So what is currently the best and easiest way to use an AMD GPU for reference I own a rx6700xt and wanted to run 13B model maybe superhot but I’m not sure if my vram is enough for that Since now I always sticked with llamacpp since it’s quiet easy to setup Does anyone have any suggestion?

  • Mechanize
    link
    fedilink
    English
    arrow-up
    3
    ·
    1 year ago

    I’ve a rx 6650 xt and I generally use llama.cpp with the ROCm patch (tested up to commit ac7876ac20124a15a44fd6317721ff1aa2538806).

    It works great with around 25 layers moved to the GPU for my 8GB card. 18, if you want to do something else GPU related (like watching a HW Accelerated video).

    To be fair it’s a long time now that I don’t update llama.cpp and it had gone through a lot of changes in the meantime, like the addition of the LLAMA_CUDA_DMMV_X, LLAMA_CUDA_DMMV_Y and LLAMA_CUDA_KQUANTS_ITER parameters, so your mileage may vary and it’s possible you’ll have to manually modify the PR before merging it in, so not really an easy one click experience for the best performance.

    It currently doesn’t support SuperHot or similar techniques, mainly because there’s a really big push on new ones each day, and they are waiting to see which will be the real winner.

    But I went a bit too much off-topic. I think the easiest, as the other commenter said, is to just go with kobold.cpp. I personally didn’t have a good experience working with text-generation-webui, but a lot of people swear by it.

    • Mixel@feddit.deOP
      link
      fedilink
      English
      arrow-up
      2
      ·
      1 year ago

      Yes thank you for the information I really appreciate it! I decided to go for kobold.cpp for the meantime with CLBlast which works just overall way better than standart CPU inference. But im looking into the ROCm LLamacpp support which I am currently trying.