Burn

lastmod 2025-11-04created 2025-11-042 edits

A next-generation Tensor library combining flexibility, efficiency, and portability

About

A Rust-based high-performance tensor library and deep learning framework for numerical computing, model inference, and training
Achieves dynamic flexibility while keeping optimization performance on par with static graph frameworks
Supports multiple GPU/CPU backends
- GPU: CUDA, ROCm, Metal, Vulkan, WebGPU, LibTorch, Candle, etc.
- CPU: based on CubeCL, NdArray, Candle, LibTorch
- Also runs in Wasm and no_std environments
Designed around the Backend trait, with a backend structure that can be swapped across various hardware and runtime environments
- Autodiff backend decorator: adds automatic differentiation to any backend
- Fusion backend decorator: kernel fusion
- Router backend: supports distributing operations across multiple hardware such as CPU and GPU
- Remote backend: enables distributed computation and remote execution over the network
Integrated support for the entire training and inference process
- Real-time training monitoring with a Ratatui-based terminal dashboard
- Deployable with the same code from embedded devices to large GPU clusters
- Supports reusing existing models through PyTorch, Safetensors, and ONNX model import
- In-browser inference based on WebAssembly and WebGPU
no_std support allows running even on embedded devices without an operating system
Distributed under a dual MIT/Apache 2.0 license

author Zer0 sitemap