burn

About burn

Burn is an open-source Rust tensor library and deep learning framework for numerical computing, training and inference. It aims to unify the two halves of a typical workflow: rather than training in one language and exporting to a separate format or runtime, the same tensor code used for training is intended to run in production, which the project says helps with workloads such as on-device personalization and federated learning. It keeps dynamic shapes and graphs while JIT-compiling streams of tensor operations with automatic kernel fusion.

Burn ships with an ecosystem of companion crates, including ONNX model import, weight storage and loading of PyTorch and Safetensors weights, vision and reinforcement learning building blocks, dataset loading, and backend benchmarking. Acceleration is provided through CubeCL backends for CUDA, ROCm, Metal, Vulkan, WebGPU and CPU, which compose with autodiff, fusion and remote-execution decorators; the pure-Rust Flex CPU/no_std backend composes with autodiff only. The repository is licensed under Apache and MIT terms. Pricing is not stated.

Key features

• Combined tensor library and deep learning framework for training and inference
• Same code path for training and production, avoiding model export steps
• JIT compilation of tensor operation streams with automatic kernel fusion
• CubeCL backends for CUDA, ROCm, Metal, Vulkan, WebGPU and CPU
• Incremental compilation said to rebuild modified model code in under 5 seconds
• Companion crates for ONNX import, weight storage, vision, RL and datasets

burn at a glance