# Why is Flux/NNlib falling back to im2col instead of MIOpen on AMD GPU?

**URL:** <https://discourse.julialang.org/t/why-is-flux-nnlib-falling-back-to-im2col-instead-of-miopen-on-amd-gpu/135867>\
**Category:** Machine Learning\
**Tags:** flux, amdgpu\
**Created:** [February 26, 2026, 3:17pm UTC](https://discourse.julialang.org/t/why-is-flux-nnlib-falling-back-to-im2col-instead-of-miopen-on-amd-gpu/135867 "2026-02-26T15:17:21Z")\
**Posts on this page:** 4\
**Page:** 1

<div class="post-metadata">

**Author:** ![ajrohr2](https://avatars.discourse-cdn.com/v4/letter/a/e36b37/32.png) [@ajrohr2](https://discourse.julialang.org/u/ajrohr2)\
**Post date:** [February 26, 2026, 3:17pm UTC](https://discourse.julialang.org/t/why-is-flux-nnlib-falling-back-to-im2col-instead-of-miopen-on-amd-gpu/135867/1 "2026-02-26T15:17:21Z")

</div>

Hello! I’ve been trying to get Flux’s convolutional layers, specifically the Flux.NNlib.conv! function, to use the MIOpen version rather than im2col or direct. I’ve verified that I have MIOpen available with `AMDGPU.functional(:MIOpen)`, and querying the version info from AMDGPU shows all of the necessary libraries, aside from rocFFT, are available. Julia recognizes my GPU, an AMD Ryzen RX 7900 XT, as well.

I’ve included the code below. Has anyone else run into this issue?

```julia-auto
using Flux, AMDGPU

ct = Conv((3,3), 4 => 32, pad=1, stride=1)
r_ct = roc(ct)
x = ROCArray(rand(Float32, 40, 40, 4, 1))

r_ct(x)

```

This always errors due to scalar indexing, because Flux is not calling the MIOpen compatible convolution function.

```julia-auto
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] errorscalar(op::String)
    @ GPUArraysCore ~/.julia/packages/GPUArraysCore/aNaXo/src/GPUArraysCore.jl:151
  [2] _assertscalar(op::String, behavior::GPUArraysCore.ScalarIndexing)
    @ GPUArraysCore ~/.julia/packages/GPUArraysCore/aNaXo/src/GPUArraysCore.jl:124
  [3] assertscalar(op::String)
    @ GPUArraysCore ~/.julia/packages/GPUArraysCore/aNaXo/src/GPUArraysCore.jl:112
  [4] getindex
    @ ~/.julia/packages/GPUArrays/3a5jB/src/host/indexing.jl:50 [inlined]
  [5] scalar_getindex
    @ ~/.julia/packages/GPUArrays/3a5jB/src/host/indexing.jl:36 [inlined]
  [6] _getindex
    @ ~/.julia/packages/GPUArrays/3a5jB/src/host/indexing.jl:19 [inlined]
  [7] getindex
    @ ~/.julia/packages/GPUArrays/3a5jB/src/host/indexing.jl:17 [inlined]
  [8] getindex
    @ ./subarray.jl:316 [inlined]
  [9] im2col!(col::ROCArray{Float32, 2, AMDGPU.Runtime.Mem.HIPBuffer}, x::SubArray{Float32, 4, ROCArray{…}, Tuple{…}, true}, cdims::DenseConvDims{3, 3, 3, 6, 3})
    @ NNlib ~/.julia/packages/NNlib/srXYX/src/impl/conv_im2col.jl:253
 [10] (::NNlib.var"#conv_part#538"{ROCArray{…}, Float32, Float32, SubArray{…}, SubArray{…}, ROCArray{…}, DenseConvDims{…}, Int64, Int64, Int64})(task_n::Int64, part::UnitRange{Int64})
    @ NNlib ~/.julia/packages/NNlib/srXYX/src/impl/conv_im2col.jl:53
 [11] conv_im2col!(y::SubArray{…}, x::SubArray{…}, w::ROCArray{…}, cdims::DenseConvDims{…}; col::ROCArray{…}, alpha::Float32, beta::Float32, ntasks::Int64)
    @ NNlib ~/.julia/packages/NNlib/srXYX/src/impl/conv_im2col.jl:69
 [12] conv_im2col!(y::SubArray{…}, x::SubArray{…}, w::ROCArray{…}, cdims::DenseConvDims{…})
    @ NNlib ~/.julia/packages/NNlib/srXYX/src/impl/conv_im2col.jl:23
 [13] (::NNlib.var"#conv_group#186"{@Kwargs{}, ROCArray{…}, ROCArray{…}, ROCArray{…}, DenseConvDims{…}})(xc::UnitRange{Int64}, wc::UnitRange{Int64})
    @ NNlib ~/.julia/packages/NNlib/srXYX/src/conv.jl:209
 [14] conv!(out::ROCArray{…}, in1::ROCArray{…}, in2::ROCArray{…}, cdims::DenseConvDims{…}; kwargs::@Kwargs{})
    @ NNlib ~/.julia/packages/NNlib/srXYX/src/conv.jl:218
 [15] conv!
    @ ~/.julia/packages/NNlib/srXYX/src/conv.jl:185 [inlined]
 [16] #conv!#143
    @ ~/.julia/packages/NNlib/srXYX/src/conv.jl:145 [inlined]
 [17] conv!
    @ ~/.julia/packages/NNlib/srXYX/src/conv.jl:140 [inlined]
 [18] conv(x::ROCArray{Float32, 4, AMDGPU.Runtime.Mem.HIPBuffer}, w::ROCArray{Float32, 4, AMDGPU.Runtime.Mem.HIPBuffer}, cdims::DenseConvDims{2, 2, 2, 4, 2}; kwargs::@Kwargs{})
    @ NNlib ~/.julia/packages/NNlib/srXYX/src/conv.jl:88
 [19] conv
    @ ~/.julia/packages/NNlib/srXYX/src/conv.jl:83 [inlined]
 [20] (::Conv{2, 4, typeof(identity), ROCArray{Float32, 4, AMDGPU.Runtime.Mem.HIPBuffer}, ROCArray{Float32, 1, AMDGPU.Runtime.Mem.HIPBuffer}})(x::ROCArray{Float32, 4, AMDGPU.Runtime.Mem.HIPBuffer})
    @ Flux ~/.julia/packages/Flux/DZYiO/src/layers/conv.jl:201
 [21] top-level scope
    @ REPL[30]:1

