llama.cpp b9941 cuts compile times with indexing, drops branches

The latest release of llama.cpp, version b9941 from the ggml-org team, focuses on two core optimizations that enhance the development and runtime experience for local large language model inference. First, the commit ‘Only index by compile times’ reworks the build process to use indexing, which avoids relying on the compiler to optimize memory layout away. This change reduces compilation times by making the build more predictable and faster, especially on large projects. Second, the ‘always multiply/add’ optimization replaces conditional control flow statements with simpler arithmetic operations (multiply and add), which are cheaper to execute on modern CPUs and GPUs. This results in more consistent inference speeds and lower latency, particularly for smaller batch sizes.

The release supports a wide range of platforms, including macOS (Apple Silicon and Intel), Linux (x64/arm64 with CPU, Vulkan, ROCm, OpenVINO, SYCL), Windows (CPU, CUDA 12/13, Vulkan, OpenCL), Android (arm64), and openEuler. The maintainers have also disabled builds for macOS Intel, iOS, and openEuler in this tag. This release is critical for developers and advanced users who run LLMs locally, as it improves both the build pipeline and runtime performance without changing model compatibility. Llama.cpp remains a key tool for running models like LLaMA, Mistral, and others efficiently on consumer hardware.