Using rusage

The rusage is one handy tool for debugging LLMs!

You prefix any command with it and you get memory usage, page faults, I/O info.

For example, below shows what happens when you load model using llama.cpp. Notice page faults with 100% memcpy which means they didn’t hit disk!