lkdvos · Nov 2, 2026 · Nov 2, 2026 · Nov 2, 2026 · Nov 2, 2026 · Nov 3, 2026
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
@@ -20,6 +20,7 @@ jobs:
     uses: "QuantumKitHub/QuantumKitHubActions/.github/workflows/TestGroups.yml@main"
     with:
       fast: ${{ github.event.pull_request.draft == true }}
+      exclude: '["testsuite"]'
       nthreads: 4
       timeout-minutes: 90
       # Enzyme plays a large role here, so also test macOS and Windows on PRs, once ubuntu passes

diff --git a/Project.toml b/Project.toml
@@ -42,56 +42,26 @@ TensorOperationsJLArraysExt = "JLArrays"
 
 [compat]
 AMDGPU = "2"
-Aqua = "0.6, 0.7, 0.8"
-Adapt = "4"
 Bumper = "0.6, 0.7"
 CUDACore = "6"
 ChainRulesCore = "1"
-ChainRulesTestUtils = "1"
-DynamicPolynomials = "0.5, 0.6"
 Enzyme = "0.13.183"
-EnzymeTestUtils = "0.2"
 GPUArrays = "11"
 JLArrays = "0.3"
 LRUCache = "1"
 LinearAlgebra = "1.6"
-Logging = "1.6"
 Mooncake = "0.5"
 PackageExtensionCompat = "1"
-ParallelTestRunner = "2"
 PrecompileTools = "1.1"
 Preferences = "1.4"
 PtrArrays = "1.2"
-Random = "1"
-Strided = "2.6"
+Strided = "2.6.7"
 StridedViews = "0.5"
 TBLIS = "0.3"
-Test = "1"
 TupleTools = "1.6"
 VectorInterface = "0.4.1, 0.5, 0.6, 0.7"
-cuRAND = "6"
 cuTENSOR = "6"
 julia = "1.10"
 
-[extras]
-Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
-Aqua = "4c88cf16-eb10-579e-8560-4a9242c79595"
-Bumper = "8ce10254-0962-460f-a3d8-1f77fea1446e"
-CUDACore = "bd0ed864-bdfe-4181-a5ed-ce625a5fdea2"
-ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
-ChainRulesTestUtils = "cdddcdb0-9152-4a09-a978-84456f9df70a"
-cuRAND = "20fd9a0b-12d5-4c2f-a8af-7c34e9e60431"
-DynamicPolynomials = "7c1d4256-1411-5781-91ec-d7bc3513ac07"
-Enzyme = "7da242da-08ed-463a-9acd-ee780be4f1d9"
-EnzymeTestUtils = "12d8515a-0907-448a-8884-5fe00fdf1c5a"
-JLArrays = "27aeb0d3-9eb9-45fb-866b-73c2ecf80fcb"
-Logging = "56ddb016-857b-54e1-b83d-db4d58db5568"
-Mooncake = "da2b9cff-9c12-43a0-ae48-6db2b0edb7d6"
-ParallelTestRunner = "d3525ed8-44d0-4b2c-a655-542cee43accc"
-Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
-TBLIS = "48530278-0828-4a49-9772-0f3830dfa1e9"
-Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
-cuTENSOR = "011b41b2-24ef-40a8-b3eb-fa098493e9e1"
-
-[targets]
-test = ["Test", "Random", "DynamicPolynomials", "ChainRulesTestUtils", "ChainRulesCore", "cuRAND", "CUDACore", "cuTENSOR", "Aqua", "Logging", "Bumper", "Mooncake", "Enzyme", "EnzymeTestUtils", "Adapt", "JLArrays", "AMDGPU", "TBLIS", "ParallelTestRunner"]
+[workspace]
+projects = ["test"]
diff --git a/ext/TensorOperationsTBLISExt.jl b/ext/TensorOperationsTBLISExt.jl
@@ -157,6 +157,8 @@ function TO.tensortrace!(
         throw(ArgumentError("output tensor must not be aliased with input tensor"))
 
