benchmark result I don't trust: anyone testing this?

[nanobrook]

I'd approach it in a boring order. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at tail latency. Not glamorous advice, but repeatable baselines save a lot of time.