知识卡片
多令牌头仍需要联合选择
Medusa等方案增加预测未来位置的头,以提出多个后续token候选。后续token之间仍有条件依赖,不能把各头独立最高分简单拼成可靠联合序列;候选树和验证等细节需按具体方法实现。
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
additional decoding heads are added to the model. These decoding heads represent subsequent tokens to be generated. For example, the standard decoding head is predicting the next (n + 1st) token in the sequence, and the additional decoding heads are predicting the n + 2nd, n + 3rd, and so on, tokens, respectively. Refer to the Medusa paper for more details on how this is
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:摘录中关于多令牌头仍需要联合选择的定义、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:additional decoding heads are added to the model. These decoding heads represent subsequent tokens to be generated. For example, the standard decoding head is predicting the next (n + 1st) token in the sequence, and the additional decoding heads are predicting the n + 2nd, n + 3rd, and so on, tokens, respectively. Refer to the Medusa paper for more details on how this is