FreakyAdy/phantom ? reverse-engineered prompt

Build me a local LLM runner that can work on a regular laptop with limited GPU memory, even when the model is too big to fit fully in VRAM.

I want it to be able to load a model, split work between GPU memory, system RAM, and disk, then generate text without crashing or freezing the machine. It should have a simple command line experience for running a prompt, benchmarking speed, and checking that results are correct. Use the llama.cpp based path as the main backend, and keep the more advanced acceleration ideas as optional extras if they are ready.

Please make it practical and honest, with clear logs, safe defaults, and benchmark outputs I can trust. If you need current docs for any library or GPU related behavior, look them up online first.