Hugging Face

Bases: Engine[PromptSignature, Result, Model, InferenceMode]

Source code in sieves/engines/huggingface_.py

class HuggingFace(Engine[PromptSignature, Result, Model, InferenceMode]):
    @property
    def inference_modes(self) -> type[InferenceMode]:
        return InferenceMode

    @property
    def supports_few_shotting(self) -> bool:
        return True

    def build_executable(
        self,
        inference_mode: InferenceMode,
        prompt_template: str | None,
        prompt_signature: type[PromptSignature] | PromptSignature,
        fewshot_examples: Iterable[pydantic.BaseModel] = (),
    ) -> Executable[Result | None]:
        cls_name = self.__class__.__name__
        assert prompt_template, ValueError(f"prompt_template has to be provided to {cls_name} engine by task.")
        assert isinstance(prompt_signature, list)

        # Render template with few-shot examples. Note that we don't use extracted document values here, as HF zero-shot
        # pipelines only support one hypothesis template per call - and we want to batch, so our hypothesis template
        # will be document-invariant.
        fewshot_examples_dict = HuggingFace._convert_fewshot_examples(fewshot_examples)
        # Render hypothesis template with everything but text.
        template = jinja2.Template(prompt_template).render(**({"examples": fewshot_examples_dict}))

        def execute(values: Iterable[dict[str, Any]]) -> Iterable[Result]:
            """Execute prompts with engine for given values.
            :param values: Values to inject into prompts.
            :return Iterable[Result]: Results for prompts.
            """
            match inference_mode:
                case InferenceMode.zeroshot_cls:
                    batch_size = self._batch_size if self._batch_size != -1 else sys.maxsize
                    # Ensure values are read as generator for standardized batch handling (otherwise we'd have to use
                    # different batch handling depending on whether lists/tuples or generators are used).
                    values = (v for v in values)

                    while batch := [vals["text"] for vals in itertools.islice(values, batch_size)]:
                        if len(batch) == 0:
                            break

                        yield from self._model(
                            batch,
                            prompt_signature,
                            hypothesis_template=template,
                            multi_label=True,
                            **self._inference_kwargs,
                        )

                case _:
                    raise ValueError(f"Inference mode {inference_mode} not supported by {cls_name} engine.")

        return execute

_attributes property

Returns attributes to serialize.

Returns:

Type	Description
dict[str, Attribute]	Dict of attributes to serialize.

model property

Return model instance.

Returns:

Type	Description
EngineModel	Model instance.

__init__(model, init_kwargs=None, inference_kwargs=None, strict_mode=False, batch_size=-1)

Parameters:

Name	Type	Description	Default
model	EngineModel	Instantiated model instance.	required
init_kwargs	dict[str, Any] | None	Optional kwargs to supply to engine executable at init time.	None
inference_kwargs	dict[str, Any] | None	Optional kwargs to supply to engine executable at inference time.	None
strict_mode	bool	If True, exception is raised if prompt response can't be parsed correctly.	False
batch_size	int	Batch size in processing prompts. -1 will batch all documents in one go. Not all engines support batching.	-1

Source code in sieves/engines/core.py

def __init__(
    self,
    model: EngineModel,
    init_kwargs: dict[str, Any] | None = None,
    inference_kwargs: dict[str, Any] | None = None,
    strict_mode: bool = False,
    batch_size: int = -1,
):
    """
    :param model: Instantiated model instance.
    :param init_kwargs: Optional kwargs to supply to engine executable at init time.
    :param inference_kwargs: Optional kwargs to supply to engine executable at inference time.
    :param strict_mode: If True, exception is raised if prompt response can't be parsed correctly.
    :param batch_size: Batch size in processing prompts. -1 will batch all documents in one go. Not all engines
        support batching.
    """
    self._model = model
    self._inference_kwargs = inference_kwargs or {}
    self._init_kwargs = init_kwargs or {}
    self._strict_mode = strict_mode
    self._batch_size = self._validate_batch_size(batch_size)

_convert_fewshot_examples(fewshot_examples) staticmethod

Convert fewshot examples from pydantic.BaseModel instance to dicts.

Parameters:

Name	Type	Description	Default
fewshot_examples	Iterable[BaseModel]	Fewshot examples to convert.	required

Returns:

Type	Description
list[dict[str, Any]]	Fewshot examples as dicts.

Source code in sieves/engines/core.py

@staticmethod
def _convert_fewshot_examples(fewshot_examples: Iterable[pydantic.BaseModel]) -> list[dict[str, Any]]:
    """
    Convert fewshot examples from pydantic.BaseModel instance to dicts.
    :param fewshot_examples: Fewshot examples to convert.
    :return: Fewshot examples as dicts.
    """
    return [fs_example.model_dump(serialize_as_any=True) for fs_example in fewshot_examples]

_execute_async_calls(calls) async staticmethod

Executes batch of async functions.

Parameters:

Name	Type	Description	Default
calls	list[Coroutine[Any, Any, Any]] | list[Awaitable[Any]]	Async calls to execute.	required

Returns:

Type	Description
Any	Parsed response objects.

Source code in sieves/engines/core.py

@staticmethod
async def _execute_async_calls(calls: list[Coroutine[Any, Any, Any]] | list[Awaitable[Any]]) -> Any:
    """Executes batch of async functions.
    :param calls: Async calls to execute.
    :return: Parsed response objects.
    """
    return await asyncio.gather(*calls)

_validate_batch_size(batch_size)

Validates batch_size. Noop by default.

Parameters:

Name	Type	Description	Default
batch_size	int	Specified batch size.	required

Returns:

Type	Description
int	Validated batch size.

Source code in sieves/engines/core.py

def _validate_batch_size(self, batch_size: int) -> int:
    """Validates batch_size. Noop by default.
    :param batch_size: Specified batch size.
    :returns int: Validated batch size.
    """
    return batch_size

deserialize(config, **kwargs) classmethod

Generate Engine instance from config.

Parameters:

Name	Type	Description	Default
config	Config	Config to generate instance from.	required
kwargs	dict[str, Any]	Values to inject into loaded config.	{}

Returns:

Type	Description
Engine[EnginePromptSignature, EngineResult, EngineModel, EngineInferenceMode]	Deserialized Engine instance.

Source code in sieves/engines/core.py

@classmethod
def deserialize(
    cls, config: Config, **kwargs: dict[str, Any]
) -> Engine[EnginePromptSignature, EngineResult, EngineModel, EngineInferenceMode]:
    """Generate Engine instance from config.
    :param config: Config to generate instance from.
    :param kwargs: Values to inject into loaded config.
    :return: Deserialized Engine instance.
    """
    return cls(**config.to_init_dict(cls, **kwargs))

serialize()

Serializes engine.

Returns:

Type	Description
Config	Config instance.

Source code in sieves/engines/core.py

def serialize(self) -> Config:
    """Serializes engine.
    :return: Config instance.
    """
    return Config.create(self.__class__, self._attributes)