3 ms·

Yep so GRPO is much more memory efficient than PPO, but other RL type algorithms can work fine as well!