[cjbprime]

Wouldn't expect that to work at all.

[hedgehog]

Ollama (which wraps llama.cpp) supports splitting a model across devices so you get some acceleration even on models too big to fit entirely in GPU memory.