The llama.cpp project has released version b9905, which includes a fix for the quantized key-value cache in DSV4. This update addresses specific issues within the model's caching mechanism to improve stability or correctness.
The release provides binaries for macOS (Apple Silicon and Intel), iOS, Linux (Ubuntu x64, arm64, s390x with CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16), Android (arm64), Windows (CPU, OpenCL Adreno, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP), and openEuler (x86 and aarch64 with ACL Graph). macOS Apple Silicon KleidiAI support is disabled in this build.
This release ensures users have access to the latest bug fixes across a wide range of hardware accelerators and operating systems.