Wang Xinyu

The CUDA environment seems not right, TRT is built on 11.5, but it's loading 11.4, and you said your CUDA is 11.1...

> 另外预处理设置的gpu_id，如果与runtime的gpu_id不符，还需处理ORT、Paddle、和TensorRT在接收到输入后的情况 @jiangjiajun @heliqi 我们是在三个backend里面加device id的判断和cudaMemcpyPeerAsync()吗？ 还是device id不一致的时候Assert？ 如果不给用户提示就直接copy到不同的device，可能会影响性能

> 2. 需要在Model(比如yolov5)里处理下: 如果用户设置了RuntimeOption的device_id却没有设置预处理，这个时候预处理默认与runtime保持一致 。 有些用户只知道runtime的设置，不知道预处理也要设置，所以不特殊设置预处理就默认与runtime保持一致 @heliqi 这个逻辑不好实现呢， 因为PaddleClasPreprocessor::UseGpu(int gpu_id=0)函数来设置device id的，Preprocessor默认是不用GPU的。 只能通过传入的gpu id来设置。 或者不给这个gpu id设置默认值，让用户显式指定。

You can implement yolov8 classification by referring to yolov5 classification, and welcome to submit a PR.

@adrianosantospb Can you try yolov8n or yolov8s?

Does (kNumClass = 1; and kInputH = 640; kInputW = 640;) work?

@Rex-LK Can you help take a look?

@adrianosantospb Seems @lindsayshuo had already fixed the issue, can you try again?

It won't be exactly one to one. The input shapes are slightly different, and also postprocessing. Can you try to compare several images from coco dataset? And also try to...

Nope