Seth Axen

> See #1320, i.e. writing a rule for `cholesky!` won't be sufficient because the stuff in `cholcopy` fails to differentiate. Seems like that's an Enzyme issue that would need to...

FWIW, I think the safest way to go is to only define the rules only for `LAPACK.potrf!`, which are necessary until Enzyme can differentiate through LAPACK code. I don't much...

I agree that since this fixes a correctness issue, it should be merged soon. But I also think that after the `cholmod` issue is fixed, it should be replaced with...

@simsurace EnzymeTestUtils v0.1.7 removes most of the limitations of the testing functions, so that should allow some of the tests here to be simplified (i.e. no problem with Hermitian arguments...

I just tested this PR locally with the following, and both forward- and reverse-mode tests fail: ```julia julia> @testset "cholesky" begin @testset for Te in (Float64,), TS in (Symmetric, Hermitian)...

> cc @sethaxen re (2) since they were playing around with enzyme rules for BLAS. Might you be interested in helping add the cudaMemcpy/cublas/NNlib/etc stuff? Interested, yes, but am still...

Either way, figuring out where the rules should go and at what level they are needed is useful for whoever takes this on.

Thanks! I'll check later. For some reason dotu and dotc started failing and not I think from changes I made. Still trying to isolate the cause.

> bumping @sethaxen if you've had a chance to isolate? No, this turned out to be much more time-consuming than anticipated, and I don't think I'll be able to finish...

`mean` doesn't like accepting a vector of numbers and instead wants a vector of vectors (see e.g. https://github.com/JuliaManifolds/Manifolds.jl/actions/runs/4829265624/jobs/8604094066?pr=594#step:5:771). Do we have a standard workaround for this for the `Circle`?