跳转至

4 最佳实践

推荐的模型组合⚓︎

  • 在默认的配置文件中,我们提供了以下模型组合

    LLM: Chatglm2-6b
    Embedding Models: m3e-base
    TextSplitter: ChineseRecursiveTextSplitter
    Kb_dataset: faiss
    

  • 我们推荐开发者根据自己的业务需求进行模型微调,如果不需要微调且配置充足,可选择以下性能较好的配置

    model_config.py
    LLM: Qwen-14B-Chat 或 Baichuan2-13B-Chat
    Embedding Models: piccolo-large-zh 或 bge-large-zh-v1.5 
    HISTORY_LEN = 20
    TEMPERATURE = 0.1
    
    使用该模型将需要更高的硬件要求
    1张 RTX A6000 或者 A40 等 48GB 显存以上的显卡。推荐 1 x A100 以上。
    (使用多张显卡拼接也能运行,但是速度非常慢,2张4090拼接运行大概为一秒一个字的速度)
    
    64GB 内存用于加载模型而不被Kill
    
    服务器级的CPU,推荐 Xeon(R) Platinum 8358P 以上
    

  • 如果开发者知识库较大,有大量文档,大文件,我们推荐开发者使用 pg 向量数据库

  • 如果开发者的知识库具有一定的关键词特征,例如:

    • 问答对文件(以Q + A 为一个组合的json文件)
    • Markdown文件
    • 并排的pdf文件
    • 具有多个表格的pdf文件

    我们推荐开发者自行开发分词器,以达到更好的效果。

  • 如果开发者想使用更全面的 Agent 功能,我们推荐开发者使用以下配置

    LLM: Qwen-14B-Chat, AgentLM-70B 或 GPT-4
    Tools 的工具控制在10个之内
    

微调模型加载实操⚓︎

非p-tuning类PEFT加载⚓︎

本项目基于 FastChat 加载 LLM 服务,故需以 FastChat 加载 PEFT 路径,针对chatglm,falcon,codet5p以外的模型,以及非p-tuning以外的peft方法,需对peft文件进行修改,步骤如下:

  1. 将config.json文件修改为adapter_config.json;
  2. 保证文件夹包含pytorch_model.bin文件;
  3. 修改文件夹名称,保证文件夹包含'peft'一词;
  4. 将peft文件夹移入项目目录下;
  5. 确保adapter_config.json文件夹中base_model_name_or_path指向基础模型;
  6. 将peft路径添加到model_config.py的llm_dict中,键为模型名,值为peft路径,注意使用相对路径,如"peft";
  7. 开启 PEFT_SHARE_BASE_WEIGHTS=true环境变量,再执行python startup.py -a

针对p-tuning和chatglm模型,需要对fastchat进行较大幅度的修改。

p-tuning加载⚓︎

P-tuning虽然是一种peft方法,但并不能于huggingface的peft python包兼容,而fastchat在多处以字符串匹配的方式进行硬编码加载模型,因此导致fastchat和chatchat不能兼容p-tuning,经langchain-chatchat开发组多次尝试,给出如下指南进行p-tuning加载。

1. peft文件夹修改⚓︎

  1. 将config.json文件修改为adapter_config.json;
  2. 保证文件夹包含pytorch_model.bin文件;
  3. 修改文件夹名称,保证文件夹包含'peft'一词;
  4. 在adapter_config.json文件中增加如下字段:
    "base_model_name_or_path": "/root/model/chatglm2-6b/"
    "task_type": "CAUSAL_LM",
    "peft_type": "PREFIX_TUNING",
    "inference_mode": true,
    "revision": "main",
    "num_virtual_tokens": 16

其中,"base_model_name_or_path"为基础模型的存在位置; 5. 将文件夹移入项目文件夹中,如Langchain-Chatchat项目文件夹目录下;

