Paddle - PArallel Distributed Deep LEarning by Baidu

Supports both dynamic and static graphs in a single framework, enabling flexible experimentation and optimized deployment, as highlighted in the unified dynamic/static graphs feature.

Requires minimal tensor partitioning annotations for automatic parallel strategy discovery, reducing development costs for large-scale models, per the automatic parallelism claim.

Provides a seamless workflow from training to inference with code reuse, specifically designed for large models, as mentioned in the integrated training and inference section.

Offers capabilities like high-order auto-diff, complex numbers, and Fourier transforms, facilitating scientific computing in fields like mathematics and meteorology, as described in the documentation.