知识卡片
草稿接受率不等于系统加速
接受率描述候选有多少被目标模型采用,速度还取决于草稿生成、验证、拒绝纠正和通信成本。即使接受率很高,草稿成本过大也可能不快;接受率为1的有限样本也不证明草稿在全部输入上已等同目标。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 推理优化
原书出处 · 1 处
- · 查看原书依据
Token acceptance rate This is the percentage of tokens generated by the draft model that are accepted. Typically, this does not reach 1, because if it did, there is no need to use the main larger model. Decoding speedup This refers to the reduction in latency between a model purely using autoregressive decoding versus one using speculative decoding. Constructing Draft Models How do we ensure
参考来源
- 位置:《Designing Large Language Model Applications》第09章《推理优化》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch09.html)
- 结论依据:摘录中关于草稿接受率不等于系统加速的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:Token acceptance rate This is the percentage of tokens generated by the draft model that are accepted. Typically, this does not reach 1, because if it did, there is no need to use the main larger model. Decoding speedup This refers to the reduction in latency between a model purely using autoregressive decoding versus one using speculative decoding. Constructing Draft Models How do we ensure