2. fastchat包代码修改⚓︎

2.1 fastchat.model.model_adapter文件修改⚓︎
  1. 将fastchat.model.model_adapter.py文件的load_model函数修改为:

def load_model(
    model_path: str,
    device: str = "cuda",
    num_gpus: int = 1,
    max_gpu_memory: Optional[str] = None,
    dtype: Optional[torch.dtype] = None,
    load_8bit: bool = False,
    cpu_offloading: bool = False,
    gptq_config: Optional[GptqConfig] = None,
    awq_config: Optional[AWQConfig] = None,
    revision: str = "main",
    debug: bool = False,
    load_kwargs = {}
):
    """Load a model from Hugging Face."""
    # get model adapter
    adapter = get_model_adapter(model_path)
    kwargs = load_kwargs
    # Handle device mapping
    cpu_offloading = raise_warning_for_incompatible_cpu_offloading_configuration(
        device, load_8bit, cpu_offloading
    )
    if device == "cpu":
        kwargs["torch_dtype"]= torch.float32
        if CPU_ISA in ["avx512_bf16", "amx"]:
            try:
                import intel_extension_for_pytorch as ipex

                kwargs ["torch_dtype"]= torch.bfloat16
            except ImportError:
                warnings.warn(
                    "Intel Extension for PyTorch is not installed, it can be installed to accelerate cpu inference"
                )
    elif device == "cuda":
        kwargs["torch_dtype"] = torch.float16
        if num_gpus != 1:
            kwargs["device_map"] = "auto"
            if max_gpu_memory is None:
                kwargs[
                    "device_map"
                ] = "sequential"  # This is important for not the same VRAM sizes
                available_gpu_memory = get_gpu_memory(num_gpus)
                kwargs["max_memory"] = {
                    i: str(int(available_gpu_memory[i] * 0.85)) + "GiB"
                    for i in range(num_gpus)
                }
            else:
                kwargs["max_memory"] = {i: max_gpu_memory for i in range(num_gpus)}
    elif device == "mps":
        kwargs["torch_dtype"] = torch.float16
        # Avoid bugs in mps backend by not using in-place operations.
        replace_llama_attn_with_non_inplace_operations()
    elif device == "xpu":
        kwargs["torch_dtype"] = torch.bfloat16
        # Try to load ipex, while it looks unused, it links into torch for xpu support
        try:
            import intel_extension_for_pytorch as ipex
        except ImportError:
            warnings.warn(
                "Intel Extension for PyTorch is not installed, but is required for xpu inference."
            )
    elif device == "npu":
        kwargs["torch_dtype"]= torch.float16
        # Try to load ipex, while it looks unused, it links into torch for xpu support
        try:
            import torch_npu
        except ImportError:
            warnings.warn("Ascend Extension for PyTorch is not installed.")
    else:
        raise ValueError(f"Invalid device: {device}")

    if cpu_offloading:
        # raises an error on incompatible platforms
        from transformers import BitsAndBytesConfig

        if "max_memory" in kwargs:
            kwargs["max_memory"]["cpu"] = (
                str(math.floor(psutil.virtual_memory().available / 2**20)) + "Mib"
            )
        kwargs["quantization_config"] = BitsAndBytesConfig(
            load_in_8bit_fp32_cpu_offload=cpu_offloading
        )
        kwargs["load_in_8bit"] = load_8bit
    elif load_8bit:
        if num_gpus != 1:
            warnings.warn(
                "8-bit quantization is not supported for multi-gpu inference."
            )
        else:
            model, tokenizer = adapter.load_compress_model(
                model_path=model_path,
                device=device,
                torch_dtype=kwargs["torch_dtype"],
                revision=revision,
            )
            if debug:
                print(model)
            return model, tokenizer
    elif awq_config and awq_config.wbits < 16:
        assert (
            awq_config.wbits == 4
        ), "Currently we only support 4-bit inference for AWQ."
        model, tokenizer = load_awq_quantized(model_path, awq_config, device)
        if num_gpus != 1:
            device_map = accelerate.infer_auto_device_map(
                model,
                max_memory=kwargs["max_memory"],
                no_split_module_classes=[
                    "OPTDecoderLayer",
                    "LlamaDecoderLayer",
                    "BloomBlock",
                    "MPTBlock",
                    "DecoderLayer",
                ],
            )
            model = accelerate.dispatch_model(
                model, device_map=device_map, offload_buffers=True
            )
        else:
            model.to(device)
        return model, tokenizer
    elif gptq_config and gptq_config.wbits < 16:
        model, tokenizer = load_gptq_quantized(model_path, gptq_config)
        if num_gpus != 1:
            device_map = accelerate.infer_auto_device_map(
                model,
                max_memory=kwargs["max_memory"],
                no_split_module_classes=["LlamaDecoderLayer"],
            )
            model = accelerate.dispatch_model(
                model, device_map=device_map, offload_buffers=True
            )
        else:
            model.to(device)
        return model, tokenizer
    kwargs["revision"] = revision

    if dtype is not None:  # Overwrite dtype if it is provided in the arguments.
        kwargs["torch_dtype"] = dtype

