A few more small fixes for upstream + CUDA

[kshyatt]

No description provided.

[lkdvos]

This change slightly confuses me, I thought that passing along the array type is what is needed to just make this work for CuArray, why is this no longer the case? Can we also set up test cases to capture this?

[kshyatt]

I think the problem was rather the reverse, that the deleted method was breaking things for ones(Float64, ...). There's still a CUDA specific specialization in the extension

[lkdvos]

I don't think this is true? There are specializations for CUDA.ones, but this is a different function !== Base.ones.

I would expect something like CUDA.ones(Float64, spaces) == Base.ones(CuVector{Float64}, spaces), but I think this latter method now is no longer defined?

[kshyatt]

The latter function (Base.ones accepting a CuArray) doesn't exist, no (neither does Base.zeros). We can add support for it in the extension, I guess? It's annoying having to work around the CUDA and Base APIs and their refusal to play nicely together...

[lkdvos]

Wait, it doesn't exist anymore because you delete it in this PR right? This is exactly why I am confused about the change, since I thought that would work previously, simply by how our generic implementation uses tensormaptype(..., TorA) which would end up with a CuTensorMap

[kshyatt]

It's now back in the extension 😇

[kshyatt]

Anyway sorry for my bad explanation, I was running into problems earlier with

Base.$fname(TorA, codomain ← domain) when TorA == Vector{Float64} (for example) because this would try to call Base.ones(Vector{Float64}, n) where n is some integer deep down. Maybe now tensormaptype fixes this?

[lkdvos]

No worries, I think I somehow just didn't expect it should end up calling that, since the endpoint should be this:

TensorKit.jl/src/tensors/tensor.jl

Lines 314 to 318 in 71d6c00

	function Base.$fname(::Type{TorA}, V::TensorMapSpace) where {TorA}
	t = tensormaptype(spacetype(V), numout(V), numin(V), TorA)(undef, V)
	fill!(t, $felt(scalartype(t)))
	return t
	end

Of course, I am just looking at code, which is notoriously a great way of debugging/reasoning about it 😉

[kshyatt]

Let me see if I can't add a bunch of tests to trigger these paths and hopefully that will illuminate things. Worst case I make a 3rd PR later.