NVIDIA DGX™ Spark产品适配性

NVIDIA DGX ™操作系统

| GPU | Blackwell Architecture |
| CPU | 20-core Arm, 10 Cortex-X925 + 10 Cortex-A725 Arm |
| OS | NVIDIA DGX™ OS NVIDIA DGX 操作系统 | ubuntu24.04

vim docker-compose.yml

version: '3.8'
services:
  Qwen3.5-122b-a10b-nvfp4:
    image: nvcr.io/nvidia/vllm:26.04-py3
    container_name: Qwen3.5-122b-a10b-nvfp4
    shm_size: '16gb'
    environment:
      - VLLM_HOST_IP=127.0.0.1
      - VLLM_USE_MODELSCOPE=True
      - NCCL_P2P_DISABLE=1
    ports:
      - "30022:30022"
    volumes:
      - ./.cache/:/root/.cache
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ["1", "2", "3", "4"]
              capabilities: ["gpu"]
    command:
      - --served-model-name=Qwen3.5-122b-a10b-nvfp4
      - --model=nv-community/Qwen3.5-122B-A10B-NVFP4
      - --host=0.0.0.0
      - --port=30022
      - --tensor-parallel-size=4
      - --max-model-len=128000
      - --reasoning-parser=qwen3
      - --enable-auto-tool-choice
      - --tool-call-parser=qwen3_coder
      - --gpu-memory-utilization=0.9
      - --speculative-config={"method":"qwen3_next_mtp","num_speculative_tokens":2}
    restart: always

docker

root@fusionxparkgb10-8bd1:/data/tools# cat /etc/docker/daemon.json 
{
  "registry-mirrors": [
    "https://docker.1panel.live",
    "https://status.anye.xyz",
    "https://docker.1ms.run",
    "https://docker.m.ixdev.cn",
    "https://hub.rat.dev",
    "https://docker.xuanyuan.me",
    "https://dockerproxy.net",
    "https://images.cloudlayer.icu"
  ],
  "runtimes": {
    "nvidia": {
      "args": [],
      "path": "nvidia-container-runtime"
    }
  }
}

管理界面

访问127.0.0.1：11000

访问后台管理界面

账号密码登陆

启动模型报错

Qwen3.5-122b-a10b-nvfp4    | ==========
Qwen3.5-122b-a10b-nvfp4    | == vLLM ==
Qwen3.5-122b-a10b-nvfp4    | ==========
Qwen3.5-122b-a10b-nvfp4    | 
Qwen3.5-122b-a10b-nvfp4    | NVIDIA Release 26.04 (build 299333414)
Qwen3.5-122b-a10b-nvfp4    | vLLM Version 0.19.0+6bc3197f
Qwen3.5-122b-a10b-nvfp4    | Container image Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
Qwen3.5-122b-a10b-nvfp4    | 
Qwen3.5-122b-a10b-nvfp4    | Various files include modifications (c) NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
Qwen3.5-122b-a10b-nvfp4    | 
Qwen3.5-122b-a10b-nvfp4    | GOVERNING TERMS: The software and materials are governed by the NVIDIA Software License Agreement
Qwen3.5-122b-a10b-nvfp4    | (found at https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-software-license-agreement/)
Qwen3.5-122b-a10b-nvfp4    | and the Product-Specific Terms for NVIDIA AI Products
Qwen3.5-122b-a10b-nvfp4    | (found at https://www.nvidia.com/en-us/agreements/enterprise-software/product-specific-terms-for-ai-products/).
Qwen3.5-122b-a10b-nvfp4    | 
Qwen3.5-122b-a10b-nvfp4    | ERROR: This container was built for NVIDIA Driver Release 595.58 or later, but
Qwen3.5-122b-a10b-nvfp4    |        version 580.159.03 was detected and compatibility mode is UNAVAILABLE.
Qwen3.5-122b-a10b-nvfp4    | 
Qwen3.5-122b-a10b-nvfp4    |        [[]]
Qwen3.5-122b-a10b-nvfp4    | 
Qwen3.5-122b-a10b-nvfp4    | /opt/nvidia/nvidia_entrypoint.sh: line 55: exec: --: invalid option
Qwen3.5-122b-a10b-nvfp4    | exec: usage: exec [-cl] [-a name] [command [argument ...]] [redirection ...]

ollama启动qwen3.5-122b

# 编写env
root@fusionxparkgb10-8bd1:/opt/1panel/apps/ollama/ollama# cat .env 
PANEL_APP_PORT_HTTP=11434
CONTAINER_NAME='1Panel-ollama-sz9a'
CPUS=0
MEMORY_LIMIT=0
HOST_IP='0.0.0.0'
format=''
collation=''

# docker-compose.yml
root@fusionxparkgb10-8bd1:/opt/1panel/apps/ollama/ollama# cat docker-compose.yml
networks:
    1panel-network:
        external: true
services:
    ollama:
        container_name: ${CONTAINER_NAME}
        deploy:
            resources:
                limits:
                    cpus: ${CPUS}
                    memory: ${MEMORY_LIMIT}
                reservations:
                    devices:
                        - capabilities:
                            - gpu
                          count: all
                          driver: nvidia
        image: ollama/ollama:0.31.2
        labels:
            createdBy: Apps
        networks:
            - 1panel-network
        ports:
            - "${HOST_IP}:${PANEL_APP_PORT_HTTP}:11434"
        restart: always
        tty: true
        volumes:
            - ./data:/root/.ollama

内存占用

、

中控

https://cyclone.feishu.cn/wiki/AecGwXlBOiBHW4kJ4HhcltaynlA

https://cyclone.feishu.cn/wiki/WSnowJDqNiscKCkEPFpcuJoXnZd