LSH-MoE：通過局部敏感哈希實(shí)現(xiàn)通信高效的專家混合模型訓(xùn)練 LSH-MoE - Communication-efficient MoE Training via Locality-Sensitive Hashing

LSH-MoE:Communication-efficientMoETrainingviaLocality-SensitiveHashing

XiaonanNie1QibinLiu1FangchengFu1ShenhanZhu1XupengMiao2

XiaoyangLi3YangZhang3ShoudaLiu3BinCui1

arXiv:2411.08446v1[cs.DC]13Nov2024

1PekingUniversity2PurdueUniversity3ByteDance

1{xiaonan.nie,2101212782,ccchengff,shenhan.zhu,bin.cui}@.cn

2xupeng@3{lixiaoyang.x,zhangyang.elfin,liushouda}@

Abstract

Largertransformermodelsalwaysperformbetteronvarioustasksbutrequiremorecoststoscaleupthemodelsize.Toefficientlyenlargemodels,themixture-of-experts(MoE)architectureiswidelyadopted,whichconsistsofagatenetworkandaseriesofexpertsandkeepthetrainingcostconstantbyroutingtheinputdatatoafixednumberofexpertsinsteadofall.Inexistinglarge-scaleMoEtrainingsystems,expertswouldbedistributedamongdifferentGPUsforparallelization,andthusinputdatarequiresadditionalall-to-allcommunicationstoaccessthetargetexpertsandconductcorrespondingcomputations.However,uponevaluatingthetrainingprocessofthreemainstreamMoEmodelsoncommonlyusedGPUclusters,wefoundthattheall-to-allcommunicationratioaveragedaround45%,whichsignificantlyhinderstheefficiencyandscalabilityoftrainingMoEmodels.Inthispaper,weproposeLSH-MoE,acommunication-efficientMoEtrainingframe-workusinglocality-sensitivehashing(LSH).WefirstpresenttheproblemsofscalingMoEtraininginexistingsystemsandhighlightthepotentialofexploitingtokensimilaritytofacilitatedatacompression.Then,weintroduceanefficientLSH-basedcompressiontechnique,whichutilizesthecross-polytopehashingforrapidclusteringandimplementsaresidual-basederrorcompensationschemetoalleviatetheadverseimpactofcompression.Toverifytheeffectivenessofourmethods,weconductexperimentsonbothlanguagemodels(e.g.,RoBERTa,GPT,andT5)andvisionmodels(e.g.,Swin)forpre-trainingandfine-tuningtasks.Theresultsdemonstratethatourmethodsubstantiallyoutperformsitscounterparts

acrossdifferenttasksby1.28×-2.2×ofspeedup.

1Introduction

Inrecentyears,large-scalepre-trainedmodelshavesignificantlyadvancedtheperformanceof

deeplearningacrossvariouscomplextasks,includingcomputervision[8,

20],naturallanguage

processing[3,

7,

28],andmulti-modallearning[19]

.Commonlyreferredtoasfoundationmodels,

thesepre-trainedmodelsareprimarilybuiltonTransformerarchitectures[34]andundergoextensive

pre-trainingonlargedatasets,utilizingsubstantialGPUresources.OpenAIhasvalidatedthescaling

lawforlargelanguagemodels[15]andsuggeststhatincreasingthemodel’sparametersize,thevolume

oftrainingdata,andthedurationoftrainingcansignificantlyenhancethemodel’sperformance.However,thisapproachresultsinaconsiderableriseintrainingcosts,makingthedevelopmentoffoundationmodelsextremelyexpensive.

XiaonanNie,QibinLiu,FangchengFu,ShenhanZhu,andBinCuiarewiththeSchoolofComputerScienceandKeyLabofHighConfidenceSoftwareTechnologies(MOE),PekingUniversity.BinCuiisalsowiththeInstituteofComputationalSocialScience,PekingUniversity(Qingdao).

38thConferenceonNeuralInformationProcessingSystems(NeurIPS2024).

2

Toreducethehighcomputationalcosts,thesparsemixture-of-experts(MoE)architectureisoftenadopted,whichcomprisesasparsegatenetworkandaseriesofexpertnetworks.Thisarchitectureroutesinputdatatoonlyasubsetofexperts,resultinginsparseactivationoftheexpertsandtherebyreducingthemodel’scomputationalFLOPs(floatpointoperations)aswellastrainingcosts.

ProminentmodelssuchasGoogle’sSwitch-Transformer[9],ST-MoE[41],Meta’sHashLayer[31]

andMistral-AI’smixtralmodels[14]havesuccessfullyimplementedthisdesign,demonstrating

improvementsinbothperformanceandefficiencywithMoEmodels.