    # Load model
    model, tokenizer = adapter.load_model(model_path, kwargs)

    if (
        device == "cpu"
        and kwargs["torch_dtype"] is torch.bfloat16
        and CPU_ISA is not None
    ):
        model = ipex.optimize(model, dtype=kwargs["torch_dtype"])

    if (device == "cuda" and num_gpus == 1 and not cpu_offloading) or device in (
        "mps",
        "xpu",
        "npu",
    ):
        model.to(device)

    if device == "xpu":
        model = torch.xpu.optimize(model, dtype=kwargs["torch_dtype"], inplace=True)

    if debug:
        print(model)

    return model, tokenizer
2. 将fastchat.model.model_adapter.py的函数修改为:

def get_generate_stream_function(model: torch.nn.Module, model_path: str):
    """Get the generate_stream function for inference."""
    from fastchat.serve.inference import generate_stream

    model_type = str(type(model)).lower()

    is_chatglm = "chatglm" in model_type 
    is_falcon = "rwforcausallm" in model_type
    is_codet5p = "codet5p" in model_type 
    is_peft = "peft" in model_type

    if is_chatglm:
        return generate_stream_chatglm
    elif is_falcon:
        return generate_stream_falcon
    elif is_codet5p:
        return generate_stream_codet5p
    elif peft_share_base_weights and is_peft:
        # Return a curried stream function that loads the right adapter
        # according to the model_name available in this context.  This ensures
        # the right weights are available.
        @torch.inference_mode()
        def generate_stream_peft(
            model,
            tokenizer,
            params: Dict,
            device: str,
            context_len: int,
            stream_interval: int = 2,
            judge_sent_end: bool = False,
        ):

            model.set_adapter(model_path)
            if "chatglm" in str(type(model.base_model)).lower():
                model.disable_adapter()
                prefix_state_dict = torch.load(os.path.join(model_path, "pytorch_model.bin"))
                new_prefix_state_dict = {}

                for k, v in prefix_state_dict.items():
                    if k.startswith("transformer.prefix_encoder."):
                        new_prefix_state_dict[k[len("transformer.prefix_encoder."):]] = v
                    elif k.startswith("transformer.prompt_encoder."):
                        new_prefix_state_dict[k[len("transformer.prompt_encoder."):]] = v
                model.transformer.prefix_encoder.load_state_dict(new_prefix_state_dict)
                for x in generate_stream_chatglm(
                    model,
                    tokenizer,
                    params,
                    device,
                    context_len,
                    stream_interval,
                    judge_sent_end,
                ):
                    yield x
            elif "rwforcausallm" in str(type(model.base_model)).lower():

                for x in generate_stream_falcon(
                    model,
                    tokenizer,
                    params,
                    device,
                    context_len,
                    stream_interval,
                    judge_sent_end,
                ):
                    yield x   
            elif "codet5p" in str(type(model.base_model)).lower():

                for x in generate_stream_codet5p(
                    model,
                    tokenizer,
                    params,
                    device,
                    context_len,
                    stream_interval,
                    judge_sent_end,
                ):
                    yield x   
            else:

                for x in generate_stream(
                    model,
                    tokenizer,
                    params,
                    device,
                    context_len,
                    stream_interval,
                    judge_sent_end,
                ):
                    yield x

        return generate_stream_peft
    else:
        return generate_stream
3. 将fastchat.model.model_adapter.py的PeftModelAdapter类的load_model方法修改为:

    def load_model(self, model_path: str, from_pretrained_kwargs: dict):
        """Loads the base model then the (peft) adapter weights"""
        from peft import PeftConfig, PeftModel

        config = PeftConfig.from_pretrained(model_path)
        base_model_path = config.base_model_name_or_path
        if "peft" in base_model_path:
            raise ValueError(
                f"PeftModelAdapter cannot load a base model with 'peft' in the name: {config.base_model_name_or_path}"
            )

