Seq2SeqTrainer with num_beams and generation_config

I am using a Seq2SeqTrainer for evaluating a translation task. I found that the Seq2SeqTrainingArguments accept the generation_num_beams argument and several other arguments for generation. Meanwhile, it has the generation_config argument, which can config the num_beams too. So I wonder what are the difference between these arguments, will one setting in Seq2SeqTrainingArguments override the counterpart setting in generation_config?