Profiling Lux Training Loops

Only for Reactant

This tutorial is applicable iff you are using Reactant.jl (AutoEnzyme with ReactantDevice) for training.

To profile the training loop, wrap the training loop with Reactant.with_profiler and pass the path to the directory where the traces should be saved. Note that this will have some overhead and hence should be used only for debugging purposes.

A simple example is shown below:

using Reactant, Lux, Random, MLUtils, Optimisers

dev = reactant_device()

x_data = rand(Float32, 32, 1024)
y_data = x_data .^ 2 .- 1

dl = DataLoader((x_data, y_data); batchsize=32, shuffle=true) |> dev;

model = Chain(Dense(32 => 64, relu), Dense(64 => 32))
ps, st = Lux.setup(Random.default_rng(), model) |> dev;

Reactant.with_profiler(joinpath(tempdir(), "lux_training_trace")) do
    train_state = Training.TrainState(model, ps, st, Adam(0.001))
    for epoch in 1:10
        for (x, y) in dl
            _, loss, _, train_state = Training.single_train_step!(
                AutoEnzyme(), MSELoss(), (x, y), train_state; return_gradients=Val(false)
            )
        end
    end
end

WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1767822922.071752    9312 profiler_session.cc:103] Profiler session initializing.
I0000 00:00:1767822922.071800    9312 profiler_session.cc:118] Profiler session started.
I0000 00:00:1767822962.793289    9312 profiler_session.cc:68] Profiler session collecting data.
I0000 00:00:1767822962.865438    9312 save_profile.cc:150] Collecting XSpace to repository: /tmp/lux_training_trace/plugins/profile/2026_01_07_21_56_02/runnervmh13bl.xplane.pb
I0000 00:00:1767822962.925005    9312 save_profile.cc:123] Creating directory: /tmp/lux_training_trace/plugins/profile/2026_01_07_21_56_02

I0000 00:00:1767822962.973125    9312 save_profile.cc:129] Dumped gzipped tool data for trace.json.gz to /tmp/lux_training_trace/plugins/profile/2026_01_07_21_56_02/runnervmh13bl.trace.json.gz
I0000 00:00:1767822962.992626    9312 profiler_session.cc:136] Profiler session tear down.

Once the run is completed, you can use xprof to analyze the traces. An example of the output is shown below: