LMStudio, but it uses llama.cpp to run inference, so yeah. This is with the vulkan backend, not ROCm.