@rumbl/laserbeak

Transformers in the Browser

Laserbeak enables developers to run transformer models in the browser/Electron using WebGPU.

It is designed to efficiently manage models by caching them in IndexedDB and sharing weights between encoder-decoder models for optimal performance.

To see what it can do, check out our:

• AI powered document editor.
• Explore all models on the playground.

🌟 Features

• Run transformer models in the browser using WebGPU
• Built on top of a custom Rust runtime for performance
• Efficient model management with caching and weight sharing
• Easy-to-use API for loading and running models

⚡️ Quick start

Install Laserbeak using npm:

npm i laserbeak

📚 Usage

Here's a simple example of how to load and run a model using Laserbeak:

import {
    SessionManager,
    AvailableModels,
    InferenceSession,
} from "@rumbl/laserbeak";

//Create a SessionManager instance
let manager = new SessionManager();

//Load a model with a callback for when it's loaded
let modelSession = await manager.loadModel(AvailableModels.FLAN_T5_BASE, () =>
    console.log("Loaded successfully!")
);

// Run the model with a prompt and handle the output
await session.run(prompt, (output: string) => {
    // Process the model output
    console.log(output);
});

🚀 Roadmap

Laserbeak is still a pre_pre_alpha project, here's our roadmap:

• F16 support
• Shader optimizations
• Expanded model support (Whisper, UNet)
• Unannounced features 🤫
• INT8, INT4 support

Stay tuned for exciting updates!

💪 Contributing

We welcome contributions to Laserbeak!