llama.cpp release b8779 adds a Vulkan Flash Attention DP4A shader for quantized KV cache computation, extending efficient attention beyond NVIDIA CUDA to AMD, Intel Arc, and mobile GPUs. Previously, Vulkan flash attention required NVIDIA's coopmat2 extension or fell back to CPU — this update enables hardware-accelerated quantized inference on any Vulkan-capable GPU. Significant for the local LLM community running models on non-NVIDIA hardware.