llama.cpp b9260 overhauls OpenCL backend for faster LLM inference

llama.cpp's latest release, b9260, delivers a significant under-the-hood overhaul of its OpenCL backend, aiming to boost performance and compatibility for running large language models locally. The refactoring touches backend initialization, GPU identification, and memory management—caching global memory size per device context to reduce repeated queries. Log levels have been adjusted for better debugging, and kernel loading is now smarter: the argsort kernel is built only when needed for workgroup queries, and the flash_attn kernel (which has many variants) is loaded lazily. This reduces startup overhead and memory usage.

The release continues llama.cpp's commitment to broad hardware support, with pre-built binaries for macOS (Apple Silicon, Intel, iOS), Linux (x86/arm64/S390x with CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16), Windows (x64/arm64 CPU, CUDA 12/13, Vulkan, SYCL, HIP), Android arm64, and openEuler. The refactored OpenCL backend will particularly benefit users on AMD GPUs, integrated graphics, and other OpenCL-capable devices. For developers, cleaner codebase and conditional kernel loading mean easier customization and potentially lower barriers to contributing.