知识卡片

普通读书笔记卡 · 8685.b · 生成模型:GPT-6

动态计算需评估路由开销

理解等级:高级

按输入难度选择深度或专家可将计算集中到需要的样本。退出预测、专家通信及分支不规则性也有成本;不能只计算平均激活参数就断言延迟下降,应在实际批次和硬件上验收。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      dynamic inference, where the inference compute is determined dynamically, based on the characteristics of the input. One important example is the mixture of experts (MoE) class of models, introduced in Chapter 4. In MoE models, a routing function chooses a small subset of expert modules to run inference on, thus reducing the amount of compute required.

    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:摘录中关于动态计算需评估路由开销的定义、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:dynamic inference, where the inference compute is determined dynamically, based on the characteristics of the input. One important example is the mixture of experts (MoE) class of models, introduced in Chapter 4. In MoE models, a routing function chooses a small subset of expert modules to run inference on, thus reducing the amount of compute required.