Hi, I found that the number of inputs for function forward_step in Megatron-LM-v1.1.5-ZeRO3/pretrain_gpt2.py is incorrect. It should accept a third param, e.g. curriculum_learning, because Megatron-LM-v1.1.5-ZeRO3/megatron/training.py calls it in this way.