Hugging Face

This section contains the reference for Hugging Face integrations:

InferenceEndpointsLLM

Bases: AsyncLLM

InferenceEndpoints LLM implementation running the async API client.

This LLM will internally use huggingface_hub.AsyncInferenceClient.

Attributes:

Name	Type	Description
model_id	Optional[str]	the model ID to use for the LLM as available in the Hugging Face Hub, which will be used to resolve the base URL for the serverless Inference Endpoints API requests. Defaults to None.
endpoint_name	Optional[RuntimeParameter[str]]	the name of the Inference Endpoint to use for the LLM. Defaults to None.
endpoint_namespace	Optional[RuntimeParameter[str]]	the namespace of the Inference Endpoint to use for the LLM. Defaults to None.
base_url	Optional[RuntimeParameter[str]]	the base URL to use for the Inference Endpoints API requests.
api_key	Optional[RuntimeParameter[SecretStr]]	the API key to authenticate the requests to the Inference Endpoints API.
tokenizer_id	Optional[str]	the tokenizer ID to use for the LLM as available in the Hugging Face Hub. Defaults to None, but defining one is recommended to properly format the prompt.
model_display_name	Optional[str]	the model display name to use for the LLM. Defaults to None.
structured_output	Optional[RuntimeParameter[StructuredOutputType]]	a dictionary containing the structured output configuration or if more fine-grained control is needed, an instance of OutlinesStructuredOutput. Defaults to None.

Icon

:hugging:

Examples:

Free serverless Inference API:

```python
from distilabel.llms.huggingface import InferenceEndpointsLLM

llm = InferenceEndpointsLLM(
    model_id="mistralai/Mistral-7B-Instruct-v0.2",
)

llm.load()

output = llm.generate(inputs=[[{"role": "user", "content": "Hello world!"}]])
```

Dedicated Inference Endpoints:

```python
from distilabel.llms.huggingface import InferenceEndpointsLLM

llm = InferenceEndpointsLLM(
    endpoint_name="<ENDPOINT_NAME>",
    api_key="<HF_API_KEY>",
    endpoint_namespace="<USER|ORG>",
)

llm.load()

output = llm.generate(inputs=[[{"role": "user", "content": "Hello world!"}]])
```

Dedicated Inference Endpoints or TGI:

```python
from distilabel.llms.huggingface import InferenceEndpointsLLM

llm = InferenceEndpointsLLM(
    api_key="<HF_API_KEY>",
    base_url="<BASE_URL>",
)

llm.load()

output = llm.generate(inputs=[[{"role": "user", "content": "Hello world!"}]])
```

Generate structured data:

```python
from pydantic import BaseModel
from distilabel.llms import InferenceEndpointsLLM

class User(BaseModel):
    name: str
    last_name: str
    id: int

llm = InferenceEndpointsLLM(
    model_id="meta-llama/Meta-Llama-3-70B-Instruct",
    tokenizer_id="meta-llama/Meta-Llama-3-70B-Instruct",
    api_key="api.key",
    structured_output={"format": "json", "schema": User.model_json_schema()}
)

llm.load()

output = llm.generate(inputs=[[{"role": "user", "content": "Create a user profile for the Tour De France"}]])
```

Source code in src/distilabel/llms/huggingface/inference_endpoints.py

class InferenceEndpointsLLM(AsyncLLM):
    """InferenceEndpoints LLM implementation running the async API client.

    This LLM will internally use `huggingface_hub.AsyncInferenceClient`.

    Attributes:
        model_id: the model ID to use for the LLM as available in the Hugging Face Hub, which
            will be used to resolve the base URL for the serverless Inference Endpoints API requests.
            Defaults to `None`.
        endpoint_name: the name of the Inference Endpoint to use for the LLM. Defaults to `None`.
        endpoint_namespace: the namespace of the Inference Endpoint to use for the LLM. Defaults to `None`.
        base_url: the base URL to use for the Inference Endpoints API requests.
        api_key: the API key to authenticate the requests to the Inference Endpoints API.
        tokenizer_id: the tokenizer ID to use for the LLM as available in the Hugging Face Hub.
            Defaults to `None`, but defining one is recommended to properly format the prompt.
        model_display_name: the model display name to use for the LLM. Defaults to `None`.
        structured_output: a dictionary containing the structured output configuration or
            if more fine-grained control is needed, an instance of `OutlinesStructuredOutput`.
            Defaults to None.

    Icon:
        `:hugging:`

    Examples:

        Free serverless Inference API:

        ```python
        from distilabel.llms.huggingface import InferenceEndpointsLLM

        llm = InferenceEndpointsLLM(
            model_id="mistralai/Mistral-7B-Instruct-v0.2",
        )

        llm.load()

        output = llm.generate(inputs=[[{"role": "user", "content": "Hello world!"}]])
        ```

        Dedicated Inference Endpoints:

        ```python
        from distilabel.llms.huggingface import InferenceEndpointsLLM

        llm = InferenceEndpointsLLM(
            endpoint_name="<ENDPOINT_NAME>",
            api_key="<HF_API_KEY>",
            endpoint_namespace="<USER|ORG>",
        )

        llm.load()

        output = llm.generate(inputs=[[{"role": "user", "content": "Hello world!"}]])
        ```

        Dedicated Inference Endpoints or TGI:

        ```python
        from distilabel.llms.huggingface import InferenceEndpointsLLM

        llm = InferenceEndpointsLLM(
            api_key="<HF_API_KEY>",
            base_url="<BASE_URL>",
        )

        llm.load()

        output = llm.generate(inputs=[[{"role": "user", "content": "Hello world!"}]])
        ```

        Generate structured data:

        ```python
        from pydantic import BaseModel
        from distilabel.llms import InferenceEndpointsLLM

        class User(BaseModel):
            name: str
            last_name: str
            id: int

        llm = InferenceEndpointsLLM(
            model_id="meta-llama/Meta-Llama-3-70B-Instruct",
            tokenizer_id="meta-llama/Meta-Llama-3-70B-Instruct",
            api_key="api.key",
            structured_output={"format": "json", "schema": User.model_json_schema()}
        )

        llm.load()

        output = llm.generate(inputs=[[{"role": "user", "content": "Create a user profile for the Tour De France"}]])
        ```
    """

    model_id: Optional[str] = None

    endpoint_name: Optional[RuntimeParameter[str]] = Field(
        default=None,
        description="The name of the Inference Endpoint to use for the LLM.",
    )
    endpoint_namespace: Optional[RuntimeParameter[str]] = Field(
        default=None,
        description="The namespace of the Inference Endpoint to use for the LLM.",
    )
    base_url: Optional[RuntimeParameter[str]] = Field(
        default=None,
        description="The base URL to use for the Inference Endpoints API requests.",
    )
    api_key: Optional[RuntimeParameter[SecretStr]] = Field(
        default=os.getenv(_INFERENCE_ENDPOINTS_API_KEY_ENV_VAR_NAME),
        description="The API key to authenticate the requests to the Inference Endpoints API.",
    )

    tokenizer_id: Optional[str] = None
    model_display_name: Optional[str] = None

    structured_output: Optional[RuntimeParameter[StructuredOutputType]] = Field(
        default=None,
        description="The structured output format to use across all the generations.",
    )

    _num_generations_param_supported = False

    _model_name: Optional[str] = PrivateAttr(default=None)
    _tokenizer: Optional["PreTrainedTokenizer"] = PrivateAttr(default=None)
    _api_key_env_var: str = PrivateAttr(_INFERENCE_ENDPOINTS_API_KEY_ENV_VAR_NAME)
    _aclient: Optional["AsyncInferenceClient"] = PrivateAttr(...)

