G. Wang - Distributed Machine Learning with Python- Accelerating model training and serving with distributed systems (2022)（英文）行業(yè)資料

DistributedMachine

Learningwith

Python

Acceleratingmodeltrainingandservingwithdistributedsystems

GuanhuaWang

BIRMINGHAM—MUMBAI

DistributedMachineLearningwithPython

Copyright?2022PacktPublishing

Allrightsreserved.Nopartofthisbookmaybereproduced,storedinaretrievalsystem,ortransmittedinanyformorbyanymeans,withoutthepriorwrittenpermissionofthepublisher,exceptinthecaseofbriefquotationsembeddedincriticalarticlesorreviews.

Everyefforthasbeenmadeinthepreparationofthisbooktoensuretheaccuracyoftheinformationpresented.However,theinformationcontainedinthisbookissoldwithoutwarranty,eitherexpressorimplied.Neithertheauthor,norPacktPublishingoritsdealersanddistributors,willbeheldliableforanydamagescausedorallegedtohavebeencauseddirectlyorindirectlybythisbook.

PacktPublishinghasendeavoredtoprovidetrademarkinformationaboutallofthecompaniesandproductsmentionedinthisbookbytheappropriateuseofcapitals.However,PacktPublishingcannotguaranteetheaccuracyofthisinformation.

PublishingProductManager:AliAbidi

SeniorEditors:RoshanKumar,NathanyaDiaz

ContentDevelopmentEditors:TazeenShaikh,ShreyaMoharir

TechnicalEditor:DevanshiAyare

CopyEditor:SafisEditing

ProjectCoordinator:AparnaRavikumarNair

Proofreader:SafisEditing

Indexer:PratikShirodkar

ProductionDesigner:AlishonMendonca

MarketingCoordinators:AbeerRiyazDawe,ShifaAnsari

Firstpublished:May2022

Productionreference:1040422

PublishedbyPacktPublishingLtd.

LiveryPlace

35LiveryStreet

Birmingham

B32PB,UK.

ISBN978-1-80181-569-7

Tomyparents,YingHanandXinWang

Tomygirlfriend,JingYuan

–GuanhuaWang

Contributors

Abouttheauthor

GuanhuaWangisafinal-yearcomputersciencePh.D.studentintheRISELabatUCBerkeley,advisedbyProfessorIonStoica.Hisresearchliesprimarilyinthemachinelearningsystemsarea,includingfastcollectivecommunication,efficientin-parallelmodeltraining,andreal-timemodelserving.Hisresearchhasgainedlotsofattentionfrombothacademiaandindustry.Hewasinvitedtogivetalkstotop-tieruniversities(MIT,Stanford,CMU,Princeton)andbigtechcompanies(Facebook/Meta,Microsoft).Hereceivedhismaster'sdegreefromHKUSTandabachelor'sdegreefromSoutheastUniversityinChina.Hehasalsodonesomecoolresearchonwirelessnetworks.Helikesplayingsoccerandhasrunmultiplehalf-marathonsintheBayAreaofCalifornia.

Aboutthereviewers

JamshaidSohailispassionateaboutdatascience,machinelearning,computervision,andnaturallanguageprocessingandhasmorethan2yearsofexperienceintheindustry.HepreviouslyworkedataSiliconValley-basedstart-up,FunnelBeam,thefoundersofwhicharefromStanfordUniversity,asadatascientist.Currently,heisworkingasadatascientistatSystemsLimited.Hehascompletedover66onlinecoursesfromdifferentplatforms.HeauthoredthebookDataWranglingwithPython3.XforPacktPublishingandhasreviewedmultiplebooksandcourses.HeisalsodevelopingacomprehensivecourseondatascienceatEducativeandisintheprocessofwritingbooksformultiplepublishers.

HiteshHindujaisanardentAIenthusiastworkingasaseniormanagerinAIatOlaElectric,whereheleadsateamof20+peopleintheareasofML,statistics,CV,NLP,andreinforcementlearning.Hehasfiled14+patentsinIndiaandtheUSandhasnumerousresearchpublicationstohisname.HiteshhasbeeninvolvedinresearchrolesatIndia'stopbusinessschools:theIndianSchoolofBusiness,Hyderabad,andtheIndianInstituteofManagement,Ahmedabad.Heisalsoactivelyinvolvedintrainingandmentoringandhasbeeninvitedtobeaguestspeakerbyvariouscorporationsandassociationsacrosstheglobe.