        # Basic proof of concept for loading peft adapters that share the base
        # weights.  This is pretty messy because Peft re-writes the underlying
        # base model and internally stores a map of adapter layers.
        # So, to make this work we:
        #  1. Cache the first peft model loaded for a given base models.
        #  2. Call `load_model` for any follow on Peft models.
        #  3. Make sure we load the adapters by the model_path.  Why? This is
        #  what's accessible during inference time.
        #  4. In get_generate_stream_function, make sure we load the right
        #  adapter before doing inference.  This *should* be safe when calls
        #  are blocked the same semaphore.
        if peft_share_base_weights:
            if base_model_path in peft_model_cache:
                model, tokenizer = peft_model_cache[base_model_path]
                # Super important: make sure we use model_path as the
                # `adapter_name`.
                model.load_adapter(model_path, adapter_name=model_path)
            else:
                base_adapter = get_model_adapter(base_model_path)
                base_model, tokenizer = base_adapter.load_model(
                    base_model_path, from_pretrained_kwargs
                )
                # Super important: make sure we use model_path as the
                # `adapter_name`.
                from peft import get_peft_model
                model = get_peft_model(base_model,config,adapter_name=model_path)
                peft_model_cache[base_model_path] = (model, tokenizer)
            return model, tokenizer

        # In the normal case, load up the base model weights again.
        base_adapter = get_model_adapter(base_model_path)
        base_model, tokenizer = base_adapter.load_model(
            base_model_path, from_pretrained_kwargs
        )
        from peft import get_peft_model
        model = get_peft_model(base_model,config,adapter_name=model_path)
        return model, tokenizer
4. 将fastchat.model.model_adapter.py的ChatglmAdapter类的load_model方法修改为:

    def load_model(self, model_path: str, from_pretrained_kwargs: dict):
        revision = from_pretrained_kwargs.get("revision", "main")
        tokenizer = AutoTokenizer.from_pretrained(
            model_path, trust_remote_code=True, revision=revision
        )
        config = AutoConfig.from_pretrained(model_path, trust_remote_code=True,**from_pretrained_kwargs)
        model = AutoModel.from_pretrained(
            model_path, trust_remote_code=True, config=config
        )
        return model, tokenizer
2.2 fastchat.serve.model_worker文件修改⚓︎
  1. 将fastchat.serve.model_worker文件的ModelWorker的__init__方法修改如下:

class ModelWorker(BaseModelWorker):
    def __init__(
        self,
        controller_addr: str,
        worker_addr: str,
        worker_id: str,
        model_path: str,
        model_names: List[str],
        limit_worker_concurrency: int,
        no_register: bool,
        device: str,
        num_gpus: int,
        max_gpu_memory: str,
        dtype: Optional[torch.dtype] = None,
        load_8bit: bool = False,
        cpu_offloading: bool = False,
        gptq_config: Optional[GptqConfig] = None,
        awq_config: Optional[AWQConfig] = None,
        stream_interval: int = 2,
        conv_template: Optional[str] = None,
        embed_in_truncate: bool = False,
        seed: Optional[int] = None,
        load_kwargs = {}, #修改点
        **kwargs,
    ):
        super().__init__(
            controller_addr,
            worker_addr,
            worker_id,
            model_path,
            model_names,
            limit_worker_concurrency,
            conv_template=conv_template,
        )

        logger.info(f"Loading the model {self.model_names} on worker {worker_id} ...")
        self.model, self.tokenizer = load_model(
            model_path,
            device=device,
            num_gpus=num_gpus,
            max_gpu_memory=max_gpu_memory,
            dtype=dtype,
            load_8bit=load_8bit,
            cpu_offloading=cpu_offloading,
            gptq_config=gptq_config,
            awq_config=awq_config,
            load_kwargs=load_kwargs #修改点
        )
        self.device = device
        if self.tokenizer.pad_token == None:
            self.tokenizer.pad_token = self.tokenizer.eos_token
        self.context_len = get_context_length(self.model.config)
        print("**"*100)
        self.generate_stream_func = get_generate_stream_function(self.model, model_path)
        print(f"self.generate_stream_func{self.generate_stream_func}")
        print("*"*100)
        self.stream_interval = stream_interval
        self.embed_in_truncate = embed_in_truncate
        self.seed = seed

