Hi. I saw your latest replied, after updating to tf-metal 0.4.0 looks like you have resolved the problem. Good for you.

Thank you, I also try to do some test, it’s interesting that if you feed in sequential data it can produce good results but not for one by one token. E.g. Feed sequential data like: (i a m g o o d) Predicted like: (a m g o o d x), the prediction for input characters are good.