llama.cpp b9929 optimizes Vulkan for small AMD GPUs

The llama.cpp project, a popular C/C++ inference engine for large language models, released version b9929 on July 8, 2025. The standout change is a Vulkan backend optimization: for small AMD GPUs, the submission threshold is now reduced based on the compute unit (CU) count. This addresses a known bottleneck where small AMD GPUs (e.g., integrated RDNA graphics or low-end discrete cards) experienced excessive command submission overhead, leading to suboptimal throughput. By dynamically adjusting the threshold, the engine balances GPU utilization and latency, particularly benefiting models like LLaMA, Mistral, and Gemma running locally.

The release ships with pre-built binaries across 25+ platform configurations, including Windows (CPU, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP), Linux (CPU, Vulkan, ROCm 7.2, SYCL), macOS (Apple Silicon and Intel), iOS XCFramework, and Android (arm64 CPU). Notably, the macOS build with KleidiAI remains disabled, and openEuler builds use specific ACL Graph versions. The Vulkan optimization targets Xilinx/AMD devices specifically, but the broader release ensures users can run LLMs efficiently on diverse hardware without compiling from source.