_text_generation.py

# coding=utf-8
# Copyright 2023-present, the HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# Original implementation taken from the `text-generation` Python client (see https://pypi.org/project/text-generation/
# and https://github.com/huggingface/text-generation-inference/tree/main/clients/python)
#
# Changes compared to original implementation:
# - use pydantic.dataclasses instead of BaseModel
# - default to Python's dataclasses if Pydantic is not installed (same implementation but no validation)
# - added default values for all parameters (not needed in BaseModel but dataclasses yes)
# - integrated in `huggingface_hub.InferenceClient``
# - added `stream: bool` and `details: bool` in the `text_generation` method instead of having different methods for each use case
import warnings
from dataclasses import field
from enum import Enum
from typing import List, NoReturn, Optional

from requests import HTTPError

from ..utils import is_pydantic_available


if is_pydantic_available():
    from pydantic import validator as pydantic_validator
    from pydantic.dataclasses import dataclass

    def validator(*args, **kwargs):
        # Pydantic v1's `@validator` is deprecated in favor of `@field_validator`. In order to support both pydantic v1
        # and v2 without changing the logic, we catch the warning message in pydantic v2 and ignore it. If we want to
        # support pydantic v3 in the future, we will drop support for pydantic v1 and use `pydantic.field_validator`
        # correctly.
        #
        # Related:
        # - https://docs.pydantic.dev/latest/migration/#changes-to-validators
        # - https://github.com/huggingface/huggingface_hub/pull/1837
        with warnings.catch_warnings():
            warnings.filterwarnings("ignore", message="Pydantic V1 style `@validator` validators are deprecated.")
            return pydantic_validator(*args, **kwargs)
else:
    # No validation if Pydantic is not installed
    from dataclasses import dataclass  # type: ignore

    def validator(x):  # type: ignore
        return lambda y: y


@dataclass
class TextGenerationParameters:
    """
    Parameters for text generation.

    Args:
        do_sample (`bool`, *optional*):
            Activate logits sampling. Defaults to False.
        max_new_tokens (`int`, *optional*):
            Maximum number of generated tokens. Defaults to 20.
        repetition_penalty (`Optional[float]`, *optional*):
            The parameter for repetition penalty. A value of 1.0 means no penalty. See [this paper](https://arxiv.org/pdf/1909.05858.pdf)
            for more details. Defaults to None.
        return_full_text (`bool`, *optional*):
            Whether to prepend the prompt to the generated text. Defaults to False.
        stop (`List[str]`, *optional*):
            Stop generating tokens if a member of `stop_sequences` is generated. Defaults to an empty list.
        seed (`Optional[int]`, *optional*):
            Random sampling seed. Defaults to None.
        temperature (`Optional[float]`, *optional*):
            The value used to modulate the logits distribution. Defaults to None.
        top_k (`Optional[int]`, *optional*):
            The number of highest probability vocabulary tokens to keep for top-k-filtering. Defaults to None.
        top_p (`Optional[float]`, *optional*):
            If set to a value less than 1, only the smallest set of most probable tokens with probabilities that add up
            to `top_p` or higher are kept for generation. Defaults to None.
        truncate (`Optional[int]`, *optional*):
            Truncate input tokens to the given size. Defaults to None.
        typical_p (`Optional[float]`, *optional*):
            Typical Decoding mass. See [Typical Decoding for Natural Language Generation](https://arxiv.org/abs/2202.00666)
            for more information. Defaults to None.
        best_of (`Optional[int]`, *optional*):
            Generate `best_of` sequences and return the one with the highest token logprobs. Defaults to None.
        watermark (`bool`, *optional*):
            Watermarking with [A Watermark for Large Language Models](https://arxiv.org/abs/2301.10226). Defaults to False.
        details (`bool`, *optional*):
            Get generation details. Defaults to False.
        decoder_input_details (`bool`, *optional*):
            Get decoder input token logprobs and ids. Defaults to False.
    """

