From 44b0b5da84b4ea49c3dbc84805310a190fa17ba2 Mon Sep 17 00:00:00 2001 From: Yan Lin Date: Fri, 30 Jan 2026 18:01:40 +0100 Subject: [PATCH] fix multi-line equations --- content/ml-tech/new-bert/index.md | 4 ++-- templates/shortcodes/math.html | 1 + 2 files changed, 3 insertions(+), 2 deletions(-) create mode 100644 templates/shortcodes/math.html diff --git a/content/ml-tech/new-bert/index.md b/content/ml-tech/new-bert/index.md index 0249bf0..8ed266d 100644 --- a/content/ml-tech/new-bert/index.md +++ b/content/ml-tech/new-bert/index.md @@ -37,9 +37,9 @@ There are limitations of this encoding, the most apparent one being it cannot ge Thus, LLMs nowadays mostly use RoPE (rotary positional embeddings) or its variants. In contrast to positional encoding which is a type of absolute position embedding, RoPE is a type of relative position embedding. Below is the formulation of RoPE in the 2D space given in the RoFormer paper. -$$ +{% math() %} f_{\{q,k\}}(x_m, m) = \begin{pmatrix} \cos m\theta & -\sin m\theta \\ \sin m\theta & \cos m\theta \end{pmatrix} \begin{pmatrix} W_{\{q,k\}}^{(11)} & W_{\{q,k\}}^{(12)} \\ W_{\{q,k\}}^{(21)} & W_{\{q,k\}}^{(22)} \end{pmatrix} \begin{pmatrix} x_m^{(1)} \\ x_m^{(2)} \end{pmatrix} -$$ +{% end %} > RoFormer: Enhanced transformer with Rotary Position Embedding (2024). Su, Jianlin and Ahmed, Murtadha and Lu, Yu and Pan, Shengfeng and Bo, Wen and Liu, Yunfeng. diff --git a/templates/shortcodes/math.html b/templates/shortcodes/math.html new file mode 100644 index 0000000..d552987 --- /dev/null +++ b/templates/shortcodes/math.html @@ -0,0 +1 @@ +$${{ body | safe }}$$ \ No newline at end of file