FNO invokes scalar indexing

[ayushinav]

Describe the bug 🐞

Calling FNO on GPUs leads to Scalar indexing is disallowed error.

Minimal Reproducible Example 👇

using CUDA, Lux

n_batch = 10
nz = 64
m_sample = cu(2 .+ randn(nz, 1, n_batch))


fno2 = FourierNeuralOperator(gelu; chs=(1, 64, 64, 128, 1), modes=(16,))
ps,st = Lux.setup(rng, fno2) |> cu
val, st_ = Lux.apply(fno2, m_sample, ps, st)

Error & Stacktrace ⚠️

ERROR: Scalar indexing is disallowed.
Invocation of getindex resulted in scalar indexing of a GPU array.
This is typically caused by calling an iterating implementation of a method.
Such implementations *do not* execute on the GPU, but very slowly on the CPU,
and therefore should be avoided.

If you want to allow scalar iteration, use `allowscalar` or `@allowscalar`
to enable scalar iteration globally or for the operations in question.
Stacktrace:
  [1] error(s::String)
    @ Base .\error.jl:35
  [2] errorscalar(op::String)
    @ GPUArraysCore C:\Users\ayush\.julia\packages\GPUArraysCore\aNaXo\src\GPUArraysCore.jl:151
  [3] _assertscalar(op::String, behavior::GPUArraysCore.ScalarIndexing)
    @ GPUArraysCore C:\Users\ayush\.julia\packages\GPUArraysCore\aNaXo\src\GPUArraysCore.jl:124
  [4] assertscalar(op::String)
    @ GPUArraysCore C:\Users\ayush\.julia\packages\GPUArraysCore\aNaXo\src\GPUArraysCore.jl:112
  [5] getindex
    @ C:\Users\ayush\.julia\packages\GPUArrays\3a5jB\src\host\indexing.jl:50 [inlined]
  [6] scalar_getindex
    @ C:\Users\ayush\.julia\packages\GPUArrays\3a5jB\src\host\indexing.jl:36 [inlined]
  [7] _getindex
    @ C:\Users\ayush\.julia\packages\GPUArrays\3a5jB\src\host\indexing.jl:19 [inlined]
  [8] getindex
    @ C:\Users\ayush\.julia\packages\GPUArrays\3a5jB\src\host\indexing.jl:17 [inlined]
  [9] getindex
    @ .\subarray.jl:320 [inlined]
 [10] im2col!(col::CuArray{…}, x::SubArray{…}, cdims::DenseConvDims{…})
    @ NNlib C:\Users\ayush\.julia\packages\NNlib\srXYX\src\impl\conv_im2col.jl:253
 [11] (::NNlib.var"#conv_part#640"{…})(task_n::Int64, part::UnitRange{…})
    @ NNlib C:\Users\ayush\.julia\packages\NNlib\srXYX\src\impl\conv_im2col.jl:53
 [12] conv_im2col!(y::SubArray{…}, x::SubArray{…}, w::CuArray{…}, cdims::DenseConvDims{…}; col::CuArray{…}, alpha::Float32, beta::Float32, ntasks::Int64)
    @ NNlib C:\Users\ayush\.julia\packages\NNlib\srXYX\src\impl\conv_im2col.jl:69
 [13] conv_im2col!(y::SubArray{…}, x::SubArray{…}, w::CuArray{…}, cdims::DenseConvDims{…})
    @ NNlib C:\Users\ayush\.julia\packages\NNlib\srXYX\src\impl\conv_im2col.jl:23
 [14] (::NNlib.var"#conv_group#301"{…})(xc::UnitRange{…}, wc::UnitRange{…})
    @ NNlib C:\Users\ayush\.julia\packages\NNlib\srXYX\src\conv.jl:209
 [15] conv!(out::CuArray{…}, in1::CuArray{…}, in2::CuArray{…}, cdims::DenseConvDims{…}; kwargs::@Kwargs{})
    @ NNlib C:\Users\ayush\.julia\packages\NNlib\srXYX\src\conv.jl:218
 [16] conv!
