從頭訓(xùn)練大型語言模型的最佳實(shí)踐

CurrentBestPractices

forTrainingLLMsfromScratch

Authors:RebeccaLi,AndreaParker,JustinTenuto

·······weights&Biases

TableofContents

Introduction03

Buildvs.BuyPre-trainedLLMModels03

TheScalingLaws05

Hardware06

Memoryvs.ComputeEfficiency06

TechniquesforParallelization06

DatasetCollection08

DatasetPre-processing08

DatasetHandling08

Tokenization09

Pre-trainingSteps13

ModelEvaluation15

BiasandToxicity16

InstructionTuning17

ReinforcementLearningthroughHumanFeedback(RLHF)19

Conclusion20

References20

Appendix21

LLMOverview21

TransformerModelArchitecture21

TheOriginalLLMScalingLaws23

www.wandb.ai?contact@wandb.ai2

·······weights&Biases

Introduction

Althoughwe’reonlyafewyearsremovedfromthetransformer

breakthrough,LLMshavealreadygrownmassivelyin

performance,cost,andpromise.AtW&B,we’vebeenfortunatetoseemoreteamstrytobuildLLMsthananyoneelse.Butmanyofthecriticaldetailsandkeydecisionpointsareoftenpasseddownbywordofmouth.

Thegoalofthiswhitepaperistodistillthebestpracticesfor

trainingyourownLLMforscratch.We’llcovereverythingfromscalingandhardwaretodatasetselectionandmodeltraining,lettingyouknowwhichtradeoffstoconsiderandflaggingsomepotentialpitfallsalongtheway.Thisismeanttobeafairly

exhaustivelookatthekeystepsandconsiderationsyou’llmakewhentraininganLLMfromscratch.

Thefirstquestionyoushouldaskyourselfiswhethertrainingonefromscratchisrightforyourorganization.Assuch,we’llstartthere:

BUILDVS.BUYPRE-TRAINEDLLMMODELS

BeforestartingLLMpre-training,thefirstquestionyouneedtoaskiswhetheryoushouldpre-trainanLLMbyyourselforuseanexistingone.Therearethreebasicapproaches:

?Option1:UsetheAPIofacommercialLLM,e.g.GPT-3(OpenAI,2020),CohereAPIs,AI21J-1

?Option2:Useanexistingopen-sourcedLLM,e.g.GPT-J(EleutherAI,2021),GPT-NeoX(EleutherAI,2022),Galactica(MetaAI),UL2(Google,2022),OPT(MetaAI,2022),BLOOM(BigScience,2022),Megatron-LM(NVIDIA,2021),CodeGen(Salesforce,2022)

?Option3:Pre-trainanLLMbyyourselforwithconsultants:

YoucaneithermanageyourowntrainingorhireLLM

consultants&platforms.Forexample,MosaicMLprovidestrainingservicesfocusingonLLMs.

Thatsaid,therearealotofdetailstoconsiderwhenmakingyourchoice.Herearethepros,cons,andapplicablescenariosforeachoption:

Option3

Pre-trainanLLMbyyourselforwithconsultants

Option2

Useanexistingopen-sourcedLLM

Option1

UsetheAPIofacommercialLLM

Pros

?RequirestheleastLLMtrainingtechnicalskills.

?Minimumupfronttraining/

explorationcost,givenmaincostincursatinferencetime.

?Theleastdata-demandingoption.

Onlyafewexamples(ornoexamples)areneededformodelstoperform

inference.

?Canleveragethebest-performingLLMsinthemarketandbuilda

superiorexperience.

?Reducetime-to-marketofyour

appsandde-riskyourprojectwithaworkingLLMmodel.

?AgoodwaytoleveragewhatLLMs

havelearnedfromavastamountofinternetdataandbuildontopofit

withoutpayingfortheIPatinference.

?Comparedtooptionone,youarelessdependentonthefuturedirectionofLLMserviceprovidersandthushavemorecontrolregardingroadmap&backwardscompatibility.

?Comparedtooptionthree,youhaveamuchfastertime-to-valuegivenyouarenotbuildingLLMsfromscratch,alsoleadingtolessdata,training

time,trainingbudgetneeded.

