Skip to content

JAX

JAX is Autograd and XLA, brought together for high-performance machine learning research.

Available

Currently supported JAX versions:

Version Module Puhti Mahti LUMI Notes
0.4.30 jax/0.4.30 default default - all packages
0.4.30 jax/0.4.30-small X X - framework only
0.4.23 jax/0.4.23-py3.9 X X default*
0.4.20 jax/0.4.20 X X X*
0.4.18 jax/0.4.18 - - X*
0.4.14 jax/0.4.14 X X -
0.4.13 jax/0.4.13 X X -
0.4.1 jax/0.4.1 X X -
0.3.13 jax/0.3.13 X X -

The modules contain JAX for Python 3.9 with GPU support via CUDA/ROCm as well as a large list of additional python packages commonly used together with JAX.

Versions in LUMI, marked as "*" are still experimental with limited support. Some of the features described below might not work for them. They are subject to change at any time without notice. Note that JAX is also available in the LUMI Software Library.

Since version 0.4.30, the JAX module comes in two flavours:

  • A "small" flavour that includes JAX with CUDA 12.2 GPU support, as well as cpu-only versions of PyTorch and Tensorflow to allow using their data loading utilities. These follow the naming scheme jax/<version>-small, or you can use jax/small to load the default (latest) version.
  • A "full" flavour that includes many commonly used machine learning packages building on JAX -- you can check pip list for a full list of all included packages. These follow the naming scheme jax/<version>, or you can simply use jax to load the default (latest) version.

Note

Since JAX releases new versions on a somewhat irregular schedule we will not make all new versions immediately available. Instead we endeavour to update the JAX version available on our systems approximately every six months, targeting February and August, on a best effort basis.

All modules are based on containers using Apptainer (previously known as Singularity). Wrapper scripts have been provided via tykky so that common commands such as python, python3, pip and pip3 and those provided by installed packages should work as normal. For other commands, you may need to prefix them with apptainer_wrapper exec. For more information, see CSC's general instructions on how to run Apptainer containers.

Additional packages

If you find that some package is missing, you can often install it yourself using pip install. It is recommended to use Python virtual environments. See our Python documentation for more information on how to install packages yourself. If you think that some important package should be included in the module provided by CSC, please contact our servicedesk.

License

JAX is licensed under Apache License 2.0.

Usage

To use the default version (most-recent) on Puhti or Mahti, initialize it with:

module load jax

To access CSC-installed JAX on LUMI:

module use /appl/local/csc/modulefiles/
module load jax

Note

In your scripts we recommend to fix the version so that changes in future upgrades do not break scripts, e.g.,: module load jax/0.4.23-py3.9

Please note that the JAX modules already include the corresponding CUDA and cuDNN or ROCm libraries, so there is no need to load any cuda, cudnn, or rocm modules separately!

This will show all available versions of JAX:

module avail jax

Note

Note that the login nodes are not intended for heavy computing, please use slurm batch jobs instead. See our instructions on how to use the batch job system.

Note

Please do not read a huge number of files from the shared file system, use fast local disk or package your data into larger files instead! See the Data storage section in our machine learning guide for more details.

More information