    @model_validator(mode="after")  # type: ignore
    def only_one_of_model_id_endpoint_name_or_base_url_provided(
        self,
    ) -> "InferenceEndpointsLLM":
        """Validates that only one of `model_id` or `endpoint_name` is provided; and if `base_url` is also
        provided, a warning will be shown informing the user that the provided `base_url` will be ignored in
        favour of the dynamically calculated one.."""

        if self.base_url and (self.model_id or self.endpoint_name):
            self._logger.warning(  # type: ignore
                f"Since the `base_url={self.base_url}` is available and either one of `model_id`"
                " or `endpoint_name` is also provided, the `base_url` will either be ignored"
                " or overwritten with the one generated from either of those args, for serverless"
                " or dedicated inference endpoints, respectively."
            )

        if (
            self.model_id
            and self.tokenizer_id is None
            and self.structured_output is not None
        ):
            self.tokenizer_id = self.model_id

        if self.base_url and not (self.model_id or self.endpoint_name):
            return self

        if self.model_id and not self.endpoint_name:
            return self

        if self.endpoint_name and not self.model_id:
            return self

        raise ValidationError(
            f"Only one of `model_id` or `endpoint_name` must be provided. If `base_url` is"
            f" provided too, it will be overwritten instead. Found `model_id`={self.model_id},"
            f" `endpoint_name`={self.endpoint_name}, and `base_url`={self.base_url}."
        )

    def load(self) -> None:  # noqa: C901
        """Loads the `AsyncInferenceClient` client to connect to the Hugging Face Inference
        Endpoint.

        Raises:
            ImportError: if the `huggingface-hub` Python client is not installed.
            ValueError: if the model is not currently deployed or is not running the TGI framework.
            ImportError: if the `transformers` Python client is not installed.
        """
        super().load()

        try:
            from huggingface_hub import (
                AsyncInferenceClient,
                InferenceClient,
                get_inference_endpoint,
            )
        except ImportError as ie:
            raise ImportError(
                "Hugging Face Hub Python client is not installed. Please install it using"
                " `pip install huggingface-hub`."
            ) from ie

        if self.api_key is None:
            self.api_key = SecretStr(get_hf_token(self.__class__.__name__, "api_key"))

        if self.model_id is not None:
            client = InferenceClient(
                model=self.model_id, token=self.api_key.get_secret_value()
            )
            status = client.get_model_status()

            if (
                status.state not in {"Loadable", "Loaded"}
                and status.framework != "text-generation-inference"
            ):
                raise ValueError(
                    f"Model {self.model_id} is not currently deployed or is not running the TGI framework"
                )

            self.base_url = client._resolve_url(
                model=self.model_id, task="text-generation"
            )

        if self.endpoint_name is not None:
            client = get_inference_endpoint(
                name=self.endpoint_name,
                namespace=self.endpoint_namespace,
                token=self.api_key.get_secret_value(),
            )
            if client.status in ["paused", "scaledToZero"]:
                client.resume().wait(timeout=300)
            elif client.status == "initializing":
                client.wait(timeout=300)

            self.base_url = client.url
            self._model_name = client.repository

        self._aclient = AsyncInferenceClient(
            model=self.base_url,
            token=self.api_key.get_secret_value(),
        )

        if self.tokenizer_id:
            try:
                from transformers import AutoTokenizer
            except ImportError as ie:
                raise ImportError(
                    "Transformers Python client is not installed. Please install it using"
                    " `pip install transformers`."
                ) from ie

            self._tokenizer = AutoTokenizer.from_pretrained(self.tokenizer_id)

    @property
    @override
    def model_name(self) -> Union[str, None]:  # type: ignore
        """Returns the model name used for the LLM."""
        return (
            self.model_display_name
            or self._model_name
            or self.model_id
            or self.endpoint_name
            or self.base_url
        )

    def prepare_input(self, input: "StandardInput") -> str:
        """Prepares the input (applying the chat template and tokenization) for the provided
        input.

        Args:
            input: the input list containing chat items.

        Returns:
            The prompt to send to the LLM.
        """

        return self._tokenizer.apply_chat_template(  # type: ignore
            conversation=input,  # type: ignore
            tokenize=False,
            add_generation_prompt=True,
        )

    def _get_structured_output(
        self, input: FormattedInput
    ) -> Union[Dict[str, Any], None]:
        """Gets the structured output (if any) for the given input.

        Args:
            input: a single input in chat format to generate responses for.

        Returns:
            The structured output that will be passed as `grammer` to the inference endpoint
            or `None` if not required.
        """
        structured_output = None

        # Specific structured output per input
        if isinstance(input, tuple):
            input, structured_output = input
            structured_output = {
                "type": structured_output["format"],  # type: ignore
                "value": structured_output["schema"],  # type: ignore
            }

        # Same structured output for all the inputs
        if structured_output is None and self.structured_output is not None:
            try:
                structured_output = {
                    "type": self.structured_output["format"],  # type: ignore
                    "value": self.structured_output["schema"],  # type: ignore
                }
            except KeyError as e:
                raise ValueError(
                    "To use the structured output you have to inform the `format` and `schema` in "
                    "the `structured_output` attribute."
                ) from e

        return structured_output

    async def _generate_with_text_generation(
        self,
        input: FormattedInput,
        max_new_tokens: int = 128,
        repetition_penalty: Optional[float] = None,
        frequency_penalty: Optional[float] = None,
        temperature: float = 1.0,
        do_sample: bool = False,
        top_k: Optional[int] = None,
        top_p: Optional[float] = None,
        typical_p: Optional[float] = None,
        stop_sequences: Union[List[str], None] = None,
        return_full_text: bool = False,
        seed: Optional[int] = None,
        watermark: bool = False,
    ) -> Union[str, None]:
        structured_output = self._get_structured_output(input)

