This paper presents SAGE, a hardware-aware framework for efficient protected vision-language inference on smart glasses. It reduces inference latency, device-sensor temperature rise, and packet size while retaining utility and prompted containment.