When I try to turn off Use CPU in settings, it says “No compatible graphics card found!”.

During the instillation, I got an error “hipErrorNoBinaryForGpu” which I looked up and found the command export HSA_OVERRIDE_GFX_VERSION=10.3.0 which got me through the instillation.

I don’t know much about GPUs, so thanks in advance for any help/advice!

  • Hiro8811@lemmy.world
    6·
    10 months ago

    Well I don’t have a Steam deck but I know it uses an APU. Which is a GPU and CPU on the same die , what every laptop has. It might that the app doesn’t recognise integrated GPUs to be powerful enough so it doesn’t use them?

  • Stampela@startrek.websiteEnglish
    4·
    10 months ago

    My understanding (potentially obsolete!) is that the integrated gpu of the Deck isn’t supported by, uh… whatever the AMD equivalent of Cuda was called, also there’s a chance you’d have to install it manually too.

  • Sharp312
    3·
    10 months ago

    You need amds proprietary ROCM driver which supports stuff like ai. However i already had issues setting up anything ai on my desktop fedora system, steam deck uses an immutable fs and afaik it gets reset after every update. If doing it on a desktop pc is hard then doing it on the deck will be like walking on lego. Good luck soldier.

    • s12@sopuli.xyzOP
      2·
      10 months ago

      Thank you so much for the information!
      Would other distros have different difficulty levels?

      Requiring proprietary drivers is somewhat unfortunate.

      • Sharp312
        2·
        10 months ago

        From every diffusion github ive seen, they all seem to have instructions for ubuntu, so that may be your easiest bet. Same with rocm, it seems easier to install on ubuntu than other distros. Especially fedora…

  • Fubarberry@sopuli.xyzMEnglish
    3·
    10 months ago

    So unfortunately I don’t think ROCM supports the deck’s APU, and it’s too different for it to work by faking that it’s a different GPU.

    You can still do stable diffusion in CPU mode, but that will take 5-20 min per image (vs 7 seconds per image on my desktop’s 6600 XT).

    Overall it’s not worth it unfortunately. As an alternative you might look into running stable diffusion through Google colab, which is a good way to get custom generation going for free.