TRT inference consumes a large amount of memory

[dengshuhai-cmd]

Using the YOLOV8n model trained by oneself (Class 2) and the official YOLOV8n model provided (Class 80), both with 640 * 640 images as input, when reasoning on Jetson (8g), the YOLOV8n model (Class 80) displays the log 'init CUDNN', which occupies a large amount of memory and is twice that of the YOLOV8n model (Class 2). However, the YOLOV8n model (Class 2) does not display 'init CUDNN' during reasoning. What is the situation? When I run these two models on a PC (Linux), they occupy similar amounts of memory and there is no occurrence of init CUDNN.