User Guide

Installation

torchgpipe is available on PyPI. Install by pip:

$ pip install torchgpipe

Python 3.6+ (CPython) is required.

PyTorch 1.0+ will be installed automatically if you don’t have a satisfied one. However, we highly recommend you to use the latest version of PyTorch.

Applying GPipe

To train a module with GPipe, simply wrap it with torchgpipe.GPipe. Your module must be nn.Sequential as GPipe will automatically split the module into partitions with consecutive layers. balance argument determines the number of layers in each partition. chunks argument specifies the number of micro-batches. Input, output, and intermediate tensors must be Tensor or Tuple[Tensor, ...]. See also Restrictions for more details.

The below example code shows how to split a module with four layers into four partitions each having a single layer. This code also splits a mini-batch into 8 micro-batches:

from torchgpipe import GPipe

model = nn.Sequential(a, b, c, d)
model = GPipe(model, balance=[1, 1, 1, 1], chunks=8)

for input in data_loader:
    output = model(input)

GPipe optimizes training using CUDA. You should not move the module to a GPU yourself, because GPipe automatically moves each partition over different devices. By default, available GPUs starting from cuda:0 are selected in order for each partition. You can also specify GPUs to select by devices parameter:

mode = GPipe(model,
             balance=[1, 1, 1, 1],
             devices=[4, 5, 6, 7],  # Specify GPUs.
             chunks=8)

Input and Output Device

Unlike a typical module, with GPipe, the input device is different from the output device except there is only one partition. This is because the first partition and last partition should be placed in different devices.

Therefore, you should move the input and target to the corresponding devices. It can be done with GPipe.devices, which holds the list of devices for each partition:

in_device = model.devices[0]
out_device = model.devices[-1]

for input, target in data_loader:
    # input on in_device
    input = input.to(in_device, non_blocking=True)

    # target on out_device
    target = target.to(out_device, non_blocking=True)

    # output on out_device
    output = model(input)
    loss = F.cross_entropy(output, target)
    loss.backward()
    ...

Automatic Balancing

It could be hard to determine the optimal balance of a model. In particular, if you are still designing a model, probably the model architecture may change over time. In this case, we highly recommend torchgpipe_balancing for automatic balancing. This library is also a part of torchgpipe package but not a part of the GPipe paper.

There are two balancing tools, balance_by_time() and balance_by_size(). Both are based on per-layer profiling. Just like PyTorch JIT, you need to feed a sample input into the model. balance_by_time() traces elapsed time of each layer, while balance_by_size() detects the CUDA memory usage of each layer. Choose a balancing tool for your needs:

from torchgpipe import GPipe
from torchgpipe_balancing import balance_by_time

sample = torch.rand(128, 3, 224, 224)
balance = balance_by_time(model, sample, partitions=4)

model = GPipe(model, balance, chunks=8)

Restrictions

If you get any errors, check the following restrictions first.

Sequential:

Your module must be nn.Sequential. For example, the models in torchvision are not sequential. They can’t be wrapped by GPipe directly:

>>> from torchvision.models.resnet import resnet101
>>> model = resnet101()
>>> type(model)
torchvision.models.resnet.ResNet
>>> GPipe(model, balance=..., chunks=...)
Traceback (most recent call last)
  ...
TypeError: non-sequential module cannot be partitioned

See the sequential ResNet example to figure out how to make a model into a nn.Sequential model.

nn.Sequential assumes that every underlying layer takes only one argument. Calling forward(x) on nn.Sequential(A(), B(), C()) is essentially the same as calling C(B(A(x))). Hence, you can’t design an underlying layer with multiple arguments:

class MyModule(nn.Module):
    def forward(self, a, b, c):
        return a + b - c

model = nn.Sequential(..., MyModule(), ...)
model(input)  # FAILS!

Tensor or Tensors:

As we discussed above, each layer must take only one argument due to nn.Sequential. There is one more restriction. Every underlying layers’ input and output must be Tensor or Tuple[Tensor, ...]:

# OK
def forward(input: Tensor) -> Tensor: ...
def forward(input: Tensor) -> Tuple[Tensor, Tensor]: ...
def forward(input: Tuple[Tensor, Tensor]) -> Tensor: ...

# Error
def forward(input1: Tensor, input2: Tensor) -> Tensor: ...
def forward(input: Tensor, label: str) -> Tensor: ...
def forward(input: Tensor) -> Dict[str, Tensor]: ...
def forward(input: Tensor) -> Tuple[Tensor, str]: ...

The reason is that GPipe can’t assume how the non-tensor inputs for a mini-batch can be split for micro-batches.

Complex Modules

This part of the documentation discusses how to implement a complex module compatible with GPipe. First, you should understand how GPipe works. See Understanding GPipe.

