This llama.cpp Update Fixes a Sneaky Hexagon Bug — And Extends AI to More Platforms Than Ever

The popular open-source LLM inference engine llama.cpp has released version b9742, a minor but important update focusing on a critical fix for Hexagon DSP support. The tag addresses issue #24470 by correcting the padded stride used for state-space model (SSM) convolution weights on Hexagon processors. This fix ensures that models using SSM-based architectures (like Mamba or similar) run correctly and efficiently on Qualcomm's Hexagon DSP, which is commonly found in mobile and edge devices.

The release is notable for its comprehensive build matrix, supporting nearly every major platform and hardware acceleration backend. Users can download prebuilt binaries for macOS (Apple Silicon, Intel, and iOS), Linux (including s390x, Vulkan, ROCm 7.2, OpenVINO, and SYCL on both x64 and arm64), Windows (CPU, CUDA 12/13, Vulkan, OpenVINO, SYCL, and HIP), Android (arm64), and even openEuler with ACL Graph support. The commit is GPG-signed and includes updated UI assets. This release underscores llama.cpp's role as a versatile, cross-platform tool for local LLM inference.