tracel-ai / burn

项目摘要

Burn is a next-generation tensor library and deep learning framework written in Rust that unifies training and inference in a single codebase. It provides PyTorch-like ergonomics with dynamic shapes and graphs while JIT-compiling tensor operations with automatic kernel fusion for high performance across multiple platforms including CUDA, Metal, ROCm, Vulkan, WebGPU, and WASM.

为什么值得关注

Burn eliminates the traditional gap between training and inference by allowing the exact same code to run in both phases, enabling novel workloads like on-device personalization and federated learning. Its Rust foundation delivers memory safety and performance, while incremental compilation provides a Python-like feedback loop with sub-5-second recompile times, making it a compelling alternative to Python-based frameworks.

优势

• Unified training and inference in a single codebase eliminates brittle export steps
• Multi-platform backend support (CUDA, Metal, ROCm, Vulkan, WebGPU, WASM) for true portability
• JIT compilation with automatic kernel fusion delivers high performance without sacrificing flexibility
• Rust's memory safety and incremental compilation provide fast edit-compile-run loops

局限性

• Rust ecosystem for ML is still maturing compared to Python's vast library support
• Relatively new project (2022) with a smaller community and fewer pre-trained models available

使用场景

• Training and deploying models in a single codebase without ONNX export
• On-device inference and personalization for mobile and edge devices
• Cross-platform deployment targeting diverse hardware (GPU, WASM, WebGPU)
• Federated learning scenarios where training happens on distributed devices