gpu(m)

Copies m to the current GPU device (using current GPU backend), if one is available. If no GPU is available, it does nothing (but prints a warning the first time).

On arrays, this calls CUDA's cu, which also changes arrays with Float64 elements to Float32 while copying them to the device (same for AMDGPU). To act on arrays within a struct, the struct type must be marked with @functor .

Use cpu to copy back to ordinary Arrays. See also f32 and f16 to change element type only.

See the CUDA.jl docs to help identify the current device.

Example

			julia> m = Dense(rand(2, 3))  # constructed with Float64 weight matrix
Dense(3 => 2)       # 8 parameters

julia> typeof(m.weight)
Matrix{Float64} (alias for Array{Float64, 2})

julia> m_gpu = gpu(m)  # can equivalently be written m_gpu = m |> gpu
Dense(3 => 2)       # 8 parameters

julia> typeof(m_gpu.weight)
CUDA.CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}

			gpu(data::DataLoader)

Transforms a given DataLoader to apply gpu to each batch of data, when iterated over. (If no GPU is available, this does nothing.)

Example

			julia> dl = Flux.DataLoader((x = ones(2,10), y='a':'j'), batchsize=3)
4-element DataLoader(::NamedTuple{(:x, :y), Tuple{Matrix{Float64}, StepRange{Char, Int64}}}, batchsize=3)
  with first element:
  (; x = 2×3 Matrix{Float64}, y = 3-element StepRange{Char, Int64})

julia> first(dl)
(x = [1.0 1.0 1.0; 1.0 1.0 1.0], y = 'a':1:'c')

julia> c_dl = gpu(dl)
4-element DataLoader(::MLUtils.MappedData{:auto, typeof(gpu), NamedTuple{(:x, :y), Tuple{Matrix{Float64}, StepRange{Char, Int64}}}}, batchsize=3)
  with first element:
  (; x = 2×3 CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}, y = 3-element StepRange{Char, Int64})

julia> first(c_dl).x
2×3 CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}:
 1.0  1.0  1.0
 1.0  1.0  1.0

For large datasets, this is preferred over moving all the data to the GPU before creating the DataLoader, like this:

			julia> Flux.DataLoader((x = ones(2,10), y=2:11) |> gpu, batchsize=3)
4-element DataLoader(::NamedTuple{(:x, :y), Tuple{CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}, UnitRange{Int64}}}, batchsize=3)
  with first element:
  (; x = 2×3 CUDA.CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}, y = 3-element UnitRange{Int64})