PyTorch

This section describes how to install and run PyTorch and how to install Horovod to perform distributed learning.

Running PyTorch on a single node

Precondition

Replace grpname with your own ABCI group.
The Python virtual environment should be created in the home or group area so that it can be referenced by interactive nodes and each compute node.
The sample program should be saved in the home or group area so that it can be referenced by interactive nodes and each compute node.

Installation

Here are the steps to create a Python virtual environment and install PyTorch into the Python virtual environment.

[username@es1 ~]$ qrsh -g grpname -l rt_G.small=1 -l h_rt=1:00:00
[username@g0001 ~]$ module load python/3.11 cuda/11.8 cudnn/8.6
[username@g0001 ~]$ python3 -m venv ~/venv/pytorch
[username@g0001 ~]$ source ~/venv/pytorch/bin/activate
(pytorch) [username@g0001 ~]$ pip3 install --upgrade pip setuptools
(pytorch) [username@g0001 ~]$ pip3 install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --extra-index-url https://download.pytorch.org/whl/cu118

With the installation, you can use PyTorch next time you want to use it by simply loading the module and activating the Python virtual environment, as follows.

[username@g0001 ~]$ module load python/3.11 cuda/11.8 cudnn/8.6
[username@g0001 ~]$ source ~/venv/pytorch/bin/activate

Execution

The following shows how to execute the PyTorch sample program main.py in the case of an interactive job and a batch job.

Run as an interactive job

[username@es1 ~]$ qrsh -g grpname -l rt_G.small=1 -l h_rt=1:00:00
[username@g0001 ~]$ module load python/3.11 cuda/11.8 cudnn/8.6
[username@g0001 ~]$ source ~/venv/pytorch/bin/activate
(pytorch) [username@g0001 ~]$ git clone https://github.com/pytorch/examples.git
(pytorch) [username@g0001 ~]$ cd examples/mnist
(pytorch) [username@g0001 ~]$ python3 main.py

Run as a batch job

Save the following job script as a run.sh file.

#!/bin/sh

#$ -l rt_G.small=1
#$ -j y
#$ -cwd

source /etc/profile.d/modules.sh
module load python/3.11 cuda/11.8 cudnn/8.6
source ~/venv/pytorch/bin/activate
git clone https://github.com/pytorch/examples.git
cd examples/mnist
python3 main.py
deactivate

Submit a saved job script run.sh as a batch job with the qsub command.

[username@es1 ~]$ qsub -g grpname run.sh
Your job 1234567 ('run.sh') has been submitted

Running PyTorch on multiple nodes

Precondition

Replace grpname with your own ABCI group.
The Python virtual environment should be created in the home or group area so that it can be referenced by interactive nodes and each compute node.
The sample program should be saved in the home or group area so that it can be referenced by interactive nodes and each compute node.

Installation

Here are the steps to create a Python virtual environment and install PyTorch and Horovod into the Python virtual environment.

[username@es1 ~]$ qrsh -g grpname -l rt_G.small=1 -l h_rt=1:00:00
[username@g0001 ~]$ module load python/3.11 cuda/11.8 cudnn/8.6 nccl/2.16 hpcx-mt/2.12
[username@g0001 ~]$ python3 -m venv ~/venv/pytorch+horovod
[username@g0001 ~]$ source ~/venv/pytorch+horovod/bin/activate
(pytorch+horovod) [username@g0001 ~]$ pip3 install --upgrade pip setuptools
(pytorch+horovod) [username@g0001 ~]$ pip3 install wheel
(pytorch+horovod) [username@g0001 ~]$ pip3 install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --extra-index-url https://download.pytorch.org/whl/cu118
(pytorch+horovod) [username@g0001 ~]$ HOROVOD_NCCL_LINK=SHARED HOROVOD_WITH_PYTORCH=1 HOROVOD_GPU_OPERATIONS=NCCL HOROVOD_NCCL_HOME=$NCCL_HOME HOROVOD_WITHOUT_GLOO=1 pip3 install --no-cache-dir horovod==0.27.0

With the installation, you can use PyTorch and Horovod next time you want to use it by simply loading the module and activating the Python virtual environment, as follows.

[username@g0001 ~]$ module load python/3.11 cuda/11.8 cudnn/8.6 nccl/2.16 hpcx-mt/2.12
[username@g0001 ~]$ source ~/venv/pytorch+horovod/bin/activate

Execution

The following shows how to execute a sample program pytorch_mnist.py of PyTorch with Horovod for distributed learning.

Run as an interactive job

In this example, using 4 GPUs in a compute node for distributed learning.

[username@es1 ~]$ qrsh -g grpname -l rt_G.large=1 -l h_rt=1:00:00
[username@g0001 ~]$ module load python/3.11 cuda/11.8 cudnn/8.6 nccl/2.16 hpcx-mt/2.12
[username@g0001 ~]$ source ~/venv/pytorch+horovod/bin/activate
(pytorch+horovod) [username@g0001 ~]$ git clone -b v0.27.0 https://github.com/horovod/horovod.git
(pytorch+horovod) [username@g0001 ~]$ mpirun -np 4 -map-by ppr:4:node -mca pml ob1 python3 horovod/examples/pytorch/pytorch_mnist.py

Run as a batch job

In this example, a total of 8 GPUs are used for distributed learning. 2 compute nodes are used, with 4 GPUs per compute node.

Save the following job script as a run.sh file.

#!/bin/sh

#$ -l rt_F=2
#$ -j y
#$ -cwd

source /etc/profile.d/modules.sh
module load python/3.11 cuda/11.8 cudnn/8.6 nccl/2.16 hpcx-mt/2.12
source ~/venv/pytorch+horovod/bin/activate

git clone -b v0.27.0 https://github.com/horovod/horovod.git

NUM_GPUS_PER_NODE=4
NUM_PROCS=$(expr ${NHOSTS} \* ${NUM_GPUS_PER_NODE})

MPIOPTS="-hostfile $SGE_JOB_HOSTLIST -np ${NUM_PROCS} -map-by ppr:${NUM_GPUS_PER_NODE}:node -mca pml ob1 -mca btl self,tcp -mca btl_tcp_if_include bond0"

mpirun ${MPIOPTS} python3 horovod/examples/pytorch/pytorch_mnist.py

deactivate

Submit a saved job script run.sh as a batch job with the qsub command.

[username@es1 ~]$ qsub -g grpname run.sh
Your job 1234567 ('run.sh') has been submitted