llama.cpp b9949 speeds up Adreno GPUs with cluster-parallel Flash Attention

The new llama.cpp release, tag b9949, brings a significant performance improvement for running large language models on Qualcomm Adreno GPUs: cluster-parallel Flash Attention (FA) decoding via OpenCL. This optimization allows the GPU to process multiple attention heads in parallel across clusters, reducing memory bandwidth bottlenecks and accelerating decode time by up to 2x on compatible hardware. The commit is signed and verified, and the release includes a wide array of build targets: Windows (CPU, CUDA, Vulkan, OpenVINO, SYCL, HIP), Linux (x64, arm64, s390x, Vulkan, ROCm, OpenVINO, SYCL), macOS (Apple Silicon with and without KleidiAI), iOS XCFramework, and Android arm64. Notably, Windows arm64 now supports OpenCL Adreno, making this the first official build for Qualcomm's mobile GPU architecture.

For practitioners and developers, this means faster local inference on smartphones, tablets, and edge devices powered by Snapdragon processors. The cluster-parallel FA approach reduces latency for real-time applications like chatbots, code assistants, and document summarization running entirely on-device. With support for multiple GPU backends, llama.cpp continues to democratize access to LLMs, enabling high-performance AI without cloud dependency. Users can expect smoother, more responsive interactions with models like Llama 3, Mistral, or Phi-3 on their own hardware, and developers can leverage the OpenCL optimizations to deploy AI in mobile apps or embedded systems.