speculative decoding in llama.cpp

[vectorport]

same here on my side.

That matches what I was seeing. That matches what I saw on GB200. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.