Outsource the Learning process of an LLM from system to Metis

Not sure you can. The Metis M.2 is designed specifically for inference, rather than training. That part still needs to run on a CPU or GPU.

As for running large models like a 4.7GB LLM, the SDK handles this by splitting the model and streaming the parts needed for each stage of inference. So the full model doesn’t have to live entirely on the Metis — it just processes chunks as required.

What is it you’re working on?

• frist to setup a raspberry pi 5 16GB with a metis device
• second DeepFaceLab for aarch64
• third run a LLM within this specs of hardware

the metis device can handle threads and graphics … sounds good

so is must be possible to run DeepFaceLAB within this device on a raspberry pi 5

the metis device can also handle threads (run in 1G of memory) … well with this specs must be possible 

to do something with it