• Asafum@lemmy.world
    link
    fedilink
    English
    arrow-up
    8
    arrow-down
    1
    ·
    27 days ago

    Yeah I’m just beginning my local AI journey on a 5080, tried Qwen3.6 27b Q4 and was getting like 1tps because of the vram overflow. Ran it over night at it was still chewing on generating a prompt for a sub agent when I got up in the middle of the night until it simply ended in some kind of “fetch failure” lol. I think I gave it something too large to tackle, but either way 1tps is kinda garbage.

    • RedstoneValley@sh.itjust.works
      link
      fedilink
      English
      arrow-up
      13
      ·
      27 days ago

      You could use the 35B MoE model, tune it a little bit and get much better results. I have a 5060 ti and 70-80 tok/s are the norm

      • Asafum@lemmy.world
        link
        fedilink
        English
        arrow-up
        2
        ·
        27 days ago

        That’s what I generally use. I wanted to see if I could use the 27b to “review” what the 35b put out. The 35b has been working pretty well, but it’s not very thorough. I asked it to make a program and then 27b was like “this is a skeleton, there are folders but no contents.” Lol

        • RedstoneValley@sh.itjust.works
          link
          fedilink
          English
          arrow-up
          5
          ·
          27 days ago

          These models generally are not capable enough to do one-shot vibe coding. They are pretty good as coding assistants if you tell them exactly what you want and let them focus on a specific aspect/part of code, not the whole thing at once.

          Using an agent framework (I like Kilo on VSCode but there are many others) you can start with a planning session to let the model find out what you want to build. Then you let it write that gist into AGENTS.md and double check if that is what you want. AGENTS.md will be loaded into the context automatically so the model has a solid base of understanding for everything you do afterwards. Once you have that, building in vertical slices on top of that skeleton is much easier. Another neat trick is to ask the model a few questions about the current code base (if there is any…) at the beginning of a session, e.g. "How does feature x work in the current code? ". This primes the model for what you are about to do. All this is obviously a bit more work than just vibe coding away, but it lets you keep in control of the code and helps in being alert for errors these models (and all LLMs in general) will inevitably produce.

          • Asafum@lemmy.world
            link
            fedilink
            English
            arrow-up
            4
            ·
            26 days ago

            Thanks for the tips! What I had started doing the other day was having one session where it reviewed the code and created a document explaining what the program did in technical detail and then in another session I asked it to review that document before attempting anything with the program. Agent programs and harnesses are the next thing I need to start learning for sure.

      • Asafum@lemmy.world
        link
        fedilink
        English
        arrow-up
        3
        ·
        27 days ago

        It’s the q4 quantization, but it requires 20+GB vram and my 5080 only has 16

        • Brett@programming.dev
          link
          fedilink
          English
          arrow-up
          1
          ·
          4 hours ago

          What are you using to run the model? Llama.cpp will automatically split the model between your system ram and graphics card’s vram.

          Qwen 3.6 is a mixture of experts model with only 3B parameters active at a time. Even without quantization your card could easily run that.

          • Asafum@lemmy.world
            link
            fedilink
            English
            arrow-up
            1
            ·
            3 hours ago

            It has been some time since my initial comment so at the time I was mainly using LM studio. Qwen 3.6 a3b is the MOE and it does work well on my card, but the dense model that is more intelligent/capable is the Qwen 3.6 27b which doesn’t fit on the card and does get offloaded, but offloading cuts the speed down to like 1/tps.

            I have since found a version of the 27b model that is “quantized,” for lack of a better term, differently and has to be run through TabbyAPI which gets back to 30ish tps. It can’t offload so it must fit fully on the card which keeps the speed high. Might be worth a look if you’re interested, the only downside is that with my 16gb card the context limit has to be kept pretty low ~40k if I remember correctly

    • scarabic@lemmy.world
      link
      fedilink
      English
      arrow-up
      3
      ·
      27 days ago

      Wow I’m starting to feel bad about that time I asked AI to make a joke about scatology & eschatology sounding similar.