llama.cpp Release b11389 Fixes Vulkan Tuning for AMD RDNA4
llama.cpp release b11389 introduces a targeted fix for matrix-vector (mat_vec) kernel tuning under the Vulkan backend for AMD's RDNA4 GPU architecture. The release updates pre-built binaries across supported operating systems and architectures. This patch improves performance and stability when running local LLM inference via Vulkan on next-generation AMD Radeon graphics cards. It reflects the project's continuous maintenance to maintain hardware compatibility across evolving GPU architectures. The release incorporates pull request #29934, specifically adjusting matrix-vector multiplication tuning parameters within the Vulkan backend for RDNA4 hardware. Updated builds are made available across Linux, Windows, macOS, Android, and iOS platforms.
## BACKGROUND
llama.cpp is an open-source C/C++ engine designed for efficient local inference of Large Language Models on modern hardware. It supports multiple compute backends, including Vulkan, which enables cross-platform hardware acceleration across various GPU vendors. RDNA4 is AMD's GPU microarchitecture powering its Radeon graphics card lineup.