llama.cpp b9504 improves build system, adds platform fixes

llama.cpp, the open-source C/C++ implementation for running large language models locally, has released version b9504. Maintained by the ggml-org team, this project has garnered over 115,000 GitHub stars and 19,200 forks, making it one of the most popular AI inference engines. The b9504 release focuses on build system refinements and platform-specific improvements.

The key change is a CMake update that now skips the cvector-generator and export-lora components when the CPU backend is disabled. This streamlines compilation for GPU-only setups (e.g., CUDA, Vulkan, ROCm) and reduces unnecessary dependencies. The release also notes that KleidiAI, a software library for optimised neural network inference, is currently disabled on macOS Apple Silicon (arm64). The full list of supported platforms includes Ubuntu (x64, arm64, s390x), Windows (x64, arm64), Android arm64, and macOS Intel/Apple Silicon. GPU backends span CUDA 12 and 13 (Windows), Vulkan, ROCm 7.2, OpenVINO, SYCL, and HIP.

For developers and AI enthusiasts, this version continues llama.cpp's tradition of efficient local inference. While no major new features are announced, the build improvements ensure smoother deployment across diverse hardware. Users can expect more reliable compilation for GPU-only environments and the usual high performance for models like Llama 3, Mistral, and Gemma.