Tests

Merge pull request #177 from isayevlab/feat/conda-no-aimnet #198

Workflow file for this run

	name: Tests

	on:
	push:
	branches: [main]
	pull_request:

	concurrency:
	group: tests-${{ github.ref }}
	cancel-in-progress: true

	jobs:
	test:
	name: pytest (py${{ matrix.python }}, ani=${{ matrix.ani }})
	runs-on: ubuntu-latest
	strategy:
	fail-fast: false
	matrix:
	python: ["3.11", "3.12"]
	ani: [false]
	include:
	- python: "3.12"
	ani: true
	steps:
	- uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4
	- uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5
	with:
	python-version: ${{ matrix.python }}
	- name: Install
	run: \|
	python -m pip install --upgrade pip
	pip install -e ".[ase,dev]" # dev provides pytest/pytest-cov
	- name: Install ani extra
	if: matrix.ani
	run: pip install -e ".[ani]"
	- name: Verify every test module imports
	# --continue-on-collection-errors on the fast run hides a module that
	# fails to import, so assert collection cleanliness separately.
	run: pytest tests/ --collect-only -q
	- name: Run fast tests
	run: pytest tests/ -q -m "not slow" --continue-on-collection-errors

	# Here rather than in `lint`, because this job is the only one that
	# installs the dependencies. mypy resolves a third-party import to `Any`
	# when the package is absent, so a dep-free run cannot see a wrong tensor,
	# array or ASE argument at all -- it reported 40 errors where a run with
	# the extras present reports 66, and the 26 it missed are precisely the
	# ones a type checker is worth having here for.
	#
	# Only on the `ani` leg: three near-identical reports are noise, and this
	# is the leg with the most complete type information. Running it under
	# 3.12 does not weaken the floor -- `python_version = "3.11"` in
	# pyproject.toml is what decides the semantics checked, not the
	# interpreter.
	- name: mypy (non-blocking until type debt is cleared)
	if: matrix.ani
	run: mypy src/Auto3D/ \|\| true

	no-aimnet:
	# The conda-forge package will ship Auto3D's dependency set minus aimnet
	# plus torchani and ase -- aimnet cannot be packaged for conda-forge (its
	# dependency nvalchemi-toolkit-ops is pip-only). conda-recipe/meta.yaml is
	# updated to that shape as part of the 3.1.1 release. This job emulates
	# that environment and is the enforcement mechanism for the no-aimnet
	# path: without it, the path rots the first time someone adds a
	# module-level `import aimnet`.
	#
	# `pip uninstall -y aimnet` below leaves aimnet's own transitive
	# dependencies installed, so this emulation cannot catch code that leans
	# on a package that only arrives via aimnet's dependency tree -- the
	# genuine conda build is the true gate for that.
	name: no-aimnet (conda package environment)
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4
	- uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5
	with:
	python-version: "3.12"
	- name: Install, then remove aimnet
	run: \|
	python -m pip install --upgrade pip
	pip install -e ".[ani,ase,dev]"
	pip uninstall -y aimnet
	- name: Bare import works without aimnet
	run: python -c "import Auto3D"
	- name: Whole-suite collection stays clean without aimnet
	run: pytest tests/ --collect-only -q
	- name: No-aimnet behavior tests
	run: pytest tests/test_no_aimnet.py -q
	- name: ANI2xt still constructs (bundled weights, no download)
	run: pytest "tests/test_model_factory.py" -q -k "ani2xt"

