llama.cpp b9973 optimizes DeepSeek V4 memory and expands platform support

The ggml-org team has released b9973 of llama.cpp, the widely used C/C++ library for running large language models locally. This incremental update focuses on optimizing memory handling for the DeepSeek V4 model architecture. The key change – clearing cache only for the current sequence instead of the entire cache – reduces memory overhead and can improve inference speed on memory-constrained devices like laptops or single-GPU setups. This is especially relevant for users running parameter-heavy DeepSeek models on consumer hardware.

Beyond the DeepSeek fix, b9973 comes with a comprehensive set of pre-compiled binaries across major operating systems and accelerators. The release includes builds for macOS Apple Silicon (with and without KleidiAI acceleration), Linux on x64/arm64/s390x with Vulkan, ROCm 7.2, OpenVINO, and SYCL (FP32/FP16), Windows x64/arm64 with CUDA 12.4 and 13.3, Vulkan, OpenVINO, SYCL, and HIP, plus Android ARM64 and OpenCL for Adreno GPUs. For developers and AI enthusiasts, this means near-instant access to the latest optimizations without needing to compile from source.