    # Activate logits sampling
    do_sample: bool = False
    # Maximum number of generated tokens
    max_new_tokens: int = 20
    # The parameter for repetition penalty. 1.0 means no penalty.
    # See [this paper](https://arxiv.org/pdf/1909.05858.pdf) for more details.
    repetition_penalty: Optional[float] = None
    # Whether to prepend the prompt to the generated text
    return_full_text: bool = False
    # Stop generating tokens if a member of `stop_sequences` is generated
    stop: List[str] = field(default_factory=lambda: [])
    # Random sampling seed
    seed: Optional[int] = None
    # The value used to module the logits distribution.
    temperature: Optional[float] = None
    # The number of highest probability vocabulary tokens to keep for top-k-filtering.
    top_k: Optional[int] = None
    # If set to < 1, only the smallest set of most probable tokens with probabilities that add up to `top_p` or
    # higher are kept for generation.
    top_p: Optional[float] = None
    # truncate inputs tokens to the given size
    truncate: Optional[int] = None
    # Typical Decoding mass
    # See [Typical Decoding for Natural Language Generation](https://arxiv.org/abs/2202.00666) for more information
    typical_p: Optional[float] = None
    # Generate best_of sequences and return the one if the highest token logprobs
    best_of: Optional[int] = None
    # Watermarking with [A Watermark for Large Language Models](https://arxiv.org/abs/2301.10226)
    watermark: bool = False
    # Get generation details
    details: bool = False
    # Get decoder input token logprobs and ids
    decoder_input_details: bool = False

    @validator("best_of")
    def valid_best_of(cls, field_value, values):
        if field_value is not None:
            if field_value <= 0:
                raise ValueError("`best_of` must be strictly positive")
            if field_value > 1 and values["seed"] is not None:
                raise ValueError("`seed` must not be set when `best_of` is > 1")
            sampling = (
                values["do_sample"]
                | (values["temperature"] is not None)
                | (values["top_k"] is not None)
                | (values["top_p"] is not None)
                | (values["typical_p"] is not None)
            )
            if field_value > 1 and not sampling:
                raise ValueError("you must use sampling when `best_of` is > 1")

        return field_value

    @validator("repetition_penalty")
    def valid_repetition_penalty(cls, v):
        if v is not None and v <= 0:
            raise ValueError("`repetition_penalty` must be strictly positive")
        return v

    @validator("seed")
    def valid_seed(cls, v):
        if v is not None and v < 0:
            raise ValueError("`seed` must be positive")
        return v

    @validator("temperature")
    def valid_temp(cls, v):
        if v is not None and v <= 0:
            raise ValueError("`temperature` must be strictly positive")
        return v

    @validator("top_k")
    def valid_top_k(cls, v):
        if v is not None and v <= 0:
            raise ValueError("`top_k` must be strictly positive")
        return v

    @validator("top_p")
    def valid_top_p(cls, v):
        if v is not None and (v <= 0 or v >= 1.0):
            raise ValueError("`top_p` must be > 0.0 and < 1.0")
        return v

    @validator("truncate")
    def valid_truncate(cls, v):
        if v is not None and v <= 0:
            raise ValueError("`truncate` must be strictly positive")
        return v

    @validator("typical_p")
    def valid_typical_p(cls, v):
        if v is not None and (v <= 0 or v >= 1.0):
            raise ValueError("`typical_p` must be > 0.0 and < 1.0")
        return v


@dataclass
class TextGenerationRequest:
    """
    Request object for text generation (only for internal use).

    Args:
        inputs (`str`):
            The prompt for text generation.
        parameters (`Optional[TextGenerationParameters]`, *optional*):
            Generation parameters.
        stream (`bool`, *optional*):
            Whether to stream output tokens. Defaults to False.
    """

    # Prompt
    inputs: str
    # Generation parameters
    parameters: Optional[TextGenerationParameters] = None
    # Whether to stream output tokens
    stream: bool = False

    @validator("inputs")
    def valid_input(cls, v):
        if not v:
            raise ValueError("`inputs` cannot be empty")
        return v

    @validator("stream")
    def valid_best_of_stream(cls, field_value, values):
        parameters = values["parameters"]
        if parameters is not None and parameters.best_of is not None and parameters.best_of > 1 and field_value:
            raise ValueError("`best_of` != 1 is not supported when `stream` == True")
        return field_value

    def __post_init__(self):
        if not is_pydantic_available():
            # If pydantic is not installed, we need to instantiate the nested dataclasses manually
            if self.parameters is not None and isinstance(self.parameters, dict):
                self.parameters = TextGenerationParameters(**self.parameters)