        completion = None
        try:
            completion = await self._aclient.text_generation(  # type: ignore
                prompt=self.prepare_input(input),  # type: ignore
                max_new_tokens=max_new_tokens,
                do_sample=do_sample,
                typical_p=typical_p,
                repetition_penalty=repetition_penalty,
                frequency_penalty=frequency_penalty,
                temperature=temperature,
                top_p=top_p,
                top_k=top_k,
                stop_sequences=stop_sequences,
                return_full_text=return_full_text,
                # NOTE: here to ensure that the cache is not used and a different response is
                # generated every time
                seed=seed or random.randint(0, sys.maxsize),
                watermark=watermark,
                grammar=structured_output,  # type: ignore
            )
        except Exception as e:
            self._logger.warning(  # type: ignore
                f"⚠️ Received no response using Inference Client (model: '{self.model_name}')."
                f" Finish reason was: {e}"
            )
        return completion

    async def _generate_with_chat_completion(
        self,
        input: "StandardInput",
        max_new_tokens: int = 128,
        frequency_penalty: Optional[float] = None,
        logit_bias: Optional[List[float]] = None,
        presence_penalty: Optional[float] = None,
        seed: Optional[int] = None,
        stop_sequences: Optional[List[str]] = None,
        temperature: float = 1.0,
        tool_choice: Optional[Union[Dict[str, str], Literal["auto"]]] = None,
        tool_prompt: Optional[str] = None,
        tools: Optional[List[Dict[str, Any]]] = None,
        top_p: Optional[float] = None,
    ) -> Union[str, None]:
        message = None
        try:
            completion = await self._aclient.chat_completion(  # type: ignore
                messages=input,  # type: ignore
                max_tokens=max_new_tokens,
                frequency_penalty=frequency_penalty,
                logit_bias=logit_bias,
                presence_penalty=presence_penalty,
                # NOTE: here to ensure that the cache is not used and a different response is
                # generated every time
                seed=seed or random.randint(0, sys.maxsize),
                stop=stop_sequences,
                temperature=temperature,
                tool_choice=tool_choice,  # type: ignore
                tool_prompt=tool_prompt,
                tools=tools,  # type: ignore
                top_p=top_p,
            )
            choice = completion.choices[0]
            if (message := choice.message.content) is None:
                self._logger.warning(  # type: ignore
                    f"⚠️ Received no response using Inference Client (model: '{self.model_name}')."
                    f" Finish reason was: {choice.finish_reason}"
                )
        except Exception as e:
            self._logger.warning(  # type: ignore
                f"⚠️ Received no response using Inference Client (model: '{self.model_name}')."
                f" Finish reason was: {e}"
            )
        return message

    def _check_stop_sequences(
        self,
        stop_sequences: Optional[Union[str, List[str]]] = None,
    ) -> Union[List[str], None]:
        """Checks that no more than 4 stop sequences are provided.

        Args:
            stop_sequences: the stop sequences to be checked.

        Returns:
            The stop sequences.
        """
        if stop_sequences is not None:
            if isinstance(stop_sequences, str):
                stop_sequences = [stop_sequences]
            if len(stop_sequences) > 4:
                warnings.warn(
                    "Only up to 4 stop sequences are allowed, so keeping the first 4 items only.",
                    UserWarning,
                    stacklevel=2,
                )
                stop_sequences = stop_sequences[:4]
        return stop_sequences

    @validate_call
    async def agenerate(  # type: ignore
        self,
        input: FormattedInput,
        max_new_tokens: int = 128,
        frequency_penalty: Optional[Annotated[float, Field(ge=-2.0, le=2.0)]] = None,
        logit_bias: Optional[List[float]] = None,
        presence_penalty: Optional[Annotated[float, Field(ge=-2.0, le=2.0)]] = None,
        seed: Optional[int] = None,
        stop_sequences: Optional[List[str]] = None,
        temperature: float = 1.0,
        tool_choice: Optional[Union[Dict[str, str], Literal["auto"]]] = None,
        tool_prompt: Optional[str] = None,
        tools: Optional[List[Dict[str, Any]]] = None,
        top_p: Optional[float] = None,
        do_sample: bool = False,
        repetition_penalty: Optional[float] = None,
        return_full_text: bool = False,
        top_k: Optional[int] = None,
        typical_p: Optional[float] = None,
        watermark: bool = False,
    ) -> GenerateOutput:
        """Generates completions for the given input using the async client. This method
        uses two methods of the `huggingface_hub.AsyncClient`: `chat_completion` and `text_generation`.
        `chat_completion` method will be used only if no `tokenizer_id` has been specified.
        Some arguments of this function are specific to the `text_generation` method, while
        some others are specific to the `chat_completion` method.

        Args:
            input: a single input in chat format to generate responses for.
            max_new_tokens: the maximum number of new tokens that the model will generate.
                Defaults to `128`.
            frequency_penalty: a value between `-2.0` and `2.0`. Positive values penalize
                new tokens based on their existing frequency in the text so far, decreasing
                model's likelihood to repeat the same line verbatim. Defauls to `None`.
            logit_bias: modify the likelihood of specified tokens appearing in the completion.
                This argument is exclusive to the `chat_completion` method and will be used
                only if `tokenizer_id` is `None`.
                Defaults to `None`.
            presence_penalty: a value between `-2.0` and `2.0`. Positive values penalize
                new tokens based on whether they appear in the text so far, increasing the
                model likelihood to talk about new topics. This argument is exclusive to
                the `chat_completion` method and will be used only if `tokenizer_id` is
                `None`. Defauls to `None`.
            seed: the seed to use for the generation. Defaults to `None`.
            stop_sequences: either a single string or a list of strings containing the sequences
                to stop the generation at. Defaults to `None`, but will be set to the
                `tokenizer.eos_token` if available.
            temperature: the temperature to use for the generation. Defaults to `1.0`.
            tool_choice: the name of the tool the model should call. It can be a dictionary
                like `{"function_name": "my_tool"}` or "auto". If not provided, then the
                model won't use any tool. This argument is exclusive to the `chat_completion`
                method and will be used only if `tokenizer_id` is `None`. Defaults to `None`.
            tool_prompt: A prompt to be appended before the tools. This argument is exclusive
                to the `chat_completion` method and will be used only if `tokenizer_id`
                is `None`. Defauls to `None`.
            tools: a list of tools definitions that the LLM can use.
                This argument is exclusive to the `chat_completion` method and will be used
                only if `tokenizer_id` is `None`. Defaults to `None`.
            top_p: the top-p value to use for the generation. Defaults to `1.0`.
            do_sample: whether to use sampling for the generation. This argument is exclusive
                of the `text_generation` method and will be only used if `tokenizer_id` is not
                `None`. Defaults to `False`.
            repetition_penalty: the repetition penalty to use for the generation. This argument
                is exclusive of the `text_generation` method and will be only used if `tokenizer_id`
                is not `None`. Defaults to `None`.
            return_full_text: whether to return the full text of the completion or just
                the generated text. Defaults to `False`, meaning that only the generated
                text will be returned. This argument is exclusive of the `text_generation`
                method and will be only used if `tokenizer_id` is not `None`.
            top_k: the top-k value to use for the generation. This argument is exclusive
                of the `text_generation` method and will be only used if `tokenizer_id`
                is not `None`. Defaults to `0.8`, since neither `0.0` nor `1.0` are valid
                values in TGI.
            typical_p: the typical-p value to use for the generation. This argument is exclusive
                of the `text_generation` method and will be only used if `tokenizer_id`
                is not `None`. Defaults to `None`.
            watermark: whether to add the watermark to the generated text. This argument
                is exclusive of the `text_generation` method and will be only used if `tokenizer_id`
                is not `None`. Defaults to `None`.