Meanwhile,effectivelyscalingthetrainingofMoEmodelsacrosshundredsoreventhousandsofGPUsremainsasignificantchallenge.ResearchersfromGooglehaveproposedtheexpertparallelism

approach[17],whichreplicatesthegatingnetworkoneachGPUsanddistributesdifferentexperts

acrossmultipleGPUsforparallelprocessing.Specifically,eachinputtokenisinitiallyprocessedbythegatingnetworktoselecttheappropriateexpert,afterwhichitisroutedtothedesignatedexpertsviapeer-to-peer(P2P)networkcommunication.Oncethedesignatedexpertscompletetheircomputation,thetokenisreturnedtotheoriginalGPUforfurtherprocessingthroughanadditionalP2Pcommunication.SinceeachGPUtypicallyneedstoexchangedatawithmanyotherGPUs,theseP2Ptransmissionsresultsinanall-to-allcommunicationpattern.Moreover,becausethecomputationoftheexpertnetworkreliesontheoutcomesofthesecommunications,thecommunicationscannotbeeffectivelyoverlappedwithongoingcomputations.ThisdependencycreatesasignificantperformancebottleneckinmodeltrainingacrossmostcommonlyusedGPUclusters.Weconductedexperimentsonthreewidely-usedMoEmodels,includingRoBERTa-MoE,GPT-MoEandSwin-MoE,onfourA100servers,eachwithacross-machinebandwidthof200Gb/s.Theresults,asshowninFigure

3,

revealthatthetimecostofall-to-allcommunicationconstitutesanaverageof45%andcanreachupto67%ofthetotalmodeltrainingtime.

ExistingmethodstoimprovedistributedMoEtrainingonbandwidth-limitedclusterstacklecommuni-cationchallengesinvariousways.

TA-MoE[4]reducescross-machinecommunicationbyadjusting

thegatingnetworktofavorexpertsonthesameserver,whilePre-gatedMoE[13]reducesdependency

betweencommunicationandcomputationthroughapre-gatingmechanismthatplanstokenroutinginadvance.However,bothapproachesrequiremodificationstothegatingmechanismandmodelstructure,limitingtheiruniversalapplicability.

DeepSpeed-MoE[29]introducesPR-MoE,which

selectsoneexpertplusasharedexpert,halvingtheall-to-allcommunicationload.

SCoMoE[40]

organizesall-to-allcommunicationbystructuringdatatransfersalongdifferentdimensionsandcontrollingdatavolumesacrossnetworklevels,andalsoclusterstokenstoimproverouting.However,noneoftheseworksconsiderreducingtheAll-to-AllcommunicationvolumeinMoEtrainingbycompressingtheforwardactivations.Therefore,theycanbeintergratedwithourmethodforfurtherimprovement.

Inthispaper,wepresentLSH-MoE,acommunication-efficientMoEtrainingframeworkthatleverageslocality-sensitivehashingtogroupsimilartokens.Ourkeycontributionsareasfollows:

?WebeginbyidentifyingkeychallengesinscalingMoEtraininginexistingsystems,notingthatall-to-allcommunicationconstitutesanaverageof45%ofthetotaltrainingtime.Addi-tionally,weinvestigatethepotentialofusingtokensimilaritytofacilitatedatacompressiontoreducecommunicationcosts.

?WeproposeanefficientLSH-basedcompressiontechniquethatemployscross-polytopehashingforrapidclustering.Thisapproachtransmitsonlytheclusteringcentroids,sig-nificantlyreducingcommunicationcosts.Tofurtherenhanceaccuracy,weimplementaresidual-basederrorcompensationschemetomitigatethenegativeeffectsofcompression.

?Throughextensiveexperimentswithlanguagemodels(RoBERTa-MoE,GPT-MoE,andT5-MoE)andvisionmodels(Swin-MoE),acrossbothpre-trainingandfine-tuningtasks,wedemonstratethatourmethodmaintainsmodelqualitywhileachievingaspeedupof1.28×-2.2×inend-to-endtrainingtime.

2Background

2.1Mixtures-of-ExpertArchitecture

ToenhancethetrainingefficiencyofTransformermodels,Williametal.

(2022)[9]introduced

aninnovativeparadigm,thesparsemixture-of-eexperts(MoE)architecture,illustratedinFigure

1.

3

fx纟ΣEix

i∈GX

1st

Expert

E1(x)E2(x)E3(x)En(x)

nth

Expert

…

GatingNetwork

t!

Figure1:Mixture-of-ExpertsonasingleGPU.

AlltoAll

AlltoAll

x0

x1

Node0

GatingNetwork

GPU0

1st

Expert

GatingNetwork

GPU1

2nd

Expert

x2

x3

Node1

GatingNetwork

GPU2

3rd

Expert

GatingNetwork

GPU3

4th

Expert

Intra-nodeComm.Inter-nodeComm.

Figure2:TrainingMixture-of-ExpertsonmultipleGPUsasexpertparallelism.

2nd

Expert

3rd

Expert

G:RM→1,NK

Thisarchitectureeffectivelybalancesparametercapacityandtrainingcosts,andcomprisestwokeycomponents:anexpertnetwork(E)andasparsegatenetwork(G).ItisevidentthatMoEmodels,withanequalnumberofactiveparametersperinput,cansignificantlysurpasstheperformanceofdensemodels.Thisbreakthroughhasalsocatalyzedfurtherresearchandtheirapplicationacross

variousindustries,ashighlightedbynumeroussubsequentstudies[5,

14,

22,

23,

25,

30,

39]

.