        if not no_register:
            self.init_heart_beat()
2. 在fastchat.serve.model_worker文件的create_model_worker增加如下args参数:

parser.add_argument("--load_kwargs",type=dict,default={})
并将如下语句:
    worker = ModelWorker(
        args.controller_address,
        args.worker_address,
        worker_id,
        args.model_path,
        args.model_names,
        args.limit_worker_concurrency,
        no_register=args.no_register,
        device=args.device,
        num_gpus=args.num_gpus,
        max_gpu_memory=args.max_gpu_memory,
        dtype=str_to_torch_dtype(args.dtype),
        load_8bit=args.load_8bit,
        cpu_offloading=args.cpu_offloading,
        gptq_config=gptq_config,
        awq_config=awq_config,
        stream_interval=args.stream_interval,
        conv_template=args.conv_template,
        embed_in_truncate=args.embed_in_truncate,
        seed=args.seed,
    )

修改为:

    worker = ModelWorker(
        args.controller_address,
        args.worker_address,
        worker_id,
        args.model_path,
        args.model_names,
        args.limit_worker_concurrency,
        no_register=args.no_register,
        device=args.device,
        num_gpus=args.num_gpus,
        max_gpu_memory=args.max_gpu_memory,
        dtype=str_to_torch_dtype(args.dtype),
        load_8bit=args.load_8bit,
        cpu_offloading=args.cpu_offloading,
        gptq_config=gptq_config,
        awq_config=awq_config,
        stream_interval=args.stream_interval,
        conv_template=args.conv_template,
        embed_in_truncate=args.embed_in_truncate,
        seed=args.seed,
        load_kwargs=args.load_kwargs
    )

至此,我们完成了fastchat加载ptuning的所有修改,在调用fastchat加载p-tuning时,可以通过加入 PEFT_SHARE_BASE_WEIGHTS=true,并以字典的形式添加--load_kwargs参数为训练ptuning时的pre_seq_len值即可,例如将2.2.2步骤中的 parser.add_argument("--load_kwargs",type=dict,default={})修改为:

parser.add_argument("--load_kwargs",type=dict,default={"pre_seq_len":16})

3 langchain-chatchat代码修改:⚓︎

  1. 在configs/serve_config.py中的FSCHAT_MODEL_WORKERS字典中增加如下字段:

"load_kwargs": {"pre_seq_len": 16} #值修改为adapter_config.json中的pre_seq_len值
2. 将startup.py中的create_model_worker_app修改为:

def create_model_worker_app(log_level: str = "INFO", **kwargs) -> FastAPI:
    """
    kwargs包含的字段如下:
    host:
    port:
    model_names:[`model_name`]
    controller_address:
    worker_address:


    对于online_api:
        online_api:True
        worker_class: `provider`
    对于离线模型:
        model_path: `model_name_or_path`,huggingface的repo-id或本地路径
        device:`LLM_DEVICE`
    """
    import fastchat.constants
    fastchat.constants.LOGDIR = LOG_PATH
    from fastchat.serve.model_worker import worker_id, logger
    import argparse
    logger.setLevel(log_level)

    parser = argparse.ArgumentParser()
    args = parser.parse_args([])

    for k, v in kwargs.items():
        setattr(args, k, v)

