1 unstable release

0.1.0 Apr 29, 2023

#576 in Machine learning

LGPL-3.0-only

19KB
358 lines

kai-rs

An experimental Rust library for using large language models via KoboldAI

Installation

[dependencies]
# ...
kai-rs = { git = "https://github.com/Pyroglyph/kai-rs" }

⚠️ Important Note ⚠️

Unless you use my fork of KoboldAI, this library will only be able to load models in CPU-mode. If you wish you can still load models onto your GPU(s) through the official GUI without my fork.

I have sumbitted a pull request to have this fixed in the official repo.

Usage

Usage should be pretty straightforward. Here is the general idea:

  1. Make a KoboldClient
  2. Load a Model and set your GPU layers. Local models must begin with ./ for now. Automatically setting GPU layers is not implemented yet, so just use the value the KoboldAI GUI gives you, or set it to vec![0] to use your CPU.
  3. Populate GenerationSettings or use the reasonable defaults.
  4. Pass a prompt and the settings into generate and await the results!

Example:

use kai_rs::prelude::*;

let kai = KoboldClient::new("http://localhost:5000", APIVersion::V1);
kai.load_model(Model::from("PygmalionAI/pygmalion-6b"), vec![28]).await?;

let settings = GenerationSettings::default();

let prompt = "You: Hi. How are you?";
let response = kai.generate(prompt, settings).await?;

// response => [ "Me: I'm good. I just got back from the gym. What about you?" ]

Contributing

Any contributions are welcome, just open an issue or a PR!

Dependencies

~6–18MB
~275K SLoC