[devmotion]

An experiment to replace the JuliaInterpreter-based test setup (from #500) with plain compiled tests run in parallel by ParallelTestRunner.jl.

Commits

1. Require Julia 1.10 and remove pre-1.10 code paths. ParallelTestRunner needs Julia ≥ 1.10, so this raises julia compat to 1.10 (the current LTS) and bumps the version to 1.74.0. It removes version checks that are now always true or always false, along with the pre-1.7 logabsdet(::UmfpackLU) backport and therefore the SuiteSparse dependency. CI tests lts instead of 1.6.
2. Run tests with ParallelTestRunner.
   • Each test file runs in its own sandbox module on a worker pool and loads the packages it uses itself.
   • The shared helpers (Multiplier, @gpu, …) move into a ChainRulesTestHelpers module that is loaded once per worker, together with the Adapt piracy for the mock GPU array.
   • fastmath_able.jl no longer needs its own @interpret wrapper.
3. Apply format suggestions to the new test/testutils.jl and to the lines changed in this PR.
4. Upload coverage to Codecov in the CI and nightly workflows, as in ChainRulesCore.jl. The README now shows a Codecov badge and a CI badge for main, and no longer has the Travis badge.
5. Check depwarn rules and wider-γ 3-arg * gradients in tests. The Base.depwarn rule tests now use @test_deprecated, so the deprecation warning no longer shows up in the test output. Three gradient comparisons in arraymath.jl that were never wrapped in @test are now actually checked.
6. Set workflow concurrency and cache permissions. CI, nightly and integration tests cancel superseded PR runs with concurrency, as in ChainRulesCore.jl, which replaces the Cancel workflow. CI and nightly grant actions: write and otherwise only contents: read, so julia-actions/cache can delete its old caches without relying on the repository's default token permissions.
7. Skip non-finite power gradients in FastMath tests. FastMath functions assume that no value is NaN or Inf, so non-finite results of the FastMath ^ rules are undefined. Since LLVM 23 (Upgrade to LLVM 23 JuliaLang/julia#63626), they are folded to arbitrary values, which made the now compiled POWERGRADS tests fail on nightly. In the FastMath variant, the tests no longer check expected values that are NaN or Inf.

Notes

• test/rulesets/Core/core.jl was never included before. It is discovered and run now. Its scalar ifelse rrule test needed check_inferred=false, because the pullback's return type depends on the value of the condition.
• Tests are now selected by path prefix rather than substring, e.g. Pkg.test("ChainRules"; test_args=["rulesets/LinearAlgebra"]). A leading ! excludes tests, and --list lists them.
• The @gpu reverse tests in test/rulesets/Base/array.jl fail because GPUArrays 12.1.1 was released before the JLArrays version it needs (reverse(::JLArray) errors with GPUArrays 12.1.1 and JLArrays 0.4.2 (JLArrays release missing) JuliaGPU/GPUArrays.jl#817). main is affected in the same way; Mark JLArray reverse tests broken with JLArrays 0.4.2 and GPUArrays 12.1.1 #849 marks these tests broken until a fixed JLArrays is released.
• On the previous run, the test jobs took 9–18 min, compared with 30–57 min on main. The default settings fit the runners: 4 workers on 4 vCPUs, and peak worker RSS of about 2 GB, well below the memory limits.

🤖 Generated with Claude Code

[github-actions]

Remaining comments which cannot be posted as a review comment to avoid GitHub Rate Limit

JuliaFormatter

---

[JuliaFormatter] _{reported by reviewdog 🐶}

ChainRules.jl/test/testutils.jl

Line 106 in 706847f

_gpu_test(frule, xdots, f, xs...; kw...)

---

[JuliaFormatter] _{reported by reviewdog 🐶}

ChainRules.jl/test/testutils.jl

Line 110 in 706847f

_gpu_test(frule, xdots, f, g, xs...; kw...)

---

[JuliaFormatter] _{reported by reviewdog 🐶}

ChainRules.jl/test/testutils.jl

Line 140 in 706847f

Multiplier_pullback(dΩ) = (Tangent{typeof(m)}(; x = dΩ * y'), m.x' * dΩ)

---

[JuliaFormatter] _{reported by reviewdog 🐶}

ChainRules.jl/test/testutils.jl

Lines 144 to 148 in 706847f

	Multiplier_pullback_3(dΩ) = (
	Tangent{typeof(m)}(; x = dΩ * (y * z)'),
	m.x' * dΩ * z',
	(m.x * y)' * dΩ,
	)

---

[JuliaFormatter] _{reported by reviewdog 🐶}

ChainRules.jl/test/testutils.jl

Line 170 in 706847f

Divider_pullback(dΩ) = (Tangent{typeof(d)}(; x = -2 * dΩ * y^2 / d.x^3), 2 * dΩ * y / d.x^2)

---

[JuliaFormatter] _{reported by reviewdog 🐶}

ChainRules.jl/test/testutils.jl

Line 191 in 706847f

(c::Counter)(x) = x * (c.n += 1)

---

[JuliaFormatter] _{reported by reviewdog 🐶}

ChainRules.jl/test/testutils.jl

Line 205 in 706847f

struct NoRules; end

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (main@5f30885). Learn more about missing BASE report.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #848   +/-   ##
=======================================
  Coverage        ?   96.57%           
=======================================
  Files           ?       28           
  Lines           ?     3619           
  Branches        ?        0           
=======================================
  Hits            ?     3495           
  Misses          ?      124           
  Partials        ?        0           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[devmotion]

I checked all changes (with whitespace changes hidden). Seems fine to me.