burn

Why we included this project

Burn is a tensor library and deep learning framework written in Rust. The idea is that the same model code you train with runs in production as-is, without an export step. Tensor operations are JIT-compiled and kernel-fused across backends, so you get PyTorch-like flexibility with dynamic shapes and the performance of a static graph. That matters for teams shipping on-device or edge workloads, where the hardware varies and you want one codebase that runs on CUDA, ROCm, Metal, Vulkan, WebGPU, or plain CPU. Rust shops also get to keep their whole ML stack in one language. The ecosystem around it, including ONNX import, weight loading from PyTorch and Safetensors, and crates for vision and reinforcement learning, makes it a practical option rather than a research prototype.