Longt5 summarization using huggingface sample code

Hello, I’m trying to summarize my own dataset using longt5 model, so I used official sample code for summarization released to huggingface notebooks here, and there is problem.
I checked that using my own dataset is not a problem so first I want to run this code with its offered example dataset “xsum”.
When I run this code with xsum dataset using the original “t5-small” model it’s working well, so I only changed the model from t5 to longt5.
The result is that training takes the proper time as if it’s training well, but the result is all 0 or nan value like this.
캡처
Can anyone give me some help to fix this problem? Is there a problem to compute_metrics?

I fixed the problem by myself. In huggingface github there was someone claiming same issue with me. He said the problem occurs when fp16 is turned on. So I turned off fp16 and now it’s working.