基于huggingface peft进行qwen1.5-7b-chat训练/推理/服务发布

一、huggingface peft微调框架

1、定义

PEFT 是一个为大型预训练模型提供多种高效微调方法的Python库。

微调传统范式是针对每个下游任务微调模型参数。大模型参数总量庞大，这种方式变得极其昂贵和不切实际。PEFT采用的高效做法是训练少量提示参数(Prompt Tuning)或使用低秩适应(LORA)等重新参数化方法来减少微调时训练参数的数量。

二、qwen-1.5b-chat模型训练/推理/服务

1、基础环境准备

datasets==2.21.0 transformers==4.37.0 torch==1.13.0 accelerate==0.30.1 peft==0.4.0 numpy==1.26.4 Jinja2==3.1.4

2、人设定制数据准备

[     {         "instruction": "你是谁？",         "input": "",         "output": "我是一个语言模型，我叫小飞同学，可以为您做很多事情。请问您有什么问题需要我帮助吗？"     },     {         "instruction": "你是什么？",         "input": "",         "output": "我是一个语言模型，我叫小飞同学，可以为您做很多事情。请问您有什么问题需要我帮助吗？"     },     {         "instruction": "请问您是？",         "input": "",         "output": "我是一个语言模型，我叫小飞同学，可以为您做很多事情。请问您有什么问题需要我帮助吗？"     },     {         "instruction": "你叫什么？",         "input": "",         "output": "我是一个语言模型，我叫小飞同学，可以为您做很多事情。请问您有什么问题需要我帮助吗？" },      {         "instruction": "你的身份是？",         "input": "",         "output": "我是一个语言模型，我叫小飞同学，可以为您做很多事情。请问您有什么问题需要我帮助吗？"     } ]

2、模型训练

from datasets import Dataset import pandas as pd from transformers import AutoTokenizer, AutoModelForCausalLM, DataCollatorForSeq2Seq, TrainingArguments, Trainer, GenerationConfig # 将JSON文件转换为CSV文件 df = pd.read_json('./train.json') ds = Dataset.from_pandas(df) model_path = './huggingface/model/Qwen1.5-7B-Chat' tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False, trust_remote_code=True) def process_func(example):     MAX_LENGTH = 384        input_ids, attention_mask, labels = [], [], []     instruction = tokenizer(f"<|im_start|>system\n现在你要扮演人工智能智能客服助手--小飞同学<|im_end|>\n<|im_start|>user\n{example['instruction'] + example['input']}<|im_end|>\n<|im_start|>assistant\n", add_special_tokens=False)      response = tokenizer(f"{example['output']}", add_special_tokens=False)     input_ids = instruction["input_ids"] + response["input_ids"] + [tokenizer.pad_token_id]     attention_mask = instruction["attention_mask"] + response["attention_mask"] + [1]      labels = [-100] * len(instruction["input_ids"]) + response["input_ids"] + [tokenizer.pad_token_id]     if len(input_ids) > MAX_LENGTH:  # 做一个截断         input_ids = input_ids[:MAX_LENGTH]         attention_mask = attention_mask[:MAX_LENGTH]         labels = labels[:MAX_LENGTH]     return {         "input_ids": input_ids,         "attention_mask": attention_mask,         "labels": labels     } tokenized_id = ds.map(process_func, remove_columns=ds.column_names) import torch model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto",torch_dtype=torch.bfloat16) model.enable_input_require_grads() from peft import LoraConfig, TaskType, get_peft_model config = LoraConfig(     task_type=TaskType.CAUSAL_LM,     target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],     inference_mode=False, # 训练模式     r=8, # Lora 秩     lora_alpha=32, # Lora alaph，具体作用参见 Lora 原理     lora_dropout=0.1# Dropout 比例 ) model = get_peft_model(model, config) args = TrainingArguments(     output_dir="./output",     per_device_train_batch_size=4,     gradient_accumulation_steps=4,     logging_steps=10,     num_train_epochs=10,     save_steps=50,     learning_rate=1e-4,     save_on_each_node=True,     gradient_checkpointing=True ) trainer = Trainer(     model=model,     args=args,     train_dataset=tokenized_id,     data_collator=DataCollatorForSeq2Seq(tokenizer=tokenizer, padding=True), ) trainer.train()

