PaddlePaddle 2.3.0 introduces the PHI operator library, a significant architectural update designed to lower development costs and improve performance. The release also enhances distributed training capabilities and adds support for new hardware like Huawei Ascend 910 and GraphCore IPU.
PaddlePaddle 2.3.1 introduces official support for CUDA 11.6, significant upgrades to quantization workflows including AMP support, and performance improvements in attention and linear operators.
PaddlePaddle 2.4.0 introduces a new dynamic graph architecture, 55 sparse computing APIs, and a GPU training engine for large-scale graph neural networks, alongside significant enhancements to distributed training and deployment.
PaddlePaddle 2.3.2 addresses CUDA 11.7 compatibility issues, improves quantization workflows, and expands inference support for ONNX Runtime and TensorRT.
PaddlePaddle 2.4.2 introduces Cutlass-based kernel fusion for sparse convolutions and fixes critical bugs in distributed training and mixed-precision inference.
PaddlePaddle 2.5.0 introduces a unified dynamic-static execution model, native 0D tensor support, and significant performance improvements for ResNet50 and BERT models. The release also deprecates legacy APIs like paddle.fluid and removes older data parallel interfaces, requiring code updates for multi-node setups.
PaddlePaddle 2.6.0 introduces a new Intermediate Representation (PIR) and integrates the CINN compiler to enhance extensibility and performance. The update focuses on large model support through Auto Parallelism, advanced distributed training strategies, and optimized inference capabilities.
PaddlePaddle 2.6.1 introduces a new graph neural network training engine (PGLBox) for large-scale models and support for Fake GroupWise quantization. The release also addresses security vulnerabilities and fixes issues in Paddle-TRT integration and Windows compilation.
PaddlePaddle 3.0 introduces significant upgrades to automatic parallelism, a unified neural network compiler (CINN), and high-order automatic differentiation for scientific computing, alongside critical changes to type promotion and hardware compatibility.
PaddlePaddle 3.1.0 introduces a refined auto-parallel architecture, FP8 low-precision training for 10-20% speedups, and a mechanism to reuse CUDA kernels for heterogeneous hardware.
Baidu’s PaddlePaddle framework updates to v3.1.1, focusing on robust large model training, FP8 optimization, and new API features.
PaddlePaddle 3.2.0 emphasizes performance for large language models through FlashMask V3, FP8 training, and enhanced fault tolerance. It also expands hardware support via a Custom Device API and improves compatibility with industry standards like Safetensors and PyTorch-style APIs.
