llama.cpp Release b11347 Fixes Qualcomm Hexagon DSP HTP Skeleton Dependencies
llama.cpp has released build b11347, providing updated binaries alongside fixes for Qualcomm Hexagon DSP dependencies. Pull request #29828 installs rebuilt Hexagon Tensor Processor (HTP) skeleton libraries and resolves catalog dependency issues. This routine maintenance release ensures stability and proper dynamic library loading for users running local LLMs on Snapdragon platforms equipped with Qualcomm NPU hardware. It maintains smooth deployment pipelines across supported operating systems including Linux, Android, Windows, macOS, and iOS. The patch rebuilds HTP skeleton dynamic libraries and corrects the skel catalog dependencies required for Hexagon NPU offloading. Updated pre-compiled binaries are provided for Linux arm64 and Android arm64 Snapdragon builds, along with standard CPU, CUDA, Vulkan, and OpenVINO release targets.
## BACKGROUND
llama.cpp is a widely used open-source framework designed for high-performance LLM inference on consumer hardware. Qualcomm Hexagon DSPs use Hexagon Tensor Processors (HTP) to accelerate neural network operations, relying on compiled skeleton ('skel') dynamic libraries to execute routines directly on the NPU.