tracel-ai / burn

项目摘要

Burn is a Rust-based deep learning framework and tensor library that unifies training and inference under a single API. It offers PyTorch-like ergonomics with JIT compilation, automatic kernel fusion, and cross-platform support including CUDA, Metal, Vulkan, WebGPU, and WASM.

为什么值得关注

Burn solves the brittle training-to-production export problem by allowing the same code to run everywhere, while achieving Python-like iteration speeds through incremental compilation in Rust. Its extensive backend support enables deployment across GPUs, mobile, and web without model conversion.

优势

• Unified training and inference codebase eliminates export friction
• Extensive cross-platform backend support (CUDA, Metal, ROCm, Vulkan, WebGPU, WASM)
• JIT compilation with automatic kernel fusion for performance
• Fast incremental compilation enabling sub-5-second research iteration
• PyTorch-like dynamic graph ergonomics with Rust's safety and speed

局限性

• Rust learning curve for Python-native ML researchers
• Younger ecosystem with fewer pre-trained models and community resources than PyTorch
• Relatively new project with evolving API stability

使用场景

• On-device ML and edge deployment without model conversion
• Web-based ML inference via WASM and WebGPU
• Research prototyping with production-ready code paths
• Cross-platform applications requiring GPU acceleration on diverse hardware