通过LLM多轮对话生成单元测试用例

代码
在采用随机生成pytorch算子测试序列且保证算子参数合法这种方法之前,曾通过本文的方法生成算子组合测试用例。目前所测LLM生成的代码均会出现BUG,且多次交互后仍不能解决.也许随着LLM的更新,这个问题会得到解决.记录备用。
代码

python 复制代码
import re
import os
import logging
import random
import numpy as np
import os
import re
import traceback
import subprocess
import tempfile
import copy
import requests
import json

import os
os.environ['MKL_THREADING_LAYER'] = 'GNU'
os.environ['MKL_SERVICE_FORCE_INTEL'] = '1'

os.environ["QIANFAN_AK"] = ""
os.environ["QIANFAN_SK"] = ""
os.environ['DASHSCOPE_API_KEY'] = 'sk-'
os.environ['MOONSHOT_API_KEY']="sk-"
os.environ['SPARKAI_APP_ID'] = ''
os.environ['SPARKAI_API_SECRET'] = ''
os.environ['SPARKAI_API_KEY'] = ''
os.environ['SPARKAI_DOMAIN'] = 'generalv3.5'
os.environ['ZhipuAI_API_KEY'] = ''
os.environ['YI_API_KEY']=""

logger = logging.getLogger('llm_logger')
logger.setLevel(logging.DEBUG)  # 设置日志级别
 
# 创建一个handler，用于写入日志文件
log_file = 'llm_opt.log'
file_handler = logging.FileHandler(log_file)
file_handler.setLevel(logging.DEBUG)
 
# 创建一个handler，用于将日志输出到控制台
console_handler = logging.StreamHandler()
console_handler.setLevel(logging.DEBUG)
 
# 设置日志格式
formatter = logging.Formatter('%(message)s')
file_handler.setFormatter(formatter)
console_handler.setFormatter(formatter)
 
# 将handlers添加到logger
logger.addHandler(file_handler)
logger.addHandler(console_handler)

system_prompt="你是一位pytorch专家,现在需要编写各种测试程序,挖掘算子的潜在BUG"

question =f'''
背景描述:
1.为了测试pytorch不同算子组合时的精度是否正常,需要构建module级别的测试用例
2.尤其需要关注unsqueeze,repeat,permute,transpose,reshape,expand,view等维度变换算子的各种组合
3.以及在这些组合之后添加其它io或计算类的算子如(contiguous,matmul,mul,concat等)

需求:
1.你一次生成一个测试用例(pytorch module及测例),只包含cpu计算
2.之后,我会从的回复中提取出python代码,执行并将结果反馈给你
3.你根据我的反馈,预测性地生成下一个测试用例
4.我们通过多次交互,最大程度地挖掘出潜在的BUG

约束:
1.所有测试用例的代码放在一个```python ```中,方便提取
2.为了防止shape不匹配,建议在forward中计算shape,并根据当前的shape合理地设置下一个算子的参数
3.你每次提供的代码都必须是完整的,不要添加任何注释
4.测试代码只输出成功、失败或抛异常,不需要输出任何多余信息
5.特别需要注意矩阵乘维度是否匹配

如果你明白我的意思,请直接输出第一个测试用例
'''

def extract_and_run_python_code(markdown_text):
    pattern = re.compile(r'```python\n([^```].*?)\n```', re.DOTALL)
    code_blocks = pattern.findall(markdown_text)
    if len(code_blocks)==0:
        return "没有找到Python代码块。"
    results = []
    for code in code_blocks:
        try:
            with tempfile.NamedTemporaryFile(delete=False, suffix=".py") as temp_file:
                temp_file.write(code.encode())
                temp_filename = temp_file.name
            result = subprocess.run(['python3', temp_filename], capture_output=True, text=True)    
            output=f"{result.stderr}{result.stdout}"
            results.append(output)
        except Exception as e:
            error_message = f"error:{traceback.format_exc()}"
            results.append(error_message)        
        finally:
            os.remove(temp_filename)
    return "".join(results)

class LLMInfer(object):
    def __init__(self, system_prompt,question,history_len=5):
        self.system_prompt=system_prompt
        self.question=question    
        self.history_len=history_len   
    def infer(self,user_input=None):
        pass    
    def reset(self):
        pass

class dashscope_llm(LLMInfer):
    def __init__(self, system_prompt, question):
        super().__init__(system_prompt, question)
        import dashscope
        dashscope.api_key=os.environ['DASHSCOPE_API_KEY'] 
        self.history=[]
        self.history.append({'role': 'system', 'content': self.system_prompt})
        self.history.append({'role': 'user', 'content': self.question})		
        
    def reset(self):
        if len(self.history)>self.history_len:
            self.history=self.history[:2] + self.history[-3:]

    def infer(self,user_input=None):
        from dashscope import Generation
        from http import HTTPStatus          
        if user_input:
            self.history.append({'role': 'user', 'content': user_input})
        response = Generation.call(model="qwen-plus", 
                                   messages=self.history,
                                   result_format='message')
        if response.status_code == HTTPStatus.OK:
            role=response.output.choices[0]['message']['role']
            content=response.output.choices[0]['message']['content']
            self.history.append({'role': role,'content': content})
            return content
        else:
            return None

class moonshot_llm(LLMInfer):
    def __init__(self, system_prompt, question):
        super().__init__(system_prompt, question)
        '''
        pip install --upgrade 'openai>=1.0'
        '''
        from openai import OpenAI
        self.client = OpenAI(
            api_key = os.environ['MOONSHOT_API_KEY'],
            base_url = "https://api.moonshot.cn/v1",
        )
        self.history=[]
        self.history.append({'role': 'system', 'content': self.system_prompt})
        self.history.append({'role': 'user', 'content': self.question})		
        
