知识卡片

普通读书笔记卡 · 8683.a · 生成模型:GPT-6

优化收益需同时报告代价

理解等级:中级

减少计算、加速解码和减少存储是不同目标。缓存以空间换计算,早退改变质量风险,量化改变数值表示;衡量时同时报告质量、首token延迟、每token时间、吞吐和峰值内存,避免把tokens/s提高当所有用户等待都缩短。

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

      Reduce compute Techniques like caching, knowledge distillation, and early exit, each of them employing distinct strategies to reduce computation. Speed up decoding Techniques for parallel and speculative decoding aim to improve the throughput of the model: the number of tokens generated per second. Reduce storage needs Quantization techniques aim to reduce the amount of storage needed for weights and activations of the model, by reducing

    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:摘录中关于优化收益需同时报告代价的定义、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:Reduce compute Techniques like caching, knowledge distillation, and early exit, each of them employing distinct strategies to reduce computation. Speed up decoding Techniques for parallel and speculative decoding aim to improve the throughput of the model: the number of tokens generated per second. Reduce storage needs Quantization techniques aim to reduce the amount of storage needed for weights and activations of the model, by reducing