QA: run_qa v1.6 form + ExplicitImports (root + sublibs)#1237
Draft
ChrisRackauckas-Claude wants to merge 4 commits into
Draft
QA: run_qa v1.6 form + ExplicitImports (root + sublibs)#1237ChrisRackauckas-Claude wants to merge 4 commits into
ChrisRackauckas-Claude wants to merge 4 commits into
Conversation
Convert the root test/qa and every lib/<Sub>/test/qa to SciMLTesting 1.6's
`run_qa(<Module>; explicit_imports = true, ...)` form and enable the six
ExplicitImports checks across the whole monorepo.
Per QA env:
* qa.jl -> `using SciMLTesting, <Module>, JET; using Test` + a single run_qa call.
* test/qa/Project.toml -> SciMLTesting compat "1.6" (+ dep); Aqua and JET kept
(both run); ExplicitImports comes transitively from SciMLTesting. [sources]
wiring preserved.
ExplicitImports: 0 hard FAIL everywhere. no_implicit_imports is the facade
`@reexport`/`using` module-name reliance and is marked ei_broken fleet-wide; the
public-API and via-owners findings are per-env `ignore`-lists of names owned by
SciMLBase/OptimizationBase/the backend/Base that are not (yet) declared public;
stale explicit imports are kept+ignored (re-surfaced API / reached as Mod.name).
PyCMA and SciPy load CodaPkg-backed Python at precompile, so their EI checks are
conservatively ei_broken (verified statically from their source).
Also clears the pre-existing Aqua reds that left the new sublibrary QA lanes red
since SciML#1223 (these fail identically under the old `Aqua.test_all` form):
* deps_compat: add the missing [compat] entries for [extras]/[deps]/[weakdeps]
across the sublibraries (stdlibs -> "1"; others to a permissive floor).
* stale_deps: drop the unused `PDMats` dep from OptimizationBase; move the
test-only `Manifolds` from [deps] to [extras]+test target in OptimizationManopt.
* piracies: each solver-wrapper intentionally extends SciMLBase's
trait/__init/__solve interface onto its backend optimizer types -> declare
those foreign types via `aqua_kwargs = (; piracies = (; treat_as_own = [...]))`.
* undefined_exports: MadNLP/Manopt (`solve!`), Metaheuristics (`summary`) and
NOMAD (`solve`) `@reexport` two packages exporting the same name, leaving it an
unresolved conflict binding -> tracked as `aqua_broken = (:undefined_exports,)`.
Verified against released SciMLTesting 1.6.0 on Julia 1.10 (QA lanes run lts+1,
never pre): root, OptimizationBase, OptimizationBBO, OptimizationMOI,
OptimizationEvolutionary, OptimizationOptimisers, OptimizationOptimJL,
OptimizationManopt, OptimizationMadNLP, OptimizationMetaheuristics and
OptimizationNOMAD all green-or-broken with 0 FAIL; the remaining Aqua sub-checks
(ambiguities/unbound_args/project_extras/persistent_tasks) verified clean across
all 24 loadable sublibraries. CI covers the rest.
Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…eased base libs SciMLTesting 1.7.0 now runs check_all_qualified_accesses_are_public and check_all_explicit_imports_are_public only on Julia >= 1.11 (skipped on the LTS), and the base-lib make-public round shipped (SciMLBase 3.24.0, OptimizationBase 5.1.3, CommonSolve 0.2.9, Symbolics 7.28.1, SymbolicUtils 4.36.0, ModelingToolkit 11.29.0, ArrayInterface 7.26.0). Many per-repo public-API ignore entries are therefore now redundant. Emptied both public-API ignore-lists in every qa env, then re-derived the minimal surviving set by running each QA group on Julia 1.12 against the registered releases (Pkg-resolved, no dev) and restoring only the names the checks still flag as non-public. The flagged names are owned by SciMLBase / OptimizationBase / Base / SciMLStructures / SymbolicUtils / the backend package and remain non-public on the released versions (imported/accessed from a module that has not declared them public); the proper fix is upstream `public` declarations, tracked against SciML/Optimization.jl. Removed entries are names the make-public round turned public (e.g. Success, Failure, MaxIters, MaxTime, build_solution, has_init, NullParameters, setup, update, ...). Verified on Julia 1.12 (public checks run): each QA group's two public checks pass with the trimmed lists. Verified on Julia 1.10 LTS (public checks skipped per SciMLTesting 1.7): root + OptimizationBase + OptimizationOptimisers + OptimizationNLPModels QA groups green-or-broken (15 pass / 1 broken / 0 fail). OptimizationPyCMA and OptimizationSciPy keep all_qualified_accesses_are_public in ei_broken (CondaPkg-gated envs, not analyzable without conda). no_implicit_imports / no_stale_explicit_imports / *_via_owners ignores and aqua/jet markers are untouched. Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com> Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…eleases (SciMLBase 3.27.0) The prior trim was derived against SciMLBase 3.24.0 / OptimizationBase 5.1.3. The full make-public round is now registered (SciMLBase 3.27.0), which turned several more accessed/imported names public — most consistently the solver-trait predicates `allowsbounds` and `allowsconstraints`, plus `build_solution` and the abstract algorithm type `AbstractDAEAlgorithm`. Those entries are now redundant in the per-repo public-API ignore-lists. Re-derived the minimal surviving set for every analyzable qa env by running ExplicitImports' check_all_qualified_accesses_are_public / check_all_explicit_imports_are_public with an empty ignore on Julia 1.12 against the Pkg-resolved registered releases (no dev), and set each ignore-list to exactly the names the checks still flag. OptimizationQuadDIRECT was resolved with its git-only `QuadDIRECT` dependency added so it could be analyzed like CI does. Removed (now public on the resolved SciMLBase 3.27.0): `allowsbounds`, `allowsconstraints` (Auglag/BBO/CMAEvolutionStrategy/Evolutionary/GCMAES/Ipopt/ LBFGSB/MadNLP/Metaheuristics/MOI/MultistartOptimization/NLopt/NOMAD/ODE/OptimJL/ PRIMA/QuadDIRECT/SpeedMapping as applicable), `build_solution` (Optimisers), `AbstractDAEAlgorithm` (ODE). OptimizationBase drops `allowsbounds`/`allowsconstraints` from all_explicit_imports_are_public. Survivors kept are genuinely still non-public on the resolved releases and are owned by SciMLBase (`__init`/`__solve`/`allowscallback`/`requires*`/`MaxSense`/`MinSense`/ `ObjSense`/`NoAD`/`Abstract*`/...), OptimizationBase (`OptimizationStats`/`OptimizationState`/ `_check_and_convert_*`/`ReInitCache`/`instantiate_function`/`supports_sense`/...), Base (`@logmsg`/`LogLevel`/`structdiff`), ForwardDiff (`gradient`), or each backend (BlackBoxOptim/CMAEvolutionStrategy/Evolutionary/Ipopt/MadNLP/Manopt/Metaheuristics/ MathOptInterface/QuadDIRECT/...). The proper fix for the SciMLBase/OptimizationBase survivors is upstream `public` declarations, tracked against SciML/Optimization.jl. OptimizationManopt is unchanged: its dependency tree resolves SciMLBase 2.155.1, where `build_solution`/`allowsbounds` remain non-public, so its list is already minimal there. Verified on Julia 1.12 (public checks run): root + OptimizationBase + OptimizationOptimisers + OptimizationMOI QA groups each pass both public-API EI checks with the trimmed lists (ExplicitImports testset 5 pass / 1 broken / 0 fail — the broken one is the intentional no_implicit_imports). The remaining `JET-test failed` in the OptimizationBase/Optimisers/MOI QA groups is a pre-existing OptimizationBase 5.1.3 source finding (undefined local `lagrangian`/`cons_oop` in OptimizationDISparseExt.jl) and is independent of this EI-only change. no_implicit_imports / no_stale_explicit_imports / *_via_owners ignores and aqua/jet markers are untouched. OptimizationPyCMA and OptimizationSciPy keep their public checks in ei_broken (CondaPkg-gated, not analyzable without conda). Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com> Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The run_qa v1.6 conversion newly runs JET typo-mode and the public-API ExplicitImports checks, surfacing issues the old hand-rolled qa.jl never exercised. Resolve each per its nature: Real fixes (verified JET-clean): * OptimizationGCMAES: hoist the gradient closure `g` out of the `!isnothing(cache.f.grad)` guard. `g` was defined only inside the guard but referenced in the `isnothing(...) ? _loss : (_loss, g)` ternary, which JET flagged as `g may be undefined`. The closure body never calls `cache.f.grad` at construction, so defining it unconditionally is behavior- preserving. * OptimizationSciPy: `safe_get_message` called `pytypeof`, which does not exist in PythonCall (the function is `pytype`). This is a genuine latent UndefVarError on the non-str/non-list message fallback path; JET