10-16-2026, 01:47 AM
Interesting, I hadn't tried that. 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.
measure twice, benchmark three times