Currently I’m using the ollama runner for messing around with the mistral 7b models (only on CPU, I have no discrete gpu >.<) - I like that it has a very simple CLI and fairly minimal configuration (the Arch Linux package even comes with a systemd service, it’s pretty neat).

However, I don’t know how sustainable it is. It hosts a database of models on it’s own here, but I don’t know how dependent the code is on a central online repository.

Ideally, I’d love if we had an AI runner (including with the ability to use LoRA modules) that can natively pull from torrentfiles or something with similar p2p architecture. I imagine this would be better for long-term sustainability and hosting/download costs of the projects ^.^

Thoughts on this, and any other suggestions/comparisons/etc?

  • midnight@kbin.social
    link
    fedilink
    arrow-up
    5
    ·
    11 months ago

    I agree that some sort of decentralized model repository would be awesome, but ollama works with local files too, so I’m not too worried about it. I’ve used many LLM backends, and ollama is my favorite so far, but given how fast everything is moving, that could change in the future.

  • silas@programming.dev
    link
    fedilink
    English
    arrow-up
    3
    ·
    11 months ago

    I’ve thought the same thing actually, but I haven’t really looked into who’s behind Ollama or how the repository is managed yet. It’s a really great project

  • rufus@discuss.tchncs.de
    link
    fedilink
    English
    arrow-up
    2
    ·
    edit-2
    11 months ago

    What does ollama add on top of llama.cpp?

    I use KoboldCPP that also works very well on CPU.

    And Oobabooga’s UI (with llama.cpp as a CPU backend) is also easy to set up.