mlflow.types

The mlflow.types module defines data types and utilities to be used by other mlflow components to describe interface independent of other frameworks or languages.

class mlflow.types.ColSpec(type: Union[mlflow.types.schema.Array, mlflow.types.schema.DataType, mlflow.types.schema.Map, mlflow.types.schema.Object, mlflow.types.schema.AnyType, str], name: Optional[str] = None, required: bool = True)

Bases: object

Specification of name and type of a single column in a dataset.

classmethod from_json_dict(**kwargs)

Deserialize from a json loaded dictionary. The dictionary is expected to contain type and optional name and required keys.

property name

The column name or None if the columns is unnamed.

property required

Whether this column is required.

Note

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

property type

The column data type.

class mlflow.types.DataType(value)

Bases: enum.Enum

MLflow data types.

binary = 7

Sequence of raw bytes.

boolean = 1

Logical data (True, False) .

datetime = 8

64b datetime data.

double = 5

64b floating point numbers.

float = 4

32b floating point numbers.

integer = 2

32b signed integer numbers.

long = 3

64b signed integer numbers.

string = 6

Text data.

to_numpy() → numpy.dtype

Get equivalent numpy data type.

to_pandas() → numpy.dtype

Get equivalent pandas data type.

to_python()

Get equivalent python data type.

class mlflow.types.ParamSchema(params: list)

Bases: object

Note

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

Specification of parameters applicable to the model. ParamSchema is represented as a list of ParamSpec.

classmethod from_json(json_str: str)

Deserialize from a json string.

property params

Representation of ParamSchema as a list of ParamSpec.

to_dict() → list

Serialize into a jsonable dictionary.

to_json() → str

Serialize into json string.

class mlflow.types.ParamSpec(name: str, dtype: Union[mlflow.types.schema.DataType, str], default: Optional[Union[mlflow.types.schema.DataType, list]], shape: Optional[tuple] = None)

Bases: object

Note

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

Specification used to represent parameters for the model.

class ParamSpecTypedDict

Bases: TypedDict

property default

Default value of the parameter.

property dtype

The parameter data type.

classmethod enforce_param_datatype(name, value, dtype: mlflow.types.schema.DataType)

Enforce the value matches the data type.

The following type conversions are allowed:

1. int -> long, float, double

2. long -> float, double

3. float -> double

4. any -> datetime (try conversion)

Any other type mismatch will raise error.

Parameters

• name – parameter name

• value – parameter value

• dtype – expected data type

classmethod from_json_dict(**kwargs)

Deserialize from a json loaded dictionary. The dictionary is expected to contain name, type and default keys.

property name

The name of the parameter.

property shape

The parameter shape. If shape is None, the parameter is a scalar.

classmethod validate_type_and_shape(spec: str, value: Optional[Union[mlflow.types.schema.DataType, list]], value_type: mlflow.types.schema.DataType, shape: Optional[tuple])

Validate that the value has the expected type and shape.

class mlflow.types.Schema(inputs: list)

Bases: object

Specification of a dataset.

Schema is represented as a list of ColSpec or TensorSpec. A combination of ColSpec and TensorSpec is not allowed.

The dataset represented by a schema can be named, with unique non empty names for every input. In the case of ColSpec, the dataset columns can be unnamed with implicit integer index defined by their list indices. Combination of named and unnamed data inputs are not allowed.

as_spark_schema()

Convert to Spark schema. If this schema is a single unnamed column, it is converted directly the corresponding spark data type, otherwise it’s returned as a struct (missing column names are filled with an integer sequence). Unsupported by TensorSpec.

classmethod from_json(json_str: str)

Deserialize from a json string.

has_input_names() → bool

Return true iff this schema declares names, false otherwise.

input_dict() → dict

Maps column names to inputs, iff this schema declares names.

input_names() → list

Get list of data names or range of indices if the schema has no names.

input_types() → list

Get types for each column in the schema.

input_types_dict() → dict

Maps column names to types, iff this schema declares names.

property inputs

Representation of a dataset that defines this schema.

is_tensor_spec() → bool

Return true iff this schema is specified using TensorSpec

numpy_types() → list

Convenience shortcut to get the datatypes as numpy types.

optional_input_names() → list

Note

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

Get list of optional data names or range of indices if schema has no names.

pandas_types() → list

Convenience shortcut to get the datatypes as pandas types. Unsupported by TensorSpec.

required_input_names() → list

Get list of required data names or range of indices if schema has no names.

to_dict() → list

Serialize into a jsonable dictionary.

to_json() → str

Serialize into json string.

class mlflow.types.TensorSpec(type: numpy.dtype, shape: Union[tuple, list], name: Optional[str] = None)

Bases: object

Specification used to represent a dataset stored as a Tensor.

classmethod from_json_dict(**kwargs)

Deserialize from a json loaded dictionary. The dictionary is expected to contain type and tensor-spec keys.

property name

The tensor name or None if the tensor is unnamed.

property required

Whether this tensor is required.

