Building GPAW in a Python venv on Niflheim

This document explains how to compile a developer version of GPAW on Niflheim. If you just want to run the pre-installed version, please read the guide Using a pre-installed GPAW on Niflheim.

Creating the venv

Download the script (use wget <url>) and run it like this:

$ python3 <venv-name>


You will need Python 3.8 or later. You can install that with:

$ module load Python

Type python3 --help for help. After a few minutes, you will have a <venv-name> folder with a GPAW installation inside.

In the following, we will assume that your venv folder is ~/venv1/.

The script does the following:

  • load relevant modules from the foss toolchain

  • create the venv

  • clone and install ASE and GPAW from gitlab

  • install some other Python packages from PyPI: sklearn, graphviz, matplotlib, pytest-xdist, myqueue, ase-ext, spglib

  • enable tab-completion for command-line tools: ase, gpaw, mq

Using the venv

The venv needs to be activated like this:

$ source venv1/bin/activate

and you can deactivate it when you no longer need to use it:

$ deactivate

You will want the activation to happen automatically for the jobs you submit to Niflheim. Here are three ways to do it (pick one, and only one):

  1. If you always want to use one venv then just put the activation command in your ~/.bashrc.

  2. If you only want jobs running inside a certain folder to use the venv, then add this to your ~/.bashrc:

    if [[ $SLURM_SUBMIT_DIR/ = $HOME/project-1* ]]; then
        source ~/venv1/bin/activate

    Now, SLURM-jobs submitted inside your ~/project-1/ folder will use the venv.

  3. Use MyQueue. Make sure you have MyQueue version 22.7.0 or later (mq --version). The venv will automatically be activated if it was activated at submit time.

    If you haven’t configured MyQueue then you can do that with this command:

    $ mq config slurm | grep -v sm3090 > ~/.myqueue/

    (skips the sm3090 GPU-enabled nodes).

Adding additional packages

In order to add more Python packages to your venv, you need to activate it and then you can pip install packages. Here is how to install ASR:

$ git clone
$ cd asr
$ git checkout old-master
$ pip install .


Pip may need co compile some code. It is therefore safest to use the thul login node to pip install software as it is the oldest CPU architcture and the other login nodes will understand its code.

Full script

#!/usr/bin/env python3
"""Install gpaw on Niflheim in a virtual environment.

Also installs ase, ase-ext, spglib, sklearn and myqueue.
import argparse
import os
import shutil
import subprocess
from pathlib import Path
from sys import version_info

if version_info < (3, 8):
    raise ValueError('Please use Python-3.8 or later')

# Python version in the venv that we are creating
version = '3.11'
fversion = 'cpython-311'

# Niflheim login hosts, with the oldest architecture as the first
nifllogin = ['sylg',  # broadwell (xeon24)
             'svol',  # skylake (xeon40)
             'surt',  # icelake (xeon56)
             'fjorm',  # epyc9004 (epyc96)
             'thul',  # skylake_el8 (xeon40el8)
             'slid2']  # broadwell_el8 (xeon24el8)

# Easybuild uses a hierarchy of toolchains for the main foss and intel
# chains.  The order in the tuples before are
#  fullchain: Full chain.
#  mathchain: Chain with math libraries but no MPI
#  compchain: Chain with full compiler suite (but no fancy libs)
#  corechain: Core compiler
# The subchain complementary to 'mathchain', with MPI but no math libs, is
# not used here.

_gcccore = 'GCCcore-12.3.0'
toolchains = {
    'foss': dict(
    'intel': dict(

# These modules are always loaded
module_cmds_all = """\
module purge
module load GPAW-setups/24.1.0
module load ELPA/2023.05.001-{fullchain}
module load Wannier90/3.1.0-{fullchain}
module load Python-bundle-PyPI/2023.06-{corechain}
module load Tkinter/3.11.3-{corechain}
module load libxc/6.2.2-{compchain}

# These modules are not loaded if --piponly is specified
module_cmds_easybuild = """\
module load matplotlib/3.7.2-{mathchain}
module load scikit-learn/1.3.1-{mathchain}
module load spglib-python/2.1.0-{mathchain}

