Speculative decoding has been treated purely as a systems optimization that speeds up large language model decoding without changing the model behavior. Unde…
机构:MIT
来源:arXiv 2611.05345 | AI4Papers 论文推荐平台