1. 硬件

CPU	GPU	内存
10 vCPU Intel® Xeon® Gold 6248 CPU @ 2.50GHz	RTX 3080 Ti 12GB	45GB

2. 软件

OS	Python	后端推理引擎
Ubuntu 22.04	3.12	vLLM

3. 部署

先安装uv。

curl -LsSf https://astral.sh/uv/install.sh | sh

这里部署的是Qwen3.5-2B模型，经过实测，跑4B模型显存溢出，只能跑2B参数量的模型。

uv venv qwen3.5
source qwen3.5/bin/activate
uv pip install vllm --torch-backend=auto --extra-index-url https://wheels.vllm.ai/nightly
uv pip install modelscope

4. 运行

VLLM_USE_MODELSCOPE=true vllm serve Qwen/Qwen3.5-2B --port 6006 --tensor-parallel-size 1 --max-model-len 128000 --reasoning-parser qwen3 --language-model-only

注意上面--language-model-only表示纯文本模式，这是多模态模型，这里只启动纯文本模式。

官方默认是--max-model-len 262144，这里修改为官方推荐最小值--max-model-len 128000。

5. 访问

接口遵循openai接口规范。

这里以python为客户端请求为例。

# !/usr/bin/env python
# _*_ coding utf-8 _*_
# @Time: 2026/3/23 1:53
# @Author: Luke Ewin
# @Blog: https://blog.lukeewin.top
import time
from openai import OpenAI

client = OpenAI(
    api_key="EMPTY",
    base_url="http://localhost:6006/v1",
    timeout=3600
)

messages = [
    {"role": "user", "content": "你给我讲一下睡前小故事吧"},
]

start = time.time()
response = client.chat.completions.create(
    model="Qwen/Qwen3.5-2B",
    messages=messages,
    max_tokens=32768,
    temperature=1.0,
    top_p=1.0,
    presence_penalty=2.0,
    extra_body={
        "top_k": 20,
        "chat_template_kwargs": {"enable_thinking": False},
    },
)
print(f"Response costs: {time.time() - start:.2f}s")
print(f"Generated text: {response.choices[0].message.content}")

6. 压测

经过实测发现这个Qwen3.5-2B占用显存将近12GB显存。

下面是压测命令。

vllm bench serve --backend openai-chat --endpoint /v1/chat/completions --model Qwen/Qwen3.5-2B --dataset-name random --random-input-len 2048 --random-output-len 512 --num-prompts 1000 --request-rate 20 --port 6006

压测结果如下：

============ Serving Benchmark Result ============
Successful requests:                     1000      
Failed requests:                         0         
Request rate configured (RPS):           20.00     
Benchmark duration (s):                  284.71    
Total input tokens:                      2048000   
Total generated tokens:                  512000    
Request throughput (req/s):              3.51      
Output token throughput (tok/s):         1798.33   
Peak output token throughput (tok/s):    5627.00   
Peak concurrent requests:                1000.00   
Total token throughput (tok/s):          8991.65   
---------------Time to First Token----------------
Mean TTFT (ms):                          122682.20 
Median TTFT (ms):                        125387.52 
P99 TTFT (ms):                           226338.11 
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          56.63     
Median TPOT (ms):                        58.24     
P99 TPOT (ms):                           78.05     
---------------Inter-token Latency----------------
Mean ITL (ms):                           56.66     
Median ITL (ms):                         22.26     
P99 ITL (ms):                            165.95    
==================================================

从上面数据可以看出这张显卡跑这个模型并发不太高，每秒钟可以处理3路并发，如果需要每秒钟处理更高的并发，需要更换显存更大算力更强的显卡。