dgx spark使用vllm部署千问模型

想用新版本建议参考此文：Nvidia DGX Spark 集群分别使用vLLM和SGLang部署 Qwen3.5-35B-A3B 技术方案-CSDN博客

用老版本模型，可以使用官方步骤：用于推理的 vLLM | DGX Spark | NVIDIA 开发者

测试推理效率

vllm

经过测试DGX Spark 128GB 用vllm运行Qwen/Qwen3.5-35B-A3B-FP8 token大概50token/s速率，占用105GB内存。推理速度OpenClaw勉强可用。

经过测试DGX Spark 128GB 用vllm运行Qwen/Qwen3.5-122B-A10B-GPTQ-Int4 token大概25token/s速率，占用118GB内存。OpenClaw基本不能用，太慢。

sglang

经过测试DGX Spark 128GB 用sglang运行Qwen/Qwen3.5-35B-A3B-FP8 token大概75token/s速率，占用105GB内存。

经过测试DGX Spark 128GB 用sglang运行Qwen/Qwen3.5-122B-A10B-GPTQ-Int4 token大概25token/s速率，占用118GB内存。

下载镜像：

docker run --gpus all --rm -v ./volumes/iios-vllm/models:/root/.cache/huggingface     -e HF_ENDPOINT=https://hf-mirror.com -e HF_HUB_ENABLE_HF_TRANSFER=1  lmsysorg/sglang:dev-arm64-cu13   huggingface-cli download Qwen/Qwen3.5-122B-A10B-GPTQ-Int4

成功部署配置：

vllm

  iios-vllm:
    container_name: iios-vllm
    hostname: iios-vllm
    image:  vllm/vllm-openai:cu130-nightly-aarch64
    ports:
      - "8000:8000"
    volumes:
      - ./volumes/iios-vllm/models:/root/.cache/huggingface:rw
    environment:
      HF_ENDPOINT: https://hf-mirror.com
      HF_HUB_ENABLE_HF_TRANSFER: 1
      SAFETENSORS_FAST_GPU: 1
      CUDA_DEVICE_MAX_CONNECTIONS: 1
    #command: 
      # -  /usr/local/bin/huggingface-cli
      # - download 
      # - Qwen/Qwen3.5-35B-A3B-FP8
    command: 
      - --model=Qwen/Qwen3.5-35B-A3B-FP8
      - --max-model-len=32768
      - --gpu-memory-utilization=0.85
      - --max-num-batched-tokens=4096
      - --kv-cache-dtype=fp8
      - --enable-prefix-caching
      - --enable-auto-tool-choice
      - --tool-call-parser=qwen3_xml
      - --reasoning-parser=qwen3 
      - --trust-remote-code
      - --enforce-eager
     # --disable-log-requests
    restart: always
    shm_size: 16g
    ipc: host
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    logging:
      options:
        max-size: "10240k"
        max-file: "10"
    networks:
         extnetwork:
            ipv4_address: 172.20.0.12

sglang

version: '3.5'

services:
  iios-sglang:
    container_name: iios-sglang
    hostname: iios-sglang
    image:  lmsysorg/sglang:dev-arm64-cu13
    ports:
      - "30000:30000"
    volumes:
      - ./volumes/iios-sglang/models:/root/.cache/huggingface:rw
    environment:
      HF_ENDPOINT: https://hf-mirror.com
      HF_HUB_ENABLE_HF_TRANSFER: 1
    #command: 
      # -  /usr/local/bin/huggingface-cli
      # - download 
      # - Qwen/Qwen3.5-35B-A3B-FP8
      # /opt/iios-model-docker/volumes/iios-sglang/models/hub/models--Qwen--Qwen3.5-122B-A10B-GPTQ-Int4/snapshots/5b9f0050d3ec98b0c81a7716776533c5eacebb64
    command: >
      /usr/bin/python3 -m sglang.launch_server
        --model-path /root/.cache/huggingface/hub/models--Qwen--Qwen3.5-35B-A3B-FP8/snapshots/0b2752837483aa34b3db6e83e151b150c0e00e49 
        --host 0.0.0.0
        --port 30000
        --tp 1
        --attention-backend triton
        --mem-fraction-static 0.80
        --context-length 32768
    restart: no
    shm_size: 16g
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    logging:
      options:
        max-size: "10240k"
        max-file: "10"
    networks:
         extnetwork:
            ipv4_address: 172.20.0.11version: '3.5'

services:
  iios-sglang:
    container_name: iios-sglang
    hostname: iios-sglang
    image:  lmsysorg/sglang:dev-arm64-cu13
    ports:
      - "30000:30000"
    volumes:
      - ./volumes/iios-sglang/models:/root/.cache/huggingface:rw
    environment:
      HF_ENDPOINT: https://hf-mirror.com
      HF_HUB_ENABLE_HF_TRANSFER: 1
    #command: 
      # -  /usr/local/bin/huggingface-cli
      # - download 
      # - Qwen/Qwen3.5-35B-A3B-FP8
      # /opt/iios-model-docker/volumes/iios-sglang/models/hub/models--Qwen--Qwen3.5-122B-A10B-GPTQ-Int4/snapshots/5b9f0050d3ec98b0c81a7716776533c5eacebb64
    command: >
      /usr/bin/python3 -m sglang.launch_server
        --model-path /root/.cache/huggingface/hub/models--Qwen--Qwen3.5-35B-A3B-FP8/snapshots/0b2752837483aa34b3db6e83e151b150c0e00e49 
        --host 0.0.0.0
        --port 30000
        --tp 1
        --attention-backend triton
        --mem-fraction-static 0.80
        --context-length 32768
    restart: no
    shm_size: 16g
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    logging:
      options:
        max-size: "10240k"
        max-file: "10"
    networks:
         extnetwork:
            ipv4_address: 172.20.0.11