?Comparedtooptionsoneandtwo,youhavethemostcontrolofyour

LLM’sperformanceandfuture

direction,givingyoulotsofflexibilitytoinnovateontechniquesand/or

customizetoyourdownstreamtasks.

?Gainfullcontroloftrainingdatasetsusedforthepre-training,which

directlyimpactsmodelquality,bias,andtoxicityissues.Incomparison,thoseissuesarelesscontrollableinoptiononeortwo.

?TrainingyourownLLMalsogives

youadeepmoat:superiorLLM

performanceeitheracrosshorizontalusecasesortailoredtoyourvertical,allowingyoutobuildasustaining

advantageespeciallyifyoucreateapositivedata/feedbackloopwithLLMdeployments.

www.wandb.ai?contact@wandb.ai3

·······weights&Biases

Option1

Option2

Option3

UsetheAPIofacommercialLLM

Useanexistingopen-sourcedLLM

Pre-trainanLLMbyyourselforwithconsultants

Cons

?CommercialLLMservicescanget

expensivewithahighvolumeoffine-tuningorinferencetasks.Itcomes

downtoLLMtotal-cost-of-ownership(TCO)amortizedtoeachinference.

?Manyindustries/usecasesforbidtheuseofcommercialLLMservicesassensitivedata/PIIdatacannotbeseenbytheserviceforcompliance(healthcareusecases,forexample).

?Ifbuildingexternalapps,you’llneedtofindothermoatsandde-riskyourbusinessifyou’rehighlyreliantonexternalLLMservicetechnology.

?Lessflexibledownstream:doesn’t

supportedgeinference,limited

abilitytocustomizethemodel(fine-tuninggetsexpensive),limitedabilityforongoingmodelimprovements.

?Notasdemandingasbuilding

yourown,butstillrequireslotsofdomainexpertskillstotrain,fine-tune,andhostanopen-sourcedLLM.LLMreproducibilityisstillasignificantissuesotheamountoftimeandworkneededcannotbeunderestimated.

?Slowertime-to-marketandlessagileifyouarebuildingdownstreamapps,duetoamoreverticaltechstack.

?Open-sourcedmodelstypically

lagperformancecomparedto

commercialmodelsbymonths/years.Ifyourcompetitorleveragescommercialmodels,theyhaveanadvantageonLLMtechandyou’llneedtofindothercompetitive

advantages.

?Veryexpensiveendeavorwith

highrisks.Needcross-domain

knowledgespanningfromNLP/ML,subjectmatterexpertise,softwareandhardwareexpertise.Ifnotdonewell,youcouldendupinasituationwhereyou’vespentthousands

orevenmillionsofdollarswith

asuboptimalmodel.Mistakes,

especiallylateintotrainingstages,arehardtofix/unwind.

?Lessefficientthanoptiontwo.

OptiontwoleveragesexistingLLMs,learningfromanentireinternet’s

worthofdataandcanprovidea

solidstartingpoint.Withoption3,youstartfromscratchandneedlotsofhigh-quality/diversedatasets

foryourmodelstogaingeneralizedcapabilities.

Whentoconsidereachoption

?BestifyoueitherhavelesstechnicalteamsbutwanttoleverageLLM

techniquestobuilddownstream

apps,oryouwanttoleveragethebest-in-classLLMsforperformancereasons(outsourcingtheLLMtech).

?Betweenoptionstwoandthree,

ifyouaren’ttryingtochangethe

modelarchitecture,itisalmost

alwaysbettertoeitherdirectlytakeanexistingpre-trainedLLMand

fine-tuneitortaketheweightsofan

?Bestifyouneedtochangemodelarchitectureortrainingdatasetfromexistingpre-trainedLLMs.Forexample,ifyouwanttouseadifferenttokenizer,changethevocabularysize,orchangethe

?Goodifyouhaveverylimitedtrainingdatasetsandwanttoleveragean

LLM’scapabilitytodozero/few-shotlearning.

existingpre-trainedLLMasastartingpointandcontinuepre-training.Thereasonisbecauseagoodpre-trainedLLMlikeGPT-NeoXhasalreadyseenavastamountofdataandthushas

numberofhiddendimensions,attentionheads,orlayers.

?Typically,inthiscasetheLLMisa

corepartofyourbusinessstrategy&

?Goodforprototypingappsand

exploringwhatispossiblewithLLMs.

learnedgeneralcapabilitiesfromthedata.Youcanleveragethatlearningespeciallyifyourtrainingdatasetisnothugeordiverse.

