Function pretrain

linear_moe/training.py:51–182

Main training program. Refer to https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/training.py This function will run the followings in the order provided: 1) initialize Megatron. 2) setup model, optimizer and lr schedule using the model_provider. 3) call train

(train_valid_test_dataset_provider,
             model_provider,
             model_type,
             forward_step_func,
             process_non_loss_data_func=None,
             extra_args_provider=None,
             args_defaults={'tokenizer_type': 'GPT2BPETokenizer'})

def pretrain(train_valid_test_dataset_provider,
             model_provider,
             model_type,
             forward_step_func,
             process_non_loss_data_func=None,
             extra_args_provider=None,
             args_defaults={'tokenizer_type': 'GPT2BPETokenizer'}):
    """Main training program.
    Refer to https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/training.py

    This function will run the followings in the order provided:
        1) initialize Megatron.
        2) setup model, optimizer and lr schedule using the model_provider.
        3) call train_val_test_data_provider to get train/val/test datasets.
        4) train the model using the forward_step_func.

    Arguments:
        train_valid_test_dataset_provider: a function that takes the size of
            train/valid/test dataset and returns `train, valid, test` datasets.
        model_provider: a function that returns a vanilla version of the
            model. By vanilla we mean
             a simple model on cpu with no fp16 or ddp.
        model_type: an enum that specifies the type of model being trained.
        forward_step_func: a function that takes a `data iterator` and `model`,
            and returns a `loss` scalar with a dictionary with key:values being
            the info we would like to monitor during training, for example
            `lm-loss: value`. We also require that this function add
            `batch generator` to the timers class.
        process_non_loss_data_func: a function to post process outputs of the
            network. It can be used for dumping output tensors (e.g images) to
            tensorboard. It takes `collected data`(list of tensors),
            `current iteration index` and `tensorboard writer` as arguments.
        extra_args_provider: a function that takes a parser and adds arguments
            to it. It is used for programs to add their own arguments.
        args_defaults: a dictionary from argument-name to argument-value. It
            to set already parse arguments.
    """

    from megatron.initialize import initialize_megatron
    initialize_megatron(extra_args_provider=extra_args_provider,
                        args_defaults=args_defaults)

    # Set pytorch JIT layer fusion options and warmup JIT functions.
    set_jit_fusion_options()

    # Adjust the startup time so it reflects the largest value.
    # This will be closer to what scheduler will see (outside of
    # image ... launches.
    global _TRAIN_START_TIME
    start_time_tensor = torch.cuda.DoubleTensor([_TRAIN_START_TIME])
    torch.distributed.all_reduce(start_time_tensor,
                                 op=torch.distributed.ReduceOp.MIN)
    _TRAIN_START_TIME = start_time_tensor.item()
    print_rank_0('time to initialize megatron (seconds): {:.3f}'.format(
        time.time() - _TRAIN_START_TIME))
    print_datetime('after megatron is initialized')

    args = get_args()