ODS 部署：Docker、GPU、端口与恢复检查

Choose the platform path

ODS 为 Linux、Windows/WSL2 和 Apple Silicon 提供不同安装路径。部署验证要同时检查 Docker、GPU、端口、模型、RAG/媒体流程、重启恢复和云模式成本。

Inspect the bootstrap source before running it, especially when using the hosted Linux/macOS endpoint that proxies repository main. Keep a local copy of the release receipt, installer hash, and support matrix so an upgrade can be explained or rolled back.

GPU, ports, and cloud fallback

The README separates NVIDIA/AMD/Intel paths on Linux, Docker GPU passthrough on Windows, and native Metal inference on macOS. Open WebUI is available on localhost:3000; inference endpoints vary by platform and can be overridden in `.env.example`. Test the effective ports from the host and from containers rather than assuming localhost means the same thing everywhere.

Machines without suitable hardware can use `./install.sh --cloud` with OpenAI, Anthropic, or Together APIs. Cloud mode changes the privacy and cost model, so keep it explicit in deployment metadata and add provider quotas and spend alerts.

Operate and recover

After install, run a small chat, model download, RAG query, and one optional voice or image flow. Then restart Docker, inspect service health, and verify that configuration and model manifests survive. Record logs, effective environment, GPU driver, model revision, and the exact ODS command when diagnosing a failure.

Test the documented uninstall path in a disposable environment and keep backups of configuration and user data. Port conflicts should be handled with environment overrides such as `WEBUI_PORT=9090`, not by editing generated compose files without a rollback plan.

Production gate

Treat ODS as an integrated appliance: restrict dashboard access to a trusted network, rotate secrets, pin release channels, monitor disk/GPU/container health, and rehearse recovery before storing sensitive prompts or documents. A green installation is not the same as a production SLA.