Calling `cpu` on `FastChain` fails.

[adrhill]

I'm writing a script that I want to run on both GPUs and CPUs by specifying a device.
While calling gpu on a FastChain works fine, cpu fails.

I'm running the lastest releases: DiffEqFlux v1.39.0 and Flux v0.12.4.

---

Here is a minimal example:

using DiffEqFlux
using Flux

device = Flux.cpu

# Works
c1 = Chain(Dense(2, 10, tanh)) |> device

# Fails
c2 = FastChain(FastDense(2, 10, tanh))  |> device

and the resulting stacktrace:

ERROR: LoadError: ArgumentError: tuple length should be ≥ 0, got -3
Stacktrace:
  [1] _ntuple(f::Adapt.var"#3#5"{FastDense{typeof(tanh), DiffEqFlux.var"#initial_params#89"{Vector{Float32}}}}, n::Int64)
    @ Base ./ntuple.jl:36
  [2] ntuple
    @ ./ntuple.jl:19 [inlined]
  [3] adapt_structure
    @ ~/.julia/packages/Adapt/RGNRk/src/base.jl:16 [inlined]
  [4] adapt
    @ ~/.julia/packages/Adapt/RGNRk/src/Adapt.jl:40 [inlined]
  [5] #1
    @ ~/.julia/packages/Adapt/RGNRk/src/base.jl:3 [inlined]
  [6] map
    @ ./tuple.jl:213 [inlined]
  [7] adapt_structure
    @ ~/.julia/packages/Adapt/RGNRk/src/base.jl:3 [inlined]
  [8] adapt
    @ ~/.julia/packages/Adapt/RGNRk/src/Adapt.jl:40 [inlined]
  [9] #4
    @ ~/.julia/packages/Adapt/RGNRk/src/base.jl:17 [inlined]
 [10] ntuple(f::Adapt.var"#4#6"{UnionAll, FastChain{Tuple{FastDense{typeof(tanh), DiffEqFlux.var"#initial_params#89"{Vector{Float32}}}}}}, n::Int64)
    @ Base ./ntuple.jl:19
 [11] adapt_structure(to::Type, f::FastChain{Tuple{FastDense{typeof(tanh), DiffEqFlux.var"#initial_params#89"{Vector{Float32}}}}})
    @ Adapt ~/.julia/packages/Adapt/RGNRk/src/base.jl:17
 [12] adapt(to::Type, x::Function)
    @ Adapt ~/.julia/packages/Adapt/RGNRk/src/Adapt.jl:40
 [13] #125
    @ ~/.julia/packages/Flux/0c9kI/src/functor.jl:67 [inlined]
 [14] fmap(f::Flux.var"#125#126", x::Function; exclude::typeof(Functors.isleaf), cache::IdDict{Any, Any})
    @ Functors ~/.julia/packages/Functors/EWaud/src/functor.jl:56
 [15] fmap
    @ ~/.julia/packages/Functors/EWaud/src/functor.jl:55 [inlined]
 [16] cpu
    @ ~/.julia/packages/Flux/0c9kI/src/functor.jl:67 [inlined]
 [17] |>(x::Function, f::typeof(cpu))
    @ Base ./operators.jl:858
 [18] top-level scope
    @ ~/Developer/lvcluster/_research/cpu_test.jl:10

[ChrisRackauckas]

I mean, we can make that a no-op. But device(FastChain(...)) doesn't do anything because there's no array in there. You have to do it to the weight vector.

[adrhill]

I looked at the GPU examples in the docs and expected FastChain to work as a drop-in replacement for Chain, so from my perspective either a no-op or a comment in the docs would be helpful.