Start by installing a suitable version of JAX and jaxlib. If you don’t plan on using
mpi4jax on GPU, the following will do:
$ pip install jax jaxlib
Much of the functionality we need has recently been added to JAX, which itself is changing frequently. Therefore,
mpi4jax has somewhat strict requirements on the supported versions of JAX and jaxlib. Be prepared to upgrade!
We recommend that you use
pip to install mpi4jax (but a distribution is also available via
conda which will work if MPI, mpi4py and mpi4jax are all installed through conda ):
$ pip install mpi4jax $ conda install -c conda-forge mpi4jax
pip requires a working installation of MPI to succeed. If you don’t already have MPI and want to get started as quickly as possible, try
conda, which bundles the MPI library (but remember not to mix pip and conda).
We advise against using the conda installation in HPC environments because it is not possible to change the MPI library
mpi4py is linked against.
And that is it! If you are familiar with MPI, you should in principle be able to get started right away. However, we recommend that you have a look at 🔪 The Sharp Bits 🔪, to make sure that you are aware of some of the pitfalls of
Selecting the MPI distribution¶
mpi4jax will use the MPI distribution with which
mpi4py was built.
mpi4py is not installed, it will be installed automatically before
To check which MPI library both libraries link to, run the following command in your prompt.
$ python -c "import mpi4py; print(mpi4py.get_config())"
If you wish to use a specific MPI library (only possible when using
pip), it is
usually sufficient to specify the
MPICC environment variable before installing
In doubt, please refer to the mpi4py documentation.
Installation with GPU support¶
To use JAX on the GPU, make sure that your
jaxlib is built with CUDA support.
mpi4jax also supports JAX arrays stored in GPU memory.
mpi4jax’s GPU extensions, we need to be able to locate the CUDA headers on your system. If they are not detected automatically, you can set the environment variable
CUDA_ROOT when installing
$ CUDA_ROOT=/usr/local/cuda pip install mpi4jax
This is sufficient for most situations. However,
mpi4jax will copy all data from GPU to CPU and back before and after invoking MPI.
If this is a bottleneck in your application, you can build MPI with CUDA support and communicate directly from GPU memory. This requires that you re-build the entire stack:
Your MPI library, e.g. OpenMPI, with CUDA support.
mpi4py, linked to your CUDA-enabled MPI installation.
mpi4jax, using the correct
Read here on how to use zero-copy GPU communication after installation.