morpheus.messages.multi_inference_message.MultiInferenceMessage

Bases: morpheus.messages.multi_tensor_message.MultiTensorMessage

This is a container class that holds the InferenceMemory container and the metadata of the data contained within it. Builds on top of the MultiTensorMessage class to add additional data for inferencing.

This class requires two separate memory blocks for a batch. One for the message metadata (i.e., start time, IP address, etc.) and another for the raw inference inputs (i.e., input_ids, seq_ids). Since there can be more inference input requests than messages (This happens when some messages get broken into multiple inference requests) this class stores two different offset and count values. mess_offset and mess_count refer to the offset and count in the message metadata batch and offset and count index into the inference batch data.

Attributes

id

Returns ID column values from morpheus.pipeline.messages.MessageMeta.df as list.

id_col

Returns ID column values from morpheus.pipeline.messages.MessageMeta.df.

inputs

Get inputs stored in the InferenceMemory container.

tensors

Get tensors stored in the TensorMemory container sliced according to offset and count.

timestamp

Returns timestamp column values from morpheus.messages.MessageMeta.df as list.

Methods

copy_meta_ranges(ranges, mask=None)[source]

Perform a copy of the underlying dataframe for the given ranges of rows.

Parameters

rangestyping.List[typing.Tuple[int, int]]

Rows to include in the copy in the form of [(`start_row, stop_row),…]` The stop_row isn’t included. For example to copy rows 1-2 & 5-7 ranges=[(1, 3), (5, 8)]

masktyping.Union[None, cupy.ndarray, numpy.ndarray]

Optionally specify rows as a cupy array (when using cudf Dataframes) or a numpy array (when using pandas Dataframes) of booleans. When not-None ranges will be ignored. This is useful as an optimization as this avoids needing to generate the mask on it’s own.

Returns

Dataframe

copy_ranges(ranges)[source]

Perform a copy of the current message, dataframe and tensors for the given ranges of rows.

Parameters

rangestyping.List[typing.Tuple[int, int]]

Rows to include in the copy in the form of [(`start_row, stop_row),…]` The stop_row isn’t included. For example to copy rows 1-2 & 5-7 ranges=[(1, 3), (5, 8)]

——-

`MultiTensorMessage`

copy_tensor_ranges(ranges, mask=None)[source]

Perform a copy of the underlying tensor tensors for the given ranges of rows.

Parameters

rangestyping.List[typing.Tuple[int, int]]

Rows to include in the copy in the form of [(`start_row, stop_row),…]` The stop_row isn’t included. For example to copy rows 1-2 & 5-7 ranges=[(1, 3), (5, 8)]

masktyping.Union[None, cupy.ndarray, numpy.ndarray]

Optionally specify rows as a cupy array (when using cudf Dataframes) or a numpy array (when using pandas Dataframes) of booleans. When not-None ranges will be ignored. This is useful as an optimization as this avoids needing to generate the mask on it’s own.

Returns

typing.Dict[str, cupy.ndarray]

classmethod from_message(message, *, meta=None, mess_offset=- 1, mess_count=- 1, memory=None, offset=- 1, count=- 1, **kwargs)[source]

Creates a new instance of a derived class from MultiMessage using an existing message as the template. This is very useful when a new message needs to be created with a single change to an existing MessageMeta.

When creating the new message, all required arguments for the class specified by cls will be pulled from message unless otherwise specified in the args or kwargs. Special handling is performed depending on whether or not a new meta object is supplied. If one is supplied, the offset and count defaults will be 0 and meta.count respectively. Otherwise offset and count will be pulled from the input message.

Parameters

clstyping.Type[Self]

The class to create

messageMultiMessage

An existing message to use as a template. Can be a base or derived from cls as long as all arguments can be pulled from message or proveded in kwargs

metaMessageMeta, optional

A new MessageMeta to use, by default None

mess_offsetint, optional

A new mess_offset to use, by default -1

mess_countint, optional

A new mess_count to use, by default -1

memoryTensorMemory, optional

A new TensorMemory to use. If supplied, offset and count default to 0 and memory.count respectively. By default None

offsetint, optional

A new offset to use, by default -1

countint, optional

A new count to use, by default -1

Returns

Self

A new instance of type cls

Raises

ValueError

If the incoming message is None

get_id_tensor()[source]

Get the tensor that holds message ID information. Equivalent to get_tensor(id_tensor_name)

Returns

cupy.ndarray

Array containing the ID information

Raises

KeyError

If self.id_tensor_name is not found in the tensors

get_input(name)[source]

Get input stored in the InferenceMemory container. Alias for MultiInferenceMessage.get_tensor.

Parameters

namestr

Input key name.

Returns

cupy.ndarray

Inference input.

Raises

KeyError

When no matching input tensor exists.

get_meta(columns=None)[source]

Return column values from morpheus.pipeline.messages.MessageMeta.df.

Parameters

columnstyping.Union[None, str, typing.List[str]]

Input column names. Returns all columns if None is specified. When a string is passed, a Series is returned. Otherwise a Dataframe is returned.

Returns

Series or Dataframe

Column values from the dataframe.

get_meta_list(col_name=None)[source]

Return a column values from morpheus.messages.MessageMeta.df as a list.

Parameters

col_namestr

Column name in the dataframe.

Returns

List[str]

Column values from the dataframe.

get_slice(start, stop)[source]

Perform a slice of the current message from start:stop (excluding stop)

For example to slice from rows 1-3 use m.get_slice(1, 4). The returned MultiTensorMessage will contain references to the same underlying Dataframe and tensor tensors, and this calling this method is reletively low cost compared to MultiTensorMessage.copy_ranges

Parameters

startint

Starting row of the slice

stopint

Stop of the slice

——-

`MultiTensorMessage`

get_tensor(name)[source]

Get tensor stored in the TensorMemory container.

Parameters

namestr

tensor key name.

Returns

cupy.ndarray

Inference tensor.

property id: List[int]

Returns ID column values from morpheus.pipeline.messages.MessageMeta.df as list.

Returns

List[int]

ID column values from the dataframe as list.

property id_col

Returns ID column values from morpheus.pipeline.messages.MessageMeta.df.

Returns

pandas.Series

ID column values from the dataframe.

property inputs

Get inputs stored in the InferenceMemory container. Alias for MultiInferenceMessage.tensors.

Returns

cupy.ndarray

Inference inputs.

set_meta(columns, value)[source]

Set column values to morpheus.pipelines.messages.MessageMeta.df.

Parameters

columnstyping.Union[None, str, typing.List[str]]

Input column names. Sets the value for the corresponding column names. If None is specified, all columns will be used. If the column does not exist, a new one will be created.

valueAny

Value to apply to the specified columns. If a single value is passed, it will be broadcast to all rows. If a Series or Dataframe is passed, rows will be matched by index.

property tensors

Get tensors stored in the TensorMemory container sliced according to offset and count.

Returns

cupy.ndarray

Inference tensors.

property timestamp: List[int]

Returns timestamp column values from morpheus.messages.MessageMeta.df as list.

Returns

List[int]

Timestamp column values from the dataframe as list.