Speculative decoding (SD) accelerates large language model inference by reducing the sequential cost of token-by-token generation. Advanced SD methods reduce…
机构:阿里
来源:arXiv 2611.02846 | AI4Papers 论文推荐平台