Comment by puilp0502

What kind of benefit does Multi-Token Prediction bring to the inference side? Is it only relevant in pretraining efficiency?