# These modules are loaded depending on the toolchain
module_cmds_tc = {
    'foss': """\
module load libvdwxc/0.4.0-{fullchain}
    'intel': ""

module_cmds_arch_dependent = """\
if [ "$CPU_ARCH" == "icelake" ];\
then module load CuPy/12.3.0-{fullchain}-CUDA-12.1.1;fi

activate_extra = """
export GPAW_SETUP_PATH=$GPAW_SETUP_PATH:{venv}/gpaw-basis-pvalence-0.9.20000

# Set matplotlib backend:
if [[ $SLURM_SUBMIT_DIR ]]; then
    export MPLBACKEND=Agg
    export PYTHONWARNINGS="ignore:Matplotlib is currently using agg"
    export MPLBACKEND=TkAgg

dftd3 = """\
mkdir {venv}/DFTD3
cd {venv}/DFTD3
wget $URL/dftd3.tgz
tar -xf dftd3.tgz
ssh {nifllogin[0]} ". {venv}/bin/activate && cd {venv}/DFTD3 && make >& d3.log"
ln -s {venv}/DFTD3/dftd3 {venv}/bin

def run(cmd: str, **kwargs) -> subprocess.CompletedProcess:
    return, shell=True, check=True, **kwargs)

def compile_gpaw_c_code(gpaw: Path, activate: Path) -> None:
    """Compile for all architectures: xeon24, xeon40, ..."""
    # Remove targets:
    for path in gpaw.glob('build/lib.linux-x86_64-*/_gpaw.*.so'):

    # Compile:
    for host in nifllogin:
        run(f'ssh {host} ". {activate} && pip install -q -e {gpaw}"')

    # Clean up:
    for path in gpaw.glob('_gpaw.*.so'):
    for path in gpaw.glob('build/temp.linux-x86_64-*'):

def fix_installed_scripts(venvdir: Path,
                          rootdir: str,
                          pythonroot: str) -> None:
    """Fix command line tools so they work in the virtual environment.

    Command line tools (pytest, sphinx-build etc) fail in virtual
    enviroments created with --system-site-packages, as the scripts
    are not copied into the virtual environment.  The scripts have
    the original Python interpreter hardcoded in the hash-bang line.

    This function copies all scripts into the virtual environment,
    and changes the hash-bang so it works.  Starting with the 2023a
    toolchains, the scripts are distributed over more than one
    EasyBuild module.

    venvdir: Path to the virtual environment
    rootdir: string holding folder of the EasyBuild package being processed
    pythondir: string holding folder of the Python package.

    assert rootdir is not None
    assert pythonroot is not None
    bindir = rootdir / Path('bin')
    print(f'Patching executable scripts from {bindir} to {venvdir}/bin')
    assert '+' not in str(pythonroot) and '+' not in str(venvdir), (
        'Script will fail with "+" in folder names!')
    sedscript = f's+{pythonroot}+{venvdir}+g'

    # Loop over potential executables
    for exe in bindir.iterdir():
        target = venvdir / 'bin' /
        # Skip files that already exist, are part of Python itself,
        # or are not a regular file or symlink to a file.
        if (not target.exists()
                and not'python')
                and exe.is_file()):
            # Check if it is a script file referring the original
            # Python executable in the hash-bang
            with open(exe) as f:
                firstline = f.readline()
            if pythonroot in firstline:
                shutil.copy2(exe, target, follow_symlinks=False)
                # Now patch the file (if not a symlink)
                if not exe.is_symlink():
                    assert not target.is_symlink()
                        f"sed -e '{sedscript}' --in-place '{target}'",

def main():
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument('venv', help='Name of venv.')
    parser.add_argument('--toolchain', default='foss',
                        choices=['foss', 'intel'],
                        help='Default is foss.')
    parser.add_argument('--dftd3', action='store_true',
                        help='Also build DFTD3.')
    parser.add_argument('--recompile', action='store_true',
                        help='Recompile the GPAW C-extensions in an '
                        'exising venv.')
    parser.add_argument('--piponly', action='store_true',
                        help='Do not use EasyBuild python modules, '
                        'install from pip (may affect performance).')
    args = parser.parse_args()

