Got NaN eval loss when I set the training batch size to 2.

[Car-pe]

Hello,

I just ran your project through your script, but when I set the batch size to 2, the eval loss is NaN. And when I increase the batch size to 4 or larger, the eval loss is not NaN.
My Transformer version is 4.44.2; Leandojo version is 2.1.2. (Leandojo version 1.1.2 cannot evaluate.)