TableofContents

Preface

Section1–DataParallelism

1

SplittingInputData

Single-nodetrainingistooslow4

Themismatchbetweendataloading

bandwidthandmodeltrainingbandwidth5

Single-nodetrainingtimeonpopular

datasets6

Acceleratingthetrainingprocesswith

dataparallelism8

Dataparallelism–the

high-levelbits9

Stochasticgradientdescent13

Modelsynchronization14

Hyperparametertuning15

Globalbatchsize16

Learningrateadjustment16

Modelsynchronizationschemes17

Summary18

2

ParameterServerandAll-Reduce

Technicalrequirements20

Parameterserverarchitecture21

Communicationbottleneckinthe

parameterserverarchitecture22

Shardingthemodelamongparameter

servers24

Implementingtheparameter

server26

Definingmodellayers26

Definingtheparameterserver27

Definingtheworker28

Passingdatabetweentheparameter

serverandworker30

Issueswiththeparameter

server32

Theparameterserverarchitecture

introducesahighcodingcomplexity

forpractitioners33

viiiTableofContents

Broadcast40

Gather41

All-Gather42

Summary

43

All-Reducearchitecture34

Reduce34

All-Reduce36

RingAll-Reduce37

Collectivecommunication40

3

BuildingaDataParallelTrainingandServingPipeline

Technicalrequirements

46

Single-machinemulti-GPU52

Thedataparalleltraining

Multi-machinemulti-GPU56

pipelineinanutshell

Inputpre-processing

Inputdatapartition

Dataloading

Training

46

48

49

50

50

Checkpointingandfault

tolerance64

Modelcheckpointing64

Loadmodelcheckpoints65

Modelsynchronization

51

Modelevaluationand

Modelupdate

52

hyperparametertuning67

Single-machinemulti-GPUsand

multi-machinemulti-GPUs

4

BottlenecksandSolutions

52

Modelservingindataparallelism71

Summary73

Communicationbottlenecksin

dataparalleltraining76

Analyzingthecommunicationworkloads76

Parameterserverarchitecture77

TheAll-Reducearchitecture80

Theinefficiencyofstate-of-the-art

communicationschemes83

Leveragingidlelinksandhost

resources85

TreeAll-Reduce85

HybriddatatransferoverPCIeand

NVLink91

On-devicememorybottlenecks93

Recomputationandquantization94

Recomputation95

Quantization98

Summary99

TableofContentsix

Section2–ModelParallelism

5

SplittingtheModel

Technicalrequirements104

Single-nodetrainingerror–out

ofmemory105

Fine-tuningBERTonasingleGPU105

Tryingtopackagiantmodelinsideone

state-of-the-artGPU107

ELMo,BERT,andGPT110

Basicconcepts110

RNN114

ELMo117

6

PipelineInputandLayerSplit

BERT

GPT

Pre-trainingandfine-tuningState-of-the-arthardware

P100,V100,andDGX-1

NVLink

A100andDGX-2

NVSwitch

Summary

119

121

122

123

123

124

125

125

125

Vanillamodelparallelismis

inefficient128

Forwardpropagation130

Backwardpropagation131

GPUidletimebetweenforwardand

backwardpropagation132

Pipelineinput137

Prosandconsofpipeline

parallelism141

Advantagesofpipelineparallelism141

Disadvantagesofpipelineparallelism142

Layersplit142

Notesonintra-layermodel

parallelism145

Summary145

xTableofContents

7

ImplementingModelParallelTrainingandServingWorkflows

Technicalrequirements148

Wrappingupthewholemodel

parallelismpipeline149

Amodelparalleltrainingoverview149

Implementingamodelparalleltraining

pipeline150

Specifyingcommunicationprotocol

amongGPUs153

Modelparallelserving158

Fine-tuningtransformers162

Hyperparametertuningin

modelparallelism163

BalancingtheworkloadamongGPUs163

Enabling/disablingpipelineparallelism164

NLPmodelserving164

Summary165

8

AchievingHigherThroughputandLowerLatency

Technicalrequirements169

Freezinglayers169

Freezinglayersduringforward

propagation171

Reducingcomputationcostduring

forwardpropagation173

Freezinglayersduringbackward

propagation174

Exploringmemoryand

storageresources177

Understandingmodel

decompositionanddistillation180

