llama.cpp

vulkan: optimize rms_norm, and allow the work to spread across multiple SMs

#15281

Merged