About the gap between Training loss and Inference loss

[T-FOUNTAIN]

I found that the loss gap between training and inference stage is very large. The loss in inference stage is 10 times that in training. Even making inference on the training set ,the situation is the same. Can you give some advice?