     T = eltype(C)
+    # TBLIS does not handle traces over dimensions of size zero, which only scale `C`
+    any(i -> iszero(size(A, i)), q[1]) && return TO._scale!(C, β)
     einA, einC = trace_labels(p, q)
     Av, Cv = SV(A), SV(C)
     lenA, strideA = tblis_dims(Av)

diff --git a/ext/TensorOperationscuTENSORExt.jl b/ext/TensorOperationscuTENSORExt.jl
@@ -59,16 +59,17 @@ end
 #-------------------------------------------------------------------------------------------
 # A Base wrapper over `CuArray` will first pass via the `select_backend` methods for
 # `AbstractArray` and be converted into a `StridedView` if it satisfies `isstrided`. Hence,
-# we only need to capture `CuStridedView` here.
-function TO.select_backend(::typeof(TO.tensoradd!), ::CuStridedView, ::CuStridedView)
+# we only need to capture `CuStridedView` here. cuTENSOR requires all arguments to have the
+# same element type; other combinations fall back to the `StridedView` methods.
+function TO.select_backend(::typeof(TO.tensoradd!), ::CuStridedView{T}, ::CuStridedView{T}) where {T}
     return cuTENSORBackend()
 end
-function TO.select_backend(::typeof(TO.tensortrace!), ::CuStridedView, ::CuStridedView)
+function TO.select_backend(::typeof(TO.tensortrace!), ::CuStridedView{T}, ::CuStridedView{T}) where {T}
     return cuTENSORBackend()
 end
 function TO.select_backend(
-        ::typeof(TO.tensorcontract!), ::CuStridedView, ::CuStridedView, ::CuStridedView
-    )
+        ::typeof(TO.tensorcontract!), ::CuStridedView{T}, ::CuStridedView{T}, ::CuStridedView{T}
+    ) where {T}
     return cuTENSORBackend()
 end
 
@@ -83,6 +84,9 @@ function TO.tensoradd!(
         α::Number, β::Number,
         backend::cuTENSORBackend, allocator = CUDAAllocator()
     )
+    TO.argcheck_tensoradd(C, A, pA)
+    TO.dimcheck_tensoradd(C, A, pA)
+    isempty(C) && return C
     C_cuda, isview = _custrided(C, allocator)
     A_cuda, = _custrided(A, allocator)
     tensoradd!(C_cuda, A_cuda, pA, conjA, α, β, backend, allocator)
@@ -97,6 +101,11 @@ function TO.tensorcontract!(
         α::Number, β::Number,
         backend::cuTENSORBackend, allocator = CUDAAllocator()
     )
+    TO.argcheck_tensorcontract(C, A, pA, B, pB, pAB)
+    TO.dimcheck_tensorcontract(C, A, pA, B, pB, pAB)
+    isempty(C) && return C
+    # cuTENSOR does not accept dimensions of size zero; contracting over them only scales `C`
+    any(i -> iszero(size(A, i)), pA[2]) && return TO._scale!(C, β)
     C_cuda, isview = _custrided(C, allocator)
     A_cuda, = _custrided(A, allocator)
     B_cuda, = _custrided(B, allocator)
@@ -112,6 +121,13 @@ function TO.tensortrace!(
         α::Number, β::Number,
         backend::cuTENSORBackend, allocator = CUDAAllocator()
     )
+    TO.argcheck_tensortrace(C, A, p, q)
+    TO.dimcheck_tensortrace(C, A, p, q)
+    isempty(C) && return C
+    # cuTENSOR does not accept dimensions of size zero; tracing over them only scales `C`
+    any(i -> iszero(size(A, i)), q[1]) && return TO._scale!(C, β)
+    # a trace without traced indices is a permutation, which cuTENSOR cannot plan as a reduction
+    isempty(q[1]) && return tensoradd!(C, A, p, conjA, α, β, backend, allocator)
     C_cuda, isview = _custrided(C, allocator)
     A_cuda, = _custrided(A, allocator)
     tensortrace!(C_cuda, A_cuda, p, q, conjA, α, β, backend, allocator)