        Returns:
            A list of lists of strings containing the generated responses for each input.
        """
        stop_sequences = self._check_stop_sequences(stop_sequences)

        if self.tokenizer_id is None:
            return [
                await self._generate_with_chat_completion(
                    input=input,  # type: ignore
                    max_new_tokens=max_new_tokens,
                    frequency_penalty=frequency_penalty,
                    logit_bias=logit_bias,
                    presence_penalty=presence_penalty,
                    seed=seed,
                    stop_sequences=stop_sequences,
                    temperature=temperature,
                    tool_choice=tool_choice,
                    tool_prompt=tool_prompt,
                    tools=tools,
                    top_p=top_p,
                )
            ]

        return [
            await self._generate_with_text_generation(
                input=input,
                max_new_tokens=max_new_tokens,
                do_sample=do_sample,
                typical_p=typical_p,
                repetition_penalty=repetition_penalty,
                frequency_penalty=frequency_penalty,
                temperature=temperature,
                top_p=top_p,
                top_k=top_k,
                stop_sequences=stop_sequences,
                return_full_text=return_full_text,
                seed=seed,
                watermark=watermark,
            )
        ]

model_name: Union[str, None] property

Returns the model name used for the LLM.

agenerate(input, max_new_tokens=128, frequency_penalty=None, logit_bias=None, presence_penalty=None, seed=None, stop_sequences=None, temperature=1.0, tool_choice=None, tool_prompt=None, tools=None, top_p=None, do_sample=False, repetition_penalty=None, return_full_text=False, top_k=None, typical_p=None, watermark=False) async

Generates completions for the given input using the async client. This method uses two methods of the huggingface_hub.AsyncClient: chat_completion and text_generation. chat_completion method will be used only if no tokenizer_id has been specified. Some arguments of this function are specific to the text_generation method, while some others are specific to the chat_completion method.

Parameters:

Name	Type	Description	Default
input	FormattedInput	a single input in chat format to generate responses for.	required
max_new_tokens	int	the maximum number of new tokens that the model will generate. Defaults to 128.	128
frequency_penalty	Optional[Annotated[float, Field(ge=-2.0, le=2.0)]]	a value between -2.0 and 2.0. Positive values penalize new tokens based on their existing frequency in the text so far, decreasing model's likelihood to repeat the same line verbatim. Defauls to None.	None
logit_bias	Optional[List[float]]	modify the likelihood of specified tokens appearing in the completion. This argument is exclusive to the chat_completion method and will be used only if tokenizer_id is None. Defaults to None.	None
presence_penalty	Optional[Annotated[float, Field(ge=-2.0, le=2.0)]]	a value between -2.0 and 2.0. Positive values penalize new tokens based on whether they appear in the text so far, increasing the model likelihood to talk about new topics. This argument is exclusive to the chat_completion method and will be used only if tokenizer_id is None. Defauls to None.	None
seed	Optional[int]	the seed to use for the generation. Defaults to None.	None
stop_sequences	Optional[List[str]]	either a single string or a list of strings containing the sequences to stop the generation at. Defaults to None, but will be set to the tokenizer.eos_token if available.	None
temperature	float	the temperature to use for the generation. Defaults to 1.0.	1.0
tool_choice	Optional[Union[Dict[str, str], Literal['auto']]]	the name of the tool the model should call. It can be a dictionary like {"function_name": "my_tool"} or "auto". If not provided, then the model won't use any tool. This argument is exclusive to the chat_completion method and will be used only if tokenizer_id is None. Defaults to None.	None
tool_prompt	Optional[str]	A prompt to be appended before the tools. This argument is exclusive to the chat_completion method and will be used only if tokenizer_id is None. Defauls to None.	None
tools	Optional[List[Dict[str, Any]]]	a list of tools definitions that the LLM can use. This argument is exclusive to the chat_completion method and will be used only if tokenizer_id is None. Defaults to None.	None
top_p	Optional[float]	the top-p value to use for the generation. Defaults to 1.0.	None
do_sample	bool	whether to use sampling for the generation. This argument is exclusive of the text_generation method and will be only used if tokenizer_id is not None. Defaults to False.	False
repetition_penalty	Optional[float]	the repetition penalty to use for the generation. This argument is exclusive of the text_generation method and will be only used if tokenizer_id is not None. Defaults to None.	None
return_full_text	bool	whether to return the full text of the completion or just the generated text. Defaults to False, meaning that only the generated text will be returned. This argument is exclusive of the text_generation method and will be only used if tokenizer_id is not None.	False
top_k	Optional[int]	the top-k value to use for the generation. This argument is exclusive of the text_generation method and will be only used if tokenizer_id is not None. Defaults to 0.8, since neither 0.0 nor 1.0 are valid values in TGI.	None
typical_p	Optional[float]	the typical-p value to use for the generation. This argument is exclusive of the text_generation method and will be only used if tokenizer_id is not None. Defaults to None.	None
watermark	bool	whether to add the watermark to the generated text. This argument is exclusive of the text_generation method and will be only used if tokenizer_id is not None. Defaults to None.	False

Returns:

Type	Description
GenerateOutput	A list of lists of strings containing the generated responses for each input.

Source code in src/distilabel/llms/huggingface/inference_endpoints.py

@validate_call
async def agenerate(  # type: ignore
    self,
    input: FormattedInput,
    max_new_tokens: int = 128,
    frequency_penalty: Optional[Annotated[float, Field(ge=-2.0, le=2.0)]] = None,
    logit_bias: Optional[List[float]] = None,
    presence_penalty: Optional[Annotated[float, Field(ge=-2.0, le=2.0)]] = None,
    seed: Optional[int] = None,
    stop_sequences: Optional[List[str]] = None,
    temperature: float = 1.0,
    tool_choice: Optional[Union[Dict[str, str], Literal["auto"]]] = None,
    tool_prompt: Optional[str] = None,
    tools: Optional[List[Dict[str, Any]]] = None,
    top_p: Optional[float] = None,
    do_sample: bool = False,
    repetition_penalty: Optional[float] = None,
    return_full_text: bool = False,
    top_k: Optional[int] = None,
    typical_p: Optional[float] = None,
    watermark: bool = False,
) -> GenerateOutput:
    """Generates completions for the given input using the async client. This method
    uses two methods of the `huggingface_hub.AsyncClient`: `chat_completion` and `text_generation`.
    `chat_completion` method will be used only if no `tokenizer_id` has been specified.
    Some arguments of this function are specific to the `text_generation` method, while
    some others are specific to the `chat_completion` method.

