llama.cpp b10829 发布：修复 GDN 归一化对齐

AI 摘要

llama.cpp 发布 b10829 版本，修复了 GDN 归一化从 max 改为 rsqrt 的问题，以与 flash-linear-attention 和 transformers 的实现对齐。该修复影响 Qwen3-Next 等模型的推理精度，并提供了多平台二进制下载。

正文节选

<details open> models : fix GDN normalization from `max` to `rsqrt` (#28068) * models: use flash-linear-attention's l2norm for gated delta net q/k The GDN q/k normalization is defined by flash-linear-attention as l2norm(x) = x * rsqrt(sum(x*x) + eps) with eps inside the root. Every GDN call site in the tree uses ggml_l2_norm instead, which is x / max(sqrt(sum(x*x)), eps), i.e. torch.nn.functional.normalize - its CUDA kernel cites that page. The clamp never engages at these magnitudes,

发布时间：2026-09-07 07:04

抓取时间：2026-09-07 07:36

来源机构：ggml-org