llama.cpp

CUDA: refactor ggml_cuda_op + lower GPU latency via quantization on main GPU and tiling

#3110

Merged