llama.cpp b9932 boosts AMD GPU performance via Vulkan optimization
New release disables FA mask_opt on GCN for faster LLM inference.
The ggml-org team has released llama.cpp version b9932, a maintenance update focused on Vulkan performance for AMD GPUs. Key change: disabling the FA mask_opt on GCN architecture, which previously caused performance regressions. This optimization specifically targets users running LLM inference on AMD graphics cards (e.g., Radeon RX 6000/7000 series) using the Vulkan backend. Additionally, the update reenables mask optimization over attention head size 256, preserving accuracy while improving throughput.
The release includes prebuilt binaries for a wide range of platforms: Windows (x64 CPU, arm64 CPU, CUDA 12/13, Vulkan, SYCL, HIP, OpenCL Adreno), Linux (Ubuntu x64/arm64/s390x with CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16), macOS (Apple Silicon with optional KleidiAI, Intel x64), iOS XCFramework, and Android arm64 CPU. OpenEuler and s390x builds are also available but disabled in this tag. The commit is verified with GitHub's GPG key. This update is critical for developers and power users running local LLMs on AMD hardware, as it directly addresses a performance bottleneck without requiring model re-quantization.
- Vulkan: disabled FA mask_opt on GCN architecture to improve inference performance on AMD GPUs
- Re-enabled mask optimization over attention head size 256 for better accuracy-throughput tradeoff
- Available across 20+ build configurations including Windows, Linux, macOS, iOS, Android, and openEuler
Why It Matters
Local LLM users on AMD GPUs get a free performance boost—no model changes needed.