speculative decoding in llama.cpp: anyone testing this?

[mangotinker73]

That matches what I was seeing. I ended up in about the same place using llama.cpp. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.