llama.cpp

vulkan: small M matrix optimizations for qwen

#28457

Merged