Installation

Prerequisites

RaNNC works only with CUDA devices (CPU only or TPU environments are not supported). RaNNC requires the following libraries and tools at runtime.

• CUDA: A CUDA runtime (Version 11) must be available in the runtime environment.

• NCCL: NCCL (Version >= 2.11.4) must be available in the runtime environment.

• MPI: A program using RaNNC must be launched with MPI. MPI libraries must also be available at runtime. RaNNC has been tested with OpenMPI v4.0.7.

• libstd++: libstd++ must support GLIBCXX_3.4.21 to use the distributed pip packages (these packages are built with gcc 7.5.0).

Installation

The current version (0.7.5) of RaNNC requires PyTorch v1.11.0. pip packages for linux_x86_64 are available for the following combinations of Python and CUDA versions.

• Python version: 3,7, 3.8, 3.9, 3.10

• CUDA version: 11.3

The following commands install PyToch and RaNNC with pip. (The package version should be specified as 0.7.5+cu[CUDA_VERSION_WITHOUT_DOT])

pip install torch==1.11.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html
pip install pyrannc==0.7.5+cu113 -f https://nict-wisdom.github.io/rannc/installation.html

Use the following links to manually download the packages.

• For Python 3.7 / CUDA 11.3

• For Python 3.8 / CUDA 11.3

• For Python 3.9 / CUDA 11.3

• For Python 3.10 / CUDA 11.3

If the above packages do not match your Python/CUDA versions, create a suitable package using Makefile in docker/. make.sh shows the commands to create wheel packages.