Llama2-Chinese项目：7-外延能力LangChain集成

  本文介绍了Llama2模型集成LangChain框架的具体实现，这样可更方便地基于Llama2开发文档检索、问答机器人和智能体应用等。

1.调用Llama2类
  针对LangChain[1]框架封装的Llama2 LLM类见examples/llama2_for_langchain.py，调用代码如下所示：

from llama2_for_langchain import Llama2
# 这里以调用4bit量化压缩的Llama2-Chinese参数FlagAlpha/Llama2-Chinese-13b-Chat-4bit为例
llm = Llama2(model_name_or_path='FlagAlpha/Llama2-Chinese-13b-Chat-4bit', bit4=True)
while True:human_input = input("Human: ")response = llm(human_input)print(f"Llama2: {response}")

2.Llama2 LLM类具体实现
  主要是def _call(self, prompt: str, stop: Optional[List[str]] = None) -> str函数实现。LangChain八股文也不难实现，如下所示：

from langchain.llms.base import LLM
from typing import Dict, List, Any, Optional
import torch,sys,os
from transformers import AutoTokenizerclass Llama2(LLM): # LLM是一个抽象类，需要实现_call方法max_token: int = 2048     # 最大token数temperature: float = 0.1  # 生成温度top_p: float = 0.95       # 生成概率tokenizer: Any            # 分词器model: Any                # 模型def __init__(self, model_name_or_path, bit4=True):super().__init__()self.tokenizer = AutoTokenizer.from_pretrained(model_name_or_path,use_fast=False)self.tokenizer.pad_token = self.tokenizer.eos_tokenif bit4==False: # 32bitfrom transformers import AutoModelForCausalLMself.model = AutoModelForCausalLM.from_pretrained(model_name_or_path,device_map='auto',torch_dtype=torch.float16,load_in_8bit=True)self.model.eval()else: # 4bitfrom auto_gptq import AutoGPTQForCausalLMself.model = AutoGPTQForCausalLM.from_quantized(model_name_or_path,low_cpu_mem_usage=True, device="cuda:0", use_triton=False,inject_fused_attention=False,inject_fused_mlp=False)if torch.__version__ >= "2" and sys.platform != "win32":self.model = torch.compile(self.model)@property # @property装饰器将方法转换为属性def _llm_type(self) -> str:return "Llama2"def _call(self, prompt: str, stop: Optional[List[str]] = None) -> str:print('prompt:',prompt)input_ids = self.tokenizer(prompt, return_tensors="pt",add_special_tokens=False).input_ids.to('cuda')generate_input = {"input_ids":input_ids,"max_new_tokens":1024,"do_sample":True,"top_k":50,"top_p":self.top_p,"temperature":self.temperature,"repetition_penalty":1.2,"eos_token_id":self.tokenizer.eos_token_id,"bos_token_id":self.tokenizer.bos_token_id,"pad_token_id":self.tokenizer.pad_token_id}generate_ids = self.model.generate(**generate_input)generate_ids = [item[len(input_ids[0]):-1] for  item in generate_ids]result_message = self.tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]return result_message # 返回生成的文本

参考文献：
[1]https://github.com/FlagAlpha/Llama2-Chinese/blob/main/examples/llama2_for_langchain.py
[2]https://github.com/langchain-ai/langchain