llama.cpp

Loading models directly into VRAM, norm calculation on GPUs, broadcasting for ggml_mul

#1483

Merged