Note

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

property shape

The tensor shape

property type

A unique character code for each of the 21 different numpy built-in types. See https://numpy.org/devdocs/reference/generated/numpy.dtype.html#numpy.dtype for details.

class mlflow.types.llm.ChatChoice(message: mlflow.types.llm.ChatMessage, index: int = 0, finish_reason: str = 'stop', logprobs: Optional[mlflow.types.llm.ChatChoiceLogProbs] = None)

A single chat response generated by the model. ref: https://platform.openai.com/docs/api-reference/chat/object

Parameters

• message (ChatMessage) – The message that was generated.

• index (int) – The index of the response in the list of responses. Defaults to 0

• finish_reason (str) – The reason why generation stopped. Optional, defaults to "stop"

• logprobs (ChatChoiceLogProbs) – Log probability information for the choice. Optional, defaults to None

class mlflow.types.llm.ChatChoiceDelta(role: Optional[str] = 'assistant', content: Optional[str] = None, refusal: Optional[str] = None, name: Optional[str] = None, tool_calls: Optional[list] = None)

A streaming message delta in a chat response.

Parameters

• role (str) – The role of the entity that sent the message (e.g. "user", "system", "assistant", "tool"). Optional defaults to "assistant" This is optional because OpenAI clients can explicitly return None for the role

• content (str) – The content of the new token being streamed Optional Can be None on the last delta chunk or if refusal or tool_calls are provided

• refusal (str) – The refusal message content. Optional Supplied if a refusal response is provided.

• name (str) – The name of the entity that sent the message. Optional.

• tool_calls (List[ToolCall]) – A list of tool calls made by the model. Optional defaults to None

class mlflow.types.llm.ChatChoiceLogProbs(content: Optional[list] = None)

Log probability information for the choice.

Parameters

content – A list of message content tokens with log probability information.

class mlflow.types.llm.ChatChunkChoice(delta: mlflow.types.llm.ChatChoiceDelta, index: int = 0, finish_reason: Optional[str] = None, logprobs: Optional[mlflow.types.llm.ChatChoiceLogProbs] = None)

A single chat response chunk generated by the model. ref: https://platform.openai.com/docs/api-reference/chat/streaming

Parameters

• index (int) – The index of the response in the list of responses. defaults to 0

• delta (ChatChoiceDelta) – The streaming chunk message that was generated.

• finish_reason (str) – The reason why generation stopped. Optional, defaults to None

• logprobs (ChatChoiceLogProbs) – Log probability information for the choice. Optional, defaults to None

class mlflow.types.llm.ChatCompletionChunk(choices: list, usage: Optional[mlflow.types.llm.TokenUsageStats] = None, id: Optional[str] = None, model: Optional[str] = None, object: str = 'chat.completion.chunk', created: int = <factory>, custom_outputs: Optional[dict] = None)

The streaming chunk returned by the chat endpoint. ref: https://platform.openai.com/docs/api-reference/chat/streaming

Parameters

• choices (List[ChatChunkChoice]) – A list of ChatChunkChoice objects containing the generated chunk of a streaming response

• usage (TokenUsageStats) – An object describing the tokens used by the request. Optional, defaults to None.

• id (str) – The ID of the response. Optional, defaults to None

• model (str) – The name of the model used. Optional, defaults to None

• object (str) – The object type. Defaults to ‘chat.completion.chunk’

• created (int) – The time the response was created. Optional, defaults to the current time.

• custom_outputs (Dict[str, Any]) – An field that can contain arbitrary additional context. The dictionary values must be JSON-serializable. Optional, defaults to None

class mlflow.types.llm.ChatCompletionRequest(temperature: float = 1.0, max_tokens: Optional[int] = None, stop: Optional[list] = None, n: int = 1, stream: bool = False, top_p: Optional[float] = None, top_k: Optional[int] = None, frequency_penalty: Optional[float] = None, presence_penalty: Optional[float] = None, custom_inputs: Optional[dict] = None, tools: Optional[list] = None, messages: list = <factory>)

Format of the request object expected by the chat endpoint.