diff --git a/src/implementation/base.jl b/src/implementation/base.jl
@@ -104,6 +104,8 @@ function tensortrace!(
     szA = size(A)
     so = TupleTools.getindices(szA, linearize(p))
     st = prod(TupleTools.getindices(szA, q[1]))
+    # traced dimensions of size zero: there is nothing to add to `β * C`
+    iszero(st) && return _scale!(C, β)
     Ã = reshape(
         PermutedDimsArray(A, (linearize(p)..., linearize(q)...)), (prod(so), st, st)
     )
@@ -141,6 +143,8 @@ function tensortrace!(
     szA = size(A)
     so = TupleTools.getindices(szA, linearize(p))
     st = prod(TupleTools.getindices(szA, q[1]))
+    # traced dimensions of size zero: there is nothing to add to `β * C`
+    iszero(st) && return _scale!(C, β)
     perm = (linearize(p)..., linearize(q)...)
     Atemp = tensoralloc_add(eltype(A), A, (perm, ()), conjA, Val(true), allocator)
     Ã = reshape(permutedims!(Atemp, A, perm), (prod(so), st, st))

diff --git a/src/implementation/strided.jl b/src/implementation/strided.jl
@@ -101,6 +101,9 @@ function tensorcontract!(
     )
     @nospecialize backend allocator
 
+    (Base.mightalias(C, A) || Base.mightalias(C, B)) &&
+        throw(ArgumentError("output tensor must not be aliased with input tensor"))
+
     # standardize input types for compilation time
     α′ = standardize_scalartype(C, α)
     β′ = standardize_scalartype(C, β)

