langchain 使用本地通义千问，7B，14B，glm4

关于 glm4-9b 下载到本地后，加载到langchain 中，也会报错，跟 qwen-7b-chat 一样的，对glm参考解决方案：self-llm/GLM-4/02-GLM-4-9B-chat langchain 接入.md at master · datawhalechina/self-llm · GitHub

from langchain.llms.base import LLM
from typing import Any, List, Optional, Dict
from langchain.callbacks.manager import CallbackManagerForLLMRun
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
 
class ChatGLM4_LLM(LLM):
    # 基于本地 ChatGLM4 自定义 LLM 类
    tokenizer: AutoTokenizer = None
    model: AutoModelForCausalLM = None
    gen_kwargs: dict = None
        
    def __init__(self, mode_name_or_path: str, gen_kwargs: dict = None):
        super().__init__()
        print("正在从本地加载模型...")
        self.tokenizer = AutoTokenizer.from_pretrained(
            mode_name_or_path, trust_remote_code=True
        )
        self.model = AutoModelForCausalLM.from_pretrained(
            mode_name_or_path,
            torch_dtype=torch.bfloat16,
            trust_remote_code=True,
            device_map="auto"
        ).eval()
        print("完成本地模型的加载")
        
        if gen_kwargs is None:
            gen_kwargs = {"max_length": 2500, "do_sample": True, "top_k": 1}
        self.gen_kwargs = gen_kwargs
        
    def _call(self, prompt: str, stop: Optional[List[str]] = None,
              run_manager: Optional[CallbackManagerForLLMRun] = None,
              **kwargs: Any) -> str:
        messages = [{"role": "user", "content": prompt}]
        model_inputs = self.tokenizer.apply_chat_template(
            messages, tokenize=True, return_tensors="pt", return_dict=True, add_generation_prompt=True
        )
        generated_ids = self.model.generate(**model_inputs, **self.gen_kwargs)
        generated_ids = [
            output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs['input_ids'], generated_ids)
        ]
        response = self.tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
        return response
    
    @property
    def _identifying_params(self) -> Dict[str, Any]:
        """返回用于识别LLM的字典,这对于缓存和跟踪目的至关重要。"""
        return {
            "model_name": "glm-4-9b-chat",
            "max_length": self.gen_kwargs.get("max_length"),
            "do_sample": self.gen_kwargs.get("do_sample"),
            "top_k": self.gen_kwargs.get("top_k"),
        }
 
    @property
    def _llm_type(self) -> str:
        return "glm-4-9b-chat"
代码解读

from LLM import ChatGLM4_LLM
gen_kwargs = {"max_length": 2500, "do_sample": True, "top_k": 1}
llm = ChatGLM4_LLM(mode_name_or_path="/root/autodl-tmp/ZhipuAI/glm-4-9b-chat", gen_kwargs=gen_kwargs)
print(llm.invoke("你是谁"))
代码解读

langchian 使用已经下载到本地的模型，我们使用通义千问

显存：24G

模型：qwen1.5-7B-Chat，qwen-7B-Chat

先使用 qwen-7B-Chat，会报错用不了：

看了下是不支持这中模型，但看列表中有一个 Qwen 字样，想着应该是支持的，就去 hugging face 搜了下这个东西 “Qwen2”找到了对应的 qwen1.5-7B-Chat 模型

https://huggingface.co/Qwen/Qwen1.5-7B-Chat

其实也就是一种公测版本，，所以总结来说目前直接导入本地 通义千问 langchaing 支持不是很好，可以使用 ollama，但这个下载非常慢，还会失败

 qwen1.5-7B-Chat 我们用这个模型，是可以加载成功的，并输出的，但是非常非常慢

from transformers import AutoTokenizer, AutoModelForCausalLM
from transformers import pipeline
from langchain import HuggingFacePipeline
from langchain_core.prompts import ChatPromptTemplate
import torch
 
device = torch.device("cuda")
 
model_path = "/root/autodl-tmp/Qwen1___5-7B-Chat"
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
            model_path,
            device_map='auto',
            trust_remote_code=True
        ).to(device).eval()
pipe = pipeline(
    "text-generation",
    model=model.to(device),
    tokenizer=tokenizer,
    # max_length=4096,
    # max_tokens=4096,
    max_new_tokens=512,
    top_p=1,
    repetition_penalty=1.15
)
llama_model = HuggingFacePipeline(pipeline=pipe)
 
prompt = ChatPromptTemplate.from_template("请编写一篇关于{topic}的中文小故事，不超过100字")
chain = prompt | llama_model
res = chain.invoke({"topic": "小白兔"})
print(res)
代码解读

qwen-14b-chat 可以运行 

指定 gpu，必须指定到开头，langchain 前面

import os
os.environ["CUDA_VISIBLE_DEVICES"] = "5,6"
 
from transformers import AutoTokenizer, AutoModelForCausalLM
from transformers import pipeline
from langchain import HuggingFacePipeline
from langchain_core.prompts import ChatPromptTemplate
 
 
tokenizer = AutoTokenizer.from_pretrained("/home/qwen-14b-chat/",
                                              trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("/home/qwen-14b-chat/",
                                                 device_map="auto",
                                                 trust_remote_code=True).eval()
 
pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    # max_length=4096,
    # max_tokens=4096,
    max_new_tokens=512,
    top_p=1,
    repetition_penalty=1.15
)
llama_model = HuggingFacePipeline(pipeline=pipe)
 
prompt = ChatPromptTemplate.from_template("请编写一篇关于{topic}的中文小故事，不超过100字")
chain = prompt | llama_model
res = chain.invoke({"topic": "小白兔"})
print(res)
代码解读