llama.cpp v9850 adds Qwen3Next support with attention normalization fix

The ggml-org/llama.cpp project has released version b9850, a significant update that brings enhanced support for the Qwen3 model family. This release registers the t_layer_inp tensor specifically for Qwen3Next models, fixes input assignment in the layer processing loop, resolves DFLASH issues for qwen-coder-next, and adds a new tensor for attention normalization in the Qwen3 architecture. These changes ensure compatibility with the latest Qwen3 variants while fixing critical bugs in the inference pipeline.

The release comes with pre-built binaries across a wide range of platforms: macOS (Apple Silicon, Intel, iOS), Linux (x64, arm64, s390x with Vulkan, ROCm 7.2, OpenVINO, SYCL), Windows (CPU, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP), and Android (arm64 CPU and OpenCL Adreno). This broad support allows developers and enthusiasts to run Qwen3 models locally on virtually any modern device. The update continues llama.cpp's mission to bring state-of-the-art LLM inference to consumer hardware without cloud dependencies.