torchgpipe

A GPipe implementation in PyTorch. It is optimized for CUDA rather than TPU.

from torchgpipe import GPipe
model = nn.Sequential(a, b, c, d)
model = GPipe(model, balance=[1, 1, 1, 1], chunks=8)
output = model(input)

What is GPipe?

GPipe is a scalable pipeline parallelism library published by Google Brain, which allows efficient training of large, memory-consuming models. According to the paper, GPipe can train a 25x larger model by using 8x devices (TPU), and train a model 3.5x faster by using 4x devices.

GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism

Google trained AmoebaNet-B with 557M parameters over GPipe. This model has achieved 84.3% top-1 and 97.0% top-5 accuracy on ImageNet classification benchmark (the state-of-the-art performance as of May 2019).

GPipe uses (a) pipeline parallelism and (b) automatic recomputation of the forward propagation during the backpropagation, hence leverages training a large model. We refer to (b) as checkpointing, following the well-known terminology in PyTorch community.

Pipeline Parallelism

GPipe splits a model into multiple partitions and places each partition on a different device to occupy more memory capacity. And it splits a mini-batch into multiple micro-batches to make the partitions work as parallel as possible.

Checkpointing

Checkpointing is applied to each partition to minimize the overall memory consumption by a model. During forward propagation, only the tensors at the boundaries between partitions are remembered. All other intermediate tensors are volatilized, and recomputed during backpropagation when necessary.

Usage

Currently, torchgpipe requires the following environments:

• Python 3.6+
• PyTorch 1.1+

To use torchgpipe, install it via PyPI:

$ pip install torchgpipe

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, ...].

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)

Documentation

Visit torchgpipe.readthedocs.io for more information including the API references.

Benchmarking

The full details and more benchmarks are available in torchgpipe.readthedocs.io.

ResNet-101 Accuracy Benchmark

GPipe should be transparent not to introduce additional hyperparameter tuning. To verify the transparency, we reproduced top-1 error rate of ResNet-101 on ImageNet, as reported in Table 2(c) of Accurate, Large Minibatch SGD by Goyal et al.

U-Net (B, C) Memory Benchmark

The table shows how GPipe facilitates scaling U-Net models. baseline denotes the baseline without pipeline parallelism nor checkpointing, and pipeline-1, -2, -4, -8 denotes that the model is trained with GPipe with the corresponding number of partitions.

Here we used a simplified U-Net architecture. The size of a model is determined by hyperparameters B and C which are proportional to the number of layers and filters, respectively.

U-Net (5, 64) Speed Benchmark

To verify efficiency with skip connections, we measured the throughput of U-Net with various number of devices. We chose to use U-Net since it has several long skip connections.

AmoebaNet-D (18, 256) Speed Benchmark

(n: number of partitions, m: number of micro-batches)

The table shows the reproduced speed benchmark on AmoebaNet-D (18, 256), as reported in Table 2 of GPipe by Huang et al. Note that we replaced K in the paper with n.

Notes

This project is functional, but the interface is not confirmed yet. All public APIs are subject to change without warning until v0.1.0.

Authors and Licensing

torchgpipe project is developed by Heungsub Lee, Myungryong Jeong, and Chiheon Kim at Kakao Brain, with Sungbin Lim, Ildoo Kim, Woonhyuk Baek, and Boogeon Yoon's help. It is distributed under the 3-clause BSD license.

Citation

If you apply this library to any project and research, please cite our code:

@article{kim2020torchgpipe,
    title={torchgpipe: On-the-fly Pipeline Parallelism for Training Giant Models},
    author={Chiheon Kim and Heungsub Lee and Myungryong Jeong and Woonhyuk Baek and Boogeon Yoon and Ildoo Kim and Sungbin Lim and Sungwoong Kim},
    year={2020},
    eprint={2004.09910},
    archivePrefix={arXiv}
}