The latest release of microsoft/onnxruntime is v1.31.0. This repository is a runtime that accelerates inference and training of machine learning models across multiple platforms. It supports models from PyTorch, TensorFlow/Keras, scikit-learn, LightGBM, XGBoost, and others. What changed in this release is not stated in the provided information.
The project focuses on enabling model inference regardless of model format or execution hardware. It improves runtime performance through graph optimization and hardware acceleration, and also supports model conversion. Mobile builds and execution providers allow ONNX models to be run on-device. The README also introduces a way to accelerate transformer training on multi-node NVIDIA GPUs by adding a single line to an existing PyTorch training script. The implementation language is C++, and the license is MIT.
That said, the provided records present v1.31.0 only as the latest release and contain no change list or performance measurements for this version. What can be confirmed is the cross-platform inference and training capability for models from multiple frameworks, plus the mobile execution path. Whether this version improves on previous ones cannot be judged from this information alone. Release notes are needed to evaluate the specific changes.