?Anothertypicalscenarioisthatyouoperateinaregulatoryenvironmentorhaveuser/sensitivedatathat

cannotbefedtocommercial

LLMservices.Oryouneededge

deploymentofthemodelforlatencyorlocationalreasons.

technologicalmoat.Youaretakingonsomeoralotofinnovations

inLLMtraining,andhavealargeinvestmentappetitetotrainandmaintainexpensivemodelsonanongoingbasis.

?Typically,youhaveorwillhavelotsofproprietarydataassociatedwithyourLLMtocreateacontinuous

modelimprovementloopfor

sustainablecompetitiveadvantage.

Itisalsoworthmentioningthatifyouonlyhaveaverytargetedsetofusecasesanddon’tneedthegeneral-purposecapabilitiesor

generativecapabilitiesfromLLMs,youmightwanttoconsidertrainingorfine-tuningamuchsmallertransformerorothermuchsimplerdeeplearningmodels.Thatcouldresultinmuchlesscomplexity,lesstrainingtime,andlessongoingcosts.

www.wandb.ai?contact@wandb.ai4

·······weights&Biases

THESCALINGLAWS

Beforeyoudiveintotraining,it’simportanttocoverhowLLMsscale.Understandingscalingletsyoueffectivelybalancethesizeandcomplexityofyourmodelandthesizeofthedatayou’llusetotrainit.

Somerelevanthistoryhere:OpenAIoriginallyintroduced“theLLMscalinglaws”in2020.Theysuggestedthatincreasingmodelsizewasmoreimportantthanscalingdatasize.Thisheldfor

abouttwoyearsbeforeDeepMindsuggestedalmostthepolaropposite:thatpreviousmodelsweresignificantlyundertrainedandthatincreasingyourfoundationaltrainingdatasetsactuallyleadstobetterperformance.

Thatchangedin2022.Specifically,DeepMindputforward

analternativeapproachintheir

TrainingCompute-Optimal

LargeLanguageModels

paper.TheyfoundthatcurrentLLMsareactuallysignificantlyundertrained.Putsimply:theselargemodelsweren’ttrainedonnearlyenoughdata.

DeepmindshowcasedthiswithamodelcalledChinchilla,whichisafourththesizeoftheGophermodelabovebuttrainedon

4.6xmoredata.Atthatreducedsizebutwithfarmoretrainingdata,ChinchillaoutperformedGopherandotherLLMs.

DeepMindclaimsthatthemodelsizeandthenumberof

trainingtokens*shouldinsteadincreaseatroughlythesameratetoachieveoptimalperformance.Ifyougeta10xincreaseincompute,youshouldmakeyourmodel3.1xtimesbiggerandthedatayoutrainover3.1xbigger;ifyougeta100xincreaseincompute,youshouldmakeyourmodel10xbiggerandyourdata10xbigger.

*Note:TokenizationinNLPisanessentialstepofseparatingapiece

oftextintosmallerunitscalledtokens.Tokenscanbeeitherwords,

characters,orsubwords.Thenumberoftrainingtokensisthesizeof

trainingdataintokenformaftertokenization.Wewilldiveintodetailedtokenizationmethodsalittlelater.

DeepMindprovidesthefollowingchartshowinghowmuch

trainingdataandcomputeyou’dneedtooptimallytrainmodelsofvarioussizes.

EstimatedoptimaltrainingFLOPsandtrainingtokensforvariousmodelsizes,

TrainingCompute-OptimalLargeLanguageModels

Thatsaid,mostexistingLLMsarestillundertrained:

Data/compute-optimal(Chinchilla)heatmap,

Chinchilla

data-optimalscalinglaws:InplainEnglish

Insummary,thecurrentbestpracticesinchoosingthesizeofyourLLMmodelsarelargelybasedontworules:

?DecideonyourdatasetandfindtheChinchilla-optimal

modelsizebasedondatasize(orclosetoChinchilla-optimalwithintheboundaryofyourdatacollectionlimitation)

?Determinethedataandmodelsizecombinationthat’sbestforyourmodel,basedonyourtrainingcomputebudgetandinferencelatencyrequirements

