llama.cpp

vulkan: do not enable MMVQ for speculative-decode steps on AMD

#25666

Open