RNNLearner predict causes mem leak in docker

I am trying to run a simple predict, but it keeps eating memory. I am not sure is it a fastai or torch issue and I have no clue why.
docker: docker Ubuntu18.04, python3.6 container
pytorch: https://download.pytorch.org/whl/cpu/torch-1.1.0-cp36-cp36m-linux_x86_64.whl
fastai: 1.0.46

from fastai.basic_train import load_learner
predictor = load_learner(path, "file.pkl")
pred_class, pred_idx, outputs = predictor.predict("string")

I cant reproduce the memory leak on my mac or linux machine, so I think it is docker related problem, that maybe some missing lib. Any thought?