llama.cpp b10284 fixes MTP memory allocation bug with 123k stars

llama.cpp, the widely used open-source C++ library for running LLaMA and other large language models locally, has released a new tagged version: b10284. Maintained by ggml-org, the project currently boasts over 123k stars and 21.4k forks on GitHub, making it one of the most popular AI inference engines available. The key change in this release is a single, targeted fix: a memory allocation issue for MTP (Multi-Token Prediction) layers, referenced in commit #26605. Multi-token prediction is a technique that lets models generate several tokens at once, speeding up decoding; the bug could cause crashes or instability when using models with MTP layers. This patch corrects the buffer sizing logic, ensuring proper memory assignment.

The release package includes a broad set of prebuilt binaries for various platforms and hardware accelerators. Users can download builds for macOS (Apple Silicon, including a KleidiAI-enabled variant, and Intel x64), iOS, Linux (x64, arm64, s390x CPU; Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16), Android arm64, and Windows (x64/arm64 CPU, CUDA 12.4 or CUDA 13.3, Vulkan, OpenVINO, SYCL, HIP, and OpenCL Adreno). OpenEuler builds for Ascend NPUs are also listed but currently disabled. The project also provides separate UI assets. This release focuses on stability rather than new features, making it a safe upgrade for developers relying on local inference pipelines that use MTP layers.