PaddlePaddle 3.2.2 introduces optimizations for distributed training, including HuggingFace checkpoint support and fixes for safetensor loading, alongside updates to the XPU memory monitoring module.
PaddlePaddle 3.3.0 introduces significant upgrades for large language model training, including a new FlashMaskV3 kernel, a virtual memory allocator to reduce fragmentation, and enhanced support for国产 hardware.
PaddlePaddle 2.2.0 introduces 100+ new APIs for scientific computing and signal processing, enhances dynamic-to-static graph conversion for mixed precision, and advances distributed training with 4D hybrid parallelism. The release also adds TensorRT 8.x support and dynamic shape auto-configuration for easier deployment.
PaddlePaddle 2.2.1 introduces experimental support for NVIDIA’s CUDA Graphs and new linear algebra APIs, alongside optimizations for training throughput and bug fixes for core tensor operations.
Baidu’s PaddlePaddle framework receives version 2.2.2, introducing the Mish activation function, expanded TensorRT support, and numerous bug fixes for training and inference workflows.
PaddlePaddle 2.3.0‑rc0 brings more than 100 new APIs, a brand‑new PHI operator library, major distributed‑training upgrades, Java and ONNX inference support, and native compatibility with a range of new AI accelerators.
PaddlePaddle 2.4.1 eliminates the need for python.so and resolves a failure that occurred in Conda and other environments when the library could not locate python.so.
