Sam Buercklin

The issue here is that `maximum` specifically attempts to convert incompatible types in this instance. Dimensions of `W` and `1/W` are inherently incompatible, there is no conversion between the two...

Would a trait system identifying numerical types as being commutative/associative/etc. be tractable to solve this?

To elaborate: I'm not familiar with just how many rules require commutativity/associativity, but the properties are inferrable at the type level. Checks like `iscommutative` and `isassociative` should essentially compile away,...

I have a `logdet` implementation inspired by @carstenbauer's code above, and it's faster than the CPU implementation for the case I care about: ```julia function logdet!(m) X_d, ipiv_d = CUSOLVER.getrf!(m)...

> ExponentialUtilities.jl's `exponential` should work with CUDA.jl array types. Which method? The recent PR only looks like it adds support for `exp(A)b` by way of `expv`, not `exp(A)` directly. None...

I'm still having issues with it, I'll see if I can narrow it down and open an issue on ExponentialUtilities

I rolled back to `0.1.0` and tested that as well ``` @btime key_collect($bigmat); # 64.064 μs (4 allocations: 78.27 KiB) (1.264 ms (10004 allocations: 390.77 KiB)) ``` So it seems...

I think this might be the only `ExponentialUtilities.jl` change needed. https://github.com/JuliaGPU/CUDA.jl/pull/1532 solves the `ldiv!` with LU decomposition We also need `LinearAlgebra.opnorm1(::CuMatrix)` which is straightforward to get something that works, but...

`sinc` looks to be fixed, thanks!

`sinc` is still not quite there, there's an issue at `0.0` ```julia julia> only(autodiff(Reverse, sinc, Active, Active(1.0))) == cosc(1.0) true julia> only(autodiff(Reverse, sinc, Active, Active(0.0))) == cosc(0.0) false julia> only(autodiff(Reverse,...