lg_loss_scale continue declining in the training step on the MODU Dataset

[CV-poo]

Thanks for your good work! When I was training the MODU dataset, I encountered a situation where lg_loss_scale kept declining and the gradient kept being unstable. Please tell me how to solve this problem. Did you not encounter this problem during the experiment?

[YusenZheng]

I've also encountered this problem. May I ask how you solved it?