Gran AI Server Monitor

Each request splits into three stages: CPU pre (JPEG decode) → GPU (the predict() call — GPU queue wait + compute, since concurrent requests share one GPU) → CPU post (result formatting / refine). Arrival / Start / Done are the ordering ranks; a highlighted Done rank differing from Arrival means the server reordered that request under load. Under concurrency the GPU column grows for later requests because they queue behind earlier ones — CPU pre/post stay small and flat.