Perhaps avoid validating input dims on AbstractVector{<:AbstractVector{<:Real}}

I would suggest either changing this, or perhaps better, skipping it all together.
https://github.com/JuliaGaussianProcesses/KernelFunctions.jl/blob/ef6d4591b36194fca069d8bc7ae8c1e2ee288080/src/utils.jl#L194

I'm developing kernels that work on pairs of Vector{Float64}, where the kernel essentially sums over all pairwise distances in x and y. For this kernel, the check above does not really make sense, and indeed passes if I construct the kernel matrix like this:

K = kernelmatrix(kernel, x,x)

However it fails if I instead do

K = kernelmatrix(kernel, x,y)

unless length(first(x))==length(first(y)). I realise that I can just create my own type essentially wrapping Vector{Vector{Float64}}, but that seems like unnecessary work, especially since the kernel machinery works beautifully as it is, as long as the inputs pass the validation.

Start at src/utils.jl line 194 and inspect the validation reached by kernelmatrix(kernel, x, y). Reproduce the difference between kernelmatrix(kernel, x, x) and kernelmatrix(kernel, x, y) with Vector{Vector{Float64}} inputs, then determine whether differing inner lengths should be accepted for kernels based on pairwise distances. Done means the intended validation behavior is implemented and covered for this case.