Fix Runic format check: add return in test/runtests.jl core_body() (#218)

* Apply Runic formatting to test/runtests.jl

The Runic Format Check on main fails because core_body()'s final
expression (the Layers Tests @safetestset) is missing the explicit
return that Runic requires. Add it to make the format check pass.

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* Restore LuxCUDA test dependency dropped during v1.2 CI conversion

The grouped-tests v1.2 conversion (#217) dropped LuxCUDA from the test
[extras]/[targets]. The self-hosted GPU lane sets BACKEND_GROUP=CUDA, which
makes test/shared_testsetup.jl do `using LuxCUDA`; without the dependency the
GPU job errors with `ArgumentError: Package LuxCUDA not found in current path`.

Re-add LuxCUDA (compat 0.3) to [compat], [extras], and the test target, using
the correct registry UUID d0bbae9a-e099-4d5b-a835-1c6931763bda (the
pre-conversion Project.toml had carried a typo'd UUID ...1c6931f17571).

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* Fix MNIST docs build: steady-state adjoint VJP shape + missing VCAB3 import

The Documentation build failed on two genuine bugs in the basic MNIST DEQ
tutorial (both reproduce on `main`, independent of the test/Project.toml
changes in this PR):

1. `DimensionMismatch` in the steady-state adjoint VJP. For a conv DEQ the
   rootfind state is a 4D array (e.g. `(13, 13, 64, 2)`). When the state has
   more than 50 elements, SciMLSensitivity's `SteadyStateAdjoint` takes the
   matrix-free `VecJacOperator` path and builds the operator from `vec(y)`, so
   it seeds the residual's pullback with a *flat* cotangent. Zygote's
   `ProjectTo` then rejects that flat cotangent when it is broadcast against the
   multi-dimensional `u` in the residual `y .- u`, raising
   `variable with size(x) == (13,13,64,2) cannot have a gradient with
   size(dx) == (21632,)`. The Core test suite never hit this because its conv
   states are <= 50 elements (dense-Jacobian path). Fix: compute the residual in
   flattened space, `reshape(vec(y) .- vec(u), size(u))` — a value-preserving
   no-op that keeps the broadcast operands consistent with the flat cotangent.

2. `UndefVarError: VCAB3`. The tutorial calls `VCAB3()` but only imports
   `OrdinaryDiffEq`, which (v7) no longer re-exports `VCAB3`. Added
   `OrdinaryDiffEqAdamsBashforthMoulton` to the docs project and an explicit
   `using OrdinaryDiffEqAdamsBashforthMoulton: VCAB3` in the tutorial.

Verified locally on Julia 1.12:
- Full `docs/make.jl` builds end-to-end (both MNIST `@example` blocks run; only
  a transient ColPrac linkcheck 429 from the sandbox differs from CI).
- `GROUP=Core Pkg.test()` passes: Utils 11/11, Layers 1538/1538.

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: ChrisRackauckas-Claude <accounts@chrisrackauckas.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: 3 people

Project.toml

@@ -1,7 +1,7 @@
 name = "DeepEquilibriumNetworks"
 uuid = "6748aba7-0e9b-415e-a410-ae3cc0ecb334"
 authors = ["Avik Pal <avikpal@mit.edu>"]
-version = "2.7.0"
+version = "2.7.1"
 
