Load pre-trained model w/o AMP and test with AMP

I have pre-trained model which is trained without AMP. (Acc:73%)
I want to load this model and train with AMP.

So I add ‘@torch.cuda.amp.autocast()’ in model.

    def forward(self, x):

And also I added with torch.cuda.amp.autocast(): in evaluation code.
But the accuracy dropped from 73% to 0.47%.

What should I do?

Could you post the model definition, please, so that we could have a look?