Parameters

• messages (List[ChatMessage]) – A list of ChatMessage that will be passed to the model. Optional, defaults to empty list ([])

• temperature (float) – A param used to control randomness and creativity during inference. Optional, defaults to 1.0

• max_tokens (int) – The maximum number of new tokens to generate. Optional, defaults to None (unlimited)

• stop (List[str]) – A list of tokens at which to stop generation. Optional, defaults to None

• n (int) – The number of responses to generate. Optional, defaults to 1

• stream (bool) – Whether to stream back responses as they are generated. Optional, defaults to False

• top_p (float) – An optional param to control sampling with temperature, the model considers the results of the tokens with top_p probability mass. E.g., 0.1 means only the tokens comprising the top 10% probability mass are considered.

• top_k (int) – An optional param for reducing the vocabulary size to top k tokens (sorted in descending order by their probabilities).

• frequency_penalty – (float): An optional param of positive or negative value, positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model’s likelihood to repeat the same line verbatim.

• presence_penalty – (float): An optional param of positive or negative value, positive values penalize new tokens based on whether they appear in the text so far, increasing the model’s likelihood to talk about new topics.

• custom_inputs (Dict[str, Any]) – An optional param to provide arbitrary additional context to the model. The dictionary values must be JSON-serializable.

• tools (List[ToolDefinition]) – An optional list of tools that can be called by the model.

Warning

In an upcoming MLflow release, default values for temperature, n and stream will be removed. Please provide these values explicitly in your code if needed.

class mlflow.types.llm.ChatCompletionResponse(choices: list, usage: Optional[mlflow.types.llm.TokenUsageStats] = None, id: Optional[str] = None, model: Optional[str] = None, object: str = 'chat.completion', created: int = <factory>, custom_outputs: Optional[dict] = None)

The full response object returned by the chat endpoint.

Parameters

• choices (List[ChatChoice]) – A list of ChatChoice objects containing the generated responses

• usage (TokenUsageStats) – An object describing the tokens used by the request. Optional, defaults to None.

• id (str) – The ID of the response. Optional, defaults to None

• model (str) – The name of the model used. Optional, defaults to None

• object (str) – The object type. Defaults to ‘chat.completion’

• created (int) – The time the response was created. Optional, defaults to the current time.

• custom_outputs (Dict[str, Any]) – An field that can contain arbitrary additional context. The dictionary values must be JSON-serializable. Optional, defaults to None

class mlflow.types.llm.ChatMessage(role: str, content: Optional[str] = None, refusal: Optional[str] = None, name: Optional[str] = None, tool_calls: Optional[list] = None, tool_call_id: Optional[str] = None)

A message in a chat request or response.

Parameters

• role (str) – The role of the entity that sent the message (e.g. "user", "system", "assistant", "tool").

• content (str) – The content of the message. Optional Can be None if refusal or tool_calls are provided.

• refusal (str) – The refusal message content. Optional Supplied if a refusal response is provided.

• name (str) – The name of the entity that sent the message. Optional.

• tool_calls (List[ToolCall]) – A list of tool calls made by the model. Optional defaults to None

• tool_call_id (str) – The ID of the tool call that this message is a response to. Optional defaults to None

class mlflow.types.llm.ChatParams(temperature: float = 1.0, max_tokens: Optional[int] = None, stop: Optional[list] = None, n: int = 1, stream: bool = False, top_p: Optional[float] = None, top_k: Optional[int] = None, frequency_penalty: Optional[float] = None, presence_penalty: Optional[float] = None, custom_inputs: Optional[dict] = None, tools: Optional[list] = None)

Common parameters used for chat inference

Parameters

• temperature (float) – A param used to control randomness and creativity during inference. Optional, defaults to 1.0

• max_tokens (int) – The maximum number of new tokens to generate. Optional, defaults to None (unlimited)

• stop (List[str]) – A list of tokens at which to stop generation. Optional, defaults to None

• n (int) – The number of responses to generate. Optional, defaults to 1

• stream (bool) – Whether to stream back responses as they are generated. Optional, defaults to False

• top_p (float) – An optional param to control sampling with temperature, the model considers the results of the tokens with top_p probability mass. E.g., 0.1 means only the tokens comprising the top 10% probability mass are considered.

• top_k (int) – An optional param for reducing the vocabulary size to top k tokens (sorted in descending order by their probabilities).

• frequency_penalty – (float): An optional param of positive or negative value, positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model’s likelihood to repeat the same line verbatim.

• presence_penalty – (float): An optional param of positive or negative value, positive values penalize new tokens based on whether they appear in the text so far, increasing the model’s likelihood to talk about new topics.

• custom_inputs (Dict[str, Any]) – An optional param to provide arbitrary additional context to the model. The dictionary values must be JSON-serializable.

• tools (List[ToolDefinition]) – An optional list of tools that can be called by the model.

