inference

Speculative decoding with draft models

Accelerating generation by drafting tokens with a small model.