Now this looks much more interesting! Is the top one input tokens and the second one output tokens?

So 38.54 t/s on 120B? Have you tested filling the context too?