How to (very slowly) run high-precision R1 quant on M2 Ultra with most of model being on SSD?

[okuvshynov]

Setup:

• 192GB M2 Ultra
• higher-precision R1 quant from https://huggingface.co/unsloth/DeepSeek-R1-GGUF

What is the right way to configure llama.cpp (either cli or server) to do roughly the following:

• always keep KV cache in memory
• keep model on SSD
• ideally keep the routing layers (which would run on every token) in memory

It's fine if it it slow, I'm ok with 1-2 token / minute.

Thank you!