Speculative decoding accelerates LLM inference by drafting candidate tokens and verifying them in parallel. Tree-attention drafters such as EAGLE-3 are widel…
机构:剑桥
来源:arXiv 2609.02897 | AI4Papers 论文推荐平台