06 · Define train_loop_config

The train_loop_config is a simple dictionary of hyperparameters that Ray passes into your training loop (train_loop_ray_train).

• It acts as the bridge between the TorchTrainer and your per-worker training code.

• Anything defined here becomes available inside the config argument of train_loop_ray_train.

In this example we define:

• num_epochs → how many full passes through the dataset to run.

• global_batch_size → the total batch size across all workers (Ray will split this evenly across GPUs).

You can add other parameters here (like learning_rate, embedding_dim, etc.) and they’ll automatically be accessible in your training loop via config["param_name"].

# 06. Define the configuration dictionary passed into the training loop

# train_loop_config is provided to TorchTrainer and injected into
# train_loop_ray_train(config) as the "config" argument.
# → Any values defined here are accessible inside the training loop.

train_loop_config = {
    "num_epochs": 2,           # Number of full passes through the dataset
    "global_batch_size": 128   # Effective batch size across ALL workers
                               # (Ray will split this evenly per worker, e.g.
                               # with 8 workers → 16 samples/worker/step)
}