2 ms·

I tested this on the latest master. Llama.cpp has had some performance improvements, although I don't know if that'd be enough to make it 3x faster.