# Decoder input tokens
@dataclass
class InputToken:
    """
    Represents an input token.

    Args:
        id (`int`):
            Token ID from the model tokenizer.
        text (`str`):
            Token text.
        logprob (`float` or `None`):
            Log probability of the token. Optional since the logprob of the first token cannot be computed.
    """

    # Token ID from the model tokenizer
    id: int
    # Token text
    text: str
    # Logprob
    # Optional since the logprob of the first token cannot be computed
    logprob: Optional[float] = None


# Generated tokens
@dataclass
class Token:
    """
    Represents a token.

    Args:
        id (`int`):
            Token ID from the model tokenizer.
        text (`str`):
            Token text.
        logprob (`float`):
            Log probability of the token.
        special (`bool`):
            Indicates whether the token is a special token. It can be used to ignore
            tokens when concatenating.
    """

    # Token ID from the model tokenizer
    id: int
    # Token text
    text: str
    # Logprob
    logprob: float
    # Is the token a special token
    # Can be used to ignore tokens when concatenating
    special: bool


# Generation finish reason
class FinishReason(str, Enum):
    # number of generated tokens == `max_new_tokens`
    Length = "length"
    # the model generated its end of sequence token
    EndOfSequenceToken = "eos_token"
    # the model generated a text included in `stop_sequences`
    StopSequence = "stop_sequence"


# Additional sequences when using the `best_of` parameter
@dataclass
class BestOfSequence:
    """
    Represents a best-of sequence generated during text generation.

    Args:
        generated_text (`str`):
            The generated text.
        finish_reason (`FinishReason`):
            The reason for the generation to finish, represented by a `FinishReason` value.
        generated_tokens (`int`):
            The number of generated tokens in the sequence.
        seed (`Optional[int]`):
            The sampling seed if sampling was activated.
        prefill (`List[InputToken]`):
            The decoder input tokens. Empty if `decoder_input_details` is False. Defaults to an empty list.
        tokens (`List[Token]`):
            The generated tokens. Defaults to an empty list.
    """

    # Generated text
    generated_text: str
    # Generation finish reason
    finish_reason: FinishReason
    # Number of generated tokens
    generated_tokens: int
    # Sampling seed if sampling was activated
    seed: Optional[int] = None
    # Decoder input tokens, empty if decoder_input_details is False
    prefill: List[InputToken] = field(default_factory=lambda: [])
    # Generated tokens
    tokens: List[Token] = field(default_factory=lambda: [])

    def __post_init__(self):
        if not is_pydantic_available():
            # If pydantic is not installed, we need to instantiate the nested dataclasses manually
            self.prefill = [
                InputToken(**input_token) if isinstance(input_token, dict) else input_token
                for input_token in self.prefill
            ]
            self.tokens = [Token(**token) if isinstance(token, dict) else token for token in self.tokens]


# `generate` details
@dataclass
class Details:
    """
    Represents details of a text generation.

    Args:
        finish_reason (`FinishReason`):
            The reason for the generation to finish, represented by a `FinishReason` value.
        generated_tokens (`int`):
            The number of generated tokens.
        seed (`Optional[int]`):
            The sampling seed if sampling was activated.
        prefill (`List[InputToken]`, *optional*):
            The decoder input tokens. Empty if `decoder_input_details` is False. Defaults to an empty list.
        tokens (`List[Token]`):
            The generated tokens. Defaults to an empty list.
        best_of_sequences (`Optional[List[BestOfSequence]]`):
            Additional sequences when using the `best_of` parameter.
    """

    # Generation finish reason
    finish_reason: FinishReason
    # Number of generated tokens
    generated_tokens: int
    # Sampling seed if sampling was activated
    seed: Optional[int] = None
    # Decoder input tokens, empty if decoder_input_details is False
    prefill: List[InputToken] = field(default_factory=lambda: [])
    # Generated tokens
    tokens: List[Token] = field(default_factory=lambda: [])
    # Additional sequences when using the `best_of` parameter
    best_of_sequences: Optional[List[BestOfSequence]] = None

    def __post_init__(self):
        if not is_pydantic_available():
            # If pydantic is not installed, we need to instantiate the nested dataclasses manually
            self.prefill = [
                InputToken(**input_token) if isinstance(input_token, dict) else input_token
                for input_token in self.prefill
            ]
            self.tokens = [Token(**token) if isinstance(token, dict) else token for token in self.tokens]
            if self.best_of_sequences is not None:
                self.best_of_sequences = [
                    BestOfSequence(**best_of_sequence) if isinstance(best_of_sequence, dict) else best_of_sequence
                    for best_of_sequence in self.best_of_sequences
                ]


# `generate` return value
@dataclass
class TextGenerationResponse:
    """
    Represents a response for text generation.

