Installation

ChainerX, or chainerx, can be installed as a top level Python package along with Chainer by configuring the environment variables below.

Note

Chainer must currently be installed from source in order to include ChainerX, but this is expected to change in the near future.

Installing from source

The following environment variables are available for building ChainerX from source.

Environment variable	Description
CHAINER_BUILD_CHAINERX	1 to build the chainerx package along with chainer. 0 to skip. Default is 0.
CHAINERX_BUILD_CUDA	1 to build chainerx with CUDA support. 0 to skip. Default is 0. See also CUDA support section below.
CHAINERX_ENABLE_BLAS	1 to enable BLAS, 0 to disable it. Default is 1. If BLAS is enabled, it is searched for and used if found. If not found, ChainerX will behave as if BLAS was disabled and use a basic implementation instead.
CHAINERX_ENABLE_LAPACK	1 to enable LAPACK, 0 to disable it. Default is 1. If LAPACK is enabled, it is searched for and used if found. If not found, ChainerX will behave as if LAPACK was disabled and may cause runtime errors.

Simply run pip install chainer after configuring the above environment variables. See Examples below.

CUDA support

When installing with the CUDA support, you also need to specify the cuDNN installation path.

You can specify either of the following environment variables to specify where to look for cuDNN installation.

Environment variable	Description
CUDNN_ROOT_DIR	Path to your cuDNN installation.
CHAINERX_CUDNN_USE_CUPY	1 to search for cuDNN library and include files in existing CuPy installation. Only applicable for CuPy installed via wheel (binary) distribution. Other variables related to cuDNN paths (such as CUDNN_ROOT_DIR) are ignored. Be warned that the resulting executable will be invalidated if CuPy is uninstalled, moved or replaced.

To support the NumPy/CuPy fallback mechanism, currently ChainerX with the CUDA support requires CuPy to be installed together.

See also

CuPy installation guide

Examples

Install ChainerX without CUDA support:

$ export CHAINER_BUILD_CHAINERX=1
$ export MAKEFLAGS=-j8  # Using 8 parallel jobs.
$ pip install chainer

Install ChainerX depending on CuPy wheel distribution:

$ pip install cupy_cuda101  # Note: Choose the proper CUDA SDK version number.
$ export CHAINER_BUILD_CHAINERX=1
$ export CHAINERX_BUILD_CUDA=1
$ export CHAINERX_CUDNN_USE_CUPY=1
$ export MAKEFLAGS=-j8  # Using 8 parallel jobs.
$ pip install chainer

Install ChainerX with CuPy built from source:

$ export CHAINER_BUILD_CHAINERX=1
$ export CHAINERX_BUILD_CUDA=1
$ export CUDNN_ROOT_DIR=path/to/cudnn
$ export MAKEFLAGS=-j8  # Using 8 parallel jobs.
$ pip install cupy
$ pip install chainer