【vLLM】使用vLLM部署Qwen3-VL-30B-A3B-Instruct

• 环境与设备配置：H20*8(96G)

MODEL_ID=Qwen/Qwen3-VL-30B-A3B-Instruct
MODEL_NAME=Qwen3-VL-30B-A3B-Instruct
python3 -m vllm.entrypoints.openai.api_server \
  --model  $MODEL_ID \
  --served-model-name $MODEL_NAME \
  --tensor-parallel-size 8 \
  --mm-encoder-tp-mode data \
  --limit-mm-per-prompt.video 0 \
  --mm-processor-cache-type shm \
  --enable-expert-parallel \
  --host 0.0.0.0 \
  --port 22002 \
  --dtype bfloat16 \
  --gpu-memory-utilization 0.75 \
  --quantization fp8 \
  --distributed-executor-backend mp

请求推理

import time
from openai import OpenAI

client = OpenAI(
    api_key="EMPTY",
    #base_url="http://127.0.0.1:22002/v1",
    base_url="http://10.0.128.211:22002/v1",
    timeout=3600
)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url": "https://ofasys-multimodal-wlcb-3-toshanghai.oss-accelerate.aliyuncs.com/wpf272043/keepme/image/receipt.png"
                }
            },
            {
                "type": "text",
                "text": "Describe the image."
            }
        ]
    }
]

start = time.time()
response = client.chat.completions.create(
    #model="Qwen3-VL-235B-A22B-Thinking",
    #model="Qwen3-30B-A3B-Instruct-2507",
    #model="/mnt/data/projects/GLM_4.5v/ckpts/Qwen/Qwen3-30B-A3B-Instruct-2507",
    model="Qwen3-VL-30B-A3B-Instruct",
    messages=messages,
    max_tokens=2048
)
print(f"Response costs: {time.time() - start:.2f}s")
print(f"Generated text: {response.choices[0].message.content}")