oneflow

OneFlow is a deep learning framework designed to be user-friendly, scalable, and efficient. It provides a PyTorch-like API for programming models, Global Tensor for scaling to n-dimensional parallel execution, and a Graph Compiler for accelerating and deploying models. ### Key Features - **PyTorch-like API**: Allows intuitive model programming similar to PyTorch. - **Global Tensor**: Facilitates scaling models across multi-dimensional parallel architectures. - **Graph Compiler**: Optimizes computational graphs to accelerate training and streamline deployment. ### System Requirements - Linux operating system. - Python 3.7, 3.8, 3.9, 3.10, or 3.11. - CUDA architecture 6.0 or above, CUDA Toolkit 10.0 or above, and Nvidia driver version 440.33 or above for GPU acceleration. ### Installation Methods - **Docker**: Prebuilt nightly Docker images with CUDA support are available. - **Pip**: Stable releases and nightly builds (CPU-only or CUDA) can be installed via pip, with domestic mirror options for users in China. - **Source Code**: Users can clone the repository and build using CMake, with configuration presets for CPU-only or CUDA environments. ### Ecosystem and Model Zoo - **Libai**: A toolbox designed for parallel training of large-scale transformer models, including BERT, GPT, T5, VisionTransformer, and SwinTransformer. - **FlowVision**: A toolbox providing computer vision datasets, state-of-the-art models, and utilities. - **OneFlow-Models**: Provides model implementations like ResNet-50 and Wide&Deep (noted as outdated). ### Community and Communication - **Chinese Channels**: QQ group (331883), WeChat (OneFlowXZS), and Zhihu organization page. - **International Channels**: Discord server, Twitter, LinkedIn, and Medium blog. - Bug reports and feature requests are managed via GitHub issues. OneFlow is originally developed by OneFlow Inc. and Zhejiang Lab, and is distributed under the Apache License 2.0.