Modeldecomposition180

Modeldistillation183

Reducingbitsinhardware184

Summary184

Section3–AdvancedParallelismParadigms

9

A

HybridofDataandModelParallelism

Technicalrequirements189

CasestudyofMegatron-LM189

Layersplitformodelparallelism189

Row-wisetrial-and-errorapproach192

Column-wisetrial-and-errorapproach196

Cross-machinefordataparallelism200

Implementationof

Megatron-LM201

Casestudyof

Mesh-TensorFlow203

TableofContentsxi

Implementationof

ProsandconsofMegatron-LM

Mesh-TensorFlow204

andMesh-TensorFlow204

Summary205

10

FederatedLearningandEdgeDevices

Technicalrequirements209

Sharingknowledgewithout

sharingdata209

Recappingthetraditionaldataparallel

modeltrainingparadigm210

Noinputsharingamongworkers211

Communicatinggradientsfor

collaborativelearning212

Casestudy:TensorFlow

Federated217

Runningedgedeviceswith

TinyML219

Casestudy:TensorFlowLite219

Summary220

11

ElasticModelTrainingandServing

Technicalrequirements223

Introducingadaptive

modeltraining223

Traditionaldataparalleltraining224

Adaptivemodeltrainingindata

parallelism226

Adaptivemodeltraining(AllReduce-

based)226

Adaptivemodeltraining(parameter

server-based)229

Traditionalmodel-parallelmodel

trainingparadigm231

Adaptivemodeltraininginmodel

parallelism232

Implementingadaptivemodel

traininginthecloud235

Elasticityinmodelinference236

Serverless238

Summary238

xiiTableofContents

12

AdvancedTechniquesforFurtherSpeed-Ups

Technicalrequirements

241

Jobmigrationandmultiplexing

249

Debuggingandperformance

Jobmigration

250

analytics

241

Jobmultiplexing

251

Generalconceptsinthe

profilingresultsCommunicationresultsanalysisComputationresultsanalysis

243

245

246

Modeltrainingina

heterogeneousenvironmentSummary

251

252

Index

OtherBooksYouMayEnjoy

Preface

Reducingtimecostsinmachinelearningleadstoashorterwaitingtimeformodeltrainingandafastermodelupdatingcycle.Distributedmachinelearningenablesmachinelearningpractitionerstoshortenmodeltrainingandinferencetimebyordersofmagnitude.Withthehelpofthispracticalguide,you'llbeabletoputyourPythondevelopmentknowledgetoworktogetupandrunningwiththeimplementationofdistributedmachinelearning,includingmulti-nodemachinelearningsystems,innotime.

You'llbeginbyexploringhowdistributedsystemsworkinthemachinelearningareaandhowdistributedmachinelearningisappliedtostate-of-the-artdeeplearningmodels.Asyouadvance,you'llseehowtousedistributedsystemstoenhancemachinelearningmodeltrainingandservingspeed.You'llalsogettogripswithapplyingdataparallelandmodelparallelapproachesbeforeoptimizingthein-parallelmodeltrainingandservingpipelineinlocalclustersorcloudenvironments.

Bytheendofthisbook,you'llhavegainedtheknowledgeandskillsneededtobuildanddeployanefficientdataprocessingpipelineformachinelearningmodeltrainingandinferenceinadistributedmanner.

Whothisbookisfor

Thisbookisfordatascientists,machinelearningengineers,andmachinelearningpractitionersinbothacademiaandindustry.AfundamentalunderstandingofmachinelearningconceptsandworkingknowledgeofPythonprogrammingisassumed.Priorexperienceimplementingmachinelearning/deeplearningmodelswithTensorFloworPyTorchwillbebeneficial.You'llfindthisbookusefulifyouareinterestedinusingdistributedsystemstoboostmachinelearningmodeltrainingandservingspeed.

xivPreface

Whatthisbookcovers

Chapter1,SplittingInputData,showshowtodistributemachinelearningtrainingorservingworkloadontheinputdatadimension,whichiscalleddataparallelism.Chapter2,ParameterServerandAll-Reduce,describestwowidely-adoptedmodelsynchronizationschemesinthedataparalleltrainingprocess.

Chapter3,BuildingaDataParallelTrainingandServingPipeline,illustrateshowtoimplementdataparalleltrainingandtheservingworkflow.

