InvalidIRError when putting a model on the GPU

[SimonCoste]

Hi, I successfully used Flux.jl on a remote machine with GPUs, but a bug suddenly appeared out of nowhere - something might have been modified on the machine.

The bug happens when I want to put a model on the gpu.
The minimal working example is as follows:

julia
using Flux, CUDA
C = Conv((3,3), 1=>1)
C = C |> gpu

results in the following error : 

ERROR: InvalidIRError: compiling kernel partial_mapreduce_grid(typeof(identity), typeof(|), Bool, CartesianIndices{4, NTuple{4, Base.OneTo{Int64}}}, CartesianIndices{4, NTuple{4, Base.OneTo{Int64}}}, Val{true}, CuDeviceArray{Bool, 5, 1}, Base.Broadcast.Broadcasted{CUDA.CuArrayStyle{4}, NTuple{4, Base.OneTo{Int64}}, Base.var"#76#77"{typeof(iszero)}, Tuple{CuDeviceArray{Float32, 4, 1}}}) resulted in invalid LLVM IR
Reason: unsupported call to an unknown function (call to julia.gpu.state_getter)

The gpu is here and available, ie has_cuda() is true, and

x = randn(10)
x = x |> gpu

works as intended.

Could you help me debug this ? 

[DhairyaLGandhi]

Hi, this seems like an issue of the setup rather than with Flux. I might point you to the JuliaLang Slack https://julialang.org/slack/ or discourse where this would be more appropriate.

As a debugging step does suppressing the output for the layer work in the REPL?

What is the output when you try to do

julia> c = c |> gpu; #notice the ; which would suppress the output

I'm going to close the issue, but in case it turns out there's anything to be fixed in Flux, I'd be happy to reopen and see this sorted.