lightrag-cn-git/lightrag/llm/lmdeploy.py

import warnings

import pipmaster as pm  # Pipmaster for dynamic library install

# install specific modules
if not pm.is_installed("lmdeploy"):
    pm.install("lmdeploy[all]")

from lightrag.exceptions import (
    APIConnectionError,
    RateLimitError,
    APITimeoutError,
)
from tenacity import (
    retry,
    stop_after_attempt,
    wait_exponential,
    retry_if_exception_type,
)


from functools import lru_cache


@lru_cache(maxsize=1)
def initialize_lmdeploy_pipeline(
    model,
    tp=1,
    chat_template=None,
    log_level="WARNING",
    model_format="hf",
    quant_policy=0,
):
    from lmdeploy import pipeline, ChatTemplateConfig, TurbomindEngineConfig

    lmdeploy_pipe = pipeline(
        model_path=model,
        backend_config=TurbomindEngineConfig(
            tp=tp, model_format=model_format, quant_policy=quant_policy
        ),
        chat_template_config=(
            ChatTemplateConfig(model_name=chat_template) if chat_template else None
        ),
        log_level="WARNING",
    )
    return lmdeploy_pipe


@retry(
    stop=stop_after_attempt(3),
    wait=wait_exponential(multiplier=1, min=4, max=10),
    retry=retry_if_exception_type(
        (RateLimitError, APIConnectionError, APITimeoutError)
    ),
)
async def lmdeploy_model_if_cache(
    model,
    prompt,
    system_prompt=None,
    history_messages=[],
    enable_cot: bool = False,
    chat_template=None,
    model_format="hf",
    quant_policy=0,
    **kwargs,
) -> str:
    """Run lmdeploy generation with LightRAG-compatible shims.

    Structured output note:
    - This adapter does not support OpenAI-style ``response_format`` JSON mode.
    - If callers pass ``response_format``, it is stripped before generation.
    - Deprecated ``keyword_extraction`` and ``entity_extraction`` booleans are
      accepted only as compatibility shims; they emit warnings and are ignored.

    Args:
        model (str): The path to the model.
            It could be one of the following options:
                    - i) A local directory path of a turbomind model which is
                        converted by `lmdeploy convert` command or download
                        from ii) and iii).
                    - ii) The model_id of a lmdeploy-quantized model hosted
                        inside a model repo on huggingface.co, such as
                        "InternLM/internlm-chat-20b-4bit",
                        "lmdeploy/llama2-chat-70b-4bit", etc.
                    - iii) The model_id of a model hosted inside a model repo
                        on huggingface.co, such as "internlm/internlm-chat-7b",
                        "Qwen/Qwen-7B-Chat ", "baichuan-inc/Baichuan2-7B-Chat"
                        and so on.
        chat_template (str): needed when model is a pytorch model on
            huggingface.co, such as "internlm-chat-7b",
            "Qwen-7B-Chat ", "Baichuan2-7B-Chat" and so on,
            and when the model name of local path did not match the original model name in HF.
        tp (int): tensor parallel
        prompt (Union[str, List[str]]): input texts to be completed.
        do_preprocess (bool): whether pre-process the messages. Default to
            True, which means chat_template will be applied.
        skip_special_tokens (bool): Whether or not to remove special tokens
            in the decoding. Default to be True.
        do_sample (bool): Whether or not to use sampling, use greedy decoding otherwise.
            Default to be False, which means greedy decoding will be applied.
    """
    if enable_cot:
        from lightrag.utils import logger

        logger.debug(
            "enable_cot=True is not supported for lmdeploy and will be ignored."
        )
    try:
        import lmdeploy
        from lmdeploy import version_info, GenerationConfig
    except Exception:
        raise ImportError("Please install lmdeploy before initialize lmdeploy backend.")
    kwargs.pop("hashing_kv", None)
    # lmdeploy has no JSON mode; drop response_format and warn when legacy
    # boolean shim flags are set.
    if kwargs.pop("keyword_extraction", False):
        warnings.warn(
            "lmdeploy_model_if_cache(keyword_extraction=True) is deprecated; "
            "pass response_format={'type': 'json_object'} instead.",
            DeprecationWarning,
            stacklevel=2,
        )
    if kwargs.pop("entity_extraction", False):
        warnings.warn(
            "lmdeploy_model_if_cache(entity_extraction=True) is deprecated; "
            "pass response_format={'type': 'json_object'} instead.",
            DeprecationWarning,
            stacklevel=2,
        )
    kwargs.pop("response_format", None)
    max_new_tokens = kwargs.pop("max_tokens", 512)
    tp = kwargs.pop("tp", 1)
    skip_special_tokens = kwargs.pop("skip_special_tokens", True)
    do_preprocess = kwargs.pop("do_preprocess", True)
    do_sample = kwargs.pop("do_sample", False)
    gen_params = kwargs

    version = version_info
    if do_sample is not None and version < (0, 6, 0):
        raise RuntimeError(
            "`do_sample` parameter is not supported by lmdeploy until "
            f"v0.6.0, but currently using lmdeloy {lmdeploy.__version__}"
        )
    else:
        do_sample = True
        gen_params.update(do_sample=do_sample)

    lmdeploy_pipe = initialize_lmdeploy_pipeline(
        model=model,
        tp=tp,
        chat_template=chat_template,
        model_format=model_format,
        quant_policy=quant_policy,
        log_level="WARNING",
    )

    messages = []
    if system_prompt:
        messages.append({"role": "system", "content": system_prompt})

    messages.extend(history_messages)
    messages.append({"role": "user", "content": prompt})

    gen_config = GenerationConfig(
        skip_special_tokens=skip_special_tokens,
        max_new_tokens=max_new_tokens,
        **gen_params,
    )

    response = ""
    async for res in lmdeploy_pipe.generate(
        messages,
        gen_config=gen_config,
        do_preprocess=do_preprocess,
        stream_response=False,
        session_id=1,
    ):
        response += res.response
    return response