    Args:
        input: a single input in chat format to generate responses for.
        max_new_tokens: the maximum number of new tokens that the model will generate.
            Defaults to `128`.
        frequency_penalty: a value between `-2.0` and `2.0`. Positive values penalize
            new tokens based on their existing frequency in the text so far, decreasing
            model's likelihood to repeat the same line verbatim. Defauls to `None`.
        logit_bias: modify the likelihood of specified tokens appearing in the completion.
            This argument is exclusive to the `chat_completion` method and will be used
            only if `tokenizer_id` is `None`.
            Defaults to `None`.
        presence_penalty: a value between `-2.0` and `2.0`. Positive values penalize
            new tokens based on whether they appear in the text so far, increasing the
            model likelihood to talk about new topics. This argument is exclusive to
            the `chat_completion` method and will be used only if `tokenizer_id` is
            `None`. Defauls to `None`.
        seed: the seed to use for the generation. Defaults to `None`.
        stop_sequences: either a single string or a list of strings containing the sequences
            to stop the generation at. Defaults to `None`, but will be set to the
            `tokenizer.eos_token` if available.
        temperature: the temperature to use for the generation. Defaults to `1.0`.
        tool_choice: the name of the tool the model should call. It can be a dictionary
            like `{"function_name": "my_tool"}` or "auto". If not provided, then the
            model won't use any tool. This argument is exclusive to the `chat_completion`
            method and will be used only if `tokenizer_id` is `None`. Defaults to `None`.
        tool_prompt: A prompt to be appended before the tools. This argument is exclusive
            to the `chat_completion` method and will be used only if `tokenizer_id`
            is `None`. Defauls to `None`.
        tools: a list of tools definitions that the LLM can use.
            This argument is exclusive to the `chat_completion` method and will be used
            only if `tokenizer_id` is `None`. Defaults to `None`.
        top_p: the top-p value to use for the generation. Defaults to `1.0`.
        do_sample: whether to use sampling for the generation. This argument is exclusive
            of the `text_generation` method and will be only used if `tokenizer_id` is not
            `None`. Defaults to `False`.
        repetition_penalty: the repetition penalty to use for the generation. This argument
            is exclusive of the `text_generation` method and will be only used if `tokenizer_id`
            is not `None`. Defaults to `None`.
        return_full_text: whether to return the full text of the completion or just
            the generated text. Defaults to `False`, meaning that only the generated
            text will be returned. This argument is exclusive of the `text_generation`
            method and will be only used if `tokenizer_id` is not `None`.
        top_k: the top-k value to use for the generation. This argument is exclusive
            of the `text_generation` method and will be only used if `tokenizer_id`
            is not `None`. Defaults to `0.8`, since neither `0.0` nor `1.0` are valid
            values in TGI.
        typical_p: the typical-p value to use for the generation. This argument is exclusive
            of the `text_generation` method and will be only used if `tokenizer_id`
            is not `None`. Defaults to `None`.
        watermark: whether to add the watermark to the generated text. This argument
            is exclusive of the `text_generation` method and will be only used if `tokenizer_id`
            is not `None`. Defaults to `None`.

    Returns:
        A list of lists of strings containing the generated responses for each input.
    """
    stop_sequences = self._check_stop_sequences(stop_sequences)

    if self.tokenizer_id is None:
        return [
            await self._generate_with_chat_completion(
                input=input,  # type: ignore
                max_new_tokens=max_new_tokens,
                frequency_penalty=frequency_penalty,
                logit_bias=logit_bias,
                presence_penalty=presence_penalty,
                seed=seed,
                stop_sequences=stop_sequences,
                temperature=temperature,
                tool_choice=tool_choice,
                tool_prompt=tool_prompt,
                tools=tools,
                top_p=top_p,
            )
        ]

    return [
        await self._generate_with_text_generation(
            input=input,
            max_new_tokens=max_new_tokens,
            do_sample=do_sample,
            typical_p=typical_p,
            repetition_penalty=repetition_penalty,
            frequency_penalty=frequency_penalty,
            temperature=temperature,
            top_p=top_p,
            top_k=top_k,
            stop_sequences=stop_sequences,
            return_full_text=return_full_text,
            seed=seed,
            watermark=watermark,
        )
    ]

load()

Loads the AsyncInferenceClient client to connect to the Hugging Face Inference Endpoint.

Raises:

Type	Description
ImportError	if the huggingface-hub Python client is not installed.
ValueError	if the model is not currently deployed or is not running the TGI framework.
ImportError	if the transformers Python client is not installed.

Source code in src/distilabel/llms/huggingface/inference_endpoints.py

def load(self) -> None:  # noqa: C901
    """Loads the `AsyncInferenceClient` client to connect to the Hugging Face Inference
    Endpoint.

    Raises:
        ImportError: if the `huggingface-hub` Python client is not installed.
        ValueError: if the model is not currently deployed or is not running the TGI framework.
        ImportError: if the `transformers` Python client is not installed.
    """
    super().load()

    try:
        from huggingface_hub import (
            AsyncInferenceClient,
            InferenceClient,
            get_inference_endpoint,
        )
    except ImportError as ie:
        raise ImportError(
            "Hugging Face Hub Python client is not installed. Please install it using"
            " `pip install huggingface-hub`."
        ) from ie

    if self.api_key is None:
        self.api_key = SecretStr(get_hf_token(self.__class__.__name__, "api_key"))

    if self.model_id is not None:
        client = InferenceClient(
            model=self.model_id, token=self.api_key.get_secret_value()
        )
        status = client.get_model_status()

        if (
            status.state not in {"Loadable", "Loaded"}
            and status.framework != "text-generation-inference"
        ):
            raise ValueError(
                f"Model {self.model_id} is not currently deployed or is not running the TGI framework"
            )

        self.base_url = client._resolve_url(
            model=self.model_id, task="text-generation"
        )

    if self.endpoint_name is not None:
        client = get_inference_endpoint(
            name=self.endpoint_name,
            namespace=self.endpoint_namespace,
            token=self.api_key.get_secret_value(),
        )
        if client.status in ["paused", "scaledToZero"]:
            client.resume().wait(timeout=300)
        elif client.status == "initializing":
            client.wait(timeout=300)

        self.base_url = client.url
        self._model_name = client.repository

    self._aclient = AsyncInferenceClient(
        model=self.base_url,
        token=self.api_key.get_secret_value(),
    )

    if self.tokenizer_id:
        try:
            from transformers import AutoTokenizer
        except ImportError as ie:
            raise ImportError(
                "Transformers Python client is not installed. Please install it using"
                " `pip install transformers`."
            ) from ie

        self._tokenizer = AutoTokenizer.from_pretrained(self.tokenizer_id)

only_one_of_model_id_endpoint_name_or_base_url_provided()

Validates that only one of model_id or endpoint_name is provided; and if base_url is also provided, a warning will be shown informing the user that the provided base_url will be ignored in favour of the dynamically calculated one..

