Flux.cpu and Flux.gpu no longer move data on views

[piever]

Currently (as of Flux 0.15), Flux.cpu and Flux.gpu do not error when fed a view, but they simply return an identical value, without moving data to / from the GPU.

MWE

julia> using Flux, CUDA, cuDNN

julia> cpu_view = view(rand(2, 2), 1, :);

julia> gpu_view = view(CUDA.rand(2, 2), 1, :);

julia> Flux.gpu(cpu_view)
2-element view(::Matrix{Float64}, 1, :) with eltype Float64:
 0.10101004098832933
 0.7891135815339817

julia> Flux.cpu(gpu_view)
2-element view(::CuArray{Float32, 2, CUDA.DeviceMemory}, 1, :) with eltype Float32:
 0.5095508
 0.93324655

Dependencies:

  [052768ef] CUDA v5.5.2
  [587475ba] Flux v0.15.2
  [02a925ec] cuDNN v1.4.0

on Julia 1.11.2

On Flux 0.14, Flux.cpu and Flux.gpu would transform views by applying the transformation to the parent array.

[CarloLucibello]

We now completely rely on MLDataDevices.jl for the data movement. gpu and cpu are simply defined as

gpu(x) = gpu_device()(x)
cpu(x) = cpu_device()(x)

Could you report the issue to the Lux repo?
Unfortunately the transition has not been entirely smooth, and some of the v0.14 behavior is not preserved, e.g. LuxDL/Lux.jl#1129

[CarloLucibello]

The issue is not specific to the CUDA backend:

julia> using Flux, Metal

julia> x = rand(2 , 2)
2×2 Matrix{Float64}:
 0.99688   0.613988
 0.363596  0.985924

julia> cpu_view = view(x, 1, :)
2-element view(::Matrix{Float64}, 1, :) with eltype Float64:
 0.9968804026540483
 0.6139877405888565

julia> gpu_view = view(gpu(x), 1, :)
2-element view(::MtlMatrix{Float32, Metal.PrivateStorage}, 1, :) with eltype Float32:
 0.9968804
 0.61398774

julia> gpu(cpu_view)
2-element view(::Matrix{Float64}, 1, :) with eltype Float64:
 0.9968804026540483
 0.6139877405888565

julia> cpu(gpu_view)
2-element view(::MtlMatrix{Float32, Metal.PrivateStorage}, 1, :) with eltype Float32:
 0.9968804
 0.61398774

[piever]

That was fast! Closing here as this was fixed in the latest MLDataDevices release before I even reported it...

Even if there are some hiccups, happy to hear that more infrastructure is being shared.