'forward' is used to compute deep learning network output for TRAINING whereas 'predict' is used to compute deep learning network output for INFERENCE.
In some cases like yours, both may give similar results, however some deep learning layers behave differently during training and inference (prediction). For example, during training, dropout layers randomly set input elements to zero to help prevent overfitting, but during inference, dropout layers do not change the input.
Please refer the below links for more details-