    @ C:\Users\ayush\.julia\packages\NNlib\srXYX\src\conv.jl:185 [inlined]
 [17] conv_bias_act!(y::CuArray{…}, x::CuArray{…}, w::CuArray{…}, cdims::DenseConvDims{…}, b::CuArray{…}, σ::Function; kwargs::@Kwargs{})
    @ NNlib C:\Users\ayush\.julia\packages\NNlib\srXYX\src\conv_bias_act.jl:10
 [18] conv_bias_act!(y::CuArray{…}, x::CuArray{…}, w::CuArray{…}, cdims::DenseConvDims{…}, b::CuArray{…}, σ::Function)
    @ NNlib C:\Users\ayush\.julia\packages\NNlib\srXYX\src\conv_bias_act.jl:8
 [19] conv_bias_act!(y::CuArray{…}, x::CuArray{…}, w::CuArray{…}, cdims::DenseConvDims{…}, b::CuArray{…}, σ::Function; kwargs::@Kwargs{})
    @ NNlib C:\Users\ayush\.julia\packages\NNlib\srXYX\src\conv_bias_act.jl:22
 [20] conv_bias_act!
    @ C:\Users\ayush\.julia\packages\NNlib\srXYX\src\conv_bias_act.jl:17 [inlined]
 [21] #conv_bias_act#404
    @ C:\Users\ayush\.julia\packages\NNlib\srXYX\src\conv_bias_act.jl:4 [inlined]
 [22] conv_bias_act
    @ C:\Users\ayush\.julia\packages\NNlib\srXYX\src\conv_bias_act.jl:1 [inlined]
 [23] conv_bias_act(::Type{…}, x::CuArray{…}, weight::CuArray{…}, cdims::DenseConvDims{…}, bias′::CuArray{…}, act::typeof(identity))
    @ LuxLib.Impl C:\Users\ayush\.julia\packages\LuxLib\ZJ3gh\src\impl\conv.jl:163
 [24] conv_bias_act
    @ C:\Users\ayush\.julia\packages\LuxLib\ZJ3gh\src\impl\conv.jl:141 [inlined]
 [25] fused_conv
    @ C:\Users\ayush\.julia\packages\LuxLib\ZJ3gh\src\impl\conv.jl:201 [inlined]
 [26] fused_conv
    @ C:\Users\ayush\.julia\packages\LuxLib\ZJ3gh\src\impl\conv.jl:177 [inlined]
 [27] fused_conv_bias_activation
    @ C:\Users\ayush\.julia\packages\LuxLib\ZJ3gh\src\api\conv.jl:37 [inlined]
 [28] ##c::Convinternal#337
    @ C:\Users\ayush\.julia\packages\Lux\GCC0y\src\layers\conv.jl:259 [inlined]
 [29] Conv
    @ C:\Users\ayush\.julia\packages\ReactantCore\SEqVX\src\ReactantCore.jl:286 [inlined]
 [30] apply
    @ C:\Users\ayush\.julia\packages\LuxCore\kQC9S\src\LuxCore.jl:155 [inlined]
 [31] macro expansion
    @ C:\Users\ayush\.julia\packages\Lux\GCC0y\src\layers\containers.jl:0 [inlined]
 [32] applychain(layers::@NamedTuple{…}, x::CuArray{…}, ps::@NamedTuple{…}, st::@NamedTuple{…})
    @ Lux C:\Users\ayush\.julia\packages\Lux\GCC0y\src\layers\containers.jl:577
 [33] Chain
    @ C:\Users\ayush\.julia\packages\Lux\GCC0y\src\layers\containers.jl:575 [inlined]
 [34] apply
    @ C:\Users\ayush\.julia\packages\LuxCore\kQC9S\src\LuxCore.jl:155 [inlined]
 [35] AbstractLuxWrapperLayer
    @ C:\Users\ayush\.julia\packages\LuxCore\kQC9S\src\LuxCore.jl:269 [inlined]
 [36] apply(model::FourierNeuralOperator{…}, x::CuArray{…}, ps::@NamedTuple{…}, st::@NamedTuple{…})
    @ LuxCore C:\Users\ayush\.julia\packages\LuxCore\kQC9S\src\LuxCore.jl:155
 [37] top-level scope
    @ c:\Users\ayush\Desktop\UQ_MT\wip\surrogate_v1.jl:1
Some type information was truncated. Use `show(err)` to see complete types.

