Prefill is 400 t/s in that hardware. Just if the prompt is very short you can't see the real speed and it will default to single token context processing.