Totheleftoftheminimaoneachcurve,modelsaretoosmall--alargermodeltrainedonlessdatawouldbeanimprovement.Totherightoftheminimaoneachcurve,modelsaretoolarge--asmallermodeltrainedonmoredatawouldbeanimprovement.Thebestmodelsareattheminima.

www.wandb.ai?contact@wandb.ai5

·······weights&Biases

HARDWARE

Itshouldcomeasnosurprisethatpre-trainingLLMsisa

hardware-intensiveeffort.Thefollowingexamplesofcurrentmodelsareagoodguidehere:

?PaLM(540B,Google):6144TPUv4chipsusedintotal,madeoftwoTPUv4Podsconnectedoverdatacenternetwork(DCN)usingacombinationofmodelanddataparallelism

?OPT(175B,MetaAI):99280GBA100GPUs,utilizingfullyshareddataparallelismwithMegatron-LMtensorparallelism

?GPT-NeoX(20B,EleutherAI):9640GBA100GPUsintotal

?Megatron-TuringNLG(530B,NVIDIA&MSFT):560DGXA100nodes,eachclusternodehas8NVIDIA80-GB

A100GPUs

TrainingLLMsischallengingfromaninfrastructureperspectivefortwobigreasons.Forstarters,itissimplynolongerpossibletofitallthemodelparametersinthememoryofeventhelargestGPU(e.g.NVIDIA80GB-A100),soyou’llneedsomeparallel

architecturehere.Theotherchallengeisthatalargenumberofcomputeoperationscanresultinunrealisticallylongtrainingtimesifyouaren’tconcurrentlyoptimizingyouralgorithms,

software,andhardwarestack(e.g.trainingGPT-3with175Bparameterswouldrequireabout288yearswithasingleV100NVIDIAGPU).

Memoryvs.ComputeEfficiency

TechniquesforParallelization

Parallelizationreferstosplittinguptasksanddistributing

themacrossmultipleprocessorsordevices,suchasGPUs,sothattheycanbecompletedsimultaneously.Thisallowsformoreefficientuseofcomputeresourcesandfastercompletiontimescomparedtorunningonasingleprocessorordevice.

ParallelizedtrainingacrossmultipleGPUsisaneffectivewaytoreducetheoveralltimeneededforthetrainingprocess.

Thereareseveraldifferentstrategiesthatcanbeusedto

parallelizetraining,includinggradientaccumulation,micro-

batching,dataparallelization,tensorparallelizationandpipelineparallelization,andmore.TypicalLLMpre-trainingemploysa

combinationofthesemethods.Let’sdefineeach:

DataParallelism

Dataparallelismisthebestandmostcommonapproachfor

dealingwithlargedatasetsthatcannotfitintoasinglemachineinadeeplearningworkflow.

Morespecifically,dataparallelismdividesthetrainingdataintomultipleshards(partitions)anddistributesthemtovarious

nodes.Eachnodefirstworkswithitslocaldatatotrainitssub-model,andthencommunicateswiththeothernodestocombinetheirresultsatcertainintervalsinordertoobtaintheglobal

model.Theparameterupdatesfordataparallelismcanbeeitherasynchronousorsynchronous.

Theadvantageofthismethodisthatitincreasescompute

efficiencyandthatitisrelativelyeasytoimplement.ThebiggestdownsideisthatduringthebackwardpassyouhavetopassthewholegradienttoallotherGPUs.Italsoreplicatesthemodelandoptimizeracrossallworkerswhichisrathermemoryinefficient.

ToachievethefullpotentialofthousandsofdistributedGPUs,itiscrucialtodesignparallelismintoyourarchitectureto

balancememoryandcomputeefficiency.

Memoryefficiency

TrainingaLLMrequiresterabytesofaggregatememoryfor

modelweights,gradients,andoptimizerstates-farbeyondwhatisavailableonasingleGPU.Onetypicalmitigationstrategyis

gradientaccumulation,inwhichthefulltrainingbatchissplitintomicro-batchesthatareprocessedinsequencewiththeirresultinggradientsaccumulatedbeforeupdatingthemodel

weights.Thatmeansyourtrainingbatchsizecanscalewithoutincreasingthepeakresidentactivationmemory.

