llama.cpp release b10012 fixes Hexagon DSP issue

llama.cpp, the flagship open-source project for running large language models locally on consumer hardware, has released version b10012. Maintained by ggml-org, this project has garnered over 120k stars and 20.6k forks on GitHub, reflecting its massive adoption for on-device AI inference. The new release primarily addresses a specific bug in Hexagon DSP support: a "hexagon: fix hmx-queue signal enum-narrowing problem" (PR #25677). This fix ensures that models running on Qualcomm's Hexagon digital signal processors—used in many mobile and edge devices—handle parallel processing queues correctly, preventing crashes or GPU hangs.

The release continues llama.cpp's tradition of broad platform compatibility. Build targets include macOS on Apple Silicon (arm64, with optional KleidiAI acceleration) and Intel x64, iOS via XCFramework, Linux on multiple architectures (x64, ARM, s390x) with GPU backends Vulkan, ROCm 7.2, OpenVINO, and SYCL (FP32/FP16), Windows on x64 and ARM with CUDA 12/13, Vulkan, OpenVINO, SYCL, and HIP, as well as Android arm64 CPU and Windows ARM with OpenCL Adreno. Notably, some builds like openEuler and macOS KleidiAI are currently disabled in this release. The fix is critical for developers deploying LLMs on Qualcomm-powered devices, ensuring reliable AI performance on edge hardware.