Got it to run, but a few bumps in the road – mostly because this is a very fresh model.
Flash Attention (FA) is not supported on the current llama.cpp (at least not for ROCm) for this model so KV cache is forced to f16 for the value (V) and I cannot lower it to q4 go get decent KV cache. But the key (K) could be set to q4, and that got me about 100k of ctx. Current benching shows ~28 tok/s decode and ~481 tok/s prefill.
It's a start, but I expected better speeds even on a non ideal setup. I'll keep experimenting!
OK, but please come back and let us know your results!