The ggml-org team has released llama.cpp b12050, with a Vulkan change aimed at improving how the software handles small workloads on AMD RDNA 3 and RDNA 4 graphics cards. The release notes do not give benchmark results or quantify any speed-up.
Downloadable builds include packages for macOS, iOS and several Linux CPU platforms.
Why it matters
People running llama.cpp on RDNA 3 or RDNA 4 GPUs may get better handling of small workloads, but the release notes do not say how much performance changes in practice.
