Comment by wirybeige

The vulkan backend for llama.cpp isn't that far behind rocm for pp and tp speeds