• joneskind@lemmy.world
    link
    fedilink
    English
    arrow-up
    2
    ·
    edit-2
    5 months ago

    Most of 7b-8b models run just fine in 4bits quant and won’t use more than 4 or 5 GB of VRAM.

    The only important metric is the amount of VRAM as the model must be loaded in VRAM for fast inference.

    You could use CPU and RAM but it is really painfully slow.

    If you got an Apple Silicon Mac it could be even simpler.

    • Venia Silente@lemm.ee
      link
      fedilink
      English
      arrow-up
      2
      ·
      5 months ago

      I have an Intel Celeron Mobile laptop with iGPU and, I think, 256MB VRAM. How many bs does that get me for the LLM?

      Only half-joking. That’s my still functional old daily driver now serving as homelab

      • joneskind@lemmy.world
        link
        fedilink
        English
        arrow-up
        2
        ·
        5 months ago

        Well, I got a good news and a bad news.

        The bad news is you won’t do shit with that my dear friend.

        The good news is that you won’t need it because the duck is back.