GPU-only will be faster if you have enough VRAM. But if you want to run a model that requires more VRAM than you have, the current approach is to use llama.cpp

GPU-only will be faster if you have enough VRAM. But if you want to run a model that requires more VRAM than you have, the current approach is to use llama.cpp and specify n_gpu_layers. That works, but is slower than GPU-only. OP claims to be 10x as fast as llama.cpp in the case when you can't fit the whole model in VRAM.