The code base on which the Mesa 25.3 release is based changes have been adopted that significantly increase the speed of the engine for executing large language models Llama.cpp when using the Vulkan backend on systems with AMD GPUs and the RADV Mesa driver. Optimized RADV driver in some llama-bench tests became faster proprietary AMDVLK driver and stack ROCm by 31% when processing requests (“pp” – prompt processing tests) and by 4% when generating tokens (“tg” – token tests generation). The optimization was performed by Rhys Perry from Valve, who is involved in the development of the Vulkan RADV driver and the ACO shader compiler.
/Reports, release notes, official announcements.