7
[help] has anyone managed to run rocm on the deck? (feddit.it)
submitted 2 years ago by to c/steamdeck@sopuli.xyz
7 comments fedilink hide all child comments
 

I've been trying to install rocm on the deck following this guide

(https://old.reddit.com/r/SteamDeckTricks/comments/102xxww/guide_how_to_install_rocm_for_gpu_julia/)

with the only exception that i used a generic ubuntu container on podman instead of using distrobox; I reach to the point where i need to use amdgpu-install to run rocm, but when i do it tries to install 30gb of files; which unfortunately it's not feasible on my metered connection and storage (╥﹏╥). From what i saw from this issue on github there should be a way to select only the needed binaries for the specific chipset;

The question is, what should i do to install only the stuff necessary for the deck?

Sorry for the long post.

all 7 comments

sorted by: hot top controversial new old
[–] 2 points 2 years ago

I had it running for image generation. As far as I recall it was slow and buggy so I didn't do more with it.

  • permalink
  • fedilink
  • source
    •
    [–] 1 point 2 years ago

    Try Vulkan in Koboldcpp https://github.com/LostRuins/koboldcpp/releases/tag/v1.81.1

    I haven't heard of anyone trying it on deck but I don't see why it wouldn't work.

  • permalink
  • fedilink
  • source
    •
    [–] 1 point 2 years ago* (last edited 2 years ago) (4 children)

    I probably can't help, I have ROCM acceleration working on my desktop linux PC for image gen and LLMs, but I don't think I had to install the massive rocm files for it to work.

    What specifically are you trying to run using ROCM? If I remember right most of the stuff I tried just needed pytorch, which I would install in a python venv with pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.3 That's still pretty heft though, my current venv is 16GB in size, but may include some duplicate/old files from repeated upgrades without cleaning.

  • permalink
  • fedilink
  • source
  • hideshow 4 child comments
    •
  • [–] [S] 2 points 2 years ago (3 children)

    I'm mostly trying to run text-generation-webui for fun; what did you install to run image generation if you never installed the rocm binaries? Hopefully i should be able to replicate it in a container

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 3 child comments
    •
  • [–] 2 points 2 years ago (2 children)

    Pytorch includes the necessary ROCm files for image gen as far as I know. My computer does have an /opt/ROCm folder but it's only a few megabytes in size.

    Unfortunately I don't see any mention of pytorch in your linked program, and it does mention a proper ROCm install as a requirement there.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] [S] 1 point 2 years ago (1 child)

    If that's the case i'm going to try to run it without installing rocm, since i saw a pytorch rocm package being downloaded by pip. I'm gonna need to reinstall though.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 1 point 2 years ago

    There are different pytorch install commands for cuda/ROCm. Usually the key to getting ROCm to work with AI GitHub projects is to edit a config file/launch option and replace the pytorch command with a ROCm one

  • permalink
  • fedilink
  • source
  • parent