Michael Abbott

I don't swear this is better, it just needs some thought. Are there any paths by which `Partials{3, ComplexF64}` can result in a Dual{Complex} instead of Complex{Dual}? That would be...

The new thing on 1.10 is methods with `AbstractVecOrMat{

Note that I can't reproduce this, I get `1.0f0`. If details of the GPU change how `mapreduce` gets done, then I would generically expect small floating-point differences. Whether 100 eps...

> `rmul!` or `sum(f, A; dims) .*= λ` Indeed. Maybe https://github.com/JuliaGPU/GPUArrays.jl/pull/443 was concerned about making more kernel launches? Never very sure what's going to be fast or not in GPU...

`Matrix(I, 3, 4)` allows non-square `I`. It's a bit unfortunate that `A = rand(3,4); [A I; I A]` already does something, but rows not blocks, it's like `vcat([A I], [I...

It's not so complicated to make `CartesianIndices` do this. ```julia tr5(A::AbstractArray; dims=:) = _mytr(dims, A) function _mytr(dims::Tuple{Integer, Vararg{Integer}}, A::AbstractArray) dimaxes = map(d -> axes(A,d), dims) allequal(dimaxes) || _tr_error(dimaxes) mask =...

That PR changed the definition to `X * Y .+ z`, with the restriction that the answer not be larger than `X * Y` alone. The motivating case (as in...

Note that the linked example has `x1 = randn32(28*28, 1)` i.e. Float32 data, which avoids the issue -- matching change 2. Edit: I can reproduce this on Enzyme v0.13.21, but...

Do any of my other tests trigger this error for you? For me it also works on 1.11: ```julia julia> reproducer() ((layers = ((weight = Float32[-0.00023009 -0.0014316047 … -0.0033742953 0.0007075156;...

I see failure again on another machine, intel linux but libLLVM-16.0.6 Somehow this must have to do with logging macro `@warn`. Reference is to [this line](https://github.com/JuliaLang/julia/blob/v1.11.2/base/logging/logging.jl#L388): ``` [3] macro expansion...