Chapter4,BottlenecksandSolutions,describeshowtoimprovedataparallelismperformancewithsomeadvancedtechniques,suchasmoreefficientcommunicationprotocols,reducingthememoryfootprint.

Chapter5,SplittingtheModel,introducesthevanillamodelparallelapproachingeneral.Chapter6,PipelineInputandLayerSplit,showshowtoimprovesystemefficiencywithpipelineparallelism.

Chapter7,ImplementingModelParallelTrainingandServingWorkflows,discusseshowtoimplementmodelparalleltrainingandservingindetail.

Chapter8,AchievingHigherThroughputandLowerLatency,coversadvancedschemestoreducecomputationandmemoryconsumptioninmodelparallelism.

Chapter9,AHybridofDataandModelParallelism,combinesdataandmodelparallelismtogetherasanadvancedin-parallelmodeltraining/servingscheme.

Chapter10,FederatedLearningandEdgeDevices,talksaboutfederatedlearningandhowedgedevicesareinvolvedinthisprocess.

Chapter11,ElasticModelTrainingandServing,describesamoreefficientschemethatcanchangethenumberofacceleratorsusedonthefly.

Chapter12,AdvancedTechniquesforFurtherSpeed-Ups,summarizesseveralusefultools,suchasaperformancedebuggingtool,jobmultiplexing,andheterogeneousmodeltraining.

Prefacexv

Togetthemostoutofthisbook

YouwillneedtoinstallPyTorch/TensorFlowsuccessfullyonyoursystem.Fordistributedworkloads,wesuggestyouatleasthavefourGPUsinhand.

WeassumeyouhaveLinux/Ubuntuasyouroperatingsystem.WeassumeyouuseNVIDIAGPUsandhaveinstalledtheproperNVIDIAdriveraswell.Wealsoassumeyouhavebasicknowledgeaboutmachinelearningingeneralandarefamiliarwithpopulardeeplearningmodels.

Ifyouareusingthedigitalversionofthisbook,weadviseyoutotypethecodeyourselforaccessthecodefromthebook'sGitHubrepository(alinkisavailableinthenextsection).Doingsowillhelpyouavoidanypotentialerrorsrelatedtothecopyingandpastingofcode.

Downloadtheexamplecodefiles

YoucandownloadtheexamplecodefilesforthisbookfromGitHubat

https://

/PacktPublishing/Distributed-Machine-Learning-with-

Python

.Ifthere'sanupdatetothecode,itwillbeupdatedintheGitHubrepository.

Wealsohaveothercodebundlesfromourrichcatalogofbooksandvideosavailableat

/PacktPublishing/

.Checkthemout!

Downloadthecolorimages

WealsoprovideaPDFfilethathascolorimagesofthescreenshotsanddiagramsusedinthisbook.Youcandownloadithere:

/

downloads/9781801815697_ColorImages.pdf

xviPreface

Conventionsused

Thereareanumberoftextconventionsusedthroughoutthisbook.

Codeintext:Indicatescodewordsintext,databasetablenames,foldernames,filenames,fileextensions,pathnames,dummyURLs,userinput,andTwitterhandles.Hereisanexample:"ReplaceYOUR_API_KEY_HEREwiththesubscriptionkeyofyourCognitiveServicesresource.Leavethequotationmarks!"

Ablockofcodeissetasfollows:

#ConnecttoAPIthroughsubscriptionkeyandendpoint

subscription_key="<your-subscription-key>"

endpoint="https://<your-cognitive-service>.cognitiveservices.

/"

#Authenticate

credential=AzureKeyCredential(subscription_key)

cog_client=TextAnalyticsClient(endpoint=endpoint,

credential=credential)

Bold:Indicatesanewterm,animportantword,orwordsthatyouseeonscreen.Forinstance,wordsinmenusordialogboxesappearinbold.Hereisanexample:"Select

Review+Create."

TipsorImportantNotes

Appearlikethis.

Getintouch

Feedbackfromourreadersisalwayswelcome.

Generalfeedback:Ifyouhavequestionsaboutanyaspectofthisbook,emailusatcustomercare@andmentionthebooktitleinthesubjectofyour

message.

Errata:Althoughwehavetakeneverycaretoensuretheaccuracyofourcontent,mistakesdohappen.Ifyouhavefoundamistakeinthisbook,wewouldbegratefulifyouwouldreportthistous.Pleasevisit

/support/errata

andfillintheform.

Prefacexvii

Piracy:Ifyo