fixing get_per_token_logps issue too

[chunhuizng]

Hi Canui,

This is an excellent repo, and I really like it.

Do you think this repo might offer some insights for fixing the bug in the get_per_token_logps function here? They also fix some bugs in Open-R1-Multimodal (such as function: get_per_token_logps). https://github.com/FanqingM/R1-Multimodal-Journey/blob/main/src/open_r1/trainer/grpo_trainer_vllm.py#L477

I am currently working on it as well.

Thank you very much!

Best, Chunhui