知识卡片
更长生成累积解码开销
自回归生成要逐步产生输出,较长答案通常增加解码步骤和等待。缓存可复用前文状态,不能把每步都理解为从头重复全部前缀;长度增长的真实成本还受模型、批量、缓存和输出条件影响。
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
- · 理解推理模型
原书出处 · 1 处
- · 查看原书依据
First, they tend to produce longer outputs because they include intermediate steps that explain how an answer is derived. As figure 1.8 illustrates, LLMs generate text one token at a time and each token requires a full forward pass. If a reasoning model's answer is twice as long, generation involves roughly twice as many forward passes, which increases compute costs.
参考来源
- 位置:《Build a Reasoning Model (From Scratch)》第01章《理解推理模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-1.html)
- 结论依据:摘录中关于更长生成累积解码开销的定义、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:First, they tend to produce longer outputs because they include intermediate steps that explain how an answer is derived. As figure 1.8 illustrates, LLMs generate text one token at a time and each token requires a full forward pass. If a reasoning model's answer is twice as long, generation involves roughly twice as many forward passes, which increases compute costs.