DeepSeek V4.1 Flash 腾讯云本地部署：Ollama 双路径实测

1# 安装依赖
2apt-get update
3apt-get install -y ca-certificates curl gnupg lsb-release
4
5# 添加腾讯云内网 Docker GPG key
6curl -fsSL https://mirrors.cloud.tencent.com/docker-ce/linux/ubuntu/gpg -o docker-keyring.gpg
7gpg --dearmor -o /etc/apt/keyrings/docker.gpg docker-keyring.gpg
8
9# 添加腾讯云内网 Docker 软件源
10echo "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/docker.gpg] https://mirrors.cloud.tencent.com/docker-ce/linux/ubuntu $(lsb_release -cs) stable" | tee /etc/apt/sources.list.d/docker.list > /dev/null
11
12# 安装 Docker Engine 与 compose 插件
13apt-get update
14apt-get install -y docker-ce docker-ce-cli containerd.io docker-compose-plugin
15
16# 配置 Docker Hub 镜像加速器（关键！否则拉 Ollama 镜像会慢到超时）
17mkdir -p /etc/docker
18tee /etc/docker/daemon.json <<-'EOF'
19{
20  "registry-mirrors": ["https://mirror.ccs.tencentyun.com"]
21}
22EOF
23systemctl daemon-reload
24systemctl restart docker
25
26# 把当前用户加入 docker 组，免去每次 sudo
27usermod -aG docker $USER

1# 创建 4GB swap 文件
2fallocate -l 4G /swapfile
3chmod 600 /swapfile
4mkswap /swapfile
5swapon /swapfile
6
7# 写入 fstab 实现开机自动挂载
8cp /etc/fstab /etc/fstab.bak
9echo '/swapfile none swap sw 0 0' | tee -a /etc/fstab
10
11# 验证
12free -h

1mkdir -p /opt/deepseek && cd /opt/deepseek
2cat > docker-compose.yml <<-'EOF'
3version: '3.8'
4
5services:
6  ollama:
7    image: ollama/ollama:latest
8    container_name: ollama
9    restart: unless-stopped
10    ports:
11      - "11434:11434"
12    volumes:
13      - ./ollama:/root/.ollama
14    environment:
15      # 监听所有地址，方便同 VPC 内网其他服务调用
16      - OLLAMA_HOST=0.0.0.0
17      # 同时保持加载的模型数，小内存机器请设为 1
18      - OLLAMA_MAX_LOADED_MODELS=1
19      # 模型在内存中保持 10 分钟，超时自动卸载省内存
20      - OLLAMA_KEEP_ALIVE=10m
21      # 并发请求数，按内存档位调整（8C16G 建议 2）
22      - OLLAMA_NUM_PARALLEL=1
23    networks:
24      - deepseek-net
25
26  open-webui:
27    image: ghcr.io/open-webui/open-webui:main
28    container_name: open-webui
29    restart: unless-stopped
30    ports:
31      - "3000:8080"
32    volumes:
33      - ./open-webui:/app/backend/data
34    environment:
35      # 注意：这里是容器名 ollama 而不是 localhost
36      - OLLAMA_BASE_URL=http://ollama:11434
37      # 关闭使用统计上报
38      - ENABLE_OTEL=false
39      # 生产环境务必设为 false，防止陌生人注册
40      - ENABLE_SIGNUP=true
41      # 替换为你自己的强密钥
42      - WEBUI_SECRET_KEY=change-this-to-a-random-string
43    depends_on:
44      - ollama
45    networks:
46      - deepseek-net
47
48networks:
49  deepseek-net:
50    driver: bridge
51EOF

1cd /opt/deepseek
2docker compose up -d
3
4# 查看容器状态，两个都应显示 Up
5docker compose ps
6
7# 跟踪启动日志
8docker compose logs -f ollama

1# 官方安装脚本（若下载慢可用下面的国内镜像方式）
2curl -fsSL https://ollama.com/install.sh | sh
3
4# 备选：下载 release 二进制手动安装（国内网络更稳）
5OLLAMA_VERSION=$(curl -s https://api.github.com/repos/ollama/ollama/releases/latest | grep tag_name | cut -d '"' -f 4)
6curl -L https://github.com/ollama/ollama/releases/download/${OLLAMA_VERSION}/ollama-linux-amd64.tgz -o ollama.tgz
7tar -C /usr -xzf ollama.tgz

1cat > /etc/systemd/system/ollama.service <<-'EOF'
2[Unit]
3Description=Ollama Service
4After=network-online.target
5
6[Service]
7ExecStart=/usr/bin/ollama serve
8User=root
9Group=root
10Restart=always
11RestartSec=3
12Environment="OLLAMA_HOST=0.0.0.0:11434"
13Environment="OLLAMA_MODELS=/root/.ollama/models"
14Environment="OLLAMA_KEEP_ALIVE=10m"
15Environment="OLLAMA_MAX_LOADED_MODELS=1"
16
17[Install]
18WantedBy=default.target
19EOF
20
21systemctl daemon-reload
22systemctl enable ollama
23systemctl start ollama
24
25# 检查服务状态
26systemctl status ollama --no-pager

1apt-get install -y python3-pip python3-venv
2python3 -m venv /opt/open-webui-env
3source /opt/open-webui-env/bin/activate
4
5# 使用腾讯云内网 PyPI 源加速
6pip config set global.index-url https://mirrors.cloud.tencent.com/pypi/simple
7pip install open-webui
8
9# 启动（监听 8080）
10ENABLE_OTEL=false OLLAMA_BASE_URL=http://127.0.0.1:11434 open-webui serve --port 8080

