llama.cpp GPU Usage Failure Due to Missing Launch Parameters

This is the most common cause for GPU underutilization. Even if the hardware is detected and the binary supports GPU acceleration, the application will default to CPU inference if the number of model layers to offload is not explicitly defined. You can confirm this is the issue by observing the console output during startup; if it shows '0/X layers offloaded' or fails to mention GPU layer offloading, the -ngl flag is missing. The system requires the '-ngl' (n-gpu-layers) parameter to be set to a value greater than zero to trigger GPU hardware acceleration.

Sources