Skip to content

anijain2305/pipedream

 
 

Repository files navigation

PipeDream: Generalized Pipeline Parallelism for DNN Training

This repository contains the source code implementation of the SOSP paper "PipeDream: Generalized Pipeline Parallelism for DNN Training". This work was done as part of Microsoft Research's Project Fiddle. This source code is available under the MIT License.

Directory Structure

graph

This contains a Python implementation of a graph, used by the PipeDream profiler and optimizer. Profiling scripts in profiler generate graph profiles, that can then be ingested by the optimizer located in optimizer to generate a partitioned model, that can then be fed to the PipeDream runtime.

profiler

Instrumented PyTorch applications which return profiles that can be ingested by the optimizer.

optimizer

A Python implementation of PipeDream's optimizer.

runtime

PipeDream's runtime, which implements model parallelism, as well as input pipelining in PyTorch. This can be fused with data parallelism to give hybrid model and data parallelism, and input pipelining.

Setup

Software Dependencies

To run PipeDream, you will need a NVIDIA GPU with CUDA 10.0, GPU driver version 418.56, nvidia-docker2, and Python 3. On a Linux server with NVIDIA GPU(s) and Ubuntu 16.04, these dependencies can be installed using,

bash setup.sh

All dependencies are in the nvcr.io/nvidia/pytorch:19.05-py3 container, which can be downloaded using,

nvidia-docker pull nvcr.io/nvidia/pytorch:19.05-py3

To run the PipeDream profiler, you will need to build a new Docker image, which can be done using the Dockerfile in this directory. Note that the Dockerfile has a dependency on the pre_hook.patch and requirements.txt files in this directory. This container can be built using,

docker build --tag <CONTAINER_NAME> .

The PyTorch Docker Container can then be run using,

nvidia-docker run -it -v /mnt:/mnt --ipc=host --net=host <CONTAINER_NAME> /bin/bash

Data

Image Classification

All image classification experiments are run using the ImageNet ILSVC 2012 dataset. This can be downloaded using the following command (within the docker container above),

cd scripts; python download_imagenet.py --data_dir <DATASET_DIR>

Note that the ImageNet dataset is about 145GB, so this download script can take some time.

Translation

All translation experiments are run using the WMT En-De dataset, also used for the MLPerf translation (RNN) task. This can be downloaded using the instructions in the MLPerf repository.

End-to-end Workflow

To run a demo, run the following commands (the optimizer and runtime have been verified to work unchanged in nvcr.io/nvidia/pytorch:19.05-py3). More detailed instructions for each of the individual components are in the corresponding directory READMEs, and more detailed instructions on how to run the main experiments in the SOSP paper are in EXPERIMENTS.md.

[from pipedream/profiler/image_classification; you will need to have the changes to PyTorch listed above] Note that the profiling step must be run with only a single GPU (hence the CUDA_VISIBLE_DEVICES=0 before the command).

CUDA_VISIBLE_DEVICES=0 python main.py -a vgg16 -b 64 --data_dir <path to ImageNet directory>

[from pipedream/optimizer]

python optimizer_graph_hierarchical.py -f ../profiler/image_classification/profiles/vgg16/graph.txt -n 4 --activation_compression_ratio 1 -o vgg16_partitioned

[from pipedream/optimizer]

python convert_graph_to_model.py -f vgg16_partitioned/gpus=4.txt -n VGG16Partitioned -a vgg16 -o ../runtime/image_classification/models/vgg16/gpus=4 --stage_to_num_ranks 0:3,1:1

[from pipedream/runtime/image_classification; run on 4 GPUs (including a single server with 4 GPUs)]

python main_with_runtime.py --module models.vgg16.gpus=4 -b 64 --data_dir <path to ImageNet> --rank 0 --local_rank 0 --master_addr <master IP address> --config_path models/vgg16/gpus=4/hybrid_conf.json --distributed_backend gloo
python main_with_runtime.py --module models.vgg16.gpus=4 -b 64 --data_dir <path to ImageNet> --rank 1 --local_rank 1 --master_addr <master IP address> --config_path models/vgg16/gpus=4/hybrid_conf.json --distributed_backend gloo
python main_with_runtime.py --module models.vgg16.gpus=4 -b 64 --data_dir <path to ImageNet> --rank 2 --local_rank 2 --master_addr <master IP address> --config_path models/vgg16/gpus=4/hybrid_conf.json --distributed_backend gloo
python main_with_runtime.py --module models.vgg16.gpus=4 -b 64 --data_dir <path to ImageNet> --rank 3 --local_rank 3 --master_addr <master IP address> --config_path models/vgg16/gpus=4/hybrid_conf.json --distributed_backend gloo

master IP address here is the IP address of the rank 0 process. On a server with 4 GPUs, localhost can be specified.

When running DP setups, please use the nccl backend for optimal performance. When running hybrid setups, please use the gloo backend.

Code of Conduct

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.

License

Copyright (c) Microsoft Corporation. All rights reserved.

Licensed under the MIT license.

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 98.9%
  • Other 1.1%