    # if args.toolchain == 'intel':
    #     raise ValueError('See:')

    venv = Path(args.venv).absolute()
    activate = venv / 'bin/activate'
    gpaw = venv / 'gpaw'

    if args.recompile:
        compile_gpaw_c_code(gpaw, activate)
        return 0

    # Sanity checks
    if args.toolchain not in ('foss', 'intel'):
        raise ValueError(f'Unsupported toolchain "{args.toolchain}"')

    module_cmds = module_cmds_all.format(**toolchains[args.toolchain])
    if not args.piponly:
        module_cmds += module_cmds_easybuild.format(
    module_cmds += module_cmds_tc[args.toolchain].format(
    module_cmds += module_cmds_arch_dependent.format(
    cmds = (' && '.join(module_cmds.splitlines()) +
            f' && python3 -m venv --system-site-packages {args.venv}')


    activate.write_text(module_cmds +

    run(f'. {activate} && pip install --upgrade pip -q')

    # Fix venv so pytest etc work
    pythonroot = None
        # Note that we need the environment variable from the newly
        # created venv, NOT from this process!
        comm = run(f'. {activate} && echo ${ebrootvar}',
                   capture_output=True, text=True)
        ebrootdir = comm.stdout.strip()
        if pythonroot is None:
            # The first module is the actual Python module.
            pythonroot = ebrootdir
        assert ebrootdir, f'Env variable {ebrootvar} appears to be unset.'

    packages = ['myqueue',
    if args.piponly:
        packages += ['matplotlib',
    run(f'. {activate} && pip install -q -U ' + ' '.join(packages))

    for name in ['ase', 'gpaw']:
        run(f'git clone -q{name}/{name}.git')

    run(f'. {activate} && pip install -q -e ase/')

    if args.dftd3:
        run(' && '.join(dftd3.format(venv=venv,

    # Compile ase-ext C-extension on old thul so that it works on
    # newer architectures
    run(f'ssh {nifllogin[0]} ". {activate} && pip install -q ase-ext"')

    if args.piponly:
        run('git clone -q')
        run(f'ssh {nifllogin[0]} ". {activate} && pip install {venv}/spglib"')

    # Install GPAW:
    siteconfig = Path(

    compile_gpaw_c_code(gpaw, activate)

    for fro, to in [('ivybridge', 'sandybridge'),
                    ('nahelem', 'icelake')]:
        f = gpaw / f'build/lib.linux-x86_64-{fro}-{fversion}'
        t = gpaw / f'build/lib.linux-x86_64-{to}-{fversion}'

    # Create .pth file to load correct .so file:
    pth = (
        'import sys, os; '
        'arch = os.environ["CPU_ARCH"]; '
        f"path = f'{venv}/gpaw/build/lib.linux-x86_64-{{arch}}-{fversion}'; "

    # Install extra basis-functions:
    run(f'. {activate} && gpaw install-data --basis --version=20000 '
        f'{venv} --no-register')

    extra = activate_extra.format(venv=venv)

    # Tab completion:
    for cmd in ['ase', 'gpaw', 'mq', 'pip']:
        if cmd == 'gpaw':
            # Currently, running the "gpaw" command writes warning message
            # to stdout, so "gpaw completion" does not work!
        txt = run(f'. {activate} && {cmd} completion' +
                  (' --bash' if cmd == 'pip' else ''),
        extra += txt

    # gpaw-hack:
    python = venv / 'bin/python3'
    complete = venv / 'gpaw/gpaw/cli/'
    extra += f'complete -o default -C "{python} {complete}" gpaw\n'

    activate.write_text(activate.read_text() + extra)

    # Run tests:
    run(f'. {activate} && ase info && gpaw test')

    return 0

if __name__ == '__main__':
    raise SystemExit(main())