The main
model then validates the candidate tokens in a single forward pass, which speeds up the decoding process.