Warning

In an upcoming MLflow release, default values for temperature, n and stream will be removed. Please provide these values explicitly in your code if needed.

classmethod keys() → set

Return the keys of the dataclass

class mlflow.types.llm.FunctionToolCallArguments(name: str, arguments: str)

The arguments of a function tool call made by the model.

Parameters

• arguments (str) – A JSON string of arguments that should be passed to the tool.

• name (str) – The name of the tool that is being called.

class mlflow.types.llm.FunctionToolDefinition(name: str, description: Optional[str] = None, parameters: Optional[mlflow.types.llm.ToolParamsSchema] = None, strict: bool = False)

Definition for function tools (currently the only supported type of tool).

Parameters

• name (str) – The name of the tool.

• description (str) – A description of what the tool does, and how it should be used. Optional, defaults to None

• parameters – A mapping of parameter names to their definitions. If not provided, this defines a function without parameters. Optional, defaults to None

• strict (bool) – A flag that represents whether or not the model should strictly follow the schema provided.

to_tool_definition()

Convenience function for wrapping this in a ToolDefinition

class mlflow.types.llm.ParamProperty(type: Literal[string, number, integer, object, array, boolean, null], description: Optional[str] = None, enum: Optional[list] = None, items: Optional[mlflow.types.llm.ParamType] = None)

A single parameter within a function definition.

Parameters

• type (str) – The type of the parameter. Possible values are “string”, “number”, “integer”, “object”, “array”, “boolean”, or “null”, conforming to the JSON Schema specification.

• description (str) – A description of the parameter. Optional, defaults to None

• enum (List[str]) – Used to constrain the possible values for the parameter. Optional, defaults to None

• items (ParamProperty) – If the param is of array type, this field can be used to specify the type of its items. Optional, defaults to None

class mlflow.types.llm.ParamType(type: Literal[string, number, integer, object, array, boolean, null])

class mlflow.types.llm.TokenLogProb(token: str, logprob: float, top_logprobs: list, bytes: Optional[list] = None)

Message content token with log probability information.

Parameters

• token – The token.

• logprob – The log probability of this token, if it is within the top 20 most likely tokens. Otherwise, the value -9999.0 is used to signify that the token is very unlikely.

• bytes – A list of integers representing the UTF-8 bytes representation of the token. Useful in instances where characters are represented by multiple tokens and their byte representations must be combined to generate the correct text representation. Can be null if there is no bytes representation for the token.

• top_logprobs – List of the most likely tokens and their log probability, at this token position. In rare cases, there may be fewer than the number of requested top_logprobs returned.

class mlflow.types.llm.TokenUsageStats(prompt_tokens: Optional[int] = None, completion_tokens: Optional[int] = None, total_tokens: Optional[int] = None)

Stats about the number of tokens used during inference.

Parameters

• prompt_tokens (int) – The number of tokens in the prompt. Optional, defaults to None

• completion_tokens (int) – The number of tokens in the generated completion. Optional, defaults to None

• total_tokens (int) – The total number of tokens used. Optional, defaults to None

class mlflow.types.llm.ToolCall(function: mlflow.types.llm.FunctionToolCallArguments, id: str = <factory>, type: str = 'function')

A tool call made by the model.

Parameters

• function (FunctionToolCallArguments) – The arguments of the function tool call.

• id (str) – The ID of the tool call. Defaults to a random UUID.

• type (str) – The type of the object. Defaults to “function”.

class mlflow.types.llm.ToolDefinition(function: mlflow.types.llm.FunctionToolDefinition, type: Literal[function] = 'function')

Definition for tools that can be called by the model.

Parameters

• function (FunctionToolDefinition) – The definition of a function tool.

• type (str) – The type of the tool. Currently only “function” is supported.

class mlflow.types.llm.ToolParamsSchema(properties: dict, type: Literal[object] = 'object', required: Optional[list] = None, additionalProperties: Optional[bool] = None)

A tool parameter definition.

Parameters

• properties (Dict[str, ParamProperty]) – A mapping of parameter names to their definitions.

• type (str) – The type of the parameter. Currently only “object” is supported.

• required (List[str]) – A list of required parameter names. Optional, defaults to None

• additionalProperties (bool) – Whether additional properties are allowed in the object. Optional, defaults to None

class mlflow.types.llm.TopTokenLogProb(token: str, logprob: float, bytes: Optional[list] = None)

Token and its log probability.

Parameters

• token – The token.

• logprob – The log probability of this token, if it is within the top 20 most likely tokens. Otherwise, the value -9999.0 is used to signify that the token is very unlikely.

