In this case they're comparing against llama.cpp because the code is literally a modification of llama.cpp. I'm not talking about using the ggml lib for matrix

In this case they're comparing against llama.cpp because the code is literally a modification of llama.cpp. I'm not talking about using the ggml lib for matrix calculations, it's literally using the llama.cpp main.cpp and other normal llama.cpp code. It's a fork. It is directly comparable. https://github.com/ggerganov/llama.cpp/pull/4543 https://github.com/ggerganov/llama.cpp/pull/4543 [Review] Merge PowerInfer with llama.cpp mainline #4543 https://github.com/ggerganov/llama.cpp/discussions/4534#discussioncomment-7900305 https://github.com/ggerganov/llama.cpp/discussions/4534#disc... "The x11 speedup is kind of cherrypicked because the llama.cpp GPU code for Falcon 40b is just not well-optimized."