(Flux) Moving data between cpu and gpu: "This intrinsic must be compiled to be called"

Specific Domains Machine Learning
1

The title and example really say it all. I’m using Flux.jl version 0.12.7 and CUDA.jl 3.5.0.

MWE:

julia> using Flux

julia> x = [1, 2]
2-element Vector{Int64}:
 1
 2

julia> Flux.pullback(() -> gpu(x))
ERROR: this intrinsic must be compiled to be called
Stacktrace:
  [1] macro expansion
    @ ~/.julia/packages/Zygote/TaBlo/src/compiler/interface2.jl:0 [inlined]
  [2] _pullback(::Zygote.Context, ::Core.IntrinsicFunction, ::String, ::Type{Int64}, ::Type{Tuple{Ptr{Int64}}}, ::Ptr{Int64})
    @ Zygote ~/.julia/packages/Zygote/TaBlo/src/compiler/interface2.jl:9
  [3] _pullback
    @ ./atomics.jl:358 [inlined]
  [4] _pullback(ctx::Zygote.Context, f::typeof(getindex), args::Base.Threads.Atomic{Int64})
    @ Zygote ~/.julia/packages/Zygote/TaBlo/src/compiler/interface2.jl:0
  [5] _pullback
    @ ~/.julia/packages/CUDA/YpW0k/lib/utils/threading.jl:46 [inlined]
  [6] _pullback(::Zygote.Context, ::CUDA.APIUtils.var"##get!#4", ::Nothing, ::typeof(get!), ::CUDA.var"#162#163", ::CUDA.APIUtils.LazyInitialized{Vector{Bool}})
    @ Zygote ~/.julia/packages/Zygote/TaBlo/src/compiler/interface2.jl:0
  [7] _pullback
    @ ~/.julia/packages/CUDA/YpW0k/lib/utils/threading.jl:46 [inlined]
  [8] _pullback
    @ ~/.julia/packages/CUDA/YpW0k/src/pool.jl:69 [inlined]
  [9] _pullback(ctx::Zygote.Context, f::typeof(CUDA.stream_ordered), args::CUDA.CuDevice)
    @ Zygote ~/.julia/packages/Zygote/TaBlo/src/compiler/interface2.jl:0
 [10] macro expansion
    @ ~/.julia/packages/CUDA/YpW0k/src/pool.jl:189 [inlined]
 [11] macro expansion
    @ ./timing.jl:287 [inlined]
 [12] _pullback
    @ ~/.julia/packages/CUDA/YpW0k/src/pool.jl:187 [inlined]
 [13] _pullback
    @ ~/.julia/packages/CUDA/YpW0k/src/pool.jl:183 [inlined]
 [14] _pullback(::Zygote.Context, ::CUDA.var"#_alloc##kw", ::NamedTuple{(:stream,), Tuple{Nothing}}, ::typeof(CUDA._alloc), ::Type{CUDA.Mem.DeviceBuffer}, ::Int64)
    @ Zygote ~/.julia/packages/Zygote/TaBlo/src/compiler/interface2.jl:0
 [15] _pullback
    @ ~/.julia/packages/CUDA/YpW0k/src/pool.jl:173 [inlined]
 [16] _pullback
    @ ~/.julia/packages/CUDA/YpW0k/src/pool.jl:169 [inlined]
 [17] _pullback
    @ ~/.julia/packages/CUDA/YpW0k/src/array.jl:44 [inlined]
 [18] _pullback(::Zygote.Context, ::Type{CUDA.CuArray{Int64, 1, CUDA.Mem.DeviceBuffer}}, ::UndefInitializer, ::Tuple{Int64})
    @ Zygote ~/.julia/packages/Zygote/TaBlo/src/compiler/interface2.jl:0
 [19] _pullback
    @ ~/.julia/packages/CUDA/YpW0k/src/array.jl:290 [inlined]
 [20] _pullback
    @ ~/.julia/packages/CUDA/YpW0k/src/array.jl:497 [inlined]
 [21] _pullback(::Zygote.Context, ::typeof(Adapt.adapt_storage), ::CUDA.CuArrayAdaptor{CUDA.Mem.DeviceBuffer}, ::Vector{Int64})
    @ Zygote ~/.julia/packages/Zygote/TaBlo/src/compiler/interface2.jl:0
 [22] _pullback
    @ ~/.julia/packages/Adapt/RGNRk/src/Adapt.jl:42 [inlined]
 [23] _pullback
    @ ~/.julia/packages/Adapt/RGNRk/src/Adapt.jl:40 [inlined]
 [24] _pullback (repeats 2 times)
    @ ~/.julia/packages/CUDA/YpW0k/src/array.jl:510 [inlined]
 [25] _pullback
    @ ~/.julia/packages/Flux/ZnXxS/src/functor.jl:66 [inlined]
 [26] _pullback
    @ ~/.julia/packages/Adapt/RGNRk/src/Adapt.jl:42 [inlined]
 [27] _pullback
    @ ~/.julia/packages/Adapt/RGNRk/src/Adapt.jl:40 [inlined]
 [28] _pullback
    @ ~/.julia/packages/Flux/ZnXxS/src/functor.jl:146 [inlined]
 [29] _pullback
    @ ~/.julia/packages/Functors/QVxVb/src/functor.jl:121 [inlined]
 [30] _pullback(::Zygote.Context, ::Functors.var"##fmap#17", ::typeof(Flux._isbitsarray), ::typeof(Functors._default_walk), ::IdDict{Any, Any}, ::typeof(fmap), ::Flux.var"#132#133", ::Vector{Int64})
    @ Zygote ~/.julia/packages/Zygote/TaBlo/src/compiler/interface2.jl:0
 [31] _pullback
    @ ~/.julia/packages/Functors/QVxVb/src/functor.jl:120 [inlined]
 [32] _pullback(::Zygote.Context, ::Functors.var"#fmap##kw", ::NamedTuple{(:exclude,), Tuple{typeof(Flux._isbitsarray)}}, ::typeof(fmap), ::Flux.var"#132#133", ::Vector{Int64})
    @ Zygote ~/.julia/packages/Zygote/TaBlo/src/compiler/interface2.jl:0
 [33] _pullback
    @ ~/.julia/packages/Flux/ZnXxS/src/functor.jl:146 [inlined]
 [34] _pullback(ctx::Zygote.Context, f::typeof(gpu), args::Vector{Int64})
    @ Zygote ~/.julia/packages/Zygote/TaBlo/src/compiler/interface2.jl:0
 [35] _pullback
    @ ./REPL[3]:1 [inlined]
 [36] _pullback(::Zygote.Context, ::var"#5#6")
    @ Zygote ~/.julia/packages/Zygote/TaBlo/src/compiler/interface2.jl:0
 [37] _pullback(::Function)
    @ Zygote ~/.julia/packages/Zygote/TaBlo/src/compiler/interface.jl:34
 [38] pullback(::Function)
    @ Zygote ~/.julia/packages/Zygote/TaBlo/src/compiler/interface.jl:40
 [39] top-level scope
    @ REPL[3]:1
2

I solved my own issue so I’m putting this here for my fellow googlers: Creating a new Pkg environment (run Julia and type ] activate testenv ) and installing JUST Flux ( ] add Flux ) allows me to run my MWE without error. I guess it has to use the CUDA and CUDNN artifacts it downloads automatically when you try to do gpu(x). I had CUDA.jl installed alongside Flux in my env originally.