Paddle Releases

A major release featuring a new dynamic graph architecture, dynamic-static unification, and a suite of sparse computing APIs. It also introduced a large-scale graph neural network GPU training engine and updated environment support for CUDA 11.7 and Ubuntu 22.04.

Nov 28, 2022

This release focused on bug fixes including profiler improvements, CUDA 11.7 compatibility, and PHI C++ API issues. Deployment features included NVCC_LAZY loading and expanded ONNX Runtime backend support.

Aug 16, 2022

Fixed issues and released CUDA 11.6 support. Enhanced quantization-aware training and DataLoader functionality while optimizing performance for fused attention and linspace operators.

Jul 12, 2022

Released the high-reusability operator library (PHI) and new dynamic/static graph executors. It introduced adaptive distributed training, mixed parallel strategies, and support for heterogeneous hardware like Ascend910 and GraphCore IPU.

May 13, 2022

Release candidate for v2.3.0, introducing the PHI operator library and adaptive distributed training architecture.

Apr 30, 2022

Focused on bug fixes and API enhancements. Added Mish activation function support and improved PReLU and ONNX Runtime capabilities.

Jan 26, 2022

Added experimental CUDA Graph support and new graph learning APIs. Enhanced heterogeneous parameter server and optimized performance for top_k and LarsMomentumOptimizer.

Dec 8, 2021

Major release with 100+ new APIs covering linear algebra and Fourier transforms. Introduced `paddle.einsum`, advanced indexing types, and Kernel Primitive API. Enhanced dynamic-static conversion and distributed training (Hybrid parallel, GPU Parameter Server).

Nov 10, 2021

Mar 19, 2019

Oct 24, 2018