llama.cpp

33f890e5 - vulkan: support flash attention GQA/split_k with small batches (#18938)