Weird pre-train loss

[BubvieyKevin]

Hello, congratulations for publishing a good paper and thank you for your excellent work.
During the pretraining and finetuning process, I noticed that the loss remained at 0 consistently. Upon checking the image tensor, I found that many tensors were also 0.
Is this normal? What are your thoughts on this phenomenon? How can I address this issue?

[xrorrim]

Thanks for report this issue and we have fixed this problem in the latest version of code.