• bytes – A list of integers representing the UTF-8 bytes representation of the token. Useful in instances where characters are represented by multiple tokens and their byte representations must be combined to generate the correct text representation. Can be null if there is no bytes representation for the token.

class mlflow.types.chat.AudioContentPart(*, type: Literal[input_audio], input_audio: mlflow.types.chat.InputAudio)

class mlflow.types.chat.BaseModel

class mlflow.types.chat.ChatTool(*, type: Literal[function, uc_function], function: Optional[mlflow.types.chat.FunctionToolDefinition] = None, uc_function: Optional[mlflow.types.chat.UnityCatalogFunctionToolDefinition] = None)

class mlflow.types.chat.ChatTools(*, tools: Optional[list] = None)

class mlflow.types.chat.Function(*, name: str, arguments: str)

class mlflow.types.chat.FunctionParams(*, properties: dict, type: Literal[object] = 'object', required: Optional[list] = None, additionalProperties: Optional[bool] = None)

class mlflow.types.chat.FunctionToolDefinition(*, name: str, description: Optional[str] = None, parameters: Optional[mlflow.types.chat.FunctionParams] = None, strict: Optional[bool] = None)

class mlflow.types.chat.ImageContentPart(*, type: Literal[image_url], image_url: mlflow.types.chat.ImageUrl)

class mlflow.types.chat.ImageUrl(*, url: str, detail: Literal[auto, low, high])

class mlflow.types.chat.InputAudio(*, data: str, format: Literal[wav, mp3])

class mlflow.types.chat.ParamProperty(*, type: Literal[string, number, integer, object, array, boolean, null], description: Optional[str] = None, enum: Optional[list] = None, items: Optional[mlflow.types.chat.ParamType] = None)

class mlflow.types.chat.ParamType(*, type: Literal[string, number, integer, object, array, boolean, null])

class mlflow.types.chat.RequestMessage(*, role: str, content: Optional[Union[str, list]] = None, tool_calls: Optional[list] = None, tool_call_id: Optional[str] = None, refusal: Optional[str] = None)

A chat request. content can be a string, or an array of content parts.

A content part is one of the following:

• TextContentPart

• ImageContentPart

• AudioContentPart

class mlflow.types.chat.TextContentPart(*, type: Literal[text], text: str)

class mlflow.types.chat.ToolCall(*, id: str, type: Literal[function], function: mlflow.types.chat.Function)

class mlflow.types.chat.UnityCatalogFunctionToolDefinition(*, name: str)

class mlflow.types.schema.AnyType

Note

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

to_dict()

Dictionary representation of the object.

class mlflow.types.schema.Array(dtype: Union[mlflow.types.schema.Array, mlflow.types.schema.DataType, mlflow.types.schema.Map, mlflow.types.schema.Object, mlflow.types.schema.AnyType, str])

Specification used to represent a json-convertible array.

property dtype

The array data type.

classmethod from_json_dict(**kwargs)

Deserialize from a json loaded dictionary. The dictionary is expected to contain type and items keys. Example: {“type”: “array”, “items”: “string”}

to_dict()

Dictionary representation of the object.

class mlflow.types.schema.Map(value_type: Union[mlflow.types.schema.Array, mlflow.types.schema.DataType, mlflow.types.schema.Map, mlflow.types.schema.Object, mlflow.types.schema.AnyType, str])

Specification used to represent a json-convertible map with string type keys.

classmethod from_json_dict(**kwargs)

Deserialize from a json loaded dictionary. The dictionary is expected to contain type and values keys. Example: {“type”: “map”, “values”: “string”}

to_dict()

Dictionary representation of the object.

property value_type

class mlflow.types.schema.Object(properties: list)

Specification used to represent a json-convertible object.

classmethod from_json_dict(**kwargs)

Deserialize from a json loaded dictionary. The dictionary is expected to contain type and properties keys. Example: {“type”: “object”, “properties”: {“property_name”: {“type”: “string”}}}

property properties

The list of object properties

to_dict()

Dictionary representation of the object.

class mlflow.types.schema.Property(name: str, dtype: Union[mlflow.types.schema.Array, mlflow.types.schema.DataType, mlflow.types.schema.Map, mlflow.types.schema.Object, mlflow.types.schema.AnyType, str], required: bool = True)

Specification used to represent a json-convertible object property.

property dtype

The property data type.

classmethod from_json_dict(**kwargs)

Deserialize from a json loaded dictionary. The dictionary is expected to contain only one key as name, and the value should be a dictionary containing type and optional required keys. Example: {“property_name”: {“type”: “string”, “required”: True}}

property name

The property name.

property required

Whether this property is required

to_dict()

Dictionary representation of the object.