    def reset(self):
        if len(self.history)>self.history_len:
            self.history=self.history[:2] + self.history[-3:]

    def infer(self,user_input=None):      
        if user_input:
            self.history.append({'role': 'user', 'content': user_input})
        completion = self.client.chat.completions.create(
            model="moonshot-v1-128k",
            messages=self.history,
            temperature=0.3,
            top_p=0.1
        )
        role="assistant"
        content=completion.choices[0].message.content
        self.history.append({'role': role,'content': content})
        return content

class qianfan_llm(LLMInfer):
    def __init__(self, system_prompt, question):
        super().__init__(system_prompt, question)
        '''
        pip3 install qianfan
        '''
        self.history=[]
        #self.history.append({'role': 'system', 'content': self.system_prompt})
        self.history.append({'role': 'user', 'content': self.question})		
        
    def reset(self):
        if len(self.history)>self.history_len:
            self.history=self.history[:1] + self.history[-2:]

    def infer(self,user_input=None):    
        import qianfan  
        if user_input:
            self.history.append({'role': 'user', 'content': user_input})
        response = qianfan.ChatCompletion().do(endpoint="completions_pro", messages=self.history,
                                                temperature=0.7, top_p=0.8, penalty_score=1,                                             
                                                disable_search=False, enable_citation=False)
        role="assistant"
        content=response.body["result"]
        self.history.append({'role': role,'content': content})
        return content

class sparkai_llm(LLMInfer):
    def __init__(self, system_prompt, question):
        super().__init__(system_prompt, question)
        '''
        pip3 install --upgrade spark_ai_python
        '''
        from sparkai.llm.llm import ChatSparkLLM
        from sparkai.core.messages import ChatMessage
        self.spark = ChatSparkLLM(
            spark_api_url='wss://spark-api.xf-yun.com/v3.5/chat',
            spark_app_id=os.environ['SPARKAI_APP_ID'],
            spark_api_key=os.environ['SPARKAI_API_KEY'],
            spark_api_secret=os.environ['SPARKAI_API_SECRET'],
            spark_llm_domain=os.environ['SPARKAI_DOMAIN'],
            streaming=False,        
            temperature=0.1
        )
        self.history=[]
        self.history.append(ChatMessage(role="system",content=self.system_prompt))
        self.history.append(ChatMessage(role="user",content=self.question))
        
    def reset(self):
        if len(self.history)>self.history_len:
            self.history=self.history[:2] + self.history[-3:]

    def infer(self,user_input=None):    
        from sparkai.core.messages import ChatMessage
        from sparkai.llm.llm import ChunkPrintHandler
        if user_input:
            self.history.append(ChatMessage(role="user",content=user_input))        
        handler = ChunkPrintHandler()
        response = self.spark.generate([self.history], callbacks=[handler])
        self.history.append(response.generations[0][0].message)
        return response.generations[0][0].text


class zhipuai_llm(LLMInfer):
    def __init__(self, system_prompt, question):
        super().__init__(system_prompt, question)
        '''
        pip install zhipuai
        '''
        from zhipuai import ZhipuAI
        self.client = ZhipuAI(api_key=os.environ['ZhipuAI_API_KEY'])
        self.history=[]
        self.history.append({'role': 'system', 'content': self.system_prompt})
        self.history.append({'role': 'user', 'content': self.question})		
        
    def reset(self):
        if len(self.history)>self.history_len:
            self.history=self.history[:2] + self.history[-3:]

    def infer(self,user_input=None):      
        if user_input:
            self.history.append({'role': 'user', 'content': user_input})
        completion = self.client.chat.completions.create(
            model="glm-4",
            messages=self.history,
            temperature=0.3,
            top_p=0.1
        )
        role="assistant"
        content=completion.choices[0].message.content
        self.history.append({'role': role,'content': content})
        return content

class yi_llm(LLMInfer):
    def __init__(self, system_prompt, question):
        super().__init__(system_prompt, question)
        '''
        pip install --upgrade 'openai>=1.0'
        '''
        from openai import OpenAI
        self.client = OpenAI(
            api_key = os.environ['YI_API_KEY'],
            base_url = "https://api.lingyiwanwu.com/v1",
        )
        self.history=[]
        self.history.append({'role': 'system', 'content': self.system_prompt})
        self.history.append({'role': 'user', 'content': self.question})		
        
    def reset(self):
        if len(self.history)>self.history_len:
            self.history=self.history[:2] + self.history[-3:]

    def infer(self,user_input=None):      
        if user_input:
            self.history.append({'role': 'user', 'content': user_input})
        completion = self.client.chat.completions.create(
            model="yi-large",
            messages=self.history,
            temperature=0.3,
            top_p=0.1
        )
        role="assistant"
        content=completion.choices[0].message.content
        self.history.append({'role': role,'content': content})
        return content

llms=[dashscope_llm,moonshot_llm,qianfan_llm,sparkai_llm,zhipuai_llm,yi_llm]
for llm in llms:
    logger.info(f" ---------------------------------- {llm.__name__} ---------------------------------- ")
    llm=llm(system_prompt,question)
    response = llm.infer()
    for i in range(15):
        llm.reset()
        logger.info(f" ---------------------------------- 第{i}轮 ---------------------------------- ")
        result=None
        logger.info("####### bot #######")
        logger.info(f"{response}")
        if response:
            result=f"{extract_and_run_python_code(response)}"     
            logger.info("####### user #######")
            logger.info(f"{result}")
        response=llm.infer(result)