diff --git a/src/utils.jl b/src/utils.jl
@@ -9,3 +9,7 @@ function _kron(Es::NTuple{N, Any}, ba) where {N}
     )
     return tensorproduct(E1, ((1, 2), ()), false, E2, p2, false, p, One(), ba...)
 end
+
+# `C = β * C`, for operations without anything to add to `C`, such as traces or contractions
+# over dimensions of size zero; `β = 0` overwrites `C`, which may contain `NaN`s
+_scale!(C::AbstractArray, β::Number) = iszero(β) ? fill!(C, zero(eltype(C))) : (C .= β .* C)
diff --git a/test/Project.toml b/test/Project.toml
@@ -0,0 +1,57 @@
+name = "TensorOperationsTests"
+
+[deps]
+AMDGPU = "21141c5a-9bdb-4563-92ae-f87d6854732e"
+Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
+Aqua = "4c88cf16-eb10-579e-8560-4a9242c79595"
+Bumper = "8ce10254-0962-460f-a3d8-1f77fea1446e"
+CUDACore = "bd0ed864-bdfe-4181-a5ed-ce625a5fdea2"
+ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
+ChainRulesTestUtils = "cdddcdb0-9152-4a09-a978-84456f9df70a"
+DynamicPolynomials = "7c1d4256-1411-5781-91ec-d7bc3513ac07"
+Enzyme = "7da242da-08ed-463a-9acd-ee780be4f1d9"
+EnzymeTestUtils = "12d8515a-0907-448a-8884-5fe00fdf1c5a"
+JLArrays = "27aeb0d3-9eb9-45fb-866b-73c2ecf80fcb"
+LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
+Logging = "56ddb016-857b-54e1-b83d-db4d58db5568"
+Mooncake = "da2b9cff-9c12-43a0-ae48-6db2b0edb7d6"
+ParallelTestRunner = "d3525ed8-44d0-4b2c-a655-542cee43accc"
+Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
+Strided = "5e0ebb24-38b0-5f93-81fe-25c709ecae67"
+TBLIS = "48530278-0828-4a49-9772-0f3830dfa1e9"
+TensorOperations = "6aa20fa7-93e2-5fca-9bc0-fbd0db3c71a2"
+Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
+TupleTools = "9d95972d-f1c8-5527-a6e0-b4b365fa01f6"
+VectorInterface = "409d34a3-91d5-4945-b6ec-7529ddf182d8"
+cuBLAS = "182d3088-87b7-4494-8cad-fc6afaa545bc"
+cuRAND = "20fd9a0b-12d5-4c2f-a8af-7c34e9e60431"
+cuTENSOR = "011b41b2-24ef-40a8-b3eb-fa098493e9e1"
+
+[sources]
+TensorOperations = {path = ".."}
+
+[compat]
+AMDGPU = "2"
+Adapt = "4"
+Aqua = "0.6, 0.7, 0.8"
+Bumper = "0.6, 0.7"
+CUDACore = "6"
+ChainRulesCore = "1"
+ChainRulesTestUtils = "1"
+DynamicPolynomials = "0.5, 0.6"
+Enzyme = "0.13.183"
+EnzymeTestUtils = "0.2"
+JLArrays = "0.3"
+LinearAlgebra = "1.6"
+Logging = "1.6"
+Mooncake = "0.5"
+ParallelTestRunner = "2"
+Random = "1"
+Strided = "2.6.7"
+TBLIS = "0.3"
+Test = "1"
+TupleTools = "1.6"
+VectorInterface = "0.4.1, 0.5, 0.6, 0.7"
+cuBLAS = "6"
+cuRAND = "6"
+cuTENSOR = "6"
diff --git a/test/core/backends/basecopy.jl b/test/core/backends/basecopy.jl
@@ -0,0 +1,15 @@
+using TensorOperations
+using TensorOperations: BaseCopy
+using Test
+
+@isdefined(TestSuite) || include("../../testsuite/TestSuite.jl")
+
+# `BaseCopy` is the reference for all other backends, and is itself checked against loops
+@testset "BaseCopy" begin
+    TestSuite.test_backend(BaseCopy(); reference = TestSuite.LoopReference(), diagonal = true)
+end
+@testset "BaseCopy with generic element types" begin
+    TestSuite.test_backend(
+        BaseCopy(); reference = TestSuite.LoopReference(), eltypes = (BigFloat, Complex{BigFloat})
+    )
+end
diff --git a/test/core/backends/baseview.jl b/test/core/backends/baseview.jl
@@ -0,0 +1,9 @@
+using TensorOperations
+using TensorOperations: BaseView
+using Test
+
+@isdefined(TestSuite) || include("../../testsuite/TestSuite.jl")
+
+@testset "BaseView" begin
+    TestSuite.test_backend(BaseView(); diagonal = true)
+end
diff --git a/test/core/backends/default.jl b/test/core/backends/default.jl
@@ -0,0 +1,9 @@
+using TensorOperations
+using TensorOperations: DefaultBackend
+using Test
+
+@isdefined(TestSuite) || include("../../testsuite/TestSuite.jl")
+
+@testset "DefaultBackend" begin
+    TestSuite.test_backend(DefaultBackend(); diagonal = true)
+end
diff --git a/test/core/backends/strided.jl b/test/core/backends/strided.jl
@@ -0,0 +1,20 @@
+using TensorOperations
+using TensorOperations: StridedBLAS, StridedNative, ManualAllocator, BufferAllocator
+using Test
+
+@isdefined(TestSuite) || include("../../testsuite/TestSuite.jl")
+
+@testset "StridedNative" begin
+    TestSuite.test_backend(StridedNative(); diagonal = true, aliasing = true)
+end
+@testset "StridedBLAS" begin
+    TestSuite.test_backend(StridedBLAS(); aliasing = true)
+end
+@testset "$backend with generic element types" for backend in (StridedNative(), StridedBLAS())
+    TestSuite.test_backend(backend; eltypes = (BigFloat, Complex{BigFloat}))
+end
+@testset "$backend with allocator $allocator" for backend in (StridedNative(), StridedBLAS()),
+        allocator in (ManualAllocator(), BufferAllocator())
+
+    TestSuite.test_tensorcontract(backend; allocator, eltypes = (Float64, ComplexF64))
+end