Practical overview
When it fits
Use Oneflow when model development requires a programmable deep-learning framework with documented distributed tensor execution and graph-compilation capabilities, and the team can provide the required Python, Linux, compiler or CUDA environment.
Who it is for
- Deep-learning developers building neural-network models
- Teams creating distributed training workflows
- CUDA-oriented AI application developers
- Engineers managing self-hosted framework installations and builds
Documented uses
- Program neural-network models with a PyTorch-like API
- Scale models through Global Tensor's n-dimensional parallel execution
- Accelerate model execution with the Graph Compiler
- Develop CUDA-based deep-learning applications
- Compile and expose the framework from source when package installation is not the chosen path
Requirements & limits
The repository is public and licensed under Apache-2.0. Documented access paths include pip packages, Docker images and source compilation. Linux and specified Python versions are listed as general requirements; CUDA has additional toolkit, architecture and driver requirements.
- The PyTorch-like API may reduce the learning gap for developers familiar with that programming style.
- Global Tensor is the relevant documented feature for n-dimensional parallel model execution.
- Graph Compiler support matters when acceleration or deployment-oriented graph processing is required.
- CUDA users must match the listed architecture, toolkit and NVIDIA driver requirements.
- Source builds require a substantial native-toolchain and library setup.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Install Oneflow with pip
Set up a stable CUDA-enabled or CPU-only Oneflow installation using the documented pip paths.
Before you start
- Linux
- Python 3.7, 3.8, 3.9, 3.10 or 3.11
- For CUDA support: CUDA architecture 60 or above, CUDA Toolkit 10.0 or above and NVIDIA driver 440.33 or above
- pip
Steps
Upgrade pip with the command below.
python3 -m pip install --upgrade pipFor the latest stable CUDA-supported release, run the command below.
python3 -m pip install oneflowFor nightly CPU-only support, use the documented pre-release pip index; for nightly CUDA support, use the documented CUDA-specific pre-release index.
Expected outcome
Oneflow is installed through one of the repository's documented pip distribution paths.
Guide limitations
- The documented stable pip command is for CUDA support; the CPU-only path is identified as a nightly release.
- The evidence does not establish successful installation on a particular machine.
Build Oneflow from source
Configure and compile Oneflow from a cloned source tree, then expose the build to Python.
Before you start
- A cloned Oneflow repository
- Linux build environment
- GCC, G++, CMake, Autoconf, Libtool, NASM, Python pip and libopenblas-dev
- For CUDA builds: CUDA Toolkit and cuDNN paths, plus a selected CUDA architecture
Steps
Clone the repository with the command below.
git clone https://github.com/Oneflow-Inc/oneflow.gitInstall the listed build dependencies.
Create and enter a
builddirectory.Run the documented CPU-only or CUDA CMake configuration for the relevant environment.
Run the command below and then
source source.shfrom the build directory.make -j$(nproc)
Expected outcome
The source tree is configured and compiled, with the build exposed through the documented Python path setup.
Guide limitations
- The
PYTHONPATHchange fromsource source.shis not permanent. - The CUDA configuration requires local paths and a selected architecture.
- The evidence documents setup and a simple doctor command, not a completed model-training or deployment run.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 2 October 2026
- Original website title
- Oneflow-Inc/oneflow · Official homepage title · retained in its source language