Working with regridding using YAC in UXARRAY

[dylannelson]

Working with some of the team on UCAR's HPC resources, specifcically CASPER in my case, I was able to get some regridding functions and workflows working using YAC. It has a comparatively more difficult install than other packages, as it needs to be downloaded separately from conda/pip, and connected to a python environment manually. Here's roughly how I was able to get a YAC friendly python environment working for me:

Before you start

export ENV_NAME=YAC_Intro
export REGRID_DIR=/glade/u/home/$USER/Regridding

Change the first line to give your future conda environment a different name, or the second to chose where all the build files will be installed. I chose to make it in my home directory to make it easier to group all my regridding work into one folder, including my scripts. Both of these lines are just setting variables that the rest of the code will use. Meaning if you close your terminal between this command and any of the upcoming commands, they won't work. Any time you see $ENV_NAME it's referring to this line, so you hopefully don't need to manually change any of the following commands. Copy + Pasting them exactly as they appear below should work.

1. Modules

Load the same modules for the build and for every later run, so the compiler and MPI runtime libraries match.

module load gcc openmpi/5.0.8 conda

2. The environment

conda create -n "$ENV_NAME" -c conda-forge python=3.14 uxarray=2026.09.0 "cython>=3.1" setuptools wheel pytest

Answer y then check some of the important packages:

conda list -n "$ENV_NAME" | grep -E "^(uxarray|numba|netcdf4|python) "

uxarray must be 2026.09.0 (or newer) and numba and netcdf4 must both be listed (this may vary in the future, but we need to make sure these packages are here). Then:

conda activate "$ENV_NAME"

You can add ipykernel if you need this env to work in JupyterHub.

3. mpi4py, built against Casper's MPI

Not from a wheel here, it must link the same MPI as YAC:

python -m pip install --no-binary=mpi4py --no-cache-dir mpi4py

Check which mpicc first. It must print a path under /glade/u/apps, not one inside the conda env. A conda path means the env has its own MPI and the build will link the wrong one.

4. YAXT

This is a long one line install for YAXT, while convoluted, it's still similar to what's on the official docs. It's one && chain, so it stops at the first failure and prints YAXT OK only if every step worked. If you see that line printed in the terminal, you should be all good.

mkdir -p "${REGRID_DIR:?set it first, see Before you start}/build" && cd "$REGRID_DIR/build" && git clone --depth 1 --branch v0.11.5.1 https://gitlab.dkrz.de/dkrz-sw/yaxt.git && cd yaxt && mkdir -p build && cd build && ../configure --prefix="$REGRID_DIR/YAC_Install" --without-regard-for-quality CC=mpicc FC=mpif90 && make -j8 && make install && echo "YAXT OK"

5. YAC

Similar to the above command, if you see "YAC OK" you should be all good and it installed correctly

cd "${REGRID_DIR:?set it first, see Before you start}/build" && git clone --depth 1 --branch v3.20.2 https://gitlab.dkrz.de/dkrz-sw/yac.git && cd yac && mkdir -p build && cd build && ../configure --prefix="$REGRID_DIR/YAC_Install" --with-yaxt-root="$REGRID_DIR/YAC_Install" --disable-mci --disable-utils --disable-examples --disable-tools --disable-netcdf --enable-python-bindings --disable-mpi-checks PYTHON="$(which python)" CC=mpicc FC=mpif90 && make -j8 && make install && echo "YAC OK"

This builds only YAC's core library and Python bindings, which is what UXarray uses.

If a block stops, the error is just above where it stopped. For a configure failure:

grep -E "^configure:[0-9]+: error" "$REGRID_DIR/build"/*/build/config.log

6. Make the env find it

YAC installs outside the env (you should now see it in a build or YAC_Install folder within Regridding if you used the same path I did. This should set the 2 variables needed to connect YAC to the conda package:

mkdir -p "$CONDA_PREFIX/etc/conda/activate.d" && printf '%s\n' '#!/bin/sh' "export LD_LIBRARY_PATH=\"${REGRID_DIR:?set it first, see Before you start}/YAC_Install/lib\${LD_LIBRARY_PATH:+:\$LD_LIBRARY_PATH}\"" "export PYTHONPATH=\"$REGRID_DIR/YAC_Install/lib/python3.14/site-packages\${PYTHONPATH:+:\$PYTHONPATH}\"" > "$CONDA_PREFIX/etc/conda/activate.d/yac.sh"

Reactivate to pick it up:

conda deactivate && conda activate "$ENV_NAME"

7. Verify

python -c "import mpi4py; from uxarray.remap.yac import _import_yac; print(_import_yac().__file__)"

Expected: a path under $REGRID_DIR/YAC_Install, i.e.
/glade/u/home/$USER/Regridding/YAC_Install/lib/python3.14/site-packages/yac/__init__.py. A path
inside the conda env instead means you are importing a PyPI yac wheel, not this build.

import mpi4py comes first on purpose: with Open MPI 5 it has to load libmpi before yac.core.

For a real test, run UXarray's YAC suite:

git clone --depth 1 --branch v2026.09.0 https://github.com/UXARRAY/uxarray.git "${REGRID_DIR:?set it first, see Before you start}/uxarray-tests" && python -m pytest "$REGRID_DIR/uxarray-tests/test/test_remap_yac.py" -q

16 passed is the expected result

At this time we were having some odd interactions with minimum versions for packages like cartopy, and this may need to be revised in the future as we are in the middle of changing how some default packages work with UXarray. If you have any concerns or different outputs, please comment below, as we would appreciate seeing how this method works for others and how it can be improved.

8. Every session after this

To open the env later on, especially in a new terminal, it's preferred to activate like this to ensure the packages are opened in the right order:

module load gcc openmpi/5.0.8 conda && conda activate YAC_Intro

Change the YAC_Intro to whatever you named the env, if you changed it