llama.cpp b11310 修复 CUDA iq4_nl 反量化越界

AI 摘要

llama.cpp 发布 b11310 版本，修复了 CUDA 后端中 iq4_nl 反量化行内核的越界问题。当行长度短于 QK_K 时，起始位置超出行尾的线程会读写越界，新代码通过跳过这些子块来避免内存越界，对 QK_K 整数倍的行无影响。该版本同时提供 macOS、Linux、Windows、Android 等多平台预编译包。

正文节选

<details open> cuda: guard the iq4_nl dequantize row kernel against short rows (#29683) dequantize_block_iq4_nl writes QK_K values per block, but a row can be shorter than that (an IQ4_NL row is only guaranteed to be a multiple of QK4_NL). Threads whose 32-value sub-block starts at or past k currently read and write past the end of the row. Skip those sub-blocks; for rows that are a multiple of QK_K the check never fires. </details> **Website:** - <https://llama.app> **Attestations:** - <ht

发布时间：2026-10-01 10:05

抓取时间：2026-10-01 10:36

来源机构：ggml-org