I want to run a 70B LLM locally with more than 1 T/s. I have a 3090 with 24GB VRAM and 64GB RAM on the system.

What I managed so far:

  • Found instructions to make 70B run on VRAM only with a 2.5B that run fast but the perplexity was unbearable. LLM was barely coherent.
  • I randomly made somehow 70B run with a variation of RAM/VRAM offloading but it run with 0.1 T/S

I saw people claiming reasonable T/s speeds. Sine I am a newbie, I barely can speak the domain language, and most instructions I found assume implicit knowledge I don’t have*.

I need explicit instructions on what 70B model to download exactly, which Model loader to use and how to set parameters that are salient in the context.

  • mrjackspade@alien.topBEnglish
    1·
    1 year ago

    If you’re only getting 0.1 then you’ve probably overshot your layer offloading.

    I can get up to 1.5 t/s with a 3090, at 5_K_M

    Try running Llama.cpp from the command line with 30 layers offloaded to the gpu, and make sure your thread count is set to match your (physical) CPU core count

    The other problem you’re likely running into is that 64gb of RAM is cutting it pretty close. Make sure your base OS usage is below 8GB if possible and try memory locking the model on load. The problem is that with that amount of system ram, its possible you have other applications running causing the OS to page the model data out to disk, which kills performance

      • BlueMetaMind@alien.topOPBEnglish
        1·
        1 year ago

        Yeah… I thought I’ll be at least “in the room “ buying my setup last year, but it turns out I’m outside in the gutter 🫣😢

    • BlueMetaMind@alien.topOPBEnglish
      1·
      1 year ago

      Thank you. What does " at 5_K_M" mean ?
      Can I use the text web UI with Llama.cpp as model loader or is this too much overhead for ?

      • mrjackspade@alien.topBEnglish
        1·
        1 year ago

        I actually don’t know how much overhead that’s going to be. I’d start by just kicking it off on the command line first as a proof of concept, its super easy,

        5_K_M is just the quantization I use. There’s almost no loss of perplexity with 5_K_M, but its also larger than 4 which is what most people use.

        Name Quant method Bits Size Max RAM required Use case
        goat-70b-storytelling.Q2_K.gguf Q2_K 2 29.28 GB 31.78 GB smallest, significant quality loss - not recommended for most purposes
        goat-70b-storytelling.Q3_K_S.gguf Q3_K_S 3 29.92 GB 32.42 GB very small, high quality loss
        goat-70b-storytelling.Q3_K_M.gguf Q3_K_M 3 33.19 GB 35.69 GB very small, high quality loss
        goat-70b-storytelling.Q3_K_L.gguf Q3_K_L 3 36.15 GB 38.65 GB small, substantial quality loss
        goat-70b-storytelling.Q4_0.gguf Q4_0 4 38.87 GB 41.37 GB legacy; small, very high quality loss - prefer using Q3_K_M
        goat-70b-storytelling.Q4_K_S.gguf Q4_K_S 4 39.07 GB 41.57 GB small, greater quality loss
        goat-70b-storytelling.Q4_K_M.gguf Q4_K_M 4 41.42 GB 43.92 GB medium, balanced quality - recommended
        goat-70b-storytelling.Q5_0.gguf Q5_0 5 47.46 GB 49.96 GB legacy; medium, balanced quality - prefer using Q4_K_M
        goat-70b-storytelling.Q5_K_S.gguf Q5_K_S 5 47.46 GB 49.96 GB large, low quality loss - recommended
        goat-70b-storytelling.Q5_K_M.gguf Q5_K_M 5 48.75 GB 51.25 GB large, very low quality loss - recommended
        goat-70b-storytelling.Q6_K.gguf Q6_K 6 56.59 GB 59.09 GB very large, extremely low quality loss
        goat-70b-storytelling.Q8_0.gguf Q8_0 8 73.29 GB 75.79 GB very large, extremely low quality loss - not recommended
  • TuuNo_@alien.topBEnglish
    1·
    1 year ago

    I would suggest you to use Koboldcpp and run GGUF. A 70B Q5 model, with around 40 layers loaded into GPU, should have more than 1t/s. At least for me, I got 1.5t/s with 4090 and 64GB ram using Q5_K_M.

    • silenceimpaired@alien.topBEnglish
      1·
      1 year ago

      I could never get up and running on Linux with Nvidia. I used Kobold on Windows, but boy is it painful on Linux.

      • TuuNo_@alien.topBEnglish
        1·
        1 year ago

        Well, I have never used Linux before since the main purpose of my pc is gaming. But I heard running LLMs on Linux is overall faster.

      • giblesnot@alien.topBEnglish
        1·
        1 year ago

        I don’t know what you were running into but I’m running Pop_OS 22.04 (a modified version of Ubuntu,) as my OS with a 3090 and everything I have tried I just follow the basic install instructions on the home page and it works. Ooga booga, Automatic1111, Tortoise TTS, Whisper STT, Bark, Kobald, etc. I just follow the “run these commands” linux instructions and everything is groovy.