Local SGD with Accelerator

You've implemented gradient accumulation and gradient checkpointing to streamline memory usage for your language translation model. Training is still a bit slow, so you decide to add local SGD to your training loop to improve communication efficiency between devices. Build the training loop with local SGD!

The model, train_dataloader, and accelerator have been pre-defined, and LocalSGD has been imported.

本练习是课程的一部分

Efficient AI Model Training with PyTorch

查看课程

练习说明

Set local_sgd_steps to synchronize gradients every eight steps.
Step the local SGD context manager.

交互式实操练习

通过完成这段示例代码来试试这个练习。

# Set up a context manager to synchronize gradients every eight steps
with LocalSGD(accelerator=accelerator, model=model, local_sgd_steps=____, enabled=True) as local_sgd:
    for batch in train_dataloader:
        with accelerator.accumulate(model):
            inputs, targets = batch["input_ids"], batch["labels"]
            outputs = model(inputs, labels=targets)
            loss = outputs.loss
            accelerator.backward(loss)
            optimizer.step()
            lr_scheduler.step()
            optimizer.zero_grad()
            # Step the local SGD context manager
            local_sgd.____()

编辑并运行代码