知识卡片
训练并行不等于生成并行
训练时完整目标序列已知,可用移位输入和因果掩码同时计算各位置的下一词元损失。生成时下一词元尚未知,通常需逐步产生并反馈进前缀。因此Transformer支持位置并行的训练计算,并不意味着一般自回归推理可一次得到所有后续内容。
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
原书出处 · 1 处
- · 查看原书依据
and so on. We can make all these predictions at once with a single matrix multiply (figure 26.18): Figure 26.18: Masked attention to make multiple causal predictions at once. Black cells are masked; they are filled with zeros. This way, one forward pass makes N predictions rather than one prediction. This is equivalent to doing a single next token prediction N times, but it all
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第26章《Transformer》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c026.xhtml)
- 结论依据:摘录中关于训练并行不等于生成并行的定义、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:and so on. We can make all these predictions at once with a single matrix multiply (figure 26.18): Figure 26.18: Masked attention to make multiple causal predictions at once. Black cells are masked; they are filled with zeros. This way, one forward pass makes N predictions rather than one prediction. This is equivalent to doing a single next token prediction N times, but it all