caught it. EI ignore-list gap: * SimpleOptimization: add the still-non-public `OptimizationStats`, `__solve`, and `_check_and_convert_maxiters` to `all_qualified_accesses_are_public`. Documented jet_broken=true (pre-existing latent issues the conversion merely surfaces; src unchanged from master, fixes belong in separate solver PRs): * OptimizationLBFGSB (`res`), OptimizationOptimisers (`x`), OptimizationSophia (`x`), OptimizationMetaheuristics (`opt_bounds`): loss/result variables assigned only inside a loop/guard and read after. * OptimizationNOMAD (`bb`/`bbcons`): closures defined in correlated if/else arms JET cannot prove correlated. * OptimizationNLopt (`thetacache`/`Jthetacache`): closure-captured-and- reassigned bindings boxed by Julia; benign closure-capture artifact. Verified locally on Julia 1.12.6 (QA lane = julia-1): each affected sublib's QA group passes with 0 fail / 0 error (broken markers as documented). Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Converts this monorepo's QA — the root
test/qaand everylib/<Sub>/test/qa(27 sublibraries) — to SciMLTesting 1.6'srun_qaform and turns on the six ExplicitImports checks fleet-wide.Per QA env
qa.jl→using SciMLTesting, <Module>, JET+using Test+ onerun_qa(<Module>; explicit_imports = true, …)call. The module under test for a sublibrary is that sublibrary's module.test/qa/Project.toml→SciMLTesting = "1.6"(added as dep + compat);AquaandJETkept (both run);ExplicitImportscomes transitively from SciMLTesting (not listed).[sources]develop wiring preserved exactly.ExplicitImports (goal: 0 hard FAIL)
no_implicit_imports: every module is a facade that@reexports SciMLBase/OptimizationBase (+usings helper modules); the implicit-import surface is module names that can't be made explicit without restructuring →ei_broken = (:no_implicit_imports,)fleet-wide.all_qualified_accesses_are_public/all_explicit_imports_are_public/all_qualified_accesses_via_owners: per-envei_kwargsignore-lists of the specific names owned by SciMLBase / OptimizationBase / the backend / Base that are not (yet) declaredpublic. (Closing those out upstream is the standardpublic-declaration follow-up.)no_stale_explicit_imports: stale names kept + ignored (intentionally re-surfaced API; e.g.Optimization.instantiate_functionis reached as a qualified name by the test suite).OptimizationPyCMA/OptimizationSciPyload a CondaPkg-backed Python env at precompile (no network in the harness), so their EI checks are conservativelyei_broken(the failing set verified statically from their source).Pre-existing Aqua reds cleared
The sublibrary QA lanes added in #1223 have been red on
masterfor reasons independent of the harness form (they fail identically under the oldAqua.test_all). This PR fixes them so the converted lanes are green-or-broken:[compat]entries for[extras]/[deps]/[weakdeps]across the sublibraries (stdlibs →"1"; others to a permissive floor consistent with existing repo compat).PDMatsfromOptimizationBase[deps]; move the test-onlyManifoldsfrom[deps]to[extras]+ the test target inOptimizationManopt.__init/__solveinterface onto its backend optimizer types → declare those foreign types viaaqua_kwargs = (; piracies = (; treat_as_own = […]))(root, MOI, Base, Evolutionary, Metaheuristics, MultistartOptimization, NLPModels, NLopt, OptimJL, Optimisers).OptimizationMadNLP/OptimizationManopt(solve!),OptimizationMetaheuristics(summary) andOptimizationNOMAD(solve)@reexporttwo packages that both export the same name, leaving it an unresolved conflict binding → tracked asaqua_broken = (:undefined_exports,)(records Broken, auto-flags if a maintainer later resolves which binding wins).Verification
Run locally against released SciMLTesting 1.6.0 on Julia 1.10 (QA lanes run on
lts+1, neverpre, so JET-on-prerelease is not a concern here). Green-or-broken with 0 FAIL for: root,OptimizationBase,OptimizationBBO,OptimizationMOI,OptimizationEvolutionary,OptimizationOptimisers,OptimizationOptimJL,OptimizationManopt,OptimizationMadNLP,OptimizationMetaheuristics,OptimizationNOMAD. The remaining Aqua sub-checks (ambiguities / unbound_args / project_extras / persistent_tasks) verified clean across all 24 loadable sublibraries. CI covers the rest.Sublibraries converted: 27 (+ root).
🤖 Generated with Claude Code