llama.cpp v9988 improves cross-platform builds and testing
New release harmonizes header use across platforms for smoother LLM inference.
The llama.cpp project, maintained by ggml-org, released version b9988 on July 13. This is a minor but important maintenance release that focuses on test integrity and build consistency across all supported platforms. The key change is the harmonization of header usage in tests, specifically ensuring that private ggml includes are used with quoted includes rather than angle-bracketed system includes. This prevents accidental mixing of shipped library headers with system-installed versions, which could cause subtle bugs during local LLM inference.
The release includes a comprehensive matrix of pre-built binaries: macOS (Apple Silicon with optional KleidiAI acceleration, Intel x64, iOS XCFramework), multiple Linux variants (x64 CPU, arm64 CPU, s390x CPU, plus GPU-accelerated builds for Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16), Android arm64 CPU, and Windows (x64 CPU, arm64 CPU, plus GPU backends like OpenCL Adreno, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP). Notably, some openEuler builds are disabled. For developers running LLMs locally, this update ensures more reliable compilation and runtime behavior across different operating systems and hardware configurations.
- Harmonized header usage in tests to prevent system header conflicts during builds
- Pre-built binaries shipped for 20+ platform/accelerator combinations including CUDA 12 & 13, Vulkan, and ROCm
- macOS Apple Silicon build now includes optional KleidiAI acceleration for improved inference performance
Why It Matters
Ensures stable cross-platform builds for local LLM inference, essential for developers deploying llama.cpp models.