本地部署大模型QPS推理测试

1、测试环境

1.1、显卡

1.2、模型

Qwen2.5-32B-Instruct

1.3、部署环境

xinference

1.3.1、docker

docker run  \
-v ~/.xinference:/root/.xinference \
-v ~/.cache/huggingface:/root/.cache/huggingface \
-v ~/.cache/modelscope:/root/.cache/modelscope \
-e XINFERENCE_MODEL_SRC=modelscope \
-p 9998:9997 \
--gpus all \
--shm-size 20g \
--name xinference \
registry.cn-hangzhou.aliyuncs.com/xprobe_xinference/xinference:latest \
 xinference-local \
-H 0.0.0.0 \
--log-level debug

1.3.2、执行命令

xinference launch --model_path ~/.cache/modelscope/hub/qwen/Qwen2___5-32B-Instruct --model-engine vLLM --model-name qwen2.5-instruct --size-in-billions 32 --model-format pytorch --quantization none --n-gpu 4  --gpu-idx 0,1,2,3

2、测试问题

以下是三个需要模型“深度思考”的问题（用ChatGTP得到）：

2.1、20字左右问题

宇宙中是否存在多重宇宙，如果存在，它们之间如何相互影响？

2.2、50字左右问题

人工智能是否可能真正理解“情感”的含义？如果可以，情感的理解会如何影响人类与 AI 的关系？

2.3、100字左右问题

在人类社会的未来发展中，科技不断进步是否会导致人类完全依赖技术？如果技术突然崩溃，人类是否具备足够的韧性重新建立自给自足的社会？如何在高速发展的同时保留这种生存能力？

3、测试代码

3.1、通用测试代码

import requests
import time
from concurrent.futures import ThreadPoolExecutor
import uuid
from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:9998/v1", api_key="not used actually")

def openai_request():
    try:
        response = client.chat.completions.create(
                model="qwen2.5-instruct",
                messages=[{"role": "user", "content": "在人类社会的未来发展中，科技不断进步是否会导致人类完全依赖技术？如果技术突然崩溃，人类是否具备足够的韧性重新建立自给自足的社会？如何在高速发展的同时保留这种生存能力？"}],
                stream=True,
            )

        for chunk in response:
            content = chunk.choices[0].delta.content
            if content:
                # print(content, end='', flush=True)
                pass
        return True
    except requests.RequestException as e:
        print(f"请求失败: {e}")
        return False


# 计算QPS
def calculate_qps(num_requests, num_threads):
    with ThreadPoolExecutor(max_workers=num_threads) as executor:
        start_time = time.time()
        futures = [executor.submit(openai_request) for _ in range(num_requests)]
        successful_requests = sum([future.result() for future in futures])
        end_time = time.time()

    duration = end_time - start_time  # 测试时长
    qps = successful_requests / duration  # 计算 QPS
    return qps, successful_requests, duration

req_test = [(10,10),(10,20),(10,50),(20,20),(20,50),(50,50),(50,100)]

for num_threads,num_requests in req_test:
    qps, successful_requests, duration = calculate_qps(num_requests, num_threads)
    print(f"并发请求数: {num_threads}")
    print(f"请求个数: {num_requests}")
    print(f"总测试时长: {duration:.2f}秒")
    print(f"QPS: {qps:.2f} 请求/秒")
    print(f"成功请求数量: {successful_requests}")
    print("*"*33)

3.2、通用测试代码（仅供参考）

import requests
import time
from concurrent.futures import ThreadPoolExecutor
import uuid
from openai import OpenAI
# 测试的 URL 和 payload
url = "http://127.0.0.1:7860/v1/chat/completions"
payload = {
    "conversation_id": "0",
    "messages": [
        {
            "role": "user",
            "content": "空调系统由哪几个系统组成"  # 你可以根据需求更换成其他问题
        },
    ],
    "stream": True,
    "temperature": 0
}

# 单次请求函数
def send_request():
    try:
        payload["conversation_id"] = str(uuid.uuid4())
        response = requests.post(url, json=payload, stream=True)
        for chunk in response.iter_content(chunk_size=512):  # 分块接收响应
            if chunk:
                chunk = chunk.decode('utf-8', errors='ignore').strip()
                # 这里可以对流式响应进行处理
                # print(chunk)  # 可选，输出流式返回的每个部分
        return True  # 如果成功接收到响应，返回 True
    except requests.RequestException as e:
        print(f"请求失败: {e}")
        return False

# 计算QPS
def calculate_qps(num_requests, num_threads):
    with ThreadPoolExecutor(max_workers=num_threads) as executor:
        start_time = time.time()
        futures = [executor.submit(send_request) for _ in range(num_requests)]
        successful_requests = sum([future.result() for future in futures])
        end_time = time.time()

    duration = end_time - start_time  # 测试时长
    qps = successful_requests / duration  # 计算 QPS
    return qps, successful_requests, duration

req_test = [(10,10),(10,20),(10,50),(20,20),(20,50),(50,50),(50,100)]

for num_threads,num_requests in req_test:
    qps, successful_requests, duration = calculate_qps(num_requests, num_threads)
    print(f"并发请求数: {num_threads}")
    print(f"请求个数: {num_requests}")
    print(f"总测试时长: {duration:.2f}秒")
    print(f"QPS: {qps:.2f} 请求/秒")
    print(f"成功请求数量: {successful_requests}")
    print("*"*33)

4、测试结果

4.1、通用测试结果

并发请求数	请求个数	时间s-20	QPS-20	时间s-50	QPS-50	时间s-100	QPS-100
10	10	25.54	0.39	28.95	0.35	30.56	0.33
10	20	47.85	0.42	57.46	0.35	64.24	0.31
10	50	122.27	0.41	135.08	0.37	151.01	0.33
20	20	34.52	0.58	35.91	0.56	44.84	0.45
20	50	83.04	0.6	93.56	0.53	106.35	0.47
50	50	49.91	1.00	54.11	0.92	66.72	0.75
50	100	101.47	0.99	110.77	0.90	123.49	0.81

4.2、RAG测试结果

并发请求数	请求个数	时间s	QPS	失败
10	1000	2968.06	0.33	0
10	100	299.89	0.33	0
10	50	178.77	0.28	0
10	20	61.24	0.33	0
10	10	32.87	0.3	0
20	20	54.89	0.36	0
20	40	108.23	0.36	1
20	50	136.97	0.35	2
50	50	120.85	0.15	32
50	100	224.15	0.08	82