AI with RWKV

go-rwkv.cpp

gowrkv.go is a wrapper around rwkv-cpp, which is an adaption of ggml.cpp.

Features

rkwv.cpp is generally faster, due to keeping the intermediate state of the model, so the entire prompt doesn't have to be reprocessed every time. For more details, see rwkv-cpp.

Also, the available models for rwkv.cpp are fully open-source, unlike llama. You can use these models commercially, and you can modify them to your heart's content.

Training may also be faster, I haven't had a chance to try that yet.

Installation

Installation is currently complex. go-rkwv.cpp does not work with go get yet (patches very welcome). You will need go, a c++ compiler(clang on Mac), and cmake.

Download

You must clone this repo /recursively/, as it contains submodules.

    git clone --recursive https://github.com/donomii/go-rwkv.cpp

Building

There is a build script, build.sh, which will build the c++ library and the go wrapper. Please file bug reports if it doesn't work for you.

    ./build-mac.sh

There is now an alternate build, which builds statically thanks to a makefile provided by @mudler.

    make example/ai

Download models

The download script will download some models, and convert them to the correct format.

    cd aimodels
    sh downloadconvert.sh

Install

go-rwkv.cpp currently builds against the dynamic library librwkv.dylib. This is not ideal, but it works for now. You will need to copy this library to a location where the system linker can find it. On Mac, this is /usr/local/lib.

    cp librwkv.dylib /usr/local/lib
    export DYLD_LIBRARY_PATH=/Users/donomii/git/go-rwkv.cpp/rwkv.cpp/

If you don't want to install it globally, you can set the DYLD_LIBRARY_PATH environment variable to the directory containing librwkv.dylib.

Use

See the example/ directory for a full working chat program. The following is a minimal example.

    package main

    import (
        "fmt"
        "github.com/donomii/go-rwkv.cpp"
    )

    func main() {
        model := LoadFiles("aimodels/small.bin", "rwkv.cpp/rwkv/20B_tokenizer.json", 8)
    model.ProcessInput("You are a chatbot that is very good at chatting.  blah blah blah")
    response := model.Generate(100, "\n")
    fmt.Println(response)

    }

You must use the tokenizer file from rwkv.cpp. go-rwkv contains a re-implementation of the tokenizer, but it is a minimal implementation that contains just enough code to work with rwkv (and there are probably bugs in it).

Packaging

To ship a working program that includes this AI, you will need to include the following files:

• librwkv.dylib
• the model file (e.g. RWKV-4-Raven-1B5-v9-Eng99%-Other1%-20230411-ctx4096_quant4.bin)
• the tokenizer file (i.e. 20B_tokenizer.json)

If you don't install librwkv.dylib globally, you will need to set the DYLD_LIBRARY_PATH environment variable to the directory containing librwkv.dylib.

License

This program is licensed under the MIT license. See LICENSE for details.

As far as I am aware, the Raven models are also open source.