Burn a Deep Learning Framework with flexibility, efficiency and portability (github.com)

Burn is a new Rust-based tensor library and deep learning framework that aims to combine the performance of static-graph systems with the flexibility of dynamic frameworks. It emphasizes portability — supporting many GPU backends (CUDA, ROCm, Metal, Vulkan, WebGPU, Candle, LibTorch) and CPU targets (NdArray, Candle, LibTorch, CubeCL), plus WebAssembly and embedded no_std environments — so you can train in the cloud and deploy on user devices without rewriting code. Burn also provides a terminal training dashboard, benchmarking tools, ONNX import (as Rust code), and direct weight loading from PyTorch/Safetensors for easier migration. Technically, Burn is built around a generic Backend trait and a set of composable backend decorators that add features without changing core code. Key decorators include Autodiff (wraps any base backend to enable safe backprop — backward is only exposed on autodiff-wrapped backends), Fusion (automatic kernel fusion for accelerated backends, enabled by feature flag), Router (beta: composes multiple backends/devices for mixed CPU/GPU execution), and Remote (beta: client/server remote tensor execution for distributed compute). The design enables backend swapping, kernel fusion + planned automatic gradient checkpointing for memory/perf trade-offs, and browser or embedded inference — all leveraging Rust optimizations to deliver efficient, portable ML workflows.