Trained ResNet doesn't work in eval mode, behaves strangely

Hello,
did anybody found a solution for this problem ?
This looks like the same issue I just posted here Model (resnet34) performs better without model.eval()

I am using the same Data Loader and pre-processing functions and I can achieve very high accuracy in model.train and very low with model.eval.

I am running with Pytorch 1.13.0+cu117