Source code in src/distilabel/llms/huggingface/inference_endpoints.py

@model_validator(mode="after")  # type: ignore
def only_one_of_model_id_endpoint_name_or_base_url_provided(
    self,
) -> "InferenceEndpointsLLM":
    """Validates that only one of `model_id` or `endpoint_name` is provided; and if `base_url` is also
    provided, a warning will be shown informing the user that the provided `base_url` will be ignored in
    favour of the dynamically calculated one.."""

    if self.base_url and (self.model_id or self.endpoint_name):
        self._logger.warning(  # type: ignore
            f"Since the `base_url={self.base_url}` is available and either one of `model_id`"
            " or `endpoint_name` is also provided, the `base_url` will either be ignored"
            " or overwritten with the one generated from either of those args, for serverless"
            " or dedicated inference endpoints, respectively."
        )

    if (
        self.model_id
        and self.tokenizer_id is None
        and self.structured_output is not None
    ):
        self.tokenizer_id = self.model_id

    if self.base_url and not (self.model_id or self.endpoint_name):
        return self

    if self.model_id and not self.endpoint_name:
        return self

    if self.endpoint_name and not self.model_id:
        return self

    raise ValidationError(
        f"Only one of `model_id` or `endpoint_name` must be provided. If `base_url` is"
        f" provided too, it will be overwritten instead. Found `model_id`={self.model_id},"
        f" `endpoint_name`={self.endpoint_name}, and `base_url`={self.base_url}."
    )

prepare_input(input)

Prepares the input (applying the chat template and tokenization) for the provided input.

Parameters:

Name	Type	Description	Default
input	StandardInput	the input list containing chat items.	required

Returns:

Type	Description
str	The prompt to send to the LLM.

Source code in src/distilabel/llms/huggingface/inference_endpoints.py

def prepare_input(self, input: "StandardInput") -> str:
    """Prepares the input (applying the chat template and tokenization) for the provided
    input.

    Args:
        input: the input list containing chat items.

    Returns:
        The prompt to send to the LLM.
    """

    return self._tokenizer.apply_chat_template(  # type: ignore
        conversation=input,  # type: ignore
        tokenize=False,
        add_generation_prompt=True,
    )

TransformersLLM

Bases: LLM, CudaDevicePlacementMixin

Hugging Face transformers library LLM implementation using the text generation pipeline.

Attributes:

Name	Type	Description
model	str	the model Hugging Face Hub repo id or a path to a directory containing the model weights and configuration files.
revision	str	if model refers to a Hugging Face Hub repository, then the revision (e.g. a branch name or a commit id) to use. Defaults to "main".
torch_dtype	str	the torch dtype to use for the model e.g. "float16", "float32", etc. Defaults to "auto".
trust_remote_code	bool	whether to trust or not remote (code in the Hugging Face Hub repository) code to load the model. Defaults to False.
model_kwargs	Optional[Dict[str, Any]]	additional dictionary of keyword arguments that will be passed to the from_pretrained method of the model.
tokenizer	Optional[str]	the tokenizer Hugging Face Hub repo id or a path to a directory containing the tokenizer config files. If not provided, the one associated to the model will be used. Defaults to None.
use_fast	bool	whether to use a fast tokenizer or not. Defaults to True.
chat_template	Optional[str]	a chat template that will be used to build the prompts before sending them to the model. If not provided, the chat template defined in the tokenizer config will be used. If not provided and the tokenizer doesn't have a chat template, then ChatML template will be used. Defaults to None.
device	Optional[Union[str, int]]	the name or index of the device where the model will be loaded. Defaults to None.
device_map	Optional[Union[str, Dict[str, Any]]]	a dictionary mapping each layer of the model to a device, or a mode like "sequential" or "auto". Defaults to None.
token	Optional[str]	the Hugging Face Hub token that will be used to authenticate to the Hugging Face Hub. If not provided, the HF_TOKEN environment or huggingface_hub package local configuration will be used. Defaults to None.
structured_output	Optional[RuntimeParameter[OutlinesStructuredOutputType]]	a dictionary containing the structured output configuration or if more fine-grained control is needed, an instance of OutlinesStructuredOutput. Defaults to None.

Icon

:hugging:

Examples:

Generate text:

```python
from distilabel.llms import TransformersLLM

llm = TransformersLLM(model="microsoft/Phi-3-mini-4k-instruct")

llm.load()

# Call the model
output = llm.generate(inputs=[[{"role": "user", "content": "Hello world!"}]])
```

Source code in src/distilabel/llms/huggingface/transformers.py

class TransformersLLM(LLM, CudaDevicePlacementMixin):
    """Hugging Face `transformers` library LLM implementation using the text generation
    pipeline.

    Attributes:
        model: the model Hugging Face Hub repo id or a path to a directory containing the
            model weights and configuration files.
        revision: if `model` refers to a Hugging Face Hub repository, then the revision
            (e.g. a branch name or a commit id) to use. Defaults to `"main"`.
        torch_dtype: the torch dtype to use for the model e.g. "float16", "float32", etc.
            Defaults to `"auto"`.
        trust_remote_code: whether to trust or not remote (code in the Hugging Face Hub
            repository) code to load the model. Defaults to `False`.
        model_kwargs: additional dictionary of keyword arguments that will be passed to
            the `from_pretrained` method of the model.
        tokenizer: the tokenizer Hugging Face Hub repo id or a path to a directory containing
            the tokenizer config files. If not provided, the one associated to the `model`
            will be used. Defaults to `None`.
        use_fast: whether to use a fast tokenizer or not. Defaults to `True`.
        chat_template: a chat template that will be used to build the prompts before
            sending them to the model. If not provided, the chat template defined in the
            tokenizer config will be used. If not provided and the tokenizer doesn't have
            a chat template, then ChatML template will be used. Defaults to `None`.
        device: the name or index of the device where the model will be loaded. Defaults
            to `None`.
        device_map: a dictionary mapping each layer of the model to a device, or a mode
            like `"sequential"` or `"auto"`. Defaults to `None`.
        token: the Hugging Face Hub token that will be used to authenticate to the Hugging
            Face Hub. If not provided, the `HF_TOKEN` environment or `huggingface_hub` package
            local configuration will be used. Defaults to `None`.
        structured_output: a dictionary containing the structured output configuration or if more
            fine-grained control is needed, an instance of `OutlinesStructuredOutput`. Defaults to None.

