llama.cpp b12221:修复 Metal Flash Attention 输出缓冲区越界写入
原标题:b12221
AI 摘要
llama.cpp 发布 b12221 版本,主要修复 Metal 后端中 Metal FA(Flash Attention)输出缓冲区越界写入的问题(PR #31662),由 Zhuomin Chen 共同提交。该版本同时提供 macOS/iOS、Linux、Android、Windows 等多平台预编译包,覆盖 CUDA、Vulkan、ROCm、SYCL、OpenVINO 等后端。此次更新属于底层推理框架的稳定性修复,可降低 Apple Silicon 设备上运行 Flash Attention 时的内存安全风险。
正文节选
<details open> metal : prevent out-of-bounds writes to the Metal FA output buffer (#31662) Co-authored-by: Zhuomin Chen <moanachen@Mac.attlocal.net> </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/attestations/65341937> **macOS/iOS:** - [macOS Apple Silicon (arm64)](https://github.com/ggml-org/llama.cpp/releases/download/b12221/llama-b12221-bin-macos-arm64.tar.gz) - macOS Apple Silicon (arm64, KleidiAI enabled) [DISABLED](https://git