icon4py 0.2.0rc2

ICON model in Python.

ICON4Py

This repository hosts a work-in-progress Python implementation of the ICON climate and weather model. Additionally, it includes icon4py.tools, a collection of command-line interfaces (CLIs), and utilities required for the integration of ICON4Py code into the ICON Fortran model. ICON4Py leverages GT4Py to ensure efficient and performance portable implementations of these components.

Project Structure

The repository is organized as a monorepo, where various ICON model components and utilities are developed as independent Python namespace packages in subfolders. An icon4py Python package is defined at the root folder with the purpose to collect specific versions of the different components as package dependencies. The component can also be installed independently, although since they are not (yet) available from a package repository, they need to be installed from their specific location within this repository.

License

ICON4Py is licensed under the terms of the BSD-3-Clause.

Installation Instructions

Since this project is still in a highly experimental state, it is not yet available as a regular Python distribution package through PyPI. The installation procedure involves cloning the ICON4Py GitHub repository and installing it in a Python virtual environment (venv).

ICON4Py uses the uv project manager for development workflow. uv is a versatile tool that consolidates functionality previously distributed across different applications into subcommands.

• The uv pip subcommand provides a fast Python package manager, emulating pip.
• The uv export | lock | sync subcommands manage dependency versions in a manner similar to the pip-tools command suite.
• The uv init | add | remove | build | publish | ... subcommands facilitate project development workflows, akin to hatch.
• The uv tool subcommand serves as a runner for Python applications in isolation, similar to pipx.
• The uv python subcommands manage different Python installations and versions, much like pyenv.

uv can be installed in various ways (see its installation instructions), with the recommended method being the standalone installer:

$ curl -LsSf https://astral.sh/uv/install.sh | sh 

ICON4Py Development Environment

Once uv is installed in your system, it is enough to clone this repository and let uv handling the installation of the development environment.

Important: the uv sync command should always be executed from the root folder of the repository, to make sure it installs all the workspace dependencies and not only the dependencies of a subproject.

The --extra distributed option installs mpi4py and ghex, which require Boost headers and an MPI implementation (e.g. OpenMPI) to be available on the system.

# Clone the repository
git clone [e-mail]:C2SM/icon4py.git
cd icon4py

# Let uv create the development environment at `.venv`.
# The `--extra all` option tells uv to install all the optional
# dependencies of icon4py, and thus it is not strictly necessary.
# Note that if no dependency groups are provided as an option,
# uv uses `--group dev` by default so the development dependencies
# are also installed. 
uv sync --extra all

# Activate the virtual environment and start writing code!
source .venv/bin/activate

The new venv is a standard Python virtual environment preconfigured with all necessary runtime and development dependencies. Additionally, all icon4py subpackages are installed in editable mode, allowing for seamless development and testing.

To install new packages, use the uv pip subcommand, which emulates the pip interface and is generally much faster. Alternatively, the standard pip command is also available within the venv, although using pip directly is slower and not recommended.

The pyproject.toml file at the root folder contains both the definition of the icon4py Python distribution package and the settings of the development tools used in this project, most notably uv, ruff, mypy and pytest. It also contains dependency groups (see PEP 735 for further reference) with the development requirements listed in different groups (build, docs, lint, test, typing, ...) and collected together in the general dev group which gets installed by default by uv.

Development Instructions

By following the installation instructions above, the source files are imported directly by the Python interpreter meaning that any code change is available and executed by the interpreter.

To add new dependencies to the project, either core/optional run-time or development-only dependencies, it is possible to use the uv cli direcly or to modify by hand the appropriate tables in the corresponding pyproject.toml (check uv documentation for more information https://docs.astral.sh/uv/concepts/projects/dependencies/).

Code Quality Checks

pre-commit is used to run several linting and checking tools. It should always be executed locally before opening a pull request. When executing pre-commit locally you can either run it for the model or tools folder:

For example to run code checks on all components in icon4py.model you can do:

# running precommit for all components in model
cd model/
pre-commit run --all-files

Testing

pytest is used for testing and it also comes with a command tool to easily run tests:

# Run all tests (verbose mode)
pytest -v

# Run only tests in `path/to/test/folder`
pytest -v path/to/test/folder

nox is recommended for running comprehensive test suites across multiple Python versions and configurations, mirroring the setup used in the CI pipeline.

