random_seed(0,use_cuda=False ) before tabular_learner and also fit_one_cycle
in V1 was helpful in reproducing results.