That is impressive. Interesting that the prefill (i am guessing this is prompt processing) is so much slower than decoding.

Its my understanding that under normal circumstances decoding is memory bandwidth bound which prompt processing isn't due to batching. Is there some quirk in your setup?