Burn is a next-generation deep learning framework built in Rust, prioritizing flexibility, efficiency, and portability. It targets researchers and engineers seeking high performance across diverse hardware without the typical friction of Python-based frameworks, offering automatic optimizations and a robust backend system.

How It Works

Burn leverages Rust's ownership system for thread-safe building blocks and employs automatic kernel fusion to optimize computations dynamically, creating custom kernels that rival handcrafted GPU implementations. It features an asynchronous execution model for improved responsiveness and framework overhead reduction. Intelligent memory management, including memory pooling and in-place mutation tracking, further enhances efficiency.

Quick Start & Requirements

• Install via cargo add burn.
• Requires Rust toolchain.
• For GPU acceleration, specific backend dependencies (CUDA, Metal, Vulkan, WGPU) are needed.
• See The Burn Book for detailed guidance.

Highlighted Details

• Automatic Kernel Fusion: Dynamically generates optimized kernels (e.g., WGSL for WGPU) for custom operations.
• Backend Agnostic Design: Supports multiple backends (CUDA, ROCm, Metal, Vulkan, WGPU, NdArray, LibTorch, Candle) via a generic Backend trait.
• Decorator Pattern: Backends can be enhanced with features like Autodiff and Fusion through composition.
• no_std Support: Core components are compatible with bare-metal embedded environments (currently via NdArray backend).

Maintenance & Community

• Active development with a Discord community (https://discord.gg/uPEBbYYDB6).
• Contribution guidelines and architecture overview are available.

Licensing & Compatibility

• Licensed under MIT and Apache 2.0.
• Compatible with commercial and closed-source applications.

Limitations & Caveats

• The project is in active development, with potential for breaking changes.
• ONNX support is limited to a subset of operators.
• WGPU backends may require increasing the Rust recursion limit (#![recursion_limit = "256"]).
• Loading model records from versions older than 0.14.0 requires the record-backward-compat feature flag.