What’s new in PaddlePaddle 2.3.0‑rc0?
Over 100 new APIs – The release adds 100 + functions that span automatic differentiation, linear algebra, probability distributions, sparse‑tensor handling, performance profiling, hardware device management, and vision‑related utilities. (vendor claim)
PHI – a high‑reusability operator library – Paddle introduces PHI (Paddle HIgh reusability operator library), offering more than 200 C++ operator APIs that mirror the Python interface and roughly 500 composable forward/backward kernel functions. This aims to lower the cost of developing native and custom operators and to enable primitive‑API kernel reuse across GPUs, XPUs and other accelerators. (vendor claim)
Distributed training revamp – The adaptive distributed‑training stack is rebuilt with elastic scaling, asynchronous pipeline execution, heterogeneous communication, and automatic parallelism. Dynamic‑graph training now supports MoE parallelism, GroupSharded strategy, and pure FP16, improving large‑model efficiency. (vendor claim)
Inference and deployment enhancements – A Java API and an ONNX Runtime CPU backend are added. TensorRT 8.0/8.2 support, including structured sparsity, is optimized for ERNIE‑style models. (vendor claim)
New hardware adapters – Paddle now ships plugins for Huawei Ascend 910, GraphCore IPU, Cambricon MLU, and Kunlun 2 chips, plus a generic plugin mechanism for future accelerators. (vendor claim)
These updates position PaddlePaddle 2.3.0‑rc0 as a more extensible, performance‑focused framework for both research and production workloads.