Gradients of broadcasted exp broken with CuArrays v2.0.1

[aterenin]

The following works on CuArrays v1.7 but doesn't work on v2.0.1.

Flux.gradient(x -> sum(exp.(x)), CuArrays.ones(1,1))

┌ Warning: calls to Base intrinsics might be GPU incompatible
│   exception =
│    You called exp(x::T) where T<:Union{Float32, Float64} in Base.Math at special/exp.jl:75, maybe you intended to call exp(x::Float64) in CUDAnative at /vol/bitbucket/at6617/dot_julia/packages/CUDAnative/cnQli/src/device/cuda/math.jl:100 instead?
│    Stacktrace:
│     [1] exp at special/exp.jl:75
│     [2] #28 at /vol/bitbucket/at6617/dot_julia/packages/GPUArrays/QDGmr/src/host/broadcast.jl:64
└ @ CUDAnative /vol/bitbucket/at6617/dot_julia/packages/CUDAnative/cnQli/src/compiler/irgen.jl:266

[DhairyaLGandhi]

Good to add to tests and see what is happening

[DhairyaLGandhi]

Ref FluxML/Flux.jl#1103

[matsueushi]

Is it just a warning, or indeed an error? cf. JuliaGPU/CuArrays.jl#615 (comment)

[aterenin]

Error with a very large stack trace directly afterwards. I can post it later, once I re-upgrade CuArrays on the GPU box again.

[ararslan]

Can you try with #565?

[ararslan]

This works for me with #565.

[aterenin]

Works for me on [e88e6eb3] Zygote v0.4.12 #04ace9 (https://github.com/FluxML/Zygote.jl.git).