Computeefficiency

WhilelargeGPUclusterscanhavethousandsofhigh-throughputGPUs,achievinghighcomputeefficiencyatthisscaleis

challenging.Alargebatchsizecanbeaneffectivewaytoincreasecomputeefficiency,becauseitincreasesthearithmeticintensityofaGPUkernelandhelpsamortizethetimespentstalledon

communicationandsynchronization.However,usingtoolargeofabatchsizecanhavenegativeeffectsonthemodelquality.

Whileparallelizationisparamount,therearemanydifferent

waystodoit.We’llgetintothemostcommoninournextsection.

www.wandb.ai?contact@wandb.ai6

·······weights&Biases

TensorParallelism

Tensorparallelismdivideslargematrixmultiplicationsintosmallersubmatrixcalculationswhicharethenexecuted

simultaneouslyusingmultipleGPUs.

Thisallowsforfastertrainingtimesduetoitsasynchronousnatureandtheabilitytoreducecommunicationoverheadbetweennodes.Thebenefitofthismethodisthatitis

memory-efficient.Thedownside,however,isthatit

introducesadditionalcommunicationofactivationsineachforward&backwardpropagation,andthereforerequireshighcommunicationbandwidthtobeefficient.

Pipelineparallelismandmodelparallelism

Pipelineparallelismimprovesboththememoryandcomputeefficiencyofdeeplearningtrainingbypartitioningthelayersofamodelintostagesthatcanbeprocessedinparallel.

Thishelpswithoverallthroughputspeedssignificantlywhile

addingthesmallestcommunicationoverhead.Youcanthinkofpipelineparallelismas“inter-layerparallelism”(wheretensor

parallelismcanbethoughtofas“intra-layerparallelism”).

Similartopipelineparallelism,modelparallelismiswhenyou

splitthemodelamongGPUsandusethesamedataforeach

model;soeachGPUworksonapartofthemodelratherthanapartofthedata.Thedownsideofpipelineandmodelparallelismisthatitcannotscaleinfinitelygiventhatthedegreeofpipelineparallelismisboundedbythedepthofthemodel.

Asmentionedatthestartofthissection,it’snotuncommonforteamstoleverageacombinationofparallelismtechniquesduringtraining.Forexample,PaLM(GoogleBrain,2022)andOPT(MetaAI,2022)bothusedacombinationoftensormodelparallelismanddataparallelism.

NVIDIAapproachedthingsalittledifferentlyinthe

Efficient

Large-ScaleLanguageModelTrainingonGPUClustersUsing

Megatron-LM

paper.TheyproposedaPTD-Ptechniquethat

combinespipeline,tensor,anddataparallelismtoachieve

state-of-the-artcomputationalperformance(52%ofpeakdevicethroughput)on1000sofGPUs.

Specifically,PTD-Pleveragesacombinationofpipeline

parallelismacrossmulti-GPUservers,tensorparallelismwithinamulti-GPUserver,anddataparallelismtopracticallytrain

modelswithatrillionparameters.Themethodalsoemploys

gracefulscalinginanoptimizedclusterenvironmentwithhigh-bandwidthlinksbetweenGPUsonthesameserverandacrossservers.

UsingthesetechniquestotrainLLMsrequiresnotonlythe

highest-performingGPUstobeefficient,butalsoneedshigh-

bandwidthnetworkingforoptimalcommunication––InfiniBandisoftenusedtomovedatabetweennodes.

Butthisofcoursecomeswithacost.Leveragingthousands

ofhigh-performingGPUsandhigh-bandwidthnetworksto

trainLLMsisinfrastructure-intensive.Forexample,aback-of-the-envelopecalculationestimatedthatthecostofthePaLMmodel(540B,Google)mightbeashighas$23MM

(seedetailed

analysis

).

Toimplementdistributeddeeplearningtrainingsystems,

softwaretoolkitssuchasDistributedTensorFlow,Torch

Distributed,Horovod,andlibrariessuchasDeepSeedand

Megatronareoftenneeded.Thereisimplementationcomplexityheresoitrequiressystemexpertiseifyou’regoingtobe

successful.

Inaddition,thefollowingtechniquesandstrategiesarecommonlyemployedtoachieveparallelism:

Gradientaccumulation

Gradientaccumulationinvolvesaddingupgradientsfrom

multiplebatchesbeforeperformingoneweightupdatesteponallaccumulatedgradientsatonce.

