accelerator.prepare() with scheduler disturbe the learning rate

[ztb-35]

Hi, guys:
Recently, I reproduce this paper. But, I found an interesting thing. The learning rate changes from 1e-3 to 4e-5 at the first step in the first epoch, and stays till the end. I am not sure if it is a bug with accelerator.prepare() or not.
Any discussion is appreciated!
Here is how I print the learning rate.

It returns 1e-3 as my setting before training.

During training, I get the learning rate always as 4e-5.

---

I also try to change the args.lradj, I tried 'type1', 'constant', and 'COS'. It is always like above said.