b8924

The llama.cpp project, a leading open-source C/C++ implementation for running large language models locally, has released version b8924. With over 106,000 stars on GitHub, this update focuses on performance enhancements for Qualcomm's Hexagon DSP architecture. Specifically, the Hexagon HMX frequency has been bumped to the maximum corner, which directly increases the clock speed of the digital signal processor for AI workloads. This change is designed to accelerate neural network inference on devices using Qualcomm's mobile and edge platforms, such as smartphones and IoT hardware.

The release also includes a fix for an error in the hex-mm logging message, improving debugging for developers working with Hexagon matrix multiplication. The b8924 release maintains llama.cpp's extensive cross-platform support, offering pre-built binaries for macOS (Apple Silicon with and without KleidiAI, Intel), Linux (x64, arm64, s390x with various backends like Vulkan, ROCm, OpenVINO, SYCL), Windows (x64, arm64, CUDA 12/13, Vulkan, SYCL, HIP), Android (arm64), and openEuler. This broad compatibility ensures developers can deploy optimized LLM inference on everything from data center GPUs to mobile CPUs.