Support Subarray{<:Adjoint{<:GPUMatrix}}

[kshyatt]

This is needed to support the blocks of an AdjointTensorMap on GPU.

[lkdvos]

Looks good to me, although I would have guessed that the ishermitian fallbacks automatically catch the non-<:StridedMatrix types?

[kshyatt]

Looks good to me, although I would have guessed that the ishermitian fallbacks automatically catch the non-<:StridedMatrix types?

Unfortunately not, at least not from my experiments :(

[lkdvos]

Any idea where that comes from? the strided_ishermitian_* implementations really are restricted to A::StridedMatrix, for example here:

MatrixAlgebraKit.jl/src/common/matrixproperties.jl

Line 81 in e72dca4

ishermitian_exact(A::StridedMatrix; kwargs...) = strided_ishermitian_exact(A, Val(false); kwargs...)

while the default is the simple one:

MatrixAlgebraKit.jl/src/common/matrixproperties.jl

Line 80 in e72dca4

ishermitian_exact(A) = A == A'

So I'm a bit confused how that fails?

[kshyatt]

Sorry, it's for the projection methods that we need this. I agree the fallbacks catch it for the is_ checkers.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
...ixAlgebraKitAMDGPUExt/MatrixAlgebraKitAMDGPUExt.jl	87.30% <100.00%> (ø)
...MatrixAlgebraKitCUDAExt/MatrixAlgebraKitCUDAExt.jl	61.81% <100.00%> (ø)

🚀 New features to boost your workflow:

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

[lkdvos]

These Mooncake tests are definitely taking quite some time, something we might actually have to worry about at some point?

[kshyatt]

Yes, we could save ourselves some hurt by testing fewer algorithms and only testing Float32 and ComplexF64 ?

…

On Wed, Dec 3, 2025 at 1:35 AM Lukas Devos ***@***.***> wrote: *lkdvos* left a comment (QuantumKitHub/MatrixAlgebraKit.jl#108) <#108 (comment)> These Mooncake tests are definitely taking quite some time, something we might actually have to worry about at some point? — Reply to this email directly, view it on GitHub <#108 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAGKJY2JLI44QYJV3B2AOZD37YV4RAVCNFSM6AAAAACN2MFU5CVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTMMBUGUZDCOJYGU> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[lkdvos]

I guess this would make sense, after all the implementations don't depend on the different algorithms anyways. We can also just reduce the size of the arrays a bit, this shouldn't affect the validity of the tests and the finite differences should be faster.

[kshyatt]

I’ll make a PR in the next day or two

…

On Wed, Dec 3, 2025 at 1:39 AM Lukas Devos ***@***.***> wrote: *lkdvos* left a comment (QuantumKitHub/MatrixAlgebraKit.jl#108) <#108 (comment)> I guess this would make sense, after all the implementations don't depend on the different algorithms anyways. We can also just reduce the size of the arrays a bit, this shouldn't affect the validity of the tests and the finite differences should be faster. — Reply to this email directly, view it on GitHub <#108 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAGKJY62NNHPFTG53EJFATL37YWLDAVCNFSM6AAAAACN2MFU5CVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTMMBUGUZDSOJTGU> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[Jutho]

I guess this would make sense, after all the implementations don't depend on the different algorithms anyways. We can also just reduce the size of the arrays a bit, this shouldn't affect the validity of the tests and the finite differences should be faster.

Do you think the size of the arrays matter at this point? Ok, some of the tests might amount to computing the full Jacobian. So for a function such as svd, that maps order(N^2) elements to order(3 * N^2) elements, that is becoming a rather big Jacobian. Still, I would assume that for these AD tests most of the time is actually compilation time?

[lkdvos]

I didn't investigate actually, nor did I check what the current sizes are so I'm mostly just guessing. It felt possible that the finite differences can quickly blow up for large Jacobians, but you are definitely right that the compilation time might just dominate

[kshyatt]

Created #110 to address the mooncake issue