```

---

<div class="post-metadata">

**Author:** ![zhzy0077](https://avatars.discourse-cdn.com/v4/letter/z/c68b51/32.png) [@zhzy0077](https://discourse.julialang.org/u/zhzy0077)\
**Post date:** [March 7, 2026, 5:27pm UTC](https://discourse.julialang.org/t/why-is-flux-nnlib-falling-back-to-im2col-instead-of-miopen-on-amd-gpu/135867/2 "2026-03-07T17:27:27Z")

</div>

I hit what looks like the same issue on AMDGPU/Flux/NNlib: `Base.get_extension(NNlib, :NNlibAMDGPUExt)` reported that the extension was loaded, but `NNlib.conv` on `ROCArray` still fell back to `src/impl/conv_im2col.jl`, causing the scalar-indexing error. In my case, `julia --compiled-modules=no --project=. repro.jl` made the same repro work immediately, which suggests the `NNlibAMDGPUExt` methods were being skipped during cached/precompiled loading even though `AMDGPU.functional(:MIOpen)` was true at runtime.

A practical workaround that fixed it for me was forcing a clean re-precompile of the environment:

```bash
 rm -rf ~/.julia/compiled/v1.12
 julia --project=. -e 'using Pkg; Pkg.instantiate(); Pkg.precompile()'

```

After clearing the compiled cache and re-precompiling, plain `julia --project=. repro.jl` started using the MIOpen path correctly. So if someone sees `NNlibAMDGPUExt` loaded but still gets `im2col`, it may be worth wiping `~/.julia/compiled/...` and rebuilding first.

---

<div class="post-metadata">

**Author:** ![Benny](https://avatars.discourse-cdn.com/v4/letter/b/49beb7/32.png) [@Benny](https://discourse.julialang.org/u/Benny)\
**Post date:** [March 9, 2026, 12:05am UTC](https://discourse.julialang.org/t/why-is-flux-nnlib-falling-back-to-im2col-instead-of-miopen-on-amd-gpu/135867/3 "2026-03-09T00:05:54Z")

</div>

> [@zhzy0077](#):
>
> which suggests the `NNlibAMDGPUExt` methods were being skipped during cached/precompiled loading even though `AMDGPU.functional(:MIOpen)` was true at runtime.

Is there an issue open for that? Sounds like there should be.

The simplest place for this to go wrong is here, but I’m not sure how `functional(:MIOpen)` could be false at precompile time. [NNlib.jl/ext/NNlibAMDGPUExt/NNlibAMDGPUExt.jl at 0c599d869216822060cd19d4f774006dcb60d2b6 · FluxML/NNlib.jl · GitHub](https://github.com/FluxML/NNlib.jl/blob/0c599d869216822060cd19d4f774006dcb60d2b6/ext/NNlibAMDGPUExt/NNlibAMDGPUExt.jl#L51)

---

<div class="post-metadata">

**Author:** ![ajrohr2](https://avatars.discourse-cdn.com/v4/letter/a/e36b37/32.png) [@ajrohr2](https://discourse.julialang.org/u/ajrohr2)\
**Post date:** [March 10, 2026, 1:30am UTC](https://discourse.julialang.org/t/why-is-flux-nnlib-falling-back-to-im2col-instead-of-miopen-on-amd-gpu/135867/4 "2026-03-10T01:30:35Z")

</div>

It looks like this worked! Thank you!