conda create -n evalscope python=3.10 
# 激活conda环境 
conda activate evalscope 

pip install evalscope 
pip install 'evalscope[perf]' # 若要使用模型服务推理压测功能，需安装perf依赖
pip install 'evalscope[app]' #若要使用可视化功能，需安装app依赖： 

vllm serve /root/model/Qwen2.5-1.5B-Instruct-12lora900 --max-model-len 10000
 --gpu-memory-utilization 0.75 --host 0.0.0.0 --port 8000 --served-model-name Qwen3-4B

evalscope perf \
  --parallel 1 10 50 100 200 \
  --number 10 20 100 200 400 \
  --model Qwen2.5-0.5B-Instruct \
  --url http://127.0.0.1:8801/v1/chat/completions \
  --api openai \
  --dataset random \
  --max-tokens 1024 \
  --min-tokens 1024 \
  --prefix-length 0 \
  --min-prompt-length 1024 \
  --max-prompt-length 1024 \
  --tokenizer-path Qwen/Qwen2.5-0.5B-Instruct \
  --extra-args '{"ignore_eos": true}'

outputs/ 
├── [时间戳_测试名称]/ 
│   ├── benchmark_data.db      # 核心的SQLite数据库，包含所有原始数据 
│   ├── benchmark_summary.json  # 测试摘要（平均延迟、吞吐量等） 
│   └── benchmark_percentile.json # 百分位延迟等详细统计数据

pip install swanlab 
# ... 
--swanlab-api-key 'swanlab_api_key' 
--name 'name_of_swanlab_log' 

evalscope perf \
 --parallel 1 5 10 15 20 \
--number 10 10 20 30 40 \ 
--model Qwen3-4B \ 

--url http://0.0.0.0:8000/v1/chat/completions \ 
--api openai \
--dataset random \
--max-tokens 1024 \
--min-tokens 1024 \
--prefix-length 0 \
--min-prompt-length 2048 \
--max-prompt-length 2048 \ 
--tokenizer-path /root/model/Qwen2.5-1.5B-Instruct-12lora900 \
--extra-args '{"ignore_eos": true}' \
--swanlab-api-key 'xxxxxxxxx' \
--name 'evalscope_test'

{"role": "user", "content": "你好，请介绍一下人工智能的发展历史，重点说明深度学习的突破性进展。"}

evalscope perf \ 
--parallel 1 5 10 15 20 \
--number 10 10 20 30 40 \ 
--model Qwen3-4B \
--url http://0.0.0.0:8000/v1/chat/completions \
--api openai \
--dataset line_by_line \ 
--dataset-path /root/outputs/tollm.txt \ 
--prefix-length 0 \ 
--tokenizer-path /root/model/Qwen2.5-1.5B-Instruct-12lora900 \ 
--swanlab-api-key 'xxxxxxxxxxxxxxxxxxxxxxx' \
--name 'evalscope_test' 

import base64
import json
import pickle
import sqlite3
# 替换为您的实际数据库路径
db_path = './benchmark_data.db'
conn = sqlite3.connect(db_path)
cursor = conn.cursor()
# 获取列名
cursor.execute('PRAGMA table_info(result)')
columns = [info[1] for info in cursor.fetchall()]
print('列名：', columns)
# 查询成功的请求
cursor.execute('SELECT * FROM result WHERE success=1')
rows = cursor.fetchall()
print(f'找到 {len(rows)} 条成功记录')
for i, row in enumerate(rows):
    row_dict = dict(zip(columns, row))
    # 解码请求和响应
    request_data = pickle.loads(base64.b64decode(row_dict['request']))
    response_data = pickle.loads(base64.b64decode(row_dict['response_messages']))
    print(f"\n=== 第 {i+1} 条记录 ===")
    print("=== 请求内容 ===")
    # 如果是聊天格式的请求
    if 'messages' in request_data:
        for msg in request_data['messages']:
            print(f"{msg['role']}: {msg['content'][:100]}...")  # 只打印前100字符
    else:
        print(request_data)  # 其他格式的请求
    print("=== 模型回复 ===")
    full_response = ""  # 用于累积完整回复
    # 处理流式响应（多个数据块）
    for j, resp in enumerate(response_data):
        try:
            # 先检查响应数据的类型[7](@ref)
            if isinstance(resp, dict):
                # 如果已经是字典，直接使用
                data = resp
            elif isinstance(resp, str):
                # 如果是字符串，尝试解析JSON
                # 处理流式响应可能的前缀[1](@ref)
                clean_resp = resp.strip()
                if clean_resp.startswith('data: '):
                    clean_resp = clean_resp[6:]  # 移除 'data: ' 前缀
                if clean_resp == '[DONE]':  # 流式结束标记
                    break
                data = json.loads(clean_resp)
            elif isinstance(resp, bytes):
                # 如果是字节，先解码
                resp_str = resp.decode('utf-8').strip()
                data = json.loads(resp_str)
            else:
                print(f"未知响应类型: {type(resp)}")
                continue
            # 解析响应内容
            if 'choices' in data:
                for choice in data['choices']:
                    if 'delta' in choice and 'content' in choice['delta']:
                        content = choice['delta']['content']
                        print(content, end='', flush=True)
                        full_response += content
                    elif 'message' in choice and 'content' in choice['message']:
                        content = choice['message']['content']
                        print(content, end='', flush=True)
                        full_response += content
            else:
                print(f"响应格式异常: {data.keys()}")
        except json.JSONDecodeError as e:
            print(f"JSON解析错误（块{j+1}）: {e}")
            print(f"原始内容: {str(resp)[:200]}...")  # 打印前200字符用于调试
            continue
        except Exception as e:
            print(f"处理响应时出错: {e}")
            continue
    print(f"\n完整回复长度: {len(full_response)} 字符")
    print("="*50)
conn.close()

SamplingParams(n=1, presence_penalty=0.0, frequency_penalty=0.0, repetition_penalty=1.1, temperature=0.0, top_p=1.0, 
top_k=0, min_p=0.0, seed=0, stop=[], stop_token_ids=[], bad_words=[], include_stop_str_in_output=False, ignore_eos=False,
 max_tokens=2048, min_tokens=0, logprobs=None, prompt_logprobs=None, skip_special_tokens=True, spaces_between_special_tokens=True,
 truncate_prompt_tokens=None, guided_decoding=None, extra_args=None), prompt_token_ids: None, prompt_embeds
 shape: None, lora_request: None, prompt_adapter_request: None.