# List all available test sessions (colored items are the default sessions)
nox -l

# Run all parametrized cases of a session
nox -s 'test_common'

# Run a test session for a specific python version and parameter value
nox -s 'test_atmosphere_advection-3.10(datatest=True)'

To run distributed tests, make sure an MPI implementation is installed and run uv sync --extra distributed or uv sync --extra all. Then run tests using mpirun, the --with-mpi pytest flag, and the -k mpi_tests filter:

mpirun -np 4 pytest -v -s --with-mpi -k mpi_tests

To avoid all ranks writing their test output to stdout, use the helper script ci/scripts/ci-mpi-wrapper.sh around the pytest command:

mpirun -np 4 ci/scripts/ci-mpi-wrapper.sh pytest -v -s --with-mpi -k mpi_tests

Distributed runs on Alps

Compiling

On Alps with Cray MPICH and GH200 or A100 GPUs, install mpi4py and GHEX as follows:

export GHEX_USE_GPU=ON
export GHEX_GPU_TYPE=NVIDIA
export GHEX_GPU_ARCH="80;90"
export GHEX_TRANSPORT_BACKEND=MPI
export MPICH_CXX=$(which g++)
export MPICH_CC=$(which gcc)
uv sync --no-binary-package mpi4py --extra all --extra distributed --extra cuda12 --python $(which python) --refresh

If you already have a broken GHEX, mpi4py, or other package in the uv cache, run the command with --no-cache after either uninstalling the broken package or wiping the virtualenv.

--no-binary-package mpi4py is required because Cray MPICH is not ABI compatible with the MPI used to build mpi4py binary wheels. If you don't do this you may get an error like:

ImportError: libmpi.so.12: cannot open shared object file: No such file or directory

when importing mpi4py. The GHEX_* options tell GHEX to build with GPU support. If you don't, you may see errors like:

AttributeError: module 'ghex.pyghex' has no attribute 'unstructured__data_descriptor_gpu_int_int_double_'

when GHEX tries to perform halo exchanges. The MPICH_* options make sure mpi4py gets built with GCC instead of NVHPC. mpi4py assumes that it can set certain compiler flags that GCC supports, but NVHPC does not support. The error message will typically look like:

      [stderr]
      nvc-Error-Unknown switch: -fwrapv
      error: Cannot compile MPI programs. Check your configuration!!!
      Installing mpi4py requires a working MPI implementation.
      If you are running on a supercomputer or cluster, check with
      the system administrator or refer to the system user guide.
      Otherwise, if you are running on a laptop or desktop computer,
      your may be missing the MPICH or Open MPI development package:

Running

When running make sure to

export MPICH_GPU_SUPPORT_ENABLED=1
export GT4PY_UNSTRUCTURED_HORIZONTAL_HAS_UNIT_STRIDE=1

Cray MPICH will otherwise segfault when communicating GPU buffers. Also see the CSCS Cray MPICH documentation for more details. The UNIT_STRIDE variable is necessary when using the DaCe backend, which otherwise fails (see PR1130).

When performance is important you may want to

export PYTHONOPTIMIZE=2

Benchmarking

We use pytest-benchmark to benchmark the execution time of stencils in icon4py. To disable benchmarking during testing you can use --benchmark-disable when invoking pytest.

Documentation

Documentation is at a very early stage given the constant state of development. Ongoing efforts to document the dycore can be viewed at c2sm.github.io/icon4py.

You can install the required packages by using the provided docs dependency group, which is included in the dev installed by default by uv sync if no dependency groups are specified.

To build the html documentation in your local repository copy starting from the most upwards directory, apply the following commands:

cd model/atmosphere/dycore/docs
make html

The local documentation could then be accessed at docs/_build/html/index.html

More Information

For more information please consult the package specific READMEs found in the model and tools folders.