turning on refit increase the inferring time of trt engine

Hi, tensorRT team! I used trt to increase the inferring time of stable diffusion about 3 times, really awesome. However, cause building a tensorrt engine is time-consuming, I'd like to use refit to replace the parameters of a tensorrt engine when I have a new stable diffusion torch model with the same structure. Then I implemented a test on refit to see how it influenced the performance of a tensorrt engine, and I found without refit, the trt engine ran 30 steps for 1.16s, while it ran 30 steps for 2.02s with refit. It seems that turning on refit will increase the inferencing time of trt engine, is that right?