    # 在线模型API
    if worker_class := kwargs.get("worker_class"):
        from fastchat.serve.model_worker import app
        worker = worker_class(model_names=args.model_names,
                              controller_addr=args.controller_address,
                              worker_addr=args.worker_address)
        sys.modules["fastchat.serve.model_worker"].worker = worker
    # 本地模型
    else:
        from configs.model_config import VLLM_MODEL_DICT
        if kwargs["model_names"][0] in VLLM_MODEL_DICT and args.infer_turbo == "vllm":
            import fastchat.serve.vllm_worker
            from fastchat.serve.vllm_worker import VLLMWorker,app
            from vllm import AsyncLLMEngine
            from vllm.engine.arg_utils import AsyncEngineArgs,EngineArgs
            args.tokenizer = args.model_path # 如果tokenizer与model_path不一致在此处添加
            args.tokenizer_mode = 'auto'
            args.trust_remote_code= True
            args.download_dir= None
            args.load_format = 'auto'
            args.dtype = 'auto'
            args.seed = 0
            args.worker_use_ray = False
            args.pipeline_parallel_size = 1
            args.tensor_parallel_size = 1
            args.block_size = 16
            args.swap_space = 4  # GiB
            args.gpu_memory_utilization = 0.90
            args.max_num_batched_tokens = 2560
            args.max_num_seqs = 256
            args.disable_log_stats = False
            args.conv_template = None
            args.limit_worker_concurrency = 5
            args.no_register = False
            args.num_gpus = 1 # vllm worker的切分是tensor并行,这里填写显卡的数量
            args.engine_use_ray = False
            args.disable_log_requests = False
            if args.model_path:
                args.model = args.model_path
            if args.num_gpus > 1:
                args.tensor_parallel_size = args.num_gpus

            for k, v in kwargs.items():
                setattr(args, k, v)

            engine_args = AsyncEngineArgs.from_cli_args(args)
            engine = AsyncLLMEngine.from_engine_args(engine_args)

            worker = VLLMWorker(
                        controller_addr = args.controller_address,
                        worker_addr = args.worker_address,
                        worker_id = worker_id,
                        model_path = args.model_path,
                        model_names = args.model_names,
                        limit_worker_concurrency = args.limit_worker_concurrency,
                        no_register = args.no_register,
                        llm_engine =  engine,
                        conv_template = args.conv_template,
                        )
            sys.modules["fastchat.serve.vllm_worker"].engine = engine
            sys.modules["fastchat.serve.vllm_worker"].worker = worker

        else:
            from fastchat.serve.model_worker import app, GptqConfig, AWQConfig, ModelWorker
            args.gpus = "0" # GPU的编号,如果有多个GPU,可以设置为"0,1,2,3"
            args.max_gpu_memory = "20GiB"
            args.num_gpus = 1  # model worker的切分是model并行,这里填写显卡的数量

            args.load_8bit = False
            args.cpu_offloading = None
            args.gptq_ckpt = None
            args.gptq_wbits = 16
            args.gptq_groupsize = -1
            args.gptq_act_order = False
            args.awq_ckpt = None
            args.awq_wbits = 16
            args.awq_groupsize = -1
            args.model_names = []
            args.conv_template = None
            args.limit_worker_concurrency = 5
            args.stream_interval = 2
            args.no_register = False
            args.embed_in_truncate = False
            args.load_kwargs = {"pre_seq_len": 16} # 改*************************
            for k, v in kwargs.items():
                setattr(args, k, v)
            if args.gpus:
                if args.num_gpus is None:
                    args.num_gpus = len(args.gpus.split(','))
                if len(args.gpus.split(",")) < args.num_gpus:
                    raise ValueError(
                        f"Larger --num-gpus ({args.num_gpus}) than --gpus {args.gpus}!"
                    )
                os.environ["CUDA_VISIBLE_DEVICES"] = args.gpus
            gptq_config = GptqConfig(
                ckpt=args.gptq_ckpt or args.model_path,
                wbits=args.gptq_wbits,
                groupsize=args.gptq_groupsize,
                act_order=args.gptq_act_order,
            )
            awq_config = AWQConfig(
                ckpt=args.awq_ckpt or args.model_path,
                wbits=args.awq_wbits,
                groupsize=args.awq_groupsize,
            )

            worker = ModelWorker(
                controller_addr=args.controller_address,
                worker_addr=args.worker_address,
                worker_id=worker_id,
                model_path=args.model_path,
                model_names=args.model_names,
                limit_worker_concurrency=args.limit_worker_concurrency,
                no_register=args.no_register,
                device=args.device,
                num_gpus=args.num_gpus,
                max_gpu_memory=args.max_gpu_memory,
                load_8bit=args.load_8bit,
                cpu_offloading=args.cpu_offloading,
                gptq_config=gptq_config,
                awq_config=awq_config,
                stream_interval=args.stream_interval,
                conv_template=args.conv_template,
                embed_in_truncate=args.embed_in_truncate,
                load_kwargs=args.load_kwargs #改*************************
            )
            sys.modules["fastchat.serve.model_worker"].args = args
            sys.modules["fastchat.serve.model_worker"].gptq_config = gptq_config

            sys.modules["fastchat.serve.model_worker"].worker = worker

    MakeFastAPIOffline(app)
    app.title = f"FastChat LLM Server ({args.model_names[0]})"
    app._worker = worker
    return app