	slow:
	name: slow tests (NNP integration)
	runs-on: ubuntu-latest
	# These tests are NNP inference on CPU, so their wall time tracks whatever
	# share of the runner we get. Three consecutive runs of the same commit
	# took 6m21s, 17m57s and 6m52s: the middle one exceeded the old 20-minute
	# ceiling and was cancelled after all 58 tests had passed, reporting a
	# red slow tier for code that was fine. A job that fails on runner luck
	# teaches people to ignore it, which is worse than a job that is slow.
	# 45 gives roughly 2.5x headroom over the worst observed run while still
	# catching a genuine hang.
	timeout-minutes: 45
	steps:
	- uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4
	- uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5
	with:
	python-version: "3.12"
	- name: Cache downloaded model weights
	uses: actions/cache@0057852bfaa89a56745cba8c7296529d2fc39830 # v4
	with:
	# AIMNet2 weights download to ~/.cache/aimnet; torchani/torch.hub
	# weights (ANI2x) land in ~/.cache/torch. Cache both so the slow
	# suite does not re-download on every run.
	path: \|
	~/.cache/aimnet
	~/.cache/torch
	key: nnp-model-cache-v1
	- name: Install
	run: \|
	python -m pip install --upgrade pip
	# ase for thermo, ani for the torchani-gated slow tests, dev for pytest
	pip install -e ".[ase,ani,dev]"
	- name: Warm the AIMNet2 model cache (hard failure if unreachable)
	# An unavailable gate must never be reportable as a passing gate. If the
	# registry or network is down, fail here loudly rather than letting the
	# suite skip or silently pass (audit: the remediation gate depends on
	# this model being present).
	run: \|
	python - <<'PY'
	import sys
	import torch
	from Auto3D.engines.model_factory import create_model
	try:
	create_model("AIMNET", torch.device("cpu"))
	except Exception as exc:
	print(f"::error::AIMNet2 model unavailable, gate cannot run: {exc}")
	sys.exit(1)
	print("AIMNet2 model cache warm")
	PY
	- name: Run slow NNP + pipeline tests
	# Previously limited to three files because the heavy end-to-end modules
	# were flaky under combined ordering. test_auto3D.py now runs every
	# pipeline test through the job_dir/isolated_input fixtures and
	# test_thermo.py stages its opt_geometry inputs in tmp_path, so those
	# two no longer write job directories or output SDFs into
	# tests/files/. The remaining three (test_SPE.py, test_isomer_engine.py,
	# test_tauto.py) still do not use the fixtures, so for them the
	# protection is unchanged: main()'s microsecond-resolution job naming,
	# strictly serial CI execution, and the autouse GPU-teardown fixture
	# applied via the `slow` marker -- assessed at 70-80% confidence, not
	# guaranteed (audit M31).
	run: >
	pytest tests/ -m slow -q

	# Executes one notebook for real. Here rather than in the docs workflow
	# because every one of the 20 needs a neural network potential, and this
	# is the only job with the model cache and the `ani` extra.
	#
	# One, and the selection was measured on a CPU box with CUDA hidden --
	# which is what a runner is -- rather than on a machine with a GPU. The
	# dividing line is not notebook size, it is whether the notebook runs an
	# optimization:
	#
	# single_point_energy (calc_spe, 4 records) 5.9 s
	# geometry_optimization (opt_geometry, 4 records) > 600 s
	#
	# The same optimization notebook takes 74 s on a GPU, which is how it
	# first looked affordable. Every other notebook either optimizes or runs
	# `main()` end to end, so all of them land on the wrong side of that line;
	# those paths are covered by the slow test tier above instead.
	#
	# What this closes: this notebook calling a function Auto3D no longer has,
	# or passing an argument it no longer accepts, now fails here. What it
	# does not: the other 19 are still parse-checked only, so the guard in
	# tests/test_notebook_properties.py remains the only thing covering all
	# 20 -- and it catches a wrong SD property name, not a wrong call.
	- name: Execute the single-point notebook
	run: \|
	pip install nbmake
	pytest --nbmake --nbmake-timeout=600 example/single_point_energy.ipynb

	lint:
	name: ruff
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4
	- uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5
	with:
	python-version: "3.12"
	# ruff needs no dependencies installed; mypy does, so it runs in the
	# `test` job instead. See the note on that step.
	# Pinned exactly, and matching the `dev` extra in pyproject.toml, because
	# `ruff format --check` is a gate: ruff promotes formatting-style changes
	# into its stable style across minor releases, with settings unchanged.
	# An unpinned install turns any such release into a red build on whichever
	# unrelated pull request happens to be open that day -- the same
	# fails-on-luck failure mode the slow job's timeout comment describes.
	#
	# Bumping it is a deliberate act: raise both pins together, and if the new
	# version reformats anything, that reformat is its own commit and its own
	# `.git-blame-ignore-revs` entry.
	- name: Install tooling
	run: pip install ruff==0.15.9
	- name: ruff check
	run: ruff check src/Auto3D/ tests/

	# CONTRIBUTING.md has told contributors to run `ruff format src/ tests/`
	# since long before anything checked. Nothing did, so the tree drifted
	# until the documented command produced a 169-file diff -- at which point
	# nobody could follow the instruction on an ordinary pull request, and the
	# drift compounded. The gate is the half that makes the instruction true.
	#
	# Unlike the mypy step, this one does NOT end in `\|\| true`. It starts from
	# zero and the fix is one command, so there is no debt to grandfather.
	- name: ruff format
	run: ruff format --check src/ tests/

Navigation Menu

Merge pull request #177 from isayevlab/feat/conda-no-aimnet #198

Workflow file

Merge pull request #177 from isayevlab/feat/conda-no-aimnet #198

Uh oh!

Workflow file for this run