Has anyone managed to get x2 AMD GPUs working with LLMs?

I know it is possible to split layers between Nvidia GPUs with cublas. But with AMD and ROCm, it is far more difficult, and maybe not implemented yet for any project?