llama.cpp

vulkan: use flops instead of weight tensor size for submission heuristic

#25005

Merged