赞
踩
LLM基座官方文档如下(科学上网):
Github: https://github.com/ymcui/Chinese-LLaMA-Alpaca-2
Huggingface: https://huggingface.co/meta-llama
Github: https://github.com/THUDM/ChatGLM3
Huggingface: https://huggingface.co/THUDM
Github: https://github.com/baichuan-inc
Huggingface: https://huggingface.co/baichuan-inc
Github: https://github.com/QwenLM
Huggingface: baichuan-inc (Baichuan Intelligent Technology) (huggingface.co)
提示:以下是本篇文章正文内容,下面案例可供参考
面LLM岗大概率会cue的内容,详见文章大模型升级与设计之道:ChatGLM、LLAMA、Baichuan及LLM结构解析 - 知乎 (zhihu.com)
该文章从原理、性能、差异、迭代版本系统地介绍了现在较受欢迎的LLM(目前ChatGLM4、Baichuan3已闭源):
1.1. 查看服务器GPU显存及占用
- # 每0.5s刷新一次
- !wathch -d -n 0.5 nvidia-smi
1.2. 模型部署所需显存查询
- # 1.2.1 安装依赖包
- !pip install accelerate transformers
- # 1.2.2 查看RicardoLee/Llama2-chat-13B-Chinese-50W显存(网络层单层最大显存、推理显存、训练显存)
- !accelerate estimate-memory RicardoLee/Llama2-chat-13B-Chinese-50W
- # 1.2.3 也可以点击在线测试链接
- https://huggingface.co/spaces/hf-accelerate/model-memory-usage
2.1. 本地部署
- # 2.1.1 执行git lfs install
- curl -s https://packagecloud.io/install/repositories/github/git-lfs/script.deb.sh | sudo bash
- sudo apt-get install git-lfs
-
- # 2.1.2 克隆模型到服务器(Llama2-chat-13B-Chinese-50W)
- git clone https://huggingface.co/RicardoLee/Llama2-chat-13B-Chinese-50W
- ### 如果遇到模型大文件无法下载,通过wget从huggingface上下载
- wget https://huggingface.co/RicardoLee/Llama2-chat-13B-Chinese-50W/resolve/main/pytorch_model-00001-of-00003.bin
- wget https://huggingface.co/RicardoLee/Llama2-chat-13B-Chinese-50W/resolve/main/pytorch_model-00002-of-00003.bin
- wget https://huggingface.co/RicardoLee/Llama2-chat-13B-Chinese-50W/resolve/main/pytorch_model-00003-of-00003.bin
2.2 网页可视化(下载并部署gradio)
从这个链接里:https://github.com/ymcui/Chinese-LLaMA-Alpaca/blob/main/scripts/inference/gradio_demo.py
里的gradio_demo.py和requirements.txt下载到服务器
- # 2.2.1 安装依赖
- !pip install -r requirements.txt
-
- # 2.2.2 把gradio.py里59、60、61行注释掉
- !pip install gradio
-
- # 2.2.3 安装其他依赖包
- !pip install bitsandbytes accelerate scipy
-
- # 2.2.4 cd Llama2路径
- !python gradio_demo.py --base_model /root/autodl-tmp/Llama2-chat-13B-Chinese-50W --tokenizer_path /root/autodl-tmp/Llama2-chat-13B-Chinese-50W --gpus 0
Llama2部署可视化
3.1. 数据预处理
- # 3.1.1 下载BelleGroup提供的50w条中文数据(注意数据量有点大)
- wget https://huggingface.co/datasets/BelleGroup/train_0.5M_CN/resolve/main/Belle_open_source_0.5M.json
-
-
- # 3.1.2 新建split_json.py文件,粘贴如下代码
- import random,json
-
- def write_txt(file_path,datas):
- with open(file_path,"w",encoding="utf8") as f:
- for d in datas:
- f.write(json.dumps(d,ensure_ascii=False)+"\n")
- f.close()
-
- with open("/root/autodl-tmp/Belle_open_source_0.5M.json","r",encoding="utf8") as f:
- lines=f.readlines()
- #拼接数据
- changed_data=[]
- for l in lines:
- l=json.loads(l)
- changed_data.append({"text":"### Human: "+l["instruction"]+" ### Assistant: "+l["output"]})
-
- #从拼好后的数据中,随机选出1000条,作为训练数据
- #为了省钱 和 演示使用,我们只用1000条,生产环境至少要使用全部50w条
- r_changed_data=random.sample(changed_data, 1000)
-
- #写到json中,root根据需求自行修改
- write_txt("/root/autodl-tmp/Belle_open_source_0.5M_changed_test.json",r_changed_data)
-
-
- # 3.1.3 新建终端运行split_json.py,切分数据集为json格式
- !python split_json.py
3.2. 运行微调文件
- # 3.2.1 安装依赖
- !pip install -q huggingface_hub
- !pip install -q -U trl transformers accelerate peft
