Warning from LuxLib when using OneHotArrays about Mixed Precision

[oxinabox]

What is going wrong with this:

julia> 
       using OneHotArrays
       using Lux
       using Random
       
       embed = Dense(5=>2)
       ps, st = Lux.setup(Xoshiro(1), embed)
       xs = onehotbatch("aabc", "abcde")
       ys, _ = embed(xs, ps, st)
┌ Warning: Mixed-Precision `matmul_cpu_fallback!` detected and Octavian.jl cannot be used for this set of inputs (C [Matrix{Float32}]: A [Matrix{Float32}] x B [OneHotMatrix{UInt32, Vector{UInt32}}]). Falling back to generic implementation. This may be slow.
└ @ LuxLib.Impl ~/.julia/packages/LuxLib/ru5RQ/src/impl/matmul.jl:145
(Float32[-0.50204086 -0.50204086 -0.5297963 -0.5723163; -0.93569875 -0.93569875 -0.50831497 0.033260167], NamedTuple())

OneHotArrays should be really fast, since it should convert multiplication into indexing.
But I guess it's overloaded matrix multiplication isn't being hit, or Lux is failing to detect that it is being hit.

It is used in the Lux docs a bunch, so i thought it would be well supported.