Output with empty tensor

[CheungBH]

Hello. Thanks for your work. I met something strange happen to me while calling your forward function.
During training, there is nan generated in the output. However, when I save the tensors and do a forward separately, the output tensor is empty.
I checked that the input activations, weights, and biases are not empty.

My code and saved tensor are attached.
nan_inference.txt
The pkl file
https://connecthkuhk-my.sharepoint.com/:u:/g/personal/bhzhang_connect_hku_hk/EdszoQrphWhIkOoKe8quikgB_85ZstY_Caf2oEKnsKXfNQ?e=ErN9Jv