至此,我们完成了langchain-chatchat加载p-tuning的全部操作,将ptuing的路径添加到model_config的llm_dict,如 ``` chatglm2-6b: 'p-tuning-peft'

即可以如下方式加载p-tuning:

```shell
PEFT_SHARE_BASE_WEIGHTS=true python startup.py -a

预处理知识库文件⚓︎

在载入知识库文件的时候,直接上传文档虽然能实现基础的问答,但是,其效果并不能发挥到最佳水平。因此,我们建议开发者对知识库文件做出以下的预处理。 以下方式的预处理如果执行了,有概率提升模型的召回率。

1. 使用TXT / Markdown 等格式化文件,并按照要点排版⚓︎

例如,以下段落应该被处理成如下内容后在嵌入知识库,会有更好的效果。

原文: PDF类型
查特查特团队荣获AGI Playground Hackathon黑客松“生产力工具的新想象”赛道季军
2023年10月16日, Founder Park在近日结束的AGI Playground Hackathon黑客松比赛中,查特查特团队展现出色的实力,荣获了“生产力工具的新想象”赛道季军。本次比赛由Founder Park主办,并由智谱、Dify、Zilliz、声网、AWS云服务等企业协办。
比赛吸引了120多支参赛团队,最终有36支队伍进入决赛,其中34支队伍成功完成了路演。比赛规定,所有参赛选手必须在短短的48小时内完成一个应用产品开发,同时要求使用智谱大模型及Zilliz向量数据库进行开发。
查特查特团队的现场参赛人员由两名项目成员组成:
来自A大学的小明负责了Agent旅游助手的开发、场地协调以及团队住宿和行程的安排;在保证团队完赛上做出了主要贡献。作为队长,栋宇坚持自信,创新,沉着的精神,不断提出改进方案并抓紧落实,遇到相关问题积极请教老师,提高了团队开发效率。
作为核心开发者的B公司小蓝,他则主管Agent智能知识库查询开发、Agent底层框架设计、相关API调整和UI调整。在最后,他代表团队在规定的时间内呈现了产品的特点和优势,并完美的展示了产品demo。为团队最终产品能够得到奖项做出了重要贡献。
修改后的Markdown文件,具有更高的召回率
# 查特查特团队荣获AGI Playground Hackathon黑客松“生产力工具的新想象”赛道季军。

## 报道简介
2023年10月16日, Founder Park在近日结束的AGI Playground Hackathon黑客松比赛中,查特查特团队展现出色的实力,荣获了“生产力工具的新想象”赛道季军。本次比赛由Founder Park主办,并由智谱、Dify、Zilliz、声网、AWS云服务等企业协办。

## 比赛介绍

比赛吸引了120多支参赛团队,最终有36支队伍进入决赛,其中34支队伍成功完成了路演。比赛规定,所有参赛选手必须在短短的48小时内完成一个应用产品开发,同时要求使用智谱大模型及Zilliz向量数据库进行开发。

## 获奖队员简介

+ 小明,A大学
  + 负责Agent旅游助手的开发、场地协调以及团队住宿和行程的安排
  + 在保证团队完赛上做出了主要贡献。作为队长,栋宇坚持自信,创新,沉着的精神,不断提出改进方案并抓紧落实,遇到相关问题积极请教老师,提高了团队开发效率。

+ 小蓝,B公司
  + 主管Agent智能知识库查询开发、Agent底层框架设计、相关API调整和UI调整。
  + 代表团队在规定的时间内呈现了产品的特点和优势,并完美的展示了产品demo。

2. 减少文件中冲突的内容,分门别类存放数据⚓︎

就像人类寻找相关点一样,如果在多份文件中存在相似的内容,可能会导致模型无法准确的搜索到相关内容。 因此,需要减少文件中相似的内容,或将其分在不同的知识库中。 例如,以下两个句子中,如果搜索外籍教师,则具有歧义,非常容易搜索到错误答案。

文件一:
