Can i use loss.item() for backward?

[BramVanroy]

You should post your training loop. Growing memory likely occurs because you do not clear the gradients anywhere (zero_grad).