- !pip install -q -U datasets bitsandbytes einops wandb
-
-
- # 3.2.2 运行微调文件
- # (1)导入相关包
- from datasets import load_dataset
- import torch,einops
- from transformers import AutoModelForCausalLM, BitsAndBytesConfig, AutoTokenizer, TrainingArguments
- from peft import LoraConfig
- from trl import SFTTrainer
-
- # (2)加载python split_json.py拼接好之后的1000条数据
- dataset = load_dataset("json",data_files="/root/autodl-tmp/Belle_open_source_0.5M_changed_test.json",split="train")
-
- # (3)模型配置
- base_model_name ="/root/autodl-tmp/Llama2-chat-13B-Chinese-50W" # 路径需要根据模型部署路径修改
- bnb_config = BitsAndBytesConfig(
- load_in_4bit=True, #在4bit上,进行量化
- bnb_4bit_use_double_quant=True, # 嵌套量化,每个参数可以多节省0.4位
- bnb_4bit_quant_type="nf4", #NF4(normalized float)或纯FP4量化 博客说推荐NF4
- bnb_4bit_compute_dtype=torch.float16)
-
- # (4)QloRA微调参数配置
- peft_config = LoraConfig(
- lora_alpha=16,
- lora_dropout=0.1,
- r=64,
- bias="none",
- task_type="CAUSAL_LM",
- )
-
- # (5)加载部署好的本地模型(Llama)
- base_model = AutoModelForCausalLM.from_pretrained(
- base_model_name,#本地模型名称
- quantization_config=bnb_config,#上面本地模型的配置
- device_map=device_map,#使用GPU的编号
- trust_remote_code=True,
- use_auth_token=True
- )
- base_model.config.use_cache = False
- base_model.config.pretraining_tp = 1
-
- # (6)长文本拆分成最小的单元词(即token)
- tokenizer = AutoTokenizer.from_pretrained(base_model_name, trust_remote_code=True)
- tokenizer.pad_token = tokenizer.eos_token
-
-
- # (7)训练参数配置
- output_dir = "./results"
- training_args = TrainingArguments(
- report_to="wandb",
- output_dir=output_dir, #训练后输出目录
- per_device_train_batch_size=4, #每个GPU的批处理数据量
- gradient_accumulation_steps=4, #在执行反向传播/更新过程之前,要累积其梯度的更新步骤数
- learning_rate=2e-4, #超参、初始学习率。太大模型不稳定,太小则模型不能收敛
- logging_steps=10, #两个日志记录之间的更新步骤数
- max_steps=100 #要执行的训练步骤总数
- )
- max_seq_length = 512
- #TrainingArguments 的参数详解:https://blog.csdn.net/qq_33293040/article/details/117376382
-
- trainer = SFTTrainer(
- model=base_model,
- train_dataset=dataset,
- peft_config=peft_config,
- dataset_text_field="text",
- max_seq_length=max_seq_length,
- tokenizer=tokenizer,
- args=training_args,
- )
-
- # (8)运行程序,进行微调
- trainer.train()
-
- # (9)保存模型
- import os
- output_dir = os.path.join(output_dir, "final_checkpoint")
- trainer.model.save_pretrained(output_dir)
3.3. 执行代码合并
新建merge_model.py的文件,把下面的代码粘贴进去, 然后然后执行上述合并代码,进行合并。终端运行python merge_model.py。
- from peft import PeftModel
- from transformers import AutoModelForCausalLM, AutoTokenizer
- import torch
-
- #设置原来本地模型的地址
- model_name_or_path = '/root/autodl-tmp/Llama2-chat-13B-Chinese-50W'
- #设置微调后模型的地址,就是上面的那个地址
- adapter_name_or_path = '/root/autodl-tmp/results/final_checkpoint'
- #设置合并后模型的导出地址
- save_path = '/root/autodl-tmp/new_model'
-
- tokenizer = AutoTokenizer.from_pretrained(
- model_name_or_path,
- trust_remote_code=True
- )
- model = AutoModelForCausalLM.from_pretrained(
- model_name_or_path,
- trust_remote_code=True,
- low_cpu_mem_usage=True,
- torch_dtype=torch.float16,
- device_map='auto'
- )
- print("load model success")
- model = PeftModel.from_pretrained(model, adapter_name_or_path)
- print("load adapter success")
- model = model.merge_and_unload()
- print("merge success")
-
- tokenizer.save_pretrained(save_path)
- model.save_pretrained(save_path)
- print("save done.")
