More tweaks

[kshyatt]

Needed to get more MPSKit examples working

[codecov]

Codecov Report

❌ Patch coverage is 45.83333% with 13 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/tensors/abstracttensor.jl	25.00%	6 Missing ⚠️
ext/TensorKitCUDAExt/cutensormap.jl	0.00%	3 Missing ⚠️
src/tensors/braidingtensor.jl	60.00%	2 Missing ⚠️
src/tensors/adjoint.jl	50.00%	1 Missing ⚠️
src/tensors/indexmanipulations.jl	75.00%	1 Missing ⚠️

Files with missing lines	Coverage Δ
src/tensors/diagonal.jl	89.76% <100.00%> (-2.80%)	⬇️
src/tensors/tensor.jl	83.33% <100.00%> (-1.73%)	⬇️
src/tensors/adjoint.jl	87.50% <50.00%> (-2.50%)	⬇️
src/tensors/indexmanipulations.jl	73.41% <75.00%> (-0.53%)	⬇️
src/tensors/braidingtensor.jl	68.78% <60.00%> (-1.61%)	⬇️
ext/TensorKitCUDAExt/cutensormap.jl	70.83% <0.00%> (-3.84%)	⬇️
src/tensors/abstracttensor.jl	54.47% <25.00%> (-0.83%)	⬇️

... and 23 files with indirect coverage changes

🚀 New features to boost your workflow:

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

[kshyatt]

Let's make this a draft too to cut down on CI thrash

[lkdvos]

Left some comment throughout, there are some things that I am not entirely convinced by but the rest looks great, thanks for working through all of this!

For the similarstoragetype(tensor, storagetype) calls that you added, this seems like something we should probably discuss over a separate PR, and it would be great if we could consolidate this one to get the remainder of the fixes in.
Would you be up for splitting these two things, and then getting this merged?

The same kind of holds for some of the comments I made too, if we can just postpone the things that are not obvious, but already get the other parts in, that would probably be helpful.

(Note that I am very much aware that none of this is your fault and this PR has lived for too long so the design shifts a bit, for which I do apologize!)

[lkdvos]

I guess this effectively means that we are deciding to promote inputs to the storagetype of the output. I'm not sure if I am fully convinced that we should solve this automatically at all, since I think that is also inconsistent with how regular matrices work (same for adding):

julia> CUDA.rand(2, 2) * rand(Float32, 2, 2)
ERROR: Scalar indexing is disallowed.

I do think that this might be the right approach, and requiring explicit conversions in the cases of mixed inputs seems like the right call to me. (Even though I can see how that is annoying for MPSKit 😉 )

[kshyatt]

No, it's for

    TTC = storagetype(C)                                                                                                                                                                                   
    # Bring A in the correct form for BLAS contraction                                                                                                                                                     
    if copyA                                                                                                                                                                                               
        Anew = TO.tensoralloc_add(TTC, A, pA, false, Val(true), allocator)                                                                                                                                 
        Anew = TO.tensoradd!(Anew, A, pA, false, One(), Zero(), backend, allocator)                                                                                                                        
        twistA && twist!(Anew, filter(!isdual ∘ Base.Fix1(space, Anew), domainind(Anew)))                                                                                                                  
    else                                                                                                                                                                                                   
        Anew = permute(A, pA)                                                                                                                                                                              
    end

Without this change, Anew will always have Vector{scalartype(T)} storage even if A was a BraidingTensor or some other object that only gets instantiated here. With the changes in #393 this won't be necessary. It's more than "annoying", with this change or #393 you have to define new tensoralloc methods for the mixed case, it's quite painful 😭

[kshyatt]

It's completely fine!! This has stayed open as I work through adding more tests for MPSKit, so I think we can pare off the simpler stuff we agree on, and then discuss things that are more contentious.

[github-actions]

Your PR requires formatting changes to meet the project's style guidelines.
Please consider running Runic (git runic main) to apply these changes.

Click here to view the suggested changes.

diff --git a/test/cuda/tensors.jl b/test/cuda/tensors.jl
index 16d1030..0bdeb60 100644
--- a/test/cuda/tensors.jl
+++ b/test/cuda/tensors.jl
@@ -568,8 +568,8 @@ end
             d4 = dim(domain(t2))
             At = convert(Array, adapt(Vector{T}, t))
             @test reshape(At, (d1, d2, d3, d4)) ≈
-            reshape(convert(Array, adapt(Vector{T}, t1)), (d1, 1, d3, 1)) .*
-            reshape(convert(Array, adapt(Vector{T}, t2)), (1, d2, 1, d4))
+                reshape(convert(Array, adapt(Vector{T}, t1)), (d1, 1, d3, 1)) .*
+                reshape(convert(Array, adapt(Vector{T}, t2)), (1, d2, 1, d4))
         end
     end
 end