llama.cpp b9789 fixes quantization for MoE with MTP models

llama.cpp, the popular open-source library for running large language models locally, has released version b9789. This minor but critical update addresses a bug in quantization handling for mixture-of-experts (MoE) models that also use multi-token prediction (MTP). MoE architectures selectively activate subsets of parameters per token to improve efficiency, while MTP allows generating multiple tokens in a single forward pass. The bug caused incorrect quantization in these combined setups, potentially degrading performance and increasing memory usage. The fix ensures that MoE models with MTP compress correctly, preserving quality while saving RAM.

The release maintains llama.cpp's broad platform support, providing prebuilt binaries for macOS (Apple Silicon and Intel, with optional KleidiAI), Linux (CPU, arm64, Vulkan, ROCm, OpenVINO, SYCL, s390x), Windows (CPU, arm64, CUDA, OpenCL Adreno, Vulkan, OpenVINO, SYCL, HIP), and Android (arm64). Also included are iOS XCFramework and openEuler builds. The commit is signed and verified. With 118k GitHub stars and 19.9k forks, llama.cpp continues to be the leading tool for on-device LLM inference, and this fix keeps it compatible with cutting-edge model architectures.