Code issue in demo

[zihaoli0629]

📚 Documentation

Hi,

In src/lightning/pytorch/demos/transformer.py, an encoder-decoder transformer is used for next token prediction. However, just as the conventional setup in encoder-decoder models, the whole src is seen by the model since there’s no src mask. And the prediction target is just the shift right of the source. Wouldn’t this result in a future leak since the model can just output the i+1-th token in the src when predicting the i-th token in the target?

cc @lantiga @Borda