把 18 篇文章里讲过的所有零件——QKV、多头注意力、causal mask、位置编码、FFN、残差、归一化——拼成一张完整的 Transfo…

How much overhead should we expect from enabling integer overflow checks…

Recently, I had tentatively announced a forthcoming result with Ben Gree…

回到 2017 年 6 月那篇论文：八位作者、Google Brain/Translate 的内部背景、LSTM 时代的工程困境、为什么这篇在当…

I’ve had this nagging feeling that the computers I use today feel slower…

I’ve joined the inaugural editorial board for a new mathematical journal…

为什么 attention 是 O(n²)，O(n²) 到底贵在哪里，5 类降复杂度方案的优劣，FlashAttention 不是 O(n) 这…

Hardware performance “obviously” affects software performance and affect…

The first Distinguished Lecture Series at UCLA of this academic year is …

自回归语言模型的核心约束是：预测 t 时刻只能用 t 之前的信息。Causal Mask 用一个上三角的 -∞ 矩阵让 softmax 之后未来…