• Eager Eagle@lemmy.world
      link
      fedilink
      English
      arrow-up
      39
      ·
      14 hours ago

      I bet he just wants a card to self host models and not give companies his data, but the amount of vram is indeed ridiculous.

      • Jeena@piefed.jeena.net
        link
        fedilink
        English
        arrow-up
        19
        ·
        14 hours ago

        Exactly, I’m in the same situation now and the 8GB in those cheaper cards don’t even let you run a 13B model. I’m trying to research if I can run a 13B one on a 3060 with 12 GB.

        • manicdave
          link
          fedilink
          English
          arrow-up
          3
          ·
          9 hours ago

          I’m running deepseek-r1:14b on a 12GB rx6700. It just about fits in memory and is pretty fast.

          • levzzz@lemmy.world
            link
            fedilink
            English
            arrow-up
            2
            ·
            8 hours ago

            You need a pretty large context window to fit all the reasoning, ollama forces 2048 by default and more uses more memory

          • Viri4thus@feddit.org
            link
            fedilink
            English
            arrow-up
            1
            ·
            10 hours ago

            I also have a 3060, can you detail which framework (sglang, ollama, etc) you are using and how you got that speed? i’m having trouble reaching that level of performance. Thx

            • The Hobbyist@lemmy.zip
              link
              fedilink
              English
              arrow-up
              4
              ·
              edit-2
              2 hours ago

              Ollama, latest version. I have it setup with Open-WebUI (though that shouldn’t matter). The 14B is around 9GB, which easily fits in the 12GB.

              I’m repeating the 28 t/s from memory, but even if I’m wrong it’s easily above 20.

              Specifically, I’m running this model: https://ollama.com/library/deepseek-r1:14b-qwen-distill-q4_K_M

              Edit: I confirmed I do get 27.9 t/s, using default ollama settings.

              • Viri4thus@feddit.org
                link
                fedilink
                English
                arrow-up
                2
                ·
                2 hours ago

                Ty. I’ll try ollama with the Q-4-M quantization. I wouldn’t expect to see a difference between ollama and SGlang.

              • Jeena@piefed.jeena.net
                link
                fedilink
                English
                arrow-up
                2
                ·
                8 hours ago

                Thanks for the additional information, that helped me to decide to get the 3060 12G instead of the 4060 8G. They have almost the same price but from what I gather when it comes to my use cases the 3060 12G seems to fit better even though it is a generation older. The memory bus is wider and it has more VRAM. Both video editing and the smaller LLMs should be working well enough.