Has anyone tried to combine a server with a moderately powerful GPU with a server with a lot of RAM to run inference? Especially with llama. Cpp where you can offload just some of the layers to GPU?

https://github.com/Juice-Labs/Juice-Labs/wiki

  • Brave-Decision-1944@alien.top
    cake
    B
    link
    fedilink
    English
    arrow-up
    1
    ·
    1 year ago

    I seen something like that in LOLLMs UI, it’s called petal, and basically it bandwidth the processing along computers connected to that network. There was also other remote “binding” from same maker as the UI. But I didn’t tired those.