Constructing the FNO via the other API atleast points out to the line in NeuralOperators, but is otherwise missing in the above error stack

fno3 = FourierNeuralOperator((16,), 1, 4, 64);
ps,st = Lux.setup(rng, fno3) |> cu;
val, st_ = Lux.apply(fno3, m_sample, ps, st);

Error stack

ERROR: Scalar indexing is disallowed.
Invocation of getindex resulted in scalar indexing of a GPU array.
This is typically caused by calling an iterating implementation of a method.
Such implementations *do not* execute on the GPU, but very slowly on the CPU,
and therefore should be avoided.

If you want to allow scalar iteration, use `allowscalar` or `@allowscalar`
to enable scalar iteration globally or for the operations in question.
Stacktrace:
  [1] error(s::String)
    @ Base .\error.jl:35
  [2] errorscalar(op::String)
    @ GPUArraysCore C:\Users\ayush\.julia\packages\GPUArraysCore\aNaXo\src\GPUArraysCore.jl:151
  [3] _assertscalar(op::String, behavior::GPUArraysCore.ScalarIndexing)
    @ GPUArraysCore C:\Users\ayush\.julia\packages\GPUArraysCore\aNaXo\src\GPUArraysCore.jl:124
  [4] assertscalar(op::String)
    @ GPUArraysCore C:\Users\ayush\.julia\packages\GPUArraysCore\aNaXo\src\GPUArraysCore.jl:112
  [5] getindex
    @ C:\Users\ayush\.julia\packages\GPUArrays\3a5jB\src\host\indexing.jl:50 [inlined]
  [6] macro expansion
    @ .\multidimensional.jl:981 [inlined]
  [7] macro expansion
    @ .\cartesian.jl:64 [inlined]
  [8] macro expansion
    @ .\multidimensional.jl:979 [inlined]
  [9] _unsafe_setindex!(::IndexLinear, ::Array{…}, ::CuArray{…}, ::UnitRange{…}, ::UnitRange{…}, ::UnitRange{…})
    @ Base .\multidimensional.jl:988
 [10] _setindex!
    @ .\multidimensional.jl:967 [inlined]
 [11] setindex!
    @ .\abstractarray.jl:1413 [inlined]
 [12] _copy_or_fill!
    @ .\abstractarray.jl:1857 [inlined]
 [13] __cat_offset1!
    @ .\abstractarray.jl:1849 [inlined]
 [14] __cat_offset!
    @ .\abstractarray.jl:1840 [inlined]
 [15] __cat_offset!(A::Array{…}, shape::Tuple{…}, catdims::Tuple{…}, offsets::Tuple{…}, x::Array{…}, X::CuArray{…})
    @ Base .\abstractarray.jl:1841
 [16] __cat(::Array{Float32, 3}, ::Tuple{Int64, Int64, Int64}, ::Tuple{Bool, Bool}, ::Array{Float32, 3}, ::Vararg{Any})
    @ Base .\abstractarray.jl:1836
 [17] _cat_t
    @ .\abstractarray.jl:1829 [inlined]
 [18] _cat(::Int64, ::Array{Float32, 3}, ::CuArray{Float32, 3, CUDA.DeviceMemory})
    @ Base .\abstractarray.jl:2086
 [19] cat
    @ .\abstractarray.jl:2084 [inlined]
 [20] (::GridEmbedding{Vector{Tuple{…}}})(x::CuArray{Float32, 3, CUDA.DeviceMemory}, ps::@NamedTuple{}, st::@NamedTuple{})
    @ NeuralOperators C:\Users\ayush\.julia\packages\NeuralOperators\IE6mQ\src\layers.jl:240
 [21] apply
    @ C:\Users\ayush\.julia\packages\LuxCore\kQC9S\src\LuxCore.jl:155 [inlined]
 [22] macro expansion
    @ C:\Users\ayush\.julia\packages\Lux\GCC0y\src\layers\containers.jl:0 [inlined]
 [23] applychain(layers::@NamedTuple{…}, x::CuArray{…}, ps::@NamedTuple{…}, st::@NamedTuple{…})
    @ Lux C:\Users\ayush\.julia\packages\Lux\GCC0y\src\layers\containers.jl:577
 [24] Chain
    @ C:\Users\ayush\.julia\packages\Lux\GCC0y\src\layers\containers.jl:575 [inlined]
 [25] apply
    @ C:\Users\ayush\.julia\packages\LuxCore\kQC9S\src\LuxCore.jl:155 [inlined]
 [26] AbstractLuxWrapperLayer
    @ C:\Users\ayush\.julia\packages\LuxCore\kQC9S\src\LuxCore.jl:269 [inlined]
 [27] apply(model::FourierNeuralOperator{…}, x::CuArray{…}, ps::@NamedTuple{…}, st::@NamedTuple{…})
    @ LuxCore C:\Users\ayush\.julia\packages\LuxCore\kQC9S\src\LuxCore.jl:155
 [28] top-level scope
    @ c:\Users\ayush\Desktop\UQ_MT\wip\surrogate_v1.jl:48

