Skip to content

Latest commit

 

History

History
209 lines (124 loc) · 10.8 KB

Rocm-ML_Deployment_User_Guide-v04.md

File metadata and controls

209 lines (124 loc) · 10.8 KB

AMDh_mb_8bit

AMD ROCm Deployment User Guide
Publication # 1.0 Revision: 4.0
Issue Date: November 2019
© 2019 Advanced Micro Devices, Inc. All rights reserved.

A close up of a newspaper Description automatically generated

Table of Contents

Revision History 4

Chapter 1 Introduction 6

Chapter 2 ROCm Setup and Installation 6

2.1 System Requirements 6

2.2 Installing ROCm 7

2.3 Installing the Applications 7

2.4 Running the Machine Learning Application 8

2.4.1 Tensorflow 8

2.4.2 PyTorch 9

2.5 Running HPC Application 10

2.5.1 NAMD 10

2.6 Known System Issues 10

Revision History

Date Revision Description
October 2019 1.0 Initial preliminary release
October 2019 2.0 Added Known System Issues
October 2019 3.0 Added Benchmark for Tensorflow and PyTorch
November 2019 4.0 Removed Benchmark for Tensorflow and PyTorch

Introduction

This guide covers the basic instructions needed to install the ROCm software suite of applications using the command line interface and verify that these Machine Learning (ML) and High-Performance Computing (HPC) applications can run on supported frameworks.

The instructions are intended to be used on a clean installation of a supported application. The document also discusses the scale-out of the High Performing Computing (HPC) and Machine Learning (ML) applications on the AMD platform.

ROCm Setup and Installation

System Requirements

To use the Machine Learning and High-Performance Computing applications on your system, you will need the following hardware and software installed:

Software Requirements

Supported Operating Systems
Ubuntu v18.04

  • CentOS v7.6

  • REHL 7.6

Note: You must install and verify the supported operating system before installing ROCm.

Hardware Requirements

You must ensure you can view the VGA/3D controllers to determine if the cards are detected prior to the installation of the ROCm framework.

To detect the cards, from the command line, enter:

sudo lshw -c video

You will see the output for each PCle device. The vendor in the output must display as "Advanced Micro Devices, Inc."

For example, see the code sample below:

*-display description: Display controller product: Vega 20 vendor: Advanced Micro Devices, Inc. [AMD/ATI] physical id: 0 bus info: pci@0000:03:00.0 version: 02 width: 64 bits clock: 33MHz capabilities: pm pciexpress msi bus_master cap_list rom configuration: driver=amdgpu latency=0

Installing ROCm

To install the ROCm application, run RET with the install command:

sudo apt -y install git cd ~/ wget https://github.com/rocmsys/RET.git cd RET sudo ./ret install rocm #see all options sudo ./ret -h

For more details, refer https://rocm.github.io/ROCmInstall.html

Installing the Applications

To install the Tensorflow, PyTorch, NAMD and other applications, enter

cd ~/RET sudo ./ret install <my application> #For example, sudo ./ret install tensorflow sudo ./ret install pytorch sudo ./ret install namd

Running the Machine Learning Application

You can run the applications using Tensorflow and PyTorch for Machine Learning.

Tensorflow

Training a Machine Learning Model Using Tensorflow

To train a machine learning model using Tensorflow, use the example below:

  1. Clone the Tensorflow test model.
git clone https://github.com/tensorflow/models
  1. Download the CIFAR-10 dataset using:
pip3 install tensorflow_datasets cd models/tutorials/image/cifar10_estimator python3 generate_cifar10_tfrecords.py --data-dir=${PWD}/cifar-10-data
  1. To run on a single node (single GPU), enter
TF_ROCM_FUSION_ENABLE=1 python3 cifar10_main.py \ --data-dir=${PWD}/cifar-10-data \ --job-dir=/tmp/cifar10 \ --num-gpus=1 \ --train-steps=100
  1. To run on a single node (multi GPUs (data parallelism)), enter
TF_ROCM_FUSION_ENABLE=1 python3 cifar10_main.py \ --data-dir=${PWD}/cifar-10-data \ --job-dir=/tmp/cifar10 \ --num-gpus=2 \ --train-steps=100
  1. To run on multi nodes

TBD

PyTorch

Training a Machine Learning Model Using PyTorch

To train a machine learning model using PyTorch, use the example below:

  1. Download the script wget
mkdir pytorch cd pytorch wget https://raw.githubusercontent.com/wiki/ROCmSoftwarePlatform/pytorch/micro_benchmarking_pytorch.py wget https://raw.githubusercontent.com/wiki/ROCmSoftwarePlatform/pytorch/fp16util.py wget https://raw.githubusercontent.com/wiki/ROCmSoftwarePlatform/pytorch/shufflenet.py wget https://raw.githubusercontent.com/wiki/ROCmSoftwarePlatform/pytorch/shufflenet_v2.py
  1. To run on a single node (single GPU), enter
python3 micro_benchmarking_pytorch.py \ --network resnet50 \ --batch-size 128 \ --fp16 1
  1. To run on a single node (multi GPU (Data Parallelism)), enter
python3 micro_benchmarking_pytorch.py \ --network resnet50 \ --batch-size 128 \ --fp16 1 \ --dataparallel \ --device_ids 0,1

Running HPC Application

NAMD

After the installation:

cd ../namd ./namd2 src/alanin -d

Known System Issues

The following table consists of the known system issues and the recommended resolutions in this release:

Known Issue Resolution
Error loading shared library libopenblas.so.3: No such file or directory >ImportError: libopenblas.so.0: cannot open shared object file: No such file or directory Run the following command to resolve this error:
sudo apt-get install libopenblas-base export LD_LIBRARY_PATH=/usr/lib/openblas-base/
Error loading module ‘torchvision’. The torchvision module consists of popular datasets, model architectures, and common image transformations for computer vision. >ImportError: No module named 'torchvision' To resolve this error, use the following command to install the ‘torchvision’ module
pip3 install torchvision
The error message indicates a problem with the locale setting. >Error: unsupported locale setting Modify the locale to fix the problem.
export LC_ALL=C