llama.cpp b9667 adds Vulkan support for gated delta networks

The ggml-org team has released llama.cpp version b9667, a minor but impactful update to the widely-adopted open-source framework for running large language models locally. The headline feature is Vulkan support for gated_delta_net with a stride of 16 (S_v=16). This targets mixture-of-experts (MoE) models, which use sparse activation patterns, by optimizing the delta network computations on Vulkan-compatible GPUs. The update ensures broader hardware compatibility and potentially faster inference for users leveraging Vulkan over CUDA or other backends.

The release also bundles pre-built binaries for a wide range of platforms: macOS (Apple Silicon and Intel), Linux (x64, arm64, s390x, with Vulkan, ROCm, OpenVINO, SYCL), Windows (CPU, CUDA 12/13, Vulkan, SYCL, HIP), and Android (arm64 CPU). While many existing users may continue using CUDA or Metal, the Vulkan addition lowers barriers for AMD GPU users and others without native CUDA support. With 117k stars and 19.7k forks on GitHub, llama.cpp remains a cornerstone of the local AI movement, enabling developers and enthusiasts to run models like LLaMA, Mistral, and Mixtral offline.