[lkdvos]

Adds a reusable test suite for TensorOperations backends, in the spirit of the GPUArrays.jl and MatrixAlgebraKit.jl test suites, and fixes the bugs it found in the backends. Stacked on #320.

Test suite

test/testsuite/TestSuite.jl tests tensoradd!, tensortrace!, tensorcontract! and their allocating versions for a given backend:

include(joinpath(pkgdir(TensorOperations), "test", "testsuite", "TestSuite.jl"))
TestSuite.test_backend(MyBackend(); arraytype = CuArray, eltypes = (Float32, ComplexF32))

• Reference results: every result is compared against a reference backend, BaseCopy() by default, run on CPU copies of the same inputs, so GPU backends are checked against CPU results. BaseCopy itself is checked against straightforward loop implementations (TestSuite.LoopReference()).
• Capabilities: the TestSuite.Options fields describe what a backend supports: views, negative strides, Adjoint/Transpose inputs, size-zero dimensions, Diagonal arguments, rejecting aliased outputs, and the combinations of mixed element types to test (() for none).
• Edge cases:
  • C filled with NaN whenever β is zero, to check strong-zero semantics, for both Zero() and numeric zeros
  • One()/Zero() and a zero α
  • zero-dimensional arrays, and dimensions of size zero and one
  • full traces and traces without traced indices
  • inner, outer and full contractions
  • layouts that cannot be fused into a matrix multiplication
  • strided views, including negative strides
  • Adjoint/Transpose inputs, combined with conjA
  • Diagonal arguments, including diagonal-times-diagonal into a Diagonal
  • mixed element types and BigFloat
  • invalid indices and sizes, and aliased outputs
  • inputs are checked to stay unchanged
• Where it runs:
  • test/core/backends/: BaseCopy, BaseView, StridedNative, StridedBLAS (also with the manual and buffer allocators) and DefaultBackend.
  • extensions/tblis.jl: TBLIS.
  • extensions/gpuarrays.jl: the Strided backends on JLArray/CuArray, and cuTENSOR and DefaultBackend on CuArray.
• Removed duplicates: the per-operation tests of methods.jl, tblis.jl and gpuarrays.jl that the suite covers.

Bugs found and fixed

Backend	Bug
BaseCopy, BaseView	trace over a size-zero dimension threw a BoundsError
TBLIS	trace over a size-zero dimension returned NaN or an unscaled result
cuTENSOR	no index or size checks: an invalid permutation was silently accepted
cuTENSOR	errors on size-zero dimensions, and on a trace without traced indices
StridedNative	contraction did not reject an output that aliases an input
DefaultBackend on CuArray	selected cuTENSOR for mixed element types, which cuTENSOR does not support; it now falls back to the Strided backends

The size-zero cases with an output that may not be read are handled through a shared _scale! helper, which overwrites C for β = 0.

Not supported, and switched off through the options: mixed element types and negative strides for cuTENSOR, mixed element types for TBLIS, and Diagonal arguments for StridedBLAS.

Other changes

• Test dependencies: moved to test/Project.toml (with [workspace] and [sources]), as in TensorKit and MatrixAlgebraKit.
• cuBLAS: added as a test dependency, since StridedBLAS on CuArray needs Strided's cuBLAS extension.
• CI: ci.yml excludes testsuite from the test groups, and runtests.jl skips the suite's own files.

Strided

Requires Strided v2.6.7 (QuantumKitHub/Strided.jl#100): before that, StridedBLAS contractions over a size-zero dimension left NaNs for β = 0 (conjugated complex and non-BLAS element types), and failed on CuArray. The compat bound is raised to 2.6.7 in both Project.toml and test/Project.toml.

Testing

All suites pass locally with Julia 1.13, including on a local GPU (the core/backends/strided.jl suite also rerun against the released Strided v2.6.7). Each file was run in its own module through a scratch script, not through runtests.jl:

Test file	Pass
core/backends/basecopy.jl	10294
core/backends/baseview.jl	6834
core/backends/strided.jl	27972
core/backends/default.jl	6834
core/methods.jl	92
extensions/tblis.jl	6417
extensions/gpuarrays.jl (JLArray and CuArray)	20398, plus 2 already marked broken

The suite is compile-heavy: the core group takes noticeably longer than before.

🤖 Generated with Claude Code