TheexpertnetworkEiscomposedofmultiplespecializedandseparatenetworks,commonlyreferredtoasexperts,denotedas{EiNrepresentsthenumberofexperts.Additionally,Ei(x)denotestheoutputproducedwhentheinputxisprocessedbythei-thexpert.Eachexpertistrainedtoexcelinaspecificsub-task,suchasinmulti-tasklearning,ortohandlespecificsegmentsofdata,asseeninlanguagemodelingandmulti-modallearning,therebyincreasingtheoverallmodelcapacity.Infoundationalmodels,theMoElayeroftenservesasasubstituteforthetraditionalfeed-forwardnetwork(FFN)layer.WithineachMoElayer,eachFFNfunctionworksasanindividualexpert,significantlyenhancingthemodel’scapabilitytoprocessdiverseandcomplexdatainputs.

ThegatingnetworkGplaysacrucialroleinthesparseMoEarchitecture.Forexample,inaK-waygatedMoEsystem,thegatingnetworkoutputsasetofintegersasEquation

1

todeterminewhichexpertsshouldbeactivated.Thisdecisionisbasedonthecharacteristicsoftheinputitself,allowingforadynamicandefficientallocationofcomputationalresources.Byonlyprocessingeachinputtokenwithaselectedsubsetoftheexpertnetwork,theMoEmodelachievescomputationsparsity,effectivelydecouplingparametercapacityfromtrainingcosts.

G:RM→[1,N]K(1)

Throughtheintegrationofmultiplespecializedexperts,asdescribedbyEquation

2,thesparseMoE

modeliscapableofdeliveringmoreaccurateandefficientpredictionsasf(x).Thisisachievedbyleveragingthespecializedknowledgeembeddedwithineachexpert,combinedwiththestrategicinputallocationmanagedbythegatingnetwork.

纟i(2)

WhileMoE’sprimaryadvantageisdecouplingparametercapacityfromnetworkcost,akeychallengeliesinlearningthegatingparameterseffectively,astheoutput’ssparsitymakesitnon-differentiable.Consequently,muchoftheresearchintheMoEfieldhascenteredondevelopingmethodsforlearninggatingfunctions.

Thesemethodsfallintothreemaincategories,asoutlinedin[6]:routingvia

learnableweighting[9,

24,

30],deterministichashrouting[31],andreinforcementlearning-based

routing[2,

32,

33]

.TheseapproachesprimarilydifferinthedesignofthegatingnetworkGratherthantheexpertnetworkE,andthereforeallencountersimilarscalingchallenges.

2.2ChallengesofScalingMoEModelTraining

WhileMoEmodelswereinitiallydevelopedtofacilitateefficientscalingduringtraining,deployingtheselarge-scalemodelsinpracticalGPU-intensiveenvironmentsposessignificantchallengesin

4

100%

80%

60%

40%

20%0%

RoBERTa-MoE

GPT-MoE

Swin-MoE-L

(16GPUs)

(16GPUs)

(16GPUs)

回All-to-All團(tuán)Others.

(a)16GPUs

100%

80%

60%

40%

20%0%

RoBERTa-MoE

(32GPUs)

GPT-MoE(32GPUs)

Swin-MoE-L(32GPUs)

All-to-AllOthers.

(b)32GPUs(double#GPUs)

100%

80%

60%

40%

20%0%

RoBERTa-MoEGPT-MoESwin-MoE-L

-Wide(16GPUs)-Wide(16GPUs)-Wide(16GPUs)

All-to-AllOthers.

(c)16GPUs(double#experts)

Figure3:Proportionofall-to-allcommunicationtimerelativetototaltrainingdurationacrossdifferentconfigurations:scalingthenumberoftrainingservers(Figure

3(b))andscalingtheparameter

sizeofmodels(Figure

3(c))

.

distributedcomputing.Specifically,theMoElayerharborsaconsiderablyhighernumberofparam-etersandrequiresadditionalmemory,yetitmaintainsalmostthesamecomputationaldemandsasthedenselayer.Thisleadstoauniquecomputedensity—definedastheratioofthelayer’sFLOPs(FloatingPointOperations)toitsnumberofparameters.Therefore,traditionalparallelismmethodssuchastensorparallelismandpipelineparallelismareinsufficientforachievingeffectiveparallelisminthescenariosofMoEtraining.

Toimprovetheefficiencyandscalabilityoftraininglarge-scaleMoEmodels,expertparallelism

[17]

hasbeenintroducedasaspecializedmodelparallelismstrategy.ThisapproachdistributesexpertswithinanMoElayeracrossmultipleGPUs,whileleveragingdataparallelismforreplicatingnon-MoElayers,thusefficientlymanagingthetrainingworkloadofMoEmodels.TheworkflowofdistributedtrainingforanMoElayerisdepictedinFigure

2.

Oncethetargetexpertforeachtokenisdetermined,anall-to-allcommunicationprocessistriggeredtodistributetokenstotheircorrespondingtargetexpertsforcomputations,denotedasEi(