Is there a bug in "torch.nn.CTCLoss"?

I just train my CRNN model for Chinese text recognition.
I meet this problem.

Maybe, it’s like a problem of “out of memory”.

But, the situation of GPUs is shown in the follow figure.
When I trained my model, the PID of the process is 4585. But, while the problem occured, the process of 4585 was replaced by the process in the figure.

I suspect that there is a bug in the “torch.nn.CTCLoss”.
Please, for help.