feat: add superderive child and open phase-2 routing

Attach the stepwise derivation toolkit as a submodule and default
SUPERPAPER_PHASE to 2 so rewrite-figure can use superderive. Notes
still keep align for short expansions without cancel or subst.
This commit is contained in:
dela
2026-08-17 10:06:39 +08:00
parent 3a322aa8dc
commit e6918ffa35
24 changed files with 659 additions and 12 deletions
@@ -0,0 +1 @@
$QK^\top \to QK^\top/\sqrt{d_k}$。缩放压方差,避免 softmax 饱和。
@@ -0,0 +1,14 @@
# Figure request F3
toolkit: superderive
language: cjk
claim: 1/√d_k 是改写,不是新算子
grammar: derivation
work_rel_dir: figures/F3
## Flow
stage: {name: D1, text: "缩放来自方差,不是装饰"}
row: {name: R1}
in_row:
- {macro: sdstep, name: s1, role: rewrite,
lhs: "QK^{\\top}", rhs: "QK^{\\top}/\\sqrt{d_k}"}
- {macro: sdreason, name: s1, text: "§3.2.1,避免 softmax 饱和"}
@@ -0,0 +1,20 @@
% Golden example -- a scale rewrite that must be seen, not just aligned.
% ../scripts/build.sh rewrite-cancel.tex
\documentclass[border=10pt]{standalone}
\usepackage[cjk]{superderive}
\begin{document}
\begin{tikzpicture}
\sdstage{D1}{缩放来自方差,不是装饰}
\sdrow{R1}
\sdstep[role=rewrite]{s1}{QK^{\top}}{QK^{\top}/\sqrt{d_k}}
\sdreason{s1}{§3.2.1,避免 softmax 饱和}
\sdrowend
\sdbbox{all}
\sdtopformula{F}{$\mathrm{Attention}(Q,K,V)=\mathrm{softmax}(QK^{\top}/\sqrt{d_k})V$}
\sdmeaningbox{mb}{120mm}{all}
{点积方差随 $d_k$ 线性涨}
{缩放是改写,不是新算子}
{没有缩放,softmax 进饱和区,梯度消失}
\end{tikzpicture}
\end{document}
Binary file not shown.

After

Width:  |  Height:  |  Size: 109 KiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 15 KiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 105 KiB

File diff suppressed because one or more lines are too long

After

Width:  |  Height:  |  Size: 238 KiB