1.什么是Dify1Dify 是一个开源的大语言模型LLM应用开发平台帮助开发者快速搭建生产级生成式AI应用2特点①支持各类应用创建聊天助手、Agent、文本生成应用、工作流等。②模型支持丰富内置数百个模型支持、直观的Prompt编排界面③易用性提供界面和API减少开发者重复工作聚焦创新与业务需求。④企业级LLMOps平台通过可视化工具和流程支持对大型语言模型的运维、监控、标注和持续优化。1.1 dify的安装1下载对应的压缩包①通过docker安装建议docker为20.0以上版本②登录github.com。搜索dify点进去即可下载对应的压缩包③点击Tags选择需要的版本进行下载即可2将dify安装包放到Linux服务器中进入 Dify 源代码的 Docker 目录cd dify/docker3复制环境配置文件cp .env.example .env4通过docker compose方式启动docker compose up -d1.2 修改启动端口1.3 设置大模型1点头像选择设置2点击模型供应商安装需要的模型然后进行配置即可1.4 dify安装ollama1上传ollama-linux-amd64.tar.zst文件到服务器sudo apt install vim -y 下载vimsudo apt install zstd -y 下载zstd2创建安装目录sudo mkdir -p /opt/ollama3进入目录并解压安装包在 ~/cd /opt/ollamasudo tar --use-compress-programzstd -xvf ~/ollama-linux-amd64.tar.zst -C /opt/ollama4为Ollama创建专用系统用户安全考虑sudo useradd -r -s /bin/false -m -d /usr/share/ollama ollama5设置目录权限sudo chown -R ollama:ollama /opt/ollama6确保Ollama用户的HOME目录存在且有正确权限sudo mkdir -p /usr/share/ollamasudo chown -R ollama:ollama /usr/share/ollamasudo chmod 755 /usr/share/ollama7提前创建模型目录避免后续权限问题sudo mkdir -p /usr/share/ollama/.ollama/modelssudo chown -R ollama:ollama /usr/share/ollama/.ollamasudo chmod 755 /usr/share/ollama/.ollama8确保Ollama主程序有执行权限sudo chmod x /opt/ollama/bin/ollama9创建 /etc/systemd/system/ollama.service注这里没有指定GPU运行①sudo vim /etc/systemd/system/ollama.service[Unit] DescriptionOllama Service Afternetwork-online.target Wantsnetwork-online.target [Service] Typeexec Userollama Groupollama WorkingDirectory/opt/ollama ExecStart/opt/ollama/bin/ollama serve EnvironmentOLLAMA_HOST0.0.0.0:11434 Restartalways RestartSec3 EnvironmentHOME/usr/share/ollama EnvironmentOLLAMA_MODELS/usr/share/ollama/.ollama/models # 允许Ollama写入模型目录 ReadWritePaths/usr/share/ollama/.ollama [Install] WantedBydefault.target②命令# 重载systemd配置 sudo systemctl daemon-reload # 启用开机自启 sudo systemctl enable ollama # 启动服务 sudo systemctl start ollama # 检查状态 sudo systemctl status ollama10验证安装成功和删除命令# 创建软链接使ollama命令全局可用 sudo ln -sf /opt/ollama/bin/ollama /usr/local/bin/ollama # 验证安装 ollama --version # 如果开启防火墙需要释放端口 sudo ufw allow 11434/tcp sudo ufw allow 8081/tcpdify是这个端口号 sudo ufw reload 重新载入防火墙 # 安装并运行模型run指运行并下载pull是下载 ollama run pielee/qwen3-4b-thinking-2507_q8 如要删除ollama # 停止并禁用服务 sudo systemctl stop ollama sudo systemctl disable ollama # 删除服务文件 sudo rm /etc/systemd/system/ollama.service sudo systemctl daemon-reload # 删除安装目录 sudo rm -rf /opt/ollama sudo rm -rf /usr/share/ollama/.ollama # 删除系统用户 sudo userdel ollama 2/dev/null rm -rf /usr/share/ollama # 删除命令软链接 sudo rm -f /usr/local/bin/ollama sudo rm -rf /usr/local/lib/ollama /usr/local/bin/ollama11指定GPU运行[Unit] DescriptionOllama Service Afternetwork-online.target Wantsnetwork-online.target [Service] Typesimple Userollama Groupollama WorkingDirectory/opt/ollama # ExecStart/opt/ollama/bin/ollama serve ExecStart/usr/bin/numactl --cpunodebind0 --membind0 /opt/ollama/bin/ollama serve EnvironmentOLLAMA_HOST0.0.0.0:11434 Restartalways RestartSec3 EnvironmentHOME/usr/share/ollama EnvironmentOLLAMA_MODELS/usr/share/ollama/.ollama/models # 隔离到特定 GPU推荐多模型独立部署时使用 EnvironmentCUDA_VISIBLE_DEVICES0 # 仅使用 GPU 0 EnvironmentOLLAMA_GPU_LAYERScuda # 确保在指定 GPU 上运行 # 允许Ollama写入模型目录 ReadWritePaths/usr/share/ollama/.ollama [Install] WantedBydefault.target12多GPU运行[Unit] DescriptionOllama Service Afternetwork-online.target Wantsnetwork-online.target [Service] Typesimple Userollama Groupollama WorkingDirectory/opt/ollama # ExecStart/opt/ollama/bin/ollama serve ExecStart/usr/bin/numactl --cpunodebind0 --membind0 /opt/ollama/bin/ollama serve EnvironmentOLLAMA_HOST0.0.0.0:11434 Restartalways RestartSec3 EnvironmentHOME/usr/share/ollama EnvironmentOLLAMA_MODELS/usr/share/ollama/.ollama/models # 隔离到特定 GPU推荐多模型独立部署时使用 EnvironmentCUDA_VISIBLE_DEVICES0,1,2 # 使用 GPU 0,1,2 EnvironmentOLLAMA_GPU_LAYERScuda # 确保在指定 GPU 上运行 # 允许Ollama写入模型目录 ReadWritePaths/usr/share/ollama/.ollama [Install] WantedBydefault.target