b8974
New release patches a critical segfault on AIX and adds more build targets.
The llama.cpp project, a popular open-source library for running large language models locally, has released version b8974. This maintenance update primarily addresses a critical bug on IBM's AIX operating system, where the tiled matrix multiplication (matmul) path caused segmentation faults. The issue arose from vec_xst operations writing near 4KB page boundaries, triggering crashes under AIX's strict memory protection. The fix, contributed by IBM's Shalini Salomi Bodapati and reviewed by Aaron Teo, disables the tiled matmul path on AIX and falls back to the more stable mnpack implementation.
Beyond the AIX fix, the release continues llama.cpp's commitment to broad platform support. The included build matrix covers macOS (Apple Silicon and Intel, with optional KleidiAI acceleration), Linux (x64, arm64, s390x), Windows (x64 and arm64 with CUDA, Vulkan, SYCL, HIP), Android (arm64), and openEuler (x86 and aarch64 with ACL Graph). This ensures developers can run LLM inference on virtually any hardware, from consumer laptops to enterprise servers. The release is signed with GitHub's verified signature for authenticity.
- Fixes a segfault on AIX caused by vec_xst operations near 4KB page boundaries in tiled matmul
- Falls back to mnpack implementation on AIX for stable execution
- Supports 20+ build targets including macOS, Linux, Windows, Android, and openEuler
Why It Matters
Ensures stable LLM inference on AIX servers, expanding llama.cpp's enterprise compatibility for critical workloads.