ThisapproachreducescommunicationoverheadbetweenGPUsbyallowingthemtoworkindependentlyontheirownlocalbatchofdatauntiltheyhavesynchronizedwitheach

otheragain,afteraccumulatingenoughgradientsforasingleoptimizationstep.

Asynchronousstochasticgradientdescentoptimization

AsynchronousstochasticgradientdescentoptimizationmethodscanalsobeemployedwhenperformingmodeloptimizationovermultipleGPUs.

www.wandb.ai?contact@wandb.ai7

·······weights&Biases

Thismethodusessmallsubsets(microbatches)ofdatafrom

eachnodeinsteadofloadingalldataatonce,whichhelpsreducememoryrequirementswhilestillallowingforfastconvergenceratesduetoitsasynchronousnature.Itworkslikethis:

?First,wefetchthemostup-to-dateparametersofthe

modelneededtoprocessthecurrentmini-batchfromtheparameterservers.

?Wethencomputegradientsofthelosswithrespecttotheseparameter

?Finally,thesegradientsaresentbacktotheparameterservers,whichthenupdatesthemodelaccordingly.

Micro-batching

Micro-batchingcombinessmallmini-batchesintolargeronessothatmorebatchescanbeprocessedinlesstimeandwithfewersynchronizationpointsbetweendevicesduringbackpropagationoperations.Ithasbecomeincreasinglypopularfortraining

verylargemodelsacrossmanyGPUsduetoitsabilitytoreducememoryconsumptionandimprovescalability.Overall,micro-

batchingisaneffectivewaytoleveragedistributeddeeplearningtechniqueswhendealingwithverylargedatasetsormodelsthatrequiresignificantamountsofprocessingpower.

Nowthatwe’vegonethroughscaling,hardware,andsome

techniquesforparallelizingyourtrainingruns,let’slookatwhatyourLLMwillactuallylearnfrom:data.

DATASETCOLLECTION

Baddataleadstobadmodels.Butcarefulprocessingofhigh-quality,high-volume,diversedatasetsdirectly

contributestomodelperformanceindownstreamtasksaswellasmodelconvergence.

DatasetdiversityisespeciallyimportantforLLMs.That’sbecausediversityimprovesthecross-domainknowledgeofthemodel,

aswellasitsdownstreamgeneralizationcapability.TrainingondiverseexampleseffectivelybroadenstheabilityofyourLLMtoperformwellonmyriadnuancedtasks.

Atypicaltrainingdatasetiscomprisedoftextualdatafrom

diversesources,suchascrawledpublicdata,onlinepublicationorbookrepositories,codedatafromGitHub,Wikipedia,news,socialmediaconversations,etc.

Forexample,considerThePile

.ThePileisapopulartextcorpuscreatedbyEleutherAIforlarge-scalelanguagemodeling.It

containsdatafrom22datasources,coarselybrokendownintofivebroadcategories:

?AcademicWriting:PubMedAbstractsandPubMedCentral,arXiv,FreeLaw,USPTOBackgrounds,PhilPapers,NIH

Exporter

?OnlineorScrapedResources:CommonCrawl,OpenWebText2,StackExchange,Wikipedia

?Prose:BookCorpus2,Bibliotik,ProjectGutenberg

?Dialog:YouTubesubtitles,UbuntuIRC,OpenSubtitles,HackerNews,Europarl

?Miscellaneous:GitHub,theDeepMindMathematicsdataset,Enronemails

NotethatThePiledatasetisoneoftheveryfewlarge-scale

textdatasetsthatisfreeforthepublic.FormostoftheexistingmodelslikeGPT-3,PaLM,andGalactica,theirtrainingand

evaluationdatasetsarenotpubliclyavailable.Giventhelargescaleeffortittakestocompileandpre-processthesedatasetsforLLMtraining,mostcompanieshavekeptthemin-house

tomaintaincompetitiveadvantage.ThatmakesdatasetslikeThePileandafewdatasetsfromAllenAIextremelyvaluableforpubliclarge-scaleNLPresearchpurposes.

Anotherthingworthmentioningisthat,duringdataset

collection,generaldatacanbecollectedbynon-expertsbut

dataforspecificdomainsnormallyne