llama.cpp

vulkan: fix wrong results when a mul_mat reads a slice of a larger cache

#28956

Merged