GPU Support

NVIDIA GPU support should work out of the box on systems with CUDA and CUDNN installed. For more details see the CUDA.jl readme.

AMD GPU support is available since Julia 1.9 on systems with ROCm and MIOpen installed. For more details refer to the AMDGPU.jl repository.

Checking GPU Availability

By default, Flux will run the checks on your system to see if it can support GPU functionality. You can check if Flux identified a valid GPU setup by typing the following:

julia> using CUDA

julia> CUDA.functional()


julia> using AMDGPU

julia> AMDGPU.functional()

julia> AMDGPU.functional(:MIOpen)

Selecting GPU backend

Available GPU backends are: CUDA, AMD.

Flux relies on Preferences.jl for selecting default GPU backend to use.

There are two ways you can specify it:

  • From the REPL/code in your project, call Flux.gpu_backend!("AMD") and restart (if needed) Julia session for the changes to take effect.
  • In LocalPreferences.toml file in you project directory specify:
gpu_backend = "AMD"

Current GPU backend can be fetched from Flux.GPU_BACKEND variable:

julia> Flux.GPU_BACKEND

Basic GPU Usage

Support for array operations on other hardware backends, like GPUs, is provided by external packages like CUDA. Flux is agnostic to array types, so we simply need to move model weights and data to the GPU and Flux will handle it.

For example, we can use CUDA.CuArray (with the cu converter) to run our basic example on an NVIDIA GPU.

(Note that you need to have CUDA available to use CUDA.CuArray – please see the CUDA.jl instructions for more details.)

using CUDA

W = cu(rand(2, 5)) # a 2×5 CuArray
b = cu(rand(2))

predict(x) = W*x .+ b
loss(x, y) = sum((predict(x) .- y).^2)

x, y = cu(rand(5)), cu(rand(2)) # Dummy data
loss(x, y) # ~ 3

Note that we convert both the parameters (W, b) and the data set (x, y) to cuda arrays. Taking derivatives and training works exactly as before.

If you define a structured model, like a Dense layer or Chain, you just need to convert the internal parameters. Flux provides fmap, which allows you to alter all parameters of a model at once.

d = Dense(10 => 5, σ)
d = fmap(cu, d)
d.weight # CuArray
d(cu(rand(10))) # CuArray output

m = Chain(Dense(10 => 5, σ), Dense(5 => 2), softmax)
m = fmap(cu, m)

As a convenience, Flux provides the gpu function to convert models and data to the GPU if one is available. By default, it'll do nothing. So, you can safely call gpu on some data or model (as shown below), and the code will not error, regardless of whether the GPU is available or not. If the GPU library (CUDA.jl) loads successfully, gpu will move data from the CPU to the GPU. As is shown below, this will change the type of something like a regular array to a CuArray.

julia> using Flux, CUDA

julia> m = Dense(10, 5) |> gpu
Dense(10 => 5)      # 55 parameters

julia> x = rand(10) |> gpu
10-element CuArray{Float32, 1, CUDA.Mem.DeviceBuffer}:

julia> m(x)
5-element CuArray{Float32, 1, CUDA.Mem.DeviceBuffer}:

The analogue cpu is also available for moving models and data back off of the GPU.

julia> x = rand(10) |> gpu
10-element CuArray{Float32, 1, CUDA.Mem.DeviceBuffer}:

julia> x |> cpu
10-element Vector{Float32}:

Transferring Training Data

In order to train the model using the GPU both model and the training data have to be transferred to GPU memory. Moving the data can be done in two different ways:

  1. Iterating over the batches in a DataLoader object transferring each one of the training batches at a time to the GPU. This is recommended for large datasets. Done by hand, it might look like this:

    train_loader = Flux.DataLoader((X, Y), batchsize=64, shuffle=true)
    # ... model definition, optimiser setup
    for epoch in 1:epochs
        for (x_cpu, y_cpu) in train_loader
            x = gpu(x_cpu)
            y = gpu(y_cpu)
            grads = gradient(m -> loss(m, x, y), model)
            Flux.update!(opt_state, model, grads[1])

    Rather than write this out every time, you can just call gpu(::DataLoader):

    gpu_train_loader = Flux.DataLoader((X, Y), batchsize=64, shuffle=true) |> gpu
    # ... model definition, optimiser setup
    for epoch in 1:epochs
        for (x, y) in gpu_train_loader
            grads = gradient(m -> loss(m, x, y), model)
            Flux.update!(opt_state, model, grads[1])

    This is equivalent to DataLoader(MLUtils.mapobs(gpu, (X, Y)); keywords...). Something similar can also be done with CUDA.CuIterator, gpu_train_loader = CUDA.CuIterator(train_loader). However, this only works with a limited number of data types: first(train_loader) should be a tuple (or NamedTuple) of arrays.

  2. Transferring all training data to the GPU at once before creating the DataLoader. This is usually performed for smaller datasets which are sure to fit in the available GPU memory.

    gpu_train_loader = Flux.DataLoader((X, Y) |> gpu, batchsize = 32)
    # ...
    for epoch in 1:epochs
        for (x, y) in gpu_train_loader
            # ...

    Here (X, Y) |> gpu applies gpu to both arrays, as it recurses into structures.

Saving GPU-Trained Models

After the training process is done, one must always transfer the trained model back to the cpu memory scope before serializing or saving to disk. This can be done, as described in the previous section, with:

model = cpu(model) # or model = model |> cpu

and then

using BSON
# ...
BSON.@save "./path/to/trained_model.bson" model

# in this approach the cpu-transferred model (referenced by the variable `model`)
# only exists inside the `let` statement
let model = cpu(model)
   # ...
   BSON.@save "./path/to/trained_model.bson" model

# is equivalent to the above, but uses `key=value` storing directive from BSON.jl
BSON.@save "./path/to/trained_model.bson" model = cpu(model)

The reason behind this is that models trained in the GPU but not transferred to the CPU memory scope will expect CuArrays as input. In other words, Flux models expect input data coming from the same kind device in which they were trained on.

In controlled scenarios in which the data fed to the loaded models is garanteed to be in the GPU there's no need to transfer them back to CPU memory scope, however in production environments, where artifacts are shared among different processes, equipments or configurations, there is no garantee that the CUDA.jl package will be available for the process performing inference on the model loaded from the disk.

Disabling CUDA or choosing which GPUs are visible to Flux

Sometimes it is required to control which GPUs are visible to julia on a system with multiple GPUs or disable GPUs entirely. This can be achieved with an environment variable CUDA_VISIBLE_DEVICES.

To disable all devices:


To select specific devices by device id:


More information for conditional use of GPUs in CUDA.jl can be found in its documentation, and information about the specific use of the variable is described in the Nvidia CUDA blog post.