OneFlow is a deep learning framework designed for user-friendliness, scalability, and efficiency. It targets researchers and engineers looking to program models with a PyTorch-like API, scale them to n-dimensional parallel execution using Global Tensor, and accelerate deployment via its Graph Compiler.

How It Works

OneFlow utilizes a Global Tensor abstraction to manage distributed data across multiple devices and nodes, enabling seamless n-dimensional parallelism. Its Graph Compiler optimizes the computation graph for efficient execution, facilitating model acceleration and deployment. This approach aims to simplify distributed training and inference compared to traditional frameworks.

Quick Start & Requirements

• Install: python3 -m pip install oneflow (stable CUDA), python3 -m pip install --pre oneflow -f https://oneflow-staging.oss-cn-beijing.aliyuncs.com/branch/master/cu118 (nightly CUDA).
• Prerequisites: Python 3.7-3.11, CUDA arch 60+, CUDA Toolkit 10.0+, Nvidia driver 440.33+. Docker images are available.
• Resources: Building from source requires libopenblas-dev, nasm, g++, gcc, python3-pip, cmake, autoconf, libtool.
• Docs: QUICKSTART, API Reference.

Highlighted Details

• PyTorch-like API for ease of use.
• N-dimensional parallelism via Global Tensor.
• Graph Compiler for deployment acceleration.
• Includes Libai for large-scale Transformer models and FlowVision for computer vision tasks.

Maintenance & Community

• Developed by OneFlow Inc and Zhejiang Lab.
• Community channels include GitHub issues, QQ group (331883), WeChat, Discord, Twitter, LinkedIn, and Medium.

Licensing & Compatibility

• License: Apache License 2.0.
• Permissive license suitable for commercial use and integration into closed-source projects.

Limitations & Caveats

• Some model zoo links (OneFlow-Models, OneFlow-Benchmark) are marked as outdated.
• Building from source requires specific system dependencies and careful CMake configuration.