 [deps]
 ADTypes = "47edcb42-4c32-4615-8424-f2b9edc5f35b"
@@ -37,6 +37,7 @@ GPUArraysCore = "0.1, 0.2"
 SafeTestsets = "0.1"
 InteractiveUtils = "<0.0.1, 1"
 Lux = "1"
+LuxCUDA = "0.3"
 LuxCore = "1"
 LuxTestUtils = "1, 2"
 MLDataDevices = "1"
@@ -67,6 +68,7 @@ Functors = "d9f16b24-f501-4c13-a1f2-28368ffc5196"
 GPUArraysCore = "46192b85-c4d5-4398-a991-12ede77f4527"
 SafeTestsets = "1bc83da4-3b8d-516f-aca4-4fe02f6d838f"
 InteractiveUtils = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
+LuxCUDA = "d0bbae9a-e099-4d5b-a835-1c6931763bda"
 LuxTestUtils = "ac9de150-d08f-4546-94fb-7472b5760531"
 MLDataDevices = "7e8f7934-dd98-4c1a-8fe8-92b47a384d40"
 NLsolve = "2774e3e8-f4cf-5e23-947b-6d7e65073b56"
@@ -80,4 +82,4 @@ Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [targets]
-test = ["Aqua", "Documenter", "ExplicitImports", "ForwardDiff", "Functors", "GPUArraysCore", "InteractiveUtils", "LuxTestUtils", "MLDataDevices", "NLsolve", "NonlinearSolve", "OrdinaryDiffEq", "OrdinaryDiffEqAdamsBashforthMoulton", "SafeTestsets", "SciMLSensitivity", "SciMLTesting", "StableRNGs", "Test", "Zygote"]
+test = ["Aqua", "Documenter", "ExplicitImports", "ForwardDiff", "Functors", "GPUArraysCore", "InteractiveUtils", "LuxCUDA", "LuxTestUtils", "MLDataDevices", "NLsolve", "NonlinearSolve", "OrdinaryDiffEq", "OrdinaryDiffEqAdamsBashforthMoulton", "SafeTestsets", "SciMLSensitivity", "SciMLTesting", "StableRNGs", "Test", "Zygote"]

docs/Project.toml

@@ -12,6 +12,7 @@ NonlinearSolve = "8913a72c-1f9b-4ce2-8d82-65094dcecaec"
 OneHotArrays = "0b1bfda6-eb8a-41d2-88d8-f5af5cad476f"
 Optimisers = "3bd65402-5787-11e9-1adc-39752487f4e2"
 OrdinaryDiffEq = "1dea7af3-3e70-54e6-95c3-0bf5283fa5ed"
+OrdinaryDiffEqAdamsBashforthMoulton = "89bda076-bce5-4f1c-845f-551c83cdda9a"
 Printf = "de0858da-6303-5e67-8744-51eddeeeb8d7"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 SciMLSensitivity = "1ed8b502-d754-442c-8d5d-10ac956f44a1"
@@ -31,6 +32,7 @@ NonlinearSolve = "3, 4"
 OneHotArrays = "0.2"
 Optimisers = "0.3, 0.4"
 OrdinaryDiffEq = "6, 7"
+OrdinaryDiffEqAdamsBashforthMoulton = "1, 2"
 Random = "1"
 SciMLSensitivity = "7"
 Setfield = "1"

docs/src/tutorials/basic_mnist_deq.md

@@ -5,6 +5,7 @@ We will train a simple Deep Equilibrium Model on MNIST. First we load a few pack
 ```@example basic_mnist_deq
 using DeepEquilibriumNetworks, SciMLSensitivity, Lux, NonlinearSolve, OrdinaryDiffEq,
       Random, Optimisers, Zygote, LinearSolve, Dates, Printf, Setfield, OneHotArrays
+using OrdinaryDiffEqAdamsBashforthMoulton: VCAB3
 using MLDatasets: MNIST
 using MLUtils: DataLoader, splitobs
 using LuxCUDA # For NVIDIA GPU support

src/layers.jl

@@ -128,7 +128,11 @@ function (deq::DEQ)(x, ps, st::NamedTuple, ::Val{false})
         # The type-assert is needed because of an upstream Lux issue with type stability of
         # conv with Dual numbers
         y = model((u, p.x), p.ps)::typeof(u)
-        return y .- u
+        # Subtract in flattened space: the matrix-free SteadyStateAdjoint VJP seeds the
+        # residual pullback with a flat cotangent (it builds the operator from `vec(y)`),
+        # which Zygote's `ProjectTo` rejects when broadcasting against a multi-dim `u`
+        # (e.g. a conv state). `vec`/`reshape` round-trip is a no-op on the value.
+        return reshape(vec(y) .- vec(u), size(u))
     end
 
     prob = construct_prob(deq.kind, ODEFunction{false}(dudt), z, (; ps = ps.model, x))

test/runtests.jl

@@ -7,7 +7,7 @@ using SafeTestsets
 # cannot be expressed as a separate folder of files. Hence explicit-args run_tests.
 function core_body()
     @safetestset "Utils Tests" include("utils_tests.jl")
-    @safetestset "Layers Tests" include("layers_tests.jl")
+    return @safetestset "Layers Tests" include("layers_tests.jl")
 end
 
 run_tests(;