4.1. 安装依赖库
!pip install -U langchain unstructured nltk sentence_transformers faiss-gpu
4.2. 外挂知识库 & 向量存储 & 问题/向量检索
- # 4.2.0 导包
- from langchain.document_loaders import UnstructuredFileLoader
- from langchain.text_splitter import RecursiveCharacterTextSplitter
- from langchain.embeddings.huggingface import HuggingFaceEmbeddings
- from langchain.vectorstores import FAISS
-
-
-
- # 4.2.1 加载外部知识库
- filepath="/root/autodl-tmp/knowledge.txt"
- loader=UnstructuredFileLoader(filepath) # 把带格式的文本,读取为无格式的纯文本
- docs=loader.load()
- print(docs) # 返回的是一个列表,列表中的元素是Document类型
-
- # 4.2.2 对读取的文档进行chunk
- text_splitter=RecursiveCharacterTextSplitter(chunk_size=20,chunk_overlap=10)
- docs=text_splitter.split_documents(docs)
-
- # 4.2.3 下载并部署embedding模型
- 执行:git lfs install
- 执行:git clone https://huggingface.co/GanymedeNil/text2vec-large-chinese
- 如果有大文件下载不下来,执行
- wget https://huggingface.co/GanymedeNil/text2vec-large-chinese/resolve/main/pytorch_model.bin
- wget https://huggingface.co/GanymedeNil/text2vec-large-chinese/resolve/main/model.safetensors
-
-
- # 4.2.4 使用text2vec-large-chinese模型,对上面chunk后的doc进行embedding。然后使用FAISS存储到向量数据库
- import os
- embeddings=HuggingFaceEmbeddings(model_name="/root/autodl-tmp/text2vec-large-chinese", model_kwargs={'device': 'cuda'})
-
- if os.path.exists("/root/autodl-tmp/my_faiss_store.faiss")==False:
- vector_store=FAISS.from_documents(docs,embeddings)
-
- else:
- vector_store=FAISS.load_local("/root/autodl-tmp/my_faiss_store.faiss",embeddings=embeddings)
- #注意:如果修改了知识库(knowledge.txt)里的内容,则需要把原来的 my_faiss_store.faiss 删除后,重新生成向量库。
- # 4.2.5 加载模型
- import torch
- from transformers import AutoTokenizer, AutoModelForCausalLM
- #先做tokenizer
- tokenizer = AutoTokenizer.from_pretrained('/root/autodl-tmp/Llama2-chat-13B-Chinese-50W',trust_remote_code=True)
- #加载本地基础模型
- base_model = AutoModelForCausalLM.from_pretrained(
- "/root/autodl-tmp/Llama2-chat-13B-Chinese-50W",
- torch_dtype=torch.float16,
- device_map='auto',
- trust_remote_code=True
- )
- model=base_model.eval()
-
- #4.2.6 向量检索:通过用户问句,到向量库中,匹配相似度高的文本
- query="小白的父亲是谁?"
- docs=vector_store.similarity_search(query)#计算相似度,并把相似度高的chunk放在前面
- context=[doc.page_content for doc in docs]#提取chunk的文本内容
- print(context)
-
- # 4.2.7 构造prompt_template
- my_input="\n".join(context)
- prompt=f"已知:\n{my_input}\n请回答:{query}"
- print(prompt)
-
- # 4.2.8 把prompt输入模型进行预测
- inputs = tokenizer([f"Human:{prompt}\nAssistant:"], return_tensors="pt")
- input_ids = inputs["input_ids"].to('cuda')
- generate_input = {
- "input_ids":input_ids,
- "max_new_tokens":1024,
- "do_sample":True,
- "top_k":50,
- "top_p":0.95,
- "temperature":0.3,
- "repetition_penalty":1.3
- }
- generate_ids = model.generate(**generate_input)
- new_tokens = tokenizer.decode(generate_ids[0], skip_special_tokens=True)
- print("new_tokens",new_tokens)
推理结果:
章节一引用《大模型升级与设计之道:ChatGLM、LLAMA、Baichuan及LLM结构解析》一文,该文章从原理、性能、差异、迭代版本系统地介绍了现在较受欢迎的LLM(目前ChatGLM4、Baichuan3已闭源)。
章节二以Llama2举例,演示了从部署环境查询、到模型部署、再到微调、最后到LangChain外挂知识库实现向量检索增强(RAG)的流程。
掌握本文流程、学习框架,后续大模型业务均可在其基础上进行延伸,其它LLM模型部署demo在前言部分的官方开源文档亦可查询。
Llama2部署教程链接:
Copyright © 2003-2013 www.wpsshop.cn 版权所有,并保留所有权利。