ULMFiT AWD LSTM predictions from pre-trained classifier vary between versions

I have a classifier trained on fastai v1.0.52 (can be earlier one but works fine on the 52nd version). When I load it and predict on a dataset in v1.0.52 the distribution of predictions is like this:
(histogram of predictions on fastai v1.0.52)

But when I take exactly the same classifier, exactly the same dataset, and the same notebook but in another virtual environment with fastai==1.0.60, predictions are distributed differently:
(histogram of predictions on fastai v1.0.60)

New environment with fastai v1.0.60 uses torch==1.4.0, and the old one (fastai 1.0.52) has pytorch 1.2.0

It looks like a bug, but what if it is not. I will appreciate any notes on this, thank you!

The two versions have a difference in the way padding is handled, so if you are using get_preds, it’s normal to see different behaviors.

Thank you!