llama.cpp b10233 improves OpenCL GPU stability for GLU operations

llama.cpp, the C/C++ library that enables running large language models locally on consumer hardware, shipped release b10233 on August 2nd. The release contains a single focused change: "opencl: limit local workgroup size for GLU operation (#26383)". GLU, or Gated Linear Unit, is an activation function used in many modern LLM architectures, including Llama and Mistral. On certain OpenCL devices, an oversized local workgroup could cause performance degradation or outright failures during matrix operations. By capping the workgroup size, this update ensures more predictable execution on GPUs that rely on OpenCL—typically Intel iGPUs, older AMD cards, and some Adreno mobile GPUs.

As with every llama.cpp release, b10233 is distributed as prebuilt binaries across an extensive matrix of platforms: macOS for Apple Silicon and Intel, Linux for x64/arm64/s390x with CPU, Vulkan, ROCm 7.2, and OpenVINO backends, plus Windows builds for CUDA 12/13, Vulkan, OpenCL Adreno, and HIP. Android arm64 and openEuler variants are also included. While this patch is small, it reflects llama.cpp's ongoing commitment to broadening GPU accessibility beyond NVIDIA's CUDA. For developers and enthusiasts running LLMs on non-CUDA hardware, this update removes a subtle stability hurdle, making local inference smoother on OpenCL-supported devices.