# Docker

Docker examples should pin images by digest for reproducibility. Do not publish
tokens, private model URLs, internal registry names, LAN IPs, hostnames, or SSH
details.

The included launcher is a sanitized single-card vLLM pattern:

- one Blackwell GPU exposed with `--gpus device=0`
- OpenAI-compatible vLLM API on container port 8000
- host bind to localhost for safety
- placeholder model path
- no secrets