/oneflow

OneFlow is a performance-centered and open-source deep learning framework.

Primary LanguageC++Apache License 2.0Apache-2.0

OneFlow is a performance-centered and open-source deep learning framework.

Install OneFlow

System Requirements

  • Python >= 3.5

  • CUDA Toolkit Linux x86_64 Driver

    OneFlow CUDA Driver Version
    oneflow_cu110 >= 450.36.06
    oneflow_cu102 >= 440.33
    oneflow_cu101 >= 418.39
    oneflow_cu100 >= 410.48
    oneflow_cu92 >= 396.26
    oneflow_cu91 >= 390.46
    oneflow_cu90 >= 384.81
    oneflow_cpu N/A

    • CUDA runtime is statically linked into OneFlow. OneFlow will work on a minimum supported driver, and any driver beyond. For more information, please refer to CUDA compatibility documentation.

    • Support for latest stable version of CUDA will be prioritized. Please upgrade your Nvidia driver to version 440.33 or above and install oneflow_cu102 if possible.

    • We are sorry that due to limits on bandwidth and other resources, we could only guarantee the efficiency and stability of oneflow_cu102. We will improve it ASAP.

Install with Pip Package

  • To install latest release of OneFlow with CUDA support:

    python3 -m pip install --find-links https://oneflow-inc.github.io/nightly oneflow_cu102 --user

  • To install latest release of CPU-only OneFlow:

    python3 -m pip install --find-links https://oneflow-inc.github.io/nightly oneflow_cpu --user

  • To install OneFlow with legacy CUDA support, run one of:

    python3 -m pip install --find-links https://oneflow-inc.github.io/nightly oneflow_cu101 --user
python3 -m pip install --find-links https://oneflow-inc.github.io/nightly oneflow_cu100 --user
python3 -m pip install --find-links https://oneflow-inc.github.io/nightly oneflow_cu92 --user
python3 -m pip install --find-links https://oneflow-inc.github.io/nightly oneflow_cu91 --user
python3 -m pip install --find-links https://oneflow-inc.github.io/nightly oneflow_cu90 --user

  • If you are in China, you could run this to have pip download packages from domestic mirror of pypi:

    python3 -m pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple

    For more information on this, please refer to pypi 镜像使用帮助

  • Releases are built with G++/GCC 4.8.5, cuDNN 7 and MKL 2020.0-088.

Build from Source

  1. System Requirements to Build OneFlow

    • Please use a newer version of CMake to build OneFlow. You could download cmake release from here.

    • Please make sure you have G++ and GCC >= 4.8.5 installed. Clang is not supported for now.

    • To install dependencies, run:

      yum-config-manager --add-repo https://yum.repos.intel.com/setup/intelproducts.repo && \
rpm --import https://yum.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS-2019.PUB && \
yum update -y && yum install -y epel-release && \
yum install -y intel-mkl-64bit-2020.0-088 nasm swig rdma-core-devel

      On CentOS, if you have MKL installed, please update the environment variable:

      export LD_LIBRARY_PATH=/opt/intel/lib/intel64_lin:/opt/intel/mkl/lib/intel64:$LD_LIBRARY_PATH

      If you don't want to build OneFlow with MKL, you could install OpenBLAS. On CentOS:

      sudo yum -y install openblas-devel

      On Ubuntu:

      sudo apt install -y libopenblas-dev

  2. Clone Source Code

    • Option 1: Clone source code from github

      git clone https://github.com/Oneflow-Inc/oneflow

    • Option 2: Download from Aliyun

      If you are in China, please download OneFlow source code from: https://oneflow-public.oss-cn-beijing.aliyuncs.com/oneflow-src.zip

      curl https://oneflow-public.oss-cn-beijing.aliyuncs.com/oneflow-src.zip -o oneflow-src.zip
unzip oneflow-src.zip

  3. Build and Install OneFlow

    • Option 1: Build in docker container (recommended)

      • In the root directory of OneFlow source code, run:

        python3 docker/package/manylinux/build_wheel.py

        This should produces .whl files in the directory wheelhouse

      • If you are in China, you might need to add these flags:

        --use_tuna --use_system_proxy --use_aliyun_mirror

      • You can choose CUDA/Python versions of wheel by adding:

        --cuda_version=10.1 --python_version=3.6,3.7

      • For more useful flags, plese run the script with flag --help or refer to the source code of the script.

    • Option 2: Build on bare metal

      • In the root directory of OneFlow source code, run:

        mkdir build
cd build
cmake ..
make -j$(nproc)
make pip_install

      • If you are in China, please add this CMake flag -DTHIRD_PARTY_MIRROR=aliyun to speed up the downloading procedure for some dependency tar files.

      • For pure CPU build, please add this CMake flag -DBUILD_CUDA=OFF.

Troubleshooting

Please refer to troubleshooting for common issues you might encounter when compiling and running OneFlow.

Advanced features

  • XRT

    You can check this doc to obtain more details about how to use XLA and TensorRT with OneFlow.

Getting Started

3 minutes to run MNIST.

  1. Clone the demo code from OneFlow documentation
git clone https://github.com/Oneflow-Inc/oneflow-documentation.git
cd oneflow-documentation/cn/docs/code/quick_start/
  1. Run it in Python
python mlp_mnist.py
  1. Oneflow is running and you got the training loss
2.7290366
0.81281316
0.50629824
0.35949975
0.35245502
...

More info on this demo, please refer to doc on quick start.

Documentation

Usage & Design Docs

API Reference

OneFlow System Design

For those who would like to understand the OneFlow internals, please read the document below:

Model Zoo and Benchmark

CNNs(ResNet-50, VGG-16, Inception-V3, AlexNet)

Wide&Deep

BERT

Communication

  • Github issues : any install, bug, feature issues.
  • www.oneflow.org : brand related information.

Contributing

The Team

OneFlow was originally developed by OneFlow Inc and Zhejiang Lab.

License

Apache License 2.0