1pip install modelscope
2
3python3 <<-'PYEOF'
4from modelscope import snapshot_download
5model_dir = snapshot_download(
6    'deepseek-ai/DeepSeek-V4.1-Flash-GGUF',
7    cache_dir='/opt/deepseek/models'
8)
9print(model_dir)
10PYEOF

1cd /opt/deepseek/models/deepseek-ai/DeepSeek-V4.1-Flash-GGUF
2
3cat > Modelfile <<-'EOF'
4FROM ./deepseek-v4.1-flash-q4_k_m.gguf
5
6PARAMETER temperature 0.6
7PARAMETER top_p 0.95
8PARAMETER num_ctx 8192
9PARAMETER repeat_penalty 1.05
10
11SYSTEM 你是 DeepSeek V4.1 Flash，一个由 DeepSeek 训练的国产大型语言模型。请用简体中文回答，回答简洁准确。
12EOF
13
14# 导入 Ollama
15ollama create deepseek-v4.1-flash -f Modelfile
16
17# 验证模型已注册
18ollama list

1# 设置 HF 镜像端点环境变量
2export HF_ENDPOINT=https://hf-mirror.com
3
4# 用 huggingface-cli 下载
5pip install -U huggingface_hub
6huggingface-cli download deepseek-ai/DeepSeek-V4.1-Flash-GGUF \
7  --include "*.gguf" \
8  --local-dir /opt/deepseek/models/hf

1# 若官方模型库已上架且网络通畅，这是最省事的
2ollama pull deepseek-v4.1-flash
3
4# 若需要指定量化版本
5ollama pull deepseek-v4.1-flash:q4_k_m

1docker exec ollama ollama list

1docker exec -it ollama ollama run deepseek-v4.1-flash "用一句话介绍 DeepSeek MoE 架构的核心优势"

1curl http://127.0.0.1:11434/v1/chat/completions \
2  -H "Content-Type: application/json" \
3  -d '{
4    "model": "deepseek-v4.1-flash",
5    "messages": [
6      {"role": "user", "content": "用 Go 写一个快速排序，只要代码"}
7    ]
8  }'

1# 看整体内存与 swap 使用
2free -h
3
4# 看容器资源占用
5docker stats --no-stream
6
7# GPU 实例请额外确认显卡可见
8nvidia-smi
9docker exec ollama nvidia-smi

1{
2  "models": [
3    {
4      "title": "自建 DeepSeek V4.1 Flash",
5      "provider": "openai",
6      "model": "deepseek-v4.1-flash",
7      "apiBase": "http://你的服务器内网IP:11434/v1",
8      "apiKey": "ollama"
9    }
10  ],
11  "tabAutocompleteModel": {
12    "title": "自建 DeepSeek Flash 补全",
13    "provider": "openai",
14    "model": "deepseek-v4.1-flash",
15    "apiBase": "http://你的服务器内网IP:11434/v1",
16    "apiKey": "ollama"
17  }
18}

1cat > /opt/deepseek/nightly_summarize.sh <<-'EOF'
2#!/bin/bash
3for f in /data/docs/*.md; do
4  summary=$(curl -s http://127.0.0.1:11434/v1/chat/completions \
5    -H "Content-Type: application/json" \
6    -d "{\"model\":\"deepseek-v4.1-flash\",\"messages\":[{\"role\":\"user\",\"content\":\"请总结以下文档的要点：$(cat $f)\"}]}" \
7    | python3 -c "import sys,json;print(json.load(sys.stdin)['choices'][0]['message']['content'])")
8  echo "$summary" > "/data/summaries/$(basename $f).summary"
9done
10EOF
11
12chmod +x /opt/deepseek/nightly_summarize.sh
13
14# 每天凌晨 2 点执行
15echo "0 2 * * * /bin/bash /opt/deepseek/nightly_summarize.sh" | crontab -

1apt-get install -y nginx certbot python3-certbot-nginx
2
3cat > /etc/nginx/sites-available/open-webui <<-'EOF'
4server {
5    listen 80;
6    server_name ai.yourcompany.com;
7
8    location / {
9        proxy_pass http://127.0.0.1:3000;
10        proxy_set_header Host $host;
11        proxy_set_header X-Real-IP $remote_addr;
12        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
13        proxy_set_header X-Forwarded-Proto $scheme;
14
15        # 下面三行是流式输出必需，缺了会导致打字机效果失效
16        proxy_http_version 1.1;
17        proxy_set_header Upgrade $http_upgrade;
18        proxy_set_header Connection "upgrade";
19    }
20}
21EOF
22
23ln -s /etc/nginx/sites-available/open-webui /etc/nginx/sites-enabled/
24nginx -t && systemctl reload nginx
25
26# 签发免费证书
27certbot --nginx -d ai.yourcompany.com

1# 若只需本机访问复查Ollama，把端口收紧到 127.0.0.1
2docker compose down
3# 修改 compose 中 ports 为：
4#   - "127.0.0.1:11434:11434"
5docker compose up -d

1# 打包模型与 WebUI 数据（会比较大，建议在 screen 里跑）
2tar czf /root/deepseek-backup-$(date +%F).tar.gz -C /opt/deepseek ollama open-webui
3
4# 同步到腾讯云 COS（内网免流量）
5# coscli cp /root/deepseek-backup-*.tar.gz cos://your-bucket/backups/