Tensort inference speed slow in 1.3.0

[wentout2007]

I am using an rtmdet model to segment small objects in a big picture.
the model has been coverted for tensorrt use

in a previous version of mmdeploy, the mask return is the same size as the detected object,
however, in version 1.30, the mask return is the same size as the origin big picture.

What is WORSE, the inference speed is extremely slow in version 1.3, as if tensorrt is not working.
Inference takes 30 times longer！！！！

anybody tell me why??????