burn Releases

A major performance release enhancing GPU utilization through new tensor transactions and mixed precision support, alongside expanded backend capabilities.

Jan 14, 2025

This release introduces support for TimesFM-2.0 models and sets the median head as the default point forecaster. It includes hyperparameter support and fixes for PyTorch decoding, while noting that cached decoding is currently unsupported.

Dec 31, 2024

Updates to the Auto Jobs Applier AI Agent focusing on better job handling, regex blacklisting, and LLM parsing improvements.

Nov 15, 2024

This release brings significant performance improvements to tensor operations and introduces multi-backend compatibility with ROCm/HIP and SPIR-V support.

Oct 28, 2024

This release introduces PyTorch support for TimesFM inference, allowing users to install dependencies separately for the Pax (JAX) or Torch versions. It also provides a PyTorch checkpoint for the 200m model on Hugging Face.

Oct 18, 2024

The Local File Organizer introduces new features for safer file management and expanded format support.

Oct 1, 2024

This release introduces the CubeCL integration for cross-platform GPU programming (CUDA and WebGPU), a new CUDA backend, and a new tensor data format for faster serialization. It also adds Quantization features and significant enhancements to ONNX support.

Aug 27, 2024

A bugfix release focused on improving autodiff graph memory management performance and fixing matmul broadcasting issues in the ndarray backend.

May 3, 2024

This release addresses autodiff memory leaks, improves inplace fused operations, and upgrades the WGPU version to 0.19.4.

Apr 26, 2024

A major update removing the `Sync` trait from most Burn types, introducing a backend bridge mechanism, refactoring the Wgpu backend into a Just-in-Time backend, and adding extensive new tensor operations and optimizers.

Apr 12, 2024

A bugfix release that reverts the WGPU version to fix performance issues and addresses problems with LibTorch batch norm and documentation.

Feb 1, 2024

Highlights include an optimized Wgpu backend, breaking changes in device management requiring explicit specification, and the addition of a PyTorch recorder for weight import.

Jan 31, 2024

Originally intended as version 1.0.0, this release introduces configuration file support for rga, switches the cache database to SQLite, and refactors the codebase to be fully asynchronous.

Jan 16, 2024

Jan 16, 2024

A bugfix release addressing concurrency issues in burn-fusion and fixing documentation typos and build errors.

Dec 4, 2023

The main feature is the introduction of automatic kernel fusion, alongside a reorganization of the organization name and backend names.

Dec 1, 2023

This release introduces the `burn-compute` crate, a new training dashboard, and WebGPU support in the browser, accompanied by several breaking changes to the tensor API.

Oct 24, 2023

Launch of the Burn Book and Model repository, introduction of three new optimizers, and numerous new tensor operations and training metrics.

Sep 6, 2023

Dec 30, 2022

Nov 20, 2022