Pipeline Parallelism

Ascend Distributed Parallel Whole Process

Overview

In recent years, the scale of neural networks has increased exponentially. Limited by the memory on a single device, the number of devices used for training large models is also increasing. Due to the low communication bandwidth between servers, the performance of the conventional hybrid parallelism (data parallel + model parallel) is poor. Therefore, pipeline parallelism needs to be introduced. Pipeline parallelism can divide a model in space based on stage. Each stage needs to execute only a part of the network, which greatly reduces memory overheads, shrinks the communication domain, and shortens the communication time. MindSpore can automatically convert a standalone model to the pipeline parallel mode based on user configurations.

Download address of the complete sample code:

https://gitee.com/mindspore/docs/tree/r1.5/docs/sample_code/distributed_training.

Preparations

Downloading the Dataset

This example uses the CIFAR-10 dataset. For details about how to download and load the dataset, visit https://www.mindspore.cn/docs/programming_guide/en/r1.5/distributed_training_ascend.html.

Configuring the Distributed Environment

Currently, pipeline parallelism supports only Ascend.

Due to the impact of HCCL, pipeline parallelism can only be performed across nodes.

For details about how to configure the distributed environment and call the HCCL, visit https://www.mindspore.cn/docs/programming_guide/en/r1.5/distributed_training_ascend.html.

Defining the Network

The network definition is the same as that of the Ascend 910 AI Processor.

For details about the definitions of the network, optimizer, and loss function, visit https://www.mindspore.cn/docs/programming_guide/en/r1.5/distributed_training_ascend.html.

To implement pipeline parallelism, you need to define the parallel strategy and call the pipeline_stage API to specify the stage on which each layer is to be executed. The granularity of the pipeline_stage API is Cell. pipeline_stage must be configured for all Cells that contain training parameters.

class ResNet(nn.Cell):
    """ResNet"""

    def __init__(self, block, num_classes=100, batch_size=32):
        """init"""
        super(ResNet, self).__init__()
        self.batch_size = batch_size
        self.num_classes = num_classes

        self.head = Head()
        self.layer1 = MakeLayer0(block, in_channels=64, out_channels=256, stride=1)
        self.layer2 = MakeLayer1(block, in_channels=256, out_channels=512, stride=2)
        self.layer3 = MakeLayer2(block, in_channels=512, out_channels=1024, stride=2)
        self.layer4 = MakeLayer3(block, in_channels=1024, out_channels=2048, stride=2)

        self.pool = ops.ReduceMean(keep_dims=True)
        self.squeeze = ops.Squeeze(axis=(2, 3))
        self.fc = fc_with_initialize(512 * block.expansion, num_classes)

        # pipeline parallel config
        self.head.pipeline_stage = 0
        self.layer1.pipeline_stage = 0
        self.layer2.pipeline_stage = 0
        self.layer3.pipeline_stage = 1
        self.layer4.pipeline_stage = 1
        self.fc.pipeline_stage = 1

    def construct(self, x):
        """construct"""
        x = self.head(x)

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        x = self.pool(x, (2, 3))
        x = self.squeeze(x)
        x = self.fc(x)
        return x

Training the Network

To enable pipeline parallelism, you need to add the following configurations to the training script:

Set pipeline_stages in set_auto_parallel_context to specify the total number of stages.
Set the SEMI_AUTO_PARALLEL mode. Currently, the pipeline parallelism supports only this mode.
Define the LossCell. In this example, the nn.WithLossCell API is called.
Pass the parameters used in this stage to the optimizer. Call the add_pipeline_stage method of Parameter to pass all stage information to Parameter if multiple stages share a parameter. Then, call the infer_param_pipeline_stage API of the Cell to obtain the training parameters of the stage.
Finally, wrap the LossCell with PipelineCell, and specify the Micro_batch size. To improve machine utilization, MindSpore divides Mini_batch into finer-grained Micro_batch to streamline the entire cluster. The final loss value is the sum of the loss values computed by all Micro_batch. The size of Micro_batch must be greater than or equal to the number of stages.

from mindspore import context, Model, nn
from mindspore.nn import Momentum
from mindspore.train.callback import LossMonitor
from mindspore.context import ParallelMode
from resnet import resnet50

def test_train_cifar(epoch_size=10):
    context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL, gradients_mean=True)
    context.set_auto_parallel_context(pipeline_stages=2, save_graphs=True)
    loss_cb = LossMonitor()
    data_path = os.getenv('DATA_PATH')
    dataset = create_dataset(data_path)
    batch_size = 32
    num_classes = 10
    net = resnet50(batch_size, num_classes)
    loss = SoftmaxCrossEntropyExpand(sparse=True)
    net_with_loss = nn.WithLossCell(net, loss)
    net_pipeline = nn.PipelineCell(net_with_loss, 2)
    opt = Momentum(net.infer_param_pipeline_stage(), 0.01, 0.9)
    model = Model(net_pipeline, optimizer=opt)
    model.train(epoch_size, dataset, callbacks=[loss_cb], dataset_sink_mode=True)

Running the Multi-node Script

Pipeline parallelism requires cross-node execution. For details about multi-node multi-device training, visit https://www.mindspore.cn/docs/programming_guide/en/r1.5/distributed_training_ascend.html.