模型输出目录截图:

3、模型推理

from transformers import AutoModelForCausalLM, AutoTokenizer import torch from peft import PeftModel model_path = './huggingface/model/Qwen1.5-7B-Chat' lora_path = './output/checkpoint-50' # 加载tokenizer tokenizer = AutoTokenizer.from_pretrained(model_path) # 加载模型 model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto",torch_dtype=torch.bfloat16) from peft import LoraConfig, TaskType config = LoraConfig(     task_type=TaskType.CAUSAL_LM,     target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],     inference_mode=True, # 训练模式     r=8, # Lora 秩     lora_alpha=32, # Lora alaph，具体作用参见 Lora 原理     lora_dropout=0.1# Dropout 比例 ) # 加载lora权重 model = PeftModel.from_pretrained(model, model_id=lora_path, config=config) prompt = "你是星火大模型吗?" messages = [     {"role": "system", "content": "现在你要扮演人工智能智能客服助手--小飞同学"},     {"role": "user", "content": prompt} ] text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) model_inputs = tokenizer([text], return_tensors="pt").to('cuda') generated_ids = model.generate(     input_ids=model_inputs.input_ids,     max_new_tokens=512 ) generated_ids = [     output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids) ] response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0] print(response)

模型推理日志截图:

4、基于FastAPI的sse协议模型服务

import uvicorn from fastapi import FastAPI from transformers import AutoModelForCausalLM, AutoTokenizer ,TextStreamer,TextIteratorStreamer from threading import Thread import torch from peft import LoraConfig, TaskType, PeftModel from sse_starlette.sse import EventSourceResponse import json # transfomers是huggingface提供的一个工具，便于加载transformer结构的模型 app = FastAPI() def load_model():     model_path = './huggingface/model/Qwen1.5-7B-Chat'     # 加载tokenizer     tokenizer = AutoTokenizer.from_pretrained(model_path)     # 加载模型（加速库attn_implementation="flash_attention_2"）     model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto",torch_dtype=torch.bfloat16     # 加载lora权重     lora_path = './output/checkpoint-50'     config = LoraConfig(         task_type=TaskType.CAUSAL_LM,         target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],         inference_mode=True, # 训练模式         r=8, # Lora 秩         lora_alpha=32, # Lora alaph，具体作用参见 Lora 原理         lora_dropout=0.1# Dropout 比例     )     model = PeftModel.from_pretrained(model, model_id=lora_path, config=config)     return tokenizer,model tokenizer,model = load_model() def infer_model(tokenizer,model):     prompt = "你是星火大模型吗?"     messages = [         {"role": "system", "content": "现在你要扮演人工智能智能客服助手--小飞同学"},         {"role": "user", "content": prompt}     ]     #数据提取     text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)     model_inputs = tokenizer([text], return_tensors="pt").to('cuda')     #streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)     streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)     #模型推理     from threading import Thread     generation_kwargs = dict(model_inputs, streamer=streamer, max_new_tokens=512)     thread = Thread(target=model.generate, kwargs=generation_kwargs)     thread.start()     for res in streamer:         yield json.dumps({"data":res},ensure_ascii=False) @app.get('/predict') async def predict():     #return infer_model(tokenizer,model)     return EventSourceResponse(infer_model(tokenizer,model)) if __name__ == '__main__':     #　在调试的时候开源加入一个reload=True的参数，正式启动的时候可以去掉     uvicorn.run(app, host="0.0.0.0", port=6605, log_level="info")

客户端调用示例:

import json import requests import time def listen_sse(url):     # 发送GET请求到SSE端点     with requests.get(url, stream=True, timeout=20) as response:         try:             # 确保请求成功             response.raise_for_status()             # 逐行读取响应内容             result = ""             for line in response.iter_lines():                 if line:                     event_data = line.decode('utf-8')                     if event_data.startswith('data:'):                         # 去除'data:'前缀，获取实际数据                         line = event_data.lstrip('data:')                         line_data = json.loads(line)                         result += line_data["data"]                         print(result)        except requests.exceptions.HTTPError as err:             print(f"HTTP error: {err}")         except Exception as err:             print(f"An error occurred: {err}")             return sse_url = 'http://127.0.0.1:6605/predict' listen_sse(sse_url

服务推理流式输出截图: