Maybe a bug in grpo_trainer.py

[Hui-design]

Hi, thanks for your amazing work!
I've found that the inputs to 'get_per_token_logps' for model and ref_model are different, which might lead to a critical bug. This bug appears in R1-mulitimodal and Open-R1-Video. I've documented my understanding in this blog Hui-design/R1-Video-fixbug. Could you please review it to see if my understanding is correct?