Burn 0.22.0 Overhauls Backend Architecture and Accelerates Model Training
The latest release of the Burn deep learning framework introduces a pure Rust backend, significant performance optimizations, and CubeCL integration to streamline hardware acceleration ahead of its 1.0 milestone.
Burn 0.22.0 lands with a complete overhaul of its API and backend architecture, setting up the framework for version 1.0.
The update introduces generic-free APIs, cutting out backend generics from model definitions. Internal benchmarks show CNN rebuild times dropped by 6.2x and Transformers by 14.7x because the compiler no longer re-processes dependency chains after a model edit.
The underlying infrastructure is now pure Rust. The team replaced MLIR and C/C++ transpilers with internal tools like Pliron.
Version 0.22.0 also shifts hardware acceleration to the CubeCL compute environment. This supports CUDA, ROCm, Metal, Vulkan, WebGPU, and CPUs. At the same time, the project deprecated the ndarray, LibTorch, and Candle backends. The new consolidation enables direct access to memory pool usage and reporting. Paired with adaptive pools, these memory changes cut peak training memory usage by up to 49% for certain projects.
Other additions in 0.22.0 include ONNX export, remote compute capabilities, and support for LoRA and QLoRA fine-tuning. Moving forward, the team will focus on upstreaming fine-grained profiling tools, deeper CubeCL integration, and interface stabilization before the 1.0 milestone.