    Only returned when `details=True`, otherwise a string is returned.

    Args:
        generated_text (`str`):
            The generated text.
        details (`Optional[Details]`):
            Generation details. Returned only if `details=True` is sent to the server.
    """

    # Generated text
    generated_text: str
    # Generation details
    details: Optional[Details] = None

    def __post_init__(self):
        if not is_pydantic_available():
            # If pydantic is not installed, we need to instantiate the nested dataclasses manually
            if self.details is not None and isinstance(self.details, dict):
                self.details = Details(**self.details)


# `generate_stream` details
@dataclass
class StreamDetails:
    """
    Represents details of a text generation stream.

    Args:
        finish_reason (`FinishReason`):
            The reason for the generation to finish, represented by a `FinishReason` value.
        generated_tokens (`int`):
            The number of generated tokens.
        seed (`Optional[int]`):
            The sampling seed if sampling was activated.
    """

    # Generation finish reason
    finish_reason: FinishReason
    # Number of generated tokens
    generated_tokens: int
    # Sampling seed if sampling was activated
    seed: Optional[int] = None


# `generate_stream` return value
@dataclass
class TextGenerationStreamResponse:
    """
    Represents a response for streaming text generation.

    Only returned when `details=True` and `stream=True`.

    Args:
        token (`Token`):
            The generated token.
        generated_text (`Optional[str]`, *optional*):
            The complete generated text. Only available when the generation is finished.
        details (`Optional[StreamDetails]`, *optional*):
            Generation details. Only available when the generation is finished.
    """

    # Generated token
    token: Token
    # Complete generated text
    # Only available when the generation is finished
    generated_text: Optional[str] = None
    # Generation details
    # Only available when the generation is finished
    details: Optional[StreamDetails] = None

    def __post_init__(self):
        if not is_pydantic_available():
            # If pydantic is not installed, we need to instantiate the nested dataclasses manually
            if isinstance(self.token, dict):
                self.token = Token(**self.token)
            if self.details is not None and isinstance(self.details, dict):
                self.details = StreamDetails(**self.details)


# TEXT GENERATION ERRORS
# ----------------------
# Text-generation errors are parsed separately to handle as much as possible the errors returned by the text generation
# inference project (https://github.com/huggingface/text-generation-inference).
# ----------------------


class TextGenerationError(HTTPError):
    """Generic error raised if text-generation went wrong."""


# Text Generation Inference Errors
class ValidationError(TextGenerationError):
    """Server-side validation error."""


class GenerationError(TextGenerationError):
    pass


class OverloadedError(TextGenerationError):
    pass


class IncompleteGenerationError(TextGenerationError):
    pass


class UnknownError(TextGenerationError):
    pass


def raise_text_generation_error(http_error: HTTPError) -> NoReturn:
    """
    Try to parse text-generation-inference error message and raise HTTPError in any case.

    Args:
        error (`HTTPError`):
            The HTTPError that have been raised.
    """
    # Try to parse a Text Generation Inference error

    try:
        # Hacky way to retrieve payload in case of aiohttp error
        payload = getattr(http_error, "response_error_payload", None) or http_error.response.json()
        error = payload.get("error")
        error_type = payload.get("error_type")
    except Exception:  # no payload
        raise http_error

    # If error_type => more information than `hf_raise_for_status`
    if error_type is not None:
        exception = _parse_text_generation_error(error, error_type)
        raise exception from http_error

    # Otherwise, fallback to default error
    raise http_error


def _parse_text_generation_error(error: Optional[str], error_type: Optional[str]) -> TextGenerationError:
    if error_type == "generation":
        return GenerationError(error)  # type: ignore
    if error_type == "incomplete_generation":
        return IncompleteGenerationError(error)  # type: ignore
    if error_type == "overloaded":
        return OverloadedError(error)  # type: ignore
    if error_type == "validation":
        return ValidationError(error)  # type: ignore
    return UnknownError(error)  # type: ignore