Wang Xinyu

意义不大，yololayer已经对输出做了过滤，nms计算的bbox很少

1ms在大多数应用场景可以忽略了 也可以尝试手动添加nms plugin试试

The tensorrt-py and the tensorrt c++ need to be the same version.

Set LD_LIBRARY_PATH, put the lib path of tensorrt into this ENV.

auto具体指哪段代码？能贴下吗？

Yes, try to use TensorRT

Seems your opencv is linking cuda12.3, you can try to use docker

@zmtttt have you check this path https://github.com/wang-xinyu/tensorrtx/blob/d4aa52db68c36d10cfcb2fd9a818faf2d82bfd00/yolov9/include/config.h#L13

@WuxinrongY Can we make yolov9 to use fp16 by default?

cc @lindsayshuo