Deep Learning

PaddleOCR is a comprehensive toolkit designed to convert images and PDF documents into structured, LLM-ready data formats like Markdown and JSON. It features state-of-the-art vision-language models and high-performance text recognition engines that support over 100 languages. The platform is widely integrated into major AI agent and RAG frameworks, offering efficient deployment options across various hardware backends.

DeepGEMM is a lightweight CUDA library designed for efficient General Matrix Multiplications, supporting FP8 and BF16 data formats. It utilizes a just-in-time compilation module to eliminate the need for pre-installation kernel compilation while maintaining performance comparable to expert-tuned libraries. The library provides specialized APIs for dense and MoE-grouped GEMMs, making it a clean resource for learning GPU kernel optimization.

DeepEP is a specialized communication library designed to optimize Mixture-of-Experts and expert parallelism through high-throughput, low-latency GPU kernels. It supports both training and inference workloads by providing advanced features like asymmetric-domain bandwidth forwarding and hook-based communication-computation overlapping. The library is highly optimized for NVLink and RDMA environments, offering significant performance gains for large-scale model deployments.

Index-AniSora is a powerful open-source framework designed specifically for high-quality anime video generation and animation production. The system integrates a comprehensive data processing pipeline, a controllable generation model with spatiotemporal masking, and a specialized evaluation benchmark. It supports diverse creative tasks including image-to-video generation, style transfer, character 3D modeling, and multimodal guidance.

Magika is an AI-powered tool that utilizes deep learning to provide highly accurate file type identification for both binary and textual formats. It features a highly optimized model capable of performing inference in milliseconds on a single CPU with approximately 99% accuracy. The project offers a versatile command-line interface and language bindings, making it suitable for large-scale security and content policy scanning.

VoxCPM2 is a tokenizer-free, 2B parameter text-to-speech system that utilizes a diffusion autoregressive architecture for highly natural audio synthesis. The model supports 30 languages and offers advanced capabilities including voice design from text descriptions and controllable voice cloning. It delivers studio-quality 48kHz audio output and is fully open-source under the Apache-2.0 license for commercial use.

ERNIE-Image is an open-source text-to-image model developed by Baidu based on a single-stream Diffusion Transformer architecture. The model is equipped with a lightweight prompt enhancer that can transform short inputs into structurally rich descriptions. With an 8B parameter scale, it excels at handling complex instructions, text rendering, and structured visual tasks, while supporting efficient deployment on consumer-grade GPUs.

RTP-LLM is a high-performance large language model inference acceleration engine developed by the Alibaba Foundation Model Inference Team. This engine has been widely applied in various Alibaba business scenarios such as Taobao and Tmall, and it supports multiple mainstream model formats and hardware backends. By integrating advanced operator optimization, quantization techniques, and distributed inference capabilities, it provides developers with efficient production-grade inference solutions.

PaddleFormers is a Transformers library built on the PaddlePaddle framework, designed to provide training interfaces for Large Language Models and Vision-Language Models equivalent to Hugging Face. By integrating tensor parallelism, pipeline parallelism, and automatic mixed precision, the project achieves training performance that surpasses Megatron-LM on key models. Furthermore, it fully supports the Safetensors format and is deeply adapted to various domestic computing chips, helping developers efficiently complete the full model training process.

PaddlePaddle is a comprehensive industrial deep learning platform that provides core frameworks, model libraries, and end-to-end development tools. It supports advanced features such as unified dynamic and static graphs, automatic parallelism, and high-order differentiation for scientific computing. The platform is designed to facilitate industrial AI commercialization across diverse sectors through its mature, heterogeneous hardware-compatible architecture.