Adding a New Program Harness

Program harnesses are for community CMS codes that can independently (or with a SCF bootstrap) compute single-point energies, derivatives, or properties or components thereof (e.g., dispersion corrections).

A single CMS code generally has one program harness. However, if there are drastically different ways of running a code (e.g., TeraChem text input file and TeraChem PBS), separate harnesses may be created. Also, if there are specialty capabilites not fitting into “single-point energies …” (e.g., GAMESS makefp task), an additional procedure harness may be created.

This guide is a coarse path through adding a new program harness.

  1. Open up communication with the QCEngine maintainers. Post an issue to GitHub and join the Slack channel (link off GH README) so you can get advice.

  2. Copy a similar harness. Choose a program that your code behaves roughly like (mostly consider parsed vs. API access) and copy that harness, renaming it as your own and commenting out all but the structure. Search for the (copied) harness name to register it.

  3. Fill in the _defaults section with program name and characteristics.

  4. Fill in the def found function using which and which_import from QCElemental. See NWChem and OpenMM for examples of handling additional dependencies.

  5. If your code’s version can be extracted short of parsing an output file, fill in def get_version next. After this, > qcengine info should show your code (provided it’s in path).

  6. To get a string output of QCSchema Molecule in your code’s format, you may need to add a dtype at qcelemental/molparse/to_string.py.

  7. If your code’s of the common translate-QCSchema-to-input, run, translate-output-to-QCSchema variety, next work on the def execute function. This is fairly simple because it calls the powerful qcengine.util.execute to handle scratch, timeout, file writing and collection, etc. The harness function needs the names of input files (hard-code a string for now), the execution command, and the names of any scratch files to return for processing. Once ready, fill in def get_version if not done above.

  8. Now fill in the short def compute entirely and def build_input and def parse_output skeletally. Set up a simple molecule-and-model AtomicInput dictionary and run it with qcng.compute(atomicinput, "yourcode") to get something to iterate on.

  9. Fill in def build_input to form your code’s usual input format from the fields of AtomicInput.

  10. Fill in def parse_output to take results and put them into AtomicResult. Most important is the return_result field. AtomicResultProperties can be populated when convenient. WavefunctionProperties is great but save for a later pass.

  11. At this point your harness can correctly run one or more QCSchema inputs of your devising. Time to put it through paces. Register your code in _programs in testing.py. Most tests will then need a @using("yourcode") decorator so that they don’t run (and fail the test suite) when your code isn’t available.

  12. Add basic tests to qcengine/tests/test_harness_canonical.py * def test_compute_energy(program, model, keywords): * def test_compute_gradient(program, model, keywords): * def test_compute_energy_qcsk_basis(program, model, keywords):

  13. Add basic failure tests to qcengine/tests/test_harness_canonical.py * def test_compute_bad_models(program, model):

  14. Add tests for the runtime config and to qcengine/programs/tests/test_canonical_config.py

  15. For QM codes, consider adding lines to the qcengine/programs/tests/test_standard_suite.py to check energies, gradients, and Hessians against other codes.

  16. For codes that can produce a Hartree–Fock, add lines to the qcengine/program/tests/test_alignment.py to check molecular and properties orientation handling.

  17. If your code is available as a visible binary (e.g., pip, conda, docker download with no or trivial build), create a testing lane by adding to devtools/conda-envs/ and .github/workflows/CI.yml. This will check your code for every PR. We’re looking into private testing for codes that aren’t available.

  18. Throughout, talk with the maintainers with questions. Error handling, especially, is intricate.