loss_kl becomes nan

[Carinazhao22]

Dear,

Thanks for the impressive work. I tried to train on my own synthetic datasets and loss_kl turns into nan after running a while. How can I fix it? turning the parameters？ Thanks!