GitHub - r2d4/react-llm: Easy-to-use headless React Hooks to run LLMs in the browser with WebGPU. Just useLLM().

3 min read Original article ↗

@react-llm/headless

Easy-to-use headless React Hooks to run LLMs in the browser with WebGPU. As simple as useLLM().

Live Demo

image

Features:

  • Supports Vicuna 7B
  • Use custom system prompts and "user:"/"assistant:" role names
  • Completion options like max tokens and stop sequences
  • No data leaves the browser. Accelerated via WebGPU.
  • Hooks built to 'Bring your own UI'
  • Persistent storage for conversations in browser storage. Hooks for loading and saving conversations.
  • Model caching for faster subsequent loads

Installation

npm install @react-llm/headless

Packages in this repository

useLLM API

Types

// Model Initialization
init: () => void;

// Model Generation
send: (msg: string, maxTokens: number, stopSequences: string[]) => void;
onMessage: (msg: GenerateTextResponse) => void;
setOnMessage: (cb: (msg: GenerateTextResponse) => void) => void;

// Model Status
loadingStatus: InitProgressReport;
isGenerating: boolean;
gpuDevice: GPUDeviceInfo;

// Model Configuration
userRoleName: string;
setUserRoleName: (roleName: string) => void;
assistantRoleName: string;
setAssistantRoleName: (roleName: string) => void;

// Conversation Management
conversation: Conversation | undefined;
allConversations: Conversation[] | undefined;
createConversation: (title?: string, prompt?: string) => void;
setConversationId: (conversationId: string) => void;
deleteConversation: (conversationId: string) => void;
deleteAllConversations: () => void;
deleteMessages: () => void;
setConversationTitle: (conversationId: string, title: string) => void;

Hooks

import useLLM from '@react-llm/headless';

const MyComponent = () => {
  const {
    conversation,
    allConversations,
    loadingStatus,
    isGenerating,
    createConversation,
    setConversationId,
    deleteConversation,
    deleteAllConversations,
    deleteMessages,
    setConversationTitle,
    onMessage,
    setOnMessage,
    userRoleName,
    setUserRoleName,
    assistantRoleName,
    setAssistantRoleName,
    gpuDevice,
    send,
    init,
  } = useLLM();

  // Component logic...

  return null;
};

Provider

import { ModelProvider } from "@react-llm/headless";

export default function Home() {
  return (
    <ModelProvider
      config={{
        kvConfig: {
          numLayers: 64,
          shape: [32, 32, 128],
          dtype: 'float32',
        },
        wasmUrl: 'https://your-custom-url.com/model.wasm',
        cacheUrl: 'https://your-custom-url.com/cache/',
        tokenizerUrl: 'https://your-custom-url.com/tokenizer.model',
        sentencePieceJsUrl: 'https://your-custom-url.com/sentencepiece.js',
        tvmRuntimeJsUrl: 'https://your-custom-url.com/tvmjs_runtime.wasi.js',
        maxWindowSize: 2048,
        persistToLocalStorage: true,
      }}
    >
      <Chat />
    </ModelProvider>
  );
}

Packages

  • @react-llm/headless - Headless React Hooks for running LLMs in the browser
  • @react-llm/retro-ui - Retro-themed UI for the hooks

How does it work?

This library is a set of React Hooks that provide a simple interface to run LLMs in the browser. It uses Vicuna 13B.

  • SentencePiece tokenizer (compiled for the browser via Emscripten)
  • Vicuna 7B (transformed to Apache TVM format)
  • Apache TVM and MLC Relax (compiled for the browser via Emscripten)
  • Off-the-main-thread WebWorker to run the model (bundled with the library)

The model, tokenizer, and TVM runtime are loaded from a CDN (huggingface). The model is cached in browser storage for faster subsequent loads.

Example

See packages/retro-ui for the full demo code. This is a simple example of how to use the hooks. To run it, after cloning the repo,

cd packages/retro-ui
pnpm install
pnpm dev

License

MIT

The code under packages/headless/worker/lib/tvm is licensed under Apache 2.0.