llama.cpp b9992 boosts CUDA kernel config for faster inferencing

ggml-org has released llama.cpp version b9992, a significant update to the open-source LLM inference runtime. The headline change is a complete refactor of the CUDA MMQ (matrix multiplication quantized) kernel configuration, which is expected to yield faster inference on NVIDIA GPUs—especially for models using quantization. The commit also removes legacy Blackwell GPU configuration code and fixes the Blackwell setup, ensuring compatibility with NVIDIA's latest architecture.

This release continues llama.cpp's tradition of broad platform support, shipping binaries for macOS (ARM64, Intel, KleidiAI), Linux (CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/16), Windows (CPU, CUDA 12/13, Vulkan, OpenCL Adreno), and Android ARM64. The change log notes no new features beyond the kernel optimization, focusing purely on performance and maintainability. For developers and AI engineers running local LLMs, this update promises lower latency and better throughput on consumer and enterprise NVIDIA GPUs.