: Union[pd.DataFrame, dict[str, Any], list[dict[str, Any]]]
):
"""
Run prediction on the input data.
Args:
context: A :class:`~PythonModelContext` instance containing artifacts that the model
can use to perform inference.
model_input: The input data for prediction, either of the following:
- Pandas DataFrame: If the default evaluator is used, input is a DF
that contains the multiple request payloads in a single column.
- A dictionary: If the model_type is "databricks-agents" and the
Databricks RAG evaluator is used, this PythonModel can be invoked
with a single dict corresponding to the ChatCompletionsRequest schema.
- A list of dictionaries: Currently we don't have any evaluator that
gives this input format, but we keep this for future use cases and
compatibility with normal pyfunc models.
Return:
The prediction result. The return type will be consistent with the model input type,
e.g., if the input is a Pandas DataFrame, the return will be a Pandas Series.
"""
if isinstance(model_input, dict):
return self._predict_single(model_input)
elif isinstance(model_input, list) and all(isinstance(data, dict) for data in model_input):
return [self._predict_single(data) for data in model_input]
elif isinstance(model_input, pd.DataFrame):
if len(model_input.columns) != 1:
raise MlflowException(
f"The number of input columns must be 1, but got {model_input.columns}. "
"Multi-column input is not supported for evaluating an MLflow Deployments "
"endpoint. Please include the input text or payload in a single column.",
error_code=INVALID_PARAMETER_VALUE,
)
input_column = model_input.columns[0]
predictions = [self._predict_single(data) for data in model_input[input_column]]
return pd.Series(predictions)
else:
raise MlflowException(
f"Invalid input data type: {type(model_input)}. The input data must be either "
"a Pandas DataFrame, a dictionary, or a list of dictionaries containing the "
"request payloads for evaluating an MLflow Deployments endpoint.",
error_code=INVALID_PARAMETER_VALUE,
)
def _predict_single(self, data: Union[str, dict[str, Any]]) -> dict[str, Any]:
"""
Send a single prediction request to the MLflow Deployments endpoint.
Args:
data: The single input data for prediction. If the input data is a string, we will
construct the request payload from it. If the input data is a dictionary, we
will directly use it as the request payload.
Returns:
The prediction result from the MLflow Deployments endpoint as a dictionary.
"""
from mlflow.metrics.genai.model_utils import call_deployments_api, get_endpoint_type
endpoint_type = get_endpoint_type(f"endpoints:/{self.endpoint}")
if isinstance(data, str):
# If the input payload is string, MLflow needs to construct the JSON
# payload based on the endpoint type. If the endpoint type is not
# set on the endpoint, we will default to chat format.
endpoint_type = endpoint_type or "llm/v1/chat"
prediction = call_deployments_api(self.endpoint, data, self.params, endpoint_type)
elif isinstance(data, dict):
# If the input is dictionary, we assume the input is already in the
# compatible format for the endpoint.
prediction = call_deployments_api(self.endpoint, data, self.params, endpoint_type)
else:
raise MlflowException(
f"Invalid input data type: {type(data)}. The feature column of the evaluation "
"dataset must contain only strings or dictionaries containing the request "
"payload for evaluating an MLflow Deployments endpoint.",
error_code=INVALID_PARAMETER_VALUE,
)
return prediction