Large Grad Norm

[pharaohcaptain]

When using coupled-GRPO, my grad_norm is often very large—usually in the tens and sometimes even dozens—and only occasionally drops below 1. Is this normal, or have I done something wrong?

[summmeer]

Hi, according to my observation, the grad_norm sometimes will be larger than AR counterparts but your case seems abnormal. Please make sure the reward function works well.