ggml‑org released llama.cpp version b12532 on 2026‑11‑19. The update lets OpenCL users adjust key‑value (KV) padding in the decomposed flash‑attention algorithm. It also ships pre‑compiled binaries for Apple Silicon, Intel Macs, iOS and a range of Linux architectures, including Ubuntu x64, arm64, s390x and Vulkan builds.
Why it matters
Developers can now fine‑tune memory layout for OpenCL GPUs, potentially improving performance on supported hardware.