    Icon:
        `:hugging:`

    Examples:

        Generate text:

        ```python
        from distilabel.llms import TransformersLLM

        llm = TransformersLLM(model="microsoft/Phi-3-mini-4k-instruct")

        llm.load()

        # Call the model
        output = llm.generate(inputs=[[{"role": "user", "content": "Hello world!"}]])
        ```
    """

    model: str
    revision: str = "main"
    torch_dtype: str = "auto"
    trust_remote_code: bool = False
    model_kwargs: Optional[Dict[str, Any]] = None
    tokenizer: Optional[str] = None
    use_fast: bool = True
    chat_template: Optional[str] = None
    device: Optional[Union[str, int]] = None
    device_map: Optional[Union[str, Dict[str, Any]]] = None
    token: Optional[str] = None
    structured_output: Optional[RuntimeParameter[OutlinesStructuredOutputType]] = Field(
        default=None,
        description="The structured output format to use across all the generations.",
    )

    _pipeline: Optional["Pipeline"] = PrivateAttr(...)
    _prefix_allowed_tokens_fn: Union[Callable, None] = PrivateAttr(default=None)

    def load(self) -> None:
        """Loads the model and tokenizer and creates the text generation pipeline. In addition,
        it will configure the tokenizer chat template."""
        if self.device == "cuda":
            CudaDevicePlacementMixin.load(self)

        try:
            from transformers import pipeline
        except ImportError as ie:
            raise ImportError(
                "Transformers is not installed. Please install it using `pip install transformers`."
            ) from ie

        self._pipeline = pipeline(
            "text-generation",
            model=self.model,
            revision=self.revision,
            torch_dtype=self.torch_dtype,
            trust_remote_code=self.trust_remote_code,
            model_kwargs=self.model_kwargs or {},
            tokenizer=self.tokenizer or self.model,
            use_fast=self.use_fast,
            device=self.device,
            device_map=self.device_map,
            token=self.token or os.getenv("HF_TOKEN"),
            return_full_text=False,
        )

        if self.chat_template is not None:
            self._pipeline.tokenizer.chat_template = self.chat_template  # type: ignore
        elif (
            self._pipeline.tokenizer.chat_template is None  # type: ignore
            and self._pipeline.tokenizer.default_chat_template is None  # type: ignore
        ):
            self._pipeline.tokenizer.chat_template = CHATML_TEMPLATE  # type: ignore

        if self.structured_output:
            self._prefix_allowed_tokens_fn = self._prepare_structured_output(
                self.structured_output
            )

        super().load()

    def unload(self) -> None:
        """Unloads the `vLLM` model."""
        CudaDevicePlacementMixin.unload(self)
        super().unload()

    @property
    def model_name(self) -> str:
        """Returns the model name used for the LLM."""
        return self.model

    def prepare_input(self, input: "StandardInput") -> str:
        """Prepares the input by applying the chat template to the input, which is formatted
        as an OpenAI conversation, and adding the generation prompt.
        """
        return self._pipeline.tokenizer.apply_chat_template(  # type: ignore
            input,  # type: ignore
            tokenize=False,
            add_generation_prompt=True,
        )

    @validate_call
    def generate(  # type: ignore
        self,
        inputs: List[StandardInput],
        num_generations: int = 1,
        max_new_tokens: int = 128,
        temperature: float = 0.1,
        repetition_penalty: float = 1.1,
        top_p: float = 1.0,
        top_k: int = 0,
        do_sample: bool = True,
    ) -> List[GenerateOutput]:
        """Generates `num_generations` responses for each input using the text generation
        pipeline.

        Args:
            inputs: a list of inputs in chat format to generate responses for.
            num_generations: the number of generations to create per input. Defaults to
                `1`.
            max_new_tokens: the maximum number of new tokens that the model will generate.
                Defaults to `128`.
            temperature: the temperature to use for the generation. Defaults to `0.1`.
            repetition_penalty: the repetition penalty to use for the generation. Defaults
                to `1.1`.
            top_p: the top-p value to use for the generation. Defaults to `1.0`.
            top_k: the top-k value to use for the generation. Defaults to `0`.
            do_sample: whether to use sampling or not. Defaults to `True`.

        Returns:
            A list of lists of strings containing the generated responses for each input.
        """
        prepared_inputs = [self.prepare_input(input=input) for input in inputs]

        outputs: List[List[Dict[str, str]]] = self._pipeline(  # type: ignore
            prepared_inputs,
            max_new_tokens=max_new_tokens,
            temperature=temperature,
            repetition_penalty=repetition_penalty,
            top_p=top_p,
            top_k=top_k,
            do_sample=do_sample,
            num_return_sequences=num_generations,
            prefix_allowed_tokens_fn=self._prefix_allowed_tokens_fn,
        )
        return [
            [generation["generated_text"] for generation in output]
            for output in outputs
        ]

    def get_last_hidden_states(
        self, inputs: List["StandardInput"]
    ) -> List["HiddenState"]:
        """Gets the last `hidden_states` of the model for the given inputs. It doesn't
        execute the task head.

        Args:
            inputs: a list of inputs in chat format to generate the embeddings for.

        Returns:
            A list containing the last hidden state for each sequence using a NumPy array
            with shape [num_tokens, hidden_size].
        """
        model: "PreTrainedModel" = (
            self._pipeline.model.model  # type: ignore
            if hasattr(self._pipeline.model, "model")  # type: ignore
            else next(self._pipeline.model.children())  # type: ignore
        )
        tokenizer: "PreTrainedTokenizer" = self._pipeline.tokenizer  # type: ignore
        input_ids = tokenizer(
            [self.prepare_input(input) for input in inputs],  # type: ignore
            return_tensors="pt",
            padding=True,
        ).to(model.device)
        last_hidden_states = model(**input_ids)["last_hidden_state"]

        return [
            seq_last_hidden_state[attention_mask.bool(), :].detach().cpu().numpy()
            for seq_last_hidden_state, attention_mask in zip(
                last_hidden_states,
                input_ids["attention_mask"],  # type: ignore
            )
        ]

    def _prepare_structured_output(
        self, structured_output: Optional[OutlinesStructuredOutputType] = None
    ) -> Union[Callable, None]:
        """Creates the appropriate function to filter tokens to generate structured outputs.

        Args:
            structured_output: the configuration dict to prepare the structured output.

        Returns:
            The callable that will be used to guide the generation of the model.
        """
        from distilabel.steps.tasks.structured_outputs.outlines import (
            prepare_guided_output,
        )

        result = prepare_guided_output(
            structured_output, "transformers", self._pipeline
        )
        if schema := result.get("schema"):
            self.structured_output["schema"] = schema
        return result["processor"]

model_name: str property

Returns the model name used for the LLM.

generate(inputs, num_generations=1, max_new_tokens=128, temperature=0.1, repetition_penalty=1.1, top_p=1.0, top_k=0, do_sample=True)

Generates num_generations responses for each input using the text generation pipeline.

