Documentation | Installation | Examples | Release Notes
ChainerMN is an additional package for Chainer, a flexible deep learning framework. ChainerMN enables multi-node distributed deep learning with the following features:
- Scalable --- it makes full use of the latest technologies such as NVIDIA NCCL and CUDA-Aware MPI,
- Flexible --- even dynamic neural networks can be trained in parallel thanks to Chainer's flexibility, and
- Easy --- minimal changes to existing user code are required.
This blog post provides our benchmark results using up to 128 GPUs.
ChainerMN can be used for both inner-node (i.e., multiple GPUs inside a node) and inter-node settings. For inter-node settings, we highly recommend to use high-speed interconnects such as InfiniBand.
In addition to Chainer, ChainerMN depends on the following software libraries: CUDA-Aware MPI, NVIDIA NCCL, and a few Python packages. After setting them up, ChainerMN can be installed via PyPI:
pip install chainermn
Please refer to the installation guide for more information.
You can invoke MNIST example with four workers by the following command:
mpiexec -n 4 python examples/mnist/train_mnist.py
- Chainer Tutorial --- If you are new to Chainer, we recommend to start from this.
- ChainerMN Tutorial --- In this tutorial, we explain how to modify your existing code using Chainer to enable distributed training with ChainerMN in a step-by-step manner.
- Examples --- The examples are based on the official examples of Chainer and the differences are highlighted.
Any contribution to ChainerMN would be highly appreciated. Please refer to Chainer Contribution Guide.
Akiba, T., Fukuda, K. and Suzuki, S., ChainerMN: Scalable Distributed Deep Learning Framework, Proceedings of Workshop on ML Systems in The Thirty-first Annual Conference on Neural Information Processing Systems (NIPS), (2017) URL, BibTex