llama.cpp b9884 fixes Vulkan integer overflow for smoother local AI

Llama.cpp b9884, released July 6, 2025 by the ggml-org team, is a maintenance release focusing on a critical bug fix in the Vulkan GPU backend. The community-developed project, which enables running large language models locally on consumer hardware, identified a 32-bit integer overflow in the CEIL_DIV macro used extensively in Vulkan compute shaders. This overflow could lead to incorrect tiling calculations, causing GPU crashes or silent data corruption during inference, particularly on GPUs with limited 32-bit integer precision.

The release delivers pre-compiled binaries for an extensive list of platforms and backends. For Windows, it covers x64 (CPU, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP) and arm64 (CPU, OpenCL Adreno). Linux builds include Ubuntu x64/arm64 for CPU, Vulkan, ROCm 7.2, OpenVINO, and SYCL FP32/FP16. macOS/iOS support spans Apple Silicon (both standard and KleidiAI-enabled builds) and Intel x64. Android arm64 CPU builds are also provided. This widens accessibility for local AI experimentation across diverse hardware, while the Vulkan fix specifically stabilizes performance on AMD Radeon, Intel Arc, and older NVIDIA GPUs using the Vulkan path.