2 ms·

This is absolutely something I was running into with LLaMA. I'm curious if this potentially extends into that particular use case...