Skip Connections

Many deep learning models, such as ResNet or AmoebaNet, contain skip connections. There are two ways to implement skip connections. Let’s assume we have to implement a skip connection like this:

latent = layer1(input)
latent = layer2(latent)
output = layer3(latent) + input  # skip connection

To make this module sequential, we will define modules for each layer. Simply, a skip connection can be implemented by making underlying layers with Tuple[Tensor, Tensor] parameter and return type:

class Layer1(nn.Module):
    #         ┌────────────────┐
    # input --│-+-> layer1 ----│--> output
    #         │ '--------------│--> skip
    #         └────────────────┘
    def forward(self, input: Tensor) -> Tuple[Tensor, Tensor]:
        return layer1(input), input

class Layer2(nn.Module):
    #         ┌────────────────┐
    # input --│---> layer2 ----│--> output
    #  skip --│----------------│--> skip
    #         └────────────────┘
    def forward(self, input_and_skip: Tuple[Tensor, Tensor]) -> Tuple[Tensor, Tensor]:
        input, skip = input_and_skip
        return layer2(input), skip

class Layer3(nn.Module):
    #         ┌────────────────┐
    # input --│---> layer3 --+-│--> output
    #  skip --│--------------' │
    #         └────────────────┘
    def forward(self, input_and_skip: Tuple[Tensor, Tensor]) -> Tensor:
        input, skip = input_and_skip
        return layer3(input) + skip

model = nn.Sequential(Layer1(), Layer2(), Layer3())

Because of the skip connection being represented as a normal parameter, GPipe can move the tensors from partition to partition:

model = GPipe(model, balance=[1, 1, 1], chunks=8)

It is the most straightforward approach to implement skip connections. But there is a disadvantage. In the above example, the skipping input tensor is copied to the second device, but it is never used at the device. Unnecessarily copied tensor wastes time and memory.

The following section introduces alternative approach for skip connection.

Long Skip Connections

The disadvantage mentioned above might be catastrophic if the unnecessarily copied tensor is very large, or it is copied over many devices. The second case often occurs when implementing long skip connections. Let’s assume now we have 8 layers between input and output:

latent = layer1(input)
latent = layer2(latent)
latent = layer3(latent)
latent = layer4(latent)
latent = layer5(latent)
latent = layer6(latent)
latent = layer7(latent)
output = layer8(latent) + input  # skip connection

With the prior approach, GPipe will copy the skipping input tensor to all devices, but 6 of them are unnecessary. The alternative approach is managing the skipping tensor manually in the module code. Now we will introduce a shared memory between Layer1 and Layer8 to toss the tensor without going through regardless layers. We might use a global variable named skip_buf for the shared memory. But actually, this approach doesn’t work:

# !!!!!!!!!!!!!!!!!!!!!!!!!
# THIS IS A FAILING EXAMPLE
# !!!!!!!!!!!!!!!!!!!!!!!!!

# The shared memory between Layer1 and Layer8.
skip_buf = None

class Layer1(nn.Module):
    def forward(self, input: Tensor) -> Tensor:
        # Remember the skipping tensor.
        global skip_buf
        skip_buf = input

        return layer1(input)

class Layer2(nn.Module):
    def forward(self, input: Tensor) -> Tensor:
        return layer2(input)

...  # Layer3-7 are similar to Layer2.

class Layer8(nn.Module):
    def forward(self, input: Tensor) -> Tensor:
        # Retrieve the skipping tensor.
        global skip_buf
        skip = skip_buf

        # Release the shared memory.
        skip_buf = None

        # The tensor should be copied to the device manually.
        skip = skip.to(input.device)

        return layer8(input) + skip

Each layer is executed several times due to the multiple micro-batches. Partitions work together concurrently, so the shared memory would be overwritten in non-deterministic order.

For example, when Layer8 processes the first micro-batch, it might receive the third (or any) micro-batch as a skipping tensor if Layer1 has just processed the latter micro-batch. We need to separate the shared memory for different micro-batches.

Therefore, the key is an identifier of each micro-batch. How do we identify which micro-batch the partition is currently processing? torchgpipe provides current_microbatch() for this purpose. If you call the function in a GPipe context, it will return some tensor. The tensor identifies the current micro-batch. You can use this as simply a dictionary key, or the target of a weak reference:

from torchgpipe import current_microbatch

# The shared memory between Layer1 and Layer8 indexed by micro-batch.
skips: Dict[Tensor, Tensor] = {}

class Layer1(nn.Module):
    def forward(self, input: Tensor) -> Tensor:
        # Remember the skipping tensor per micro-batch.
        skips[current_microbatch()] = input

        return layer1(input)

...  # Layer2-7 are folded.

class Layer8(nn.Module):
    def forward(self, input: Tensor) -> Tensor:
        # Retrieve the skipping tensor for the current micro-batch.
        skip = skips.pop(current_microbatch())

        # The tensor should be copied to the device manually.
        skip = skip.to(input.device)

        return layer8(input) + skip

This approach is not required to everyone. Furthermore, we didn’t intend to modify user’s module to apply GPipe. However, long skip connections are one of the common building blocks in modern CNN models. That is why we have provided this functionality.

Detecting Recomputation

Checkpointing in GPipe performs forward propagations twice. The second forward propagation is called recomputation. This may cause a problem when a module such as nn.BatchNorm2d updates its buffers on each forward propagation. It should not update the buffers again during the recomputation. To achieve it, modules’ forward method should be able to detect that is the recomputation or the first forward progagation.

It can be done by is_recomputing(). This function returns True if the code is running on the recomputation:

class Counter(nn.Module):
    def __init__(self):
        super().__init__()
        self.counter = 0

    def forward(self, input):
        if not is_recomputing():
            self.counter += 1
        return input

Note

deferred_batch_norm=True on GPipe will prevent updating the running statistics twice.