Gradient Synchronization

PyTorch’s distributed module operates by communicating back and forth between all of the GPUs in your system. This communication takes time, and ensuring all processes know the states of each other happens at particular triggerpoints when using the ddp module.

These triggerpoints are added to the PyTorch model, specifically their forward() and backward() methods. This happens when the model is wrapped with DistributedDataParallel:

import torch.nn as nn
from torch.nn.parallel import DistributedDataParallel

model = nn.Linear(10, 10)
ddp_model = DistributedDataParallel(model)

In 🤗 Accelerate this conversion happens automatically when calling prepare() and passing in your model.

+ from accelerate import Accelerator
+ accelerator = Accelerator()
  import torch.nn as nn
- from torch.nn.parallel import DistributedDataParallel

  model = nn.Linear(10,10)
+ model = accelerator.prepare(model)

The slowdown in gradient accumulation

You now understand that PyTorch adds hooks to the forward and backward method of your PyTorch model when training in a distributed setup. But how does this risk slowing down your code?

In DDP (distributed data parallel), the specific order in which processes are performed and ran are expected at specific points and these must also occur at roughly the same time before moving on.

The most direct example is when you update all of the parameters in a model through .backward(). All instances of the model need to have updated their gradients, collated, and updated again before moving onto the next batch of data. But when performing gradient accumulation, you accumulate n losses and skip .backward() until n batches have been reached. This can cause a significant slowdown since all the processes need to communicate with them more times than needed. How can you avoid this overhead?

Solving the slowdown problem

Since you are skipping these batches, their gradients do not need to be synchronized until the point where .backward() is actually called. PyTorch cannot automagically tell when you need to do this, but they do provide a tool to help through the no_sync context manager that is added to your model after converting it to DDP.

Under this context manager, PyTorch will skip synchronizing the gradients when .backward() is called, and the first call to .backward() outside this context manager will trigger the synchronization. See an example below:

ddp_model, dataloader = accelerator.prepare(model, dataloader)

for index, batch in enumerate(dataloader):
    inputs, targets = batch
    # Trigger gradient synchronization on the last batch
    if index != (len(dataloader) - 1):
        with ddp_model.no_sync():
            # Gradients only accumulate
            outputs = ddp_model(inputs)
            loss = loss_func(outputs)
            accelerator.backward(loss)
    else:
        # Gradients finally sync
        outputs = ddp_model(inputs)
        loss = loss_func(outputs)
        accelerator.backward(loss)

In 🤗 Accelerate to make this an API that can be called no matter the training device (though it may not do anything if you are not in a distributed system!), ddp_model.no_sync gets replaced with no_sync() and operates the same way:

  ddp_model, dataloader = accelerator.prepare(model, dataloader)

  for index, batch in enumerate(dataloader):
      inputs, targets = batch
      # Trigger gradient synchronization on the last batch
      if index != (len(dataloader)-1):
-         with ddp_model.no_sync():
+         with accelerator.no_sync(model):
              # Gradients only accumulate
              outputs = ddp_model(inputs)
              loss = loss_func(outputs, targets)
              accelerator.backward(loss)
      else:
          # Gradients finally sync
          outputs = ddp_model(inputs)
          loss = loss_func(outputs)
          accelerator.backward(loss)

As you may expect, the accumulate() function wraps around this conditional check by keeping track of the current batch number, leaving you with the final gradient accumulation API:

ddp_model, dataloader = accelerator.prepare(model, dataloader)

for batch in dataloader:
    with accelerator.accumulate(model):
        optimizer.zero_grad()
        inputs, targets = batch
        outputs = model(inputs)
        loss = loss_function(outputs, targets)
        accelerator.backward(loss)

As a result, you should either use accelerator.accumulate or accelerator.no_sync when it comes to API choice.

Just how much of a slowdown is there, and easy mistakes you can make

To setup a realistic example, consider the following setup:

• Two single-GPU T4 nodes and one node with two GPUs
• Each GPU is a T4, and are hosted on GCP
• The script used is a modification of the NLP Example script
• Batch size per GPU is 16, and gradients are accumulated every 4 steps

All scripts are available in this repository.

If not careful about gradient synchronization and GPU communication, a large amount of time can be wasted from when these GPUs communicate to each other during unnessisary periods.

By how much?

Reference:

• Baseline: uses no synchronization practices discussed here
• no_sync improperly: no_sync only around the backward call, not the forward
• no_sync: using the no_sync pattern properly
• accumulate: using accumulate() properly

Below are the average seconds per batch iterating over 29 batches of data for each setup on both a single node and on the dual-node setup:

As you can see, if you are not careful about how you setup your gradient synchronization, you can get upwards of more than a 2x slowdown during training!

If you are worried about making sure everything is done properly, we highly recommend utilizing the accumulate() function and passing in gradient_accumulation_steps to the Accelerator object so Accelerate can handle this for you.