在大数据专业中,我们已经拥有超过1/3的外籍博士和教师。

文件二:

本专业具有40%的外籍教师比例,
本专业有博士生10人,研究生12人。

3. 减少具有歧义的句子⚓︎

知识库中应该减少具有歧义的句子和段落,或者汉语的高级用法,例如

1. 他说他会杀了那个人。
2. 你说啥子?
3. 我喜欢你的头发。
4. 地板真的滑,我差点没摔倒。
在相似度模型对比的时候,仅仅能搜索句子的表面意思,因此,使用有歧义的句子和段落可能导致搜索错误。

4. 减少单个文件的大小,减少文件中的特殊符号⚓︎

  • 上传知识库的单个文件不建议超过5MB,以免出现向量化中断卡死等情况。同时,上传大文件不要使用faiss数据库。
  • 减少上传文件中的中文符号,特殊符号,无意义空格等。

自定义的关键词调整Embedding模型⚓︎

1.首先准备一个关键字的文本文件,每一行是一个关键字。例如:

文件key_words.txt:
iphone13pro
中石油
2. 配置kb_config.py
EMBEDDING_KEYWORD_FILE = "embedding_keywords.txt"
3. 运行embeddings/add_embedding_keywords.py
输入的文本(这里只是一个没分隔的一串字符):iphone13pro
生成的token id序列:[101, 21128, 102]
token到token id的映射:
[CLS]->101
iphone13pro->21128
[SEP]->102

输入的文本:中石油
生成的token id序列:[101, 21129, 102]
token到token id的映射:
[CLS]->101
中石油->21129
[SEP]->102
这样,你就获得了一个新的带有关键词调整的Embedding模型

实际使用效果⚓︎

在这里,我们放置了一些成功调用的效果图,方便开发者进行查看自己是否成功运行了框架。

检查是否成功上传/管理自己的知识库⚓︎

在WebUI界面上传知识库,则必须保证知识库进行向量化,成功之后,文件会被切分并在向量位置打钩。 下图展示了成功上传知识库的画面

成功上传知识库

请确保所有知识库都已经进行了向量化。

检查是否成功开启LLM对话⚓︎

若打开webui后,在该模式下能成功跟大模型对话即成功调用。

下图为成功调用LLM的效果图:

LLM对话

检查是否成功调用知识库/搜索⚓︎

若成功调用知识库,则你应该能看到,在大模型回答的下方有一个知识库匹配结果的展开框,并且内部显示了相关的匹配结果。 如果没有搜索到相关内容,则会提示根据已知信息无法回答问题,并且下拉框中没有任何内容。

下图为成功调用知识库效果图:

成功调用知识库

在这个案例中,第一次用户的提问无法在知识库中寻找到合适的答案,因此,大模型回答了根据已知信息无法回答问题

第二次用户的提问能在知识库中寻找到合适的答案,因此,大模型给出了一个正确的回答。

注意: 知识库的搜索情况取决于嵌入模型的准度,分词器的设置,知识库的排版和大模型的数量,提示词设定等多个因素。因此,需要开发者进行深度的优化和调试。

检查是否成功调用Agent工具⚓︎

若成功调用Agent工具,则你应该看到大模型完整的思维过程,这会在思考过程下拉框中显示出来。如果成功调用Agent工具,则你应该看到Markdown引用效果的工具使用情况。 在Agent对话模式中,思考过程中显示的是大模型的思考过程,而下拉框之前的内容为大模型的Final Answer,缺乏中间的运算过程。

下图展现了一个成功调用Agent工具的效果图:

成功调用单个Agent工具

本框架支持模型连续掉用多个Agent工具,下图展示了一个一个提问中大模型连续调用多个Agent工具的效果图:

连续调用多个Agent工具

在这个案例中,3900是大模型的最终答案,其余都是思考过程。


最后更新: December 6, 2023
创建日期: December 6, 2023