Good point, that should theoretically be double the decode speed, though not sure how the prefill would differ since it’s mainly compute-bound.