Parameters:

Name	Type	Description	Default
inputs	List[StandardInput]	a list of inputs in chat format to generate responses for.	required
num_generations	int	the number of generations to create per input. Defaults to 1.	1
max_new_tokens	int	the maximum number of new tokens that the model will generate. Defaults to 128.	128
temperature	float	the temperature to use for the generation. Defaults to 0.1.	0.1
repetition_penalty	float	the repetition penalty to use for the generation. Defaults to 1.1.	1.1
top_p	float	the top-p value to use for the generation. Defaults to 1.0.	1.0
top_k	int	the top-k value to use for the generation. Defaults to 0.	0
do_sample	bool	whether to use sampling or not. Defaults to True.	True

Returns:

Type	Description
List[GenerateOutput]	A list of lists of strings containing the generated responses for each input.

Source code in src/distilabel/llms/huggingface/transformers.py

@validate_call
def generate(  # type: ignore
    self,
    inputs: List[StandardInput],
    num_generations: int = 1,
    max_new_tokens: int = 128,
    temperature: float = 0.1,
    repetition_penalty: float = 1.1,
    top_p: float = 1.0,
    top_k: int = 0,
    do_sample: bool = True,
) -> List[GenerateOutput]:
    """Generates `num_generations` responses for each input using the text generation
    pipeline.

    Args:
        inputs: a list of inputs in chat format to generate responses for.
        num_generations: the number of generations to create per input. Defaults to
            `1`.
        max_new_tokens: the maximum number of new tokens that the model will generate.
            Defaults to `128`.
        temperature: the temperature to use for the generation. Defaults to `0.1`.
        repetition_penalty: the repetition penalty to use for the generation. Defaults
            to `1.1`.
        top_p: the top-p value to use for the generation. Defaults to `1.0`.
        top_k: the top-k value to use for the generation. Defaults to `0`.
        do_sample: whether to use sampling or not. Defaults to `True`.

    Returns:
        A list of lists of strings containing the generated responses for each input.
    """
    prepared_inputs = [self.prepare_input(input=input) for input in inputs]

    outputs: List[List[Dict[str, str]]] = self._pipeline(  # type: ignore
        prepared_inputs,
        max_new_tokens=max_new_tokens,
        temperature=temperature,
        repetition_penalty=repetition_penalty,
        top_p=top_p,
        top_k=top_k,
        do_sample=do_sample,
        num_return_sequences=num_generations,
        prefix_allowed_tokens_fn=self._prefix_allowed_tokens_fn,
    )
    return [
        [generation["generated_text"] for generation in output]
        for output in outputs
    ]

get_last_hidden_states(inputs)

Gets the last hidden_states of the model for the given inputs. It doesn't execute the task head.

Parameters:

Name	Type	Description	Default
inputs	List[StandardInput]	a list of inputs in chat format to generate the embeddings for.	required

Returns:

Type	Description
List[HiddenState]	A list containing the last hidden state for each sequence using a NumPy array
List[HiddenState]	with shape [num_tokens, hidden_size].

Source code in src/distilabel/llms/huggingface/transformers.py

def get_last_hidden_states(
    self, inputs: List["StandardInput"]
) -> List["HiddenState"]:
    """Gets the last `hidden_states` of the model for the given inputs. It doesn't
    execute the task head.

    Args:
        inputs: a list of inputs in chat format to generate the embeddings for.

    Returns:
        A list containing the last hidden state for each sequence using a NumPy array
        with shape [num_tokens, hidden_size].
    """
    model: "PreTrainedModel" = (
        self._pipeline.model.model  # type: ignore
        if hasattr(self._pipeline.model, "model")  # type: ignore
        else next(self._pipeline.model.children())  # type: ignore
    )
    tokenizer: "PreTrainedTokenizer" = self._pipeline.tokenizer  # type: ignore
    input_ids = tokenizer(
        [self.prepare_input(input) for input in inputs],  # type: ignore
        return_tensors="pt",
        padding=True,
    ).to(model.device)
    last_hidden_states = model(**input_ids)["last_hidden_state"]

    return [
        seq_last_hidden_state[attention_mask.bool(), :].detach().cpu().numpy()
        for seq_last_hidden_state, attention_mask in zip(
            last_hidden_states,
            input_ids["attention_mask"],  # type: ignore
        )
    ]

load()

Loads the model and tokenizer and creates the text generation pipeline. In addition, it will configure the tokenizer chat template.

Source code in src/distilabel/llms/huggingface/transformers.py

def load(self) -> None:
    """Loads the model and tokenizer and creates the text generation pipeline. In addition,
    it will configure the tokenizer chat template."""
    if self.device == "cuda":
        CudaDevicePlacementMixin.load(self)

    try:
        from transformers import pipeline
    except ImportError as ie:
        raise ImportError(
            "Transformers is not installed. Please install it using `pip install transformers`."
        ) from ie

    self._pipeline = pipeline(
        "text-generation",
        model=self.model,
        revision=self.revision,
        torch_dtype=self.torch_dtype,
        trust_remote_code=self.trust_remote_code,
        model_kwargs=self.model_kwargs or {},
        tokenizer=self.tokenizer or self.model,
        use_fast=self.use_fast,
        device=self.device,
        device_map=self.device_map,
        token=self.token or os.getenv("HF_TOKEN"),
        return_full_text=False,
    )

    if self.chat_template is not None:
        self._pipeline.tokenizer.chat_template = self.chat_template  # type: ignore
    elif (
        self._pipeline.tokenizer.chat_template is None  # type: ignore
        and self._pipeline.tokenizer.default_chat_template is None  # type: ignore
    ):
        self._pipeline.tokenizer.chat_template = CHATML_TEMPLATE  # type: ignore

    if self.structured_output:
        self._prefix_allowed_tokens_fn = self._prepare_structured_output(
            self.structured_output
        )

    super().load()

prepare_input(input)

Prepares the input by applying the chat template to the input, which is formatted as an OpenAI conversation, and adding the generation prompt.

Source code in src/distilabel/llms/huggingface/transformers.py

def prepare_input(self, input: "StandardInput") -> str:
    """Prepares the input by applying the chat template to the input, which is formatted
    as an OpenAI conversation, and adding the generation prompt.
    """
    return self._pipeline.tokenizer.apply_chat_template(  # type: ignore
        input,  # type: ignore
        tokenize=False,
        add_generation_prompt=True,
    )

unload()

Unloads the vLLM model.

Source code in src/distilabel/llms/huggingface/transformers.py

def unload(self) -> None:
    """Unloads the `vLLM` model."""
    CudaDevicePlacementMixin.unload(self)
    super().unload()