megatron.training#

Description

Pretrain utilities.

Functions

build_train_valid_test_data_iterators(...[, ...])

cyclic_iter(iter)

evaluate(forward_step_func, data_iterator, ...)

Evaluation.

evaluate_and_print_results(prefix, ...[, ...])

Helper function to evaluate and dump results on screen.

get_model(model_provider_func[, model_type, ...])

Build the model.

pretrain(args, ...[, ...])

Main training program.

print_datetime(string)

Note that this call will sync across all ranks.

save_checkpoint_and_time(iteration, model, ...)

train_step(forward_step_func, data_iterator, ...)

Single training step.

training_log(loss_dict, total_loss_dict, ...)

Log training information such as losses, timing, ....