llama.cpp

ggml-webgpu: only use subgroup-matrix path when head dims are divisib…

#23020

Merged