Environment (please complete the following information):

• Output of using Pkg; Pkg.status()

Status `C:\Users\ayush\Desktop\UQ_MT\Project.toml`
  [052768ef] CUDA v5.10.1
  [b2108857] Lux v1.31.3
  [872c559c] NNlib v0.9.33
  [ea5c82af] NeuralOperators v0.6.2
  [3c362404] Reactant v0.2.231

• Output of versioninfo()

Julia Version 1.11.4
Commit 8561cc3d68 (2025-03-10 11:36 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: 16 × 13th Gen Intel(R) Core(TM) i7-13620H
  WORD_SIZE: 64
  LLVM: libLLVM-16.0.6 (ORCJIT, goldmont)
Threads: 1 default, 0 interactive, 1 GC (on 16 virtual cores)
Environment:
  JULIA_EDITOR = code
  JULIA_VSCODE_REPL = 1

Additional context

Add any other context about the problem here.

[ayushinav]

I see that the tests use Reactant, but I had reactant_device() return (::ReactantDevice{Missing, Missing, Missing, Missing, Union{}}) (generic function with 1 method), which I think implies that GPU is not loaded. I had gpu_device() also returning (::CPUDevice{Missing}) (generic function with 1 method), which points to a different problem. To preface, I had a similar problem as JuliaGPU/CUDA.jl#3000 but then reinstalling everything, it ran fine. Not sure if there's a connection. Anyways, the FNO call in the MWE should be able to execute without errors.

[ayushinav]

I found out the harder way (through another issue: LuxDL/Lux.jl#1245) that cuDNN is also required to run code on GPUs. https://github.com/JuliaAttic/CUDNN.jl says that the package is being phased out with the same functionalities being available in CUDA. It's not obvious at the first glance that cuDNN needs to be explicitly loaded after CUDA given the former is being deprecated (soon?). It might be worth putting it on the Lux.jl docs @avik-pal.

Anyways, the second API for FNOs, the one compatible with neuraloperators API does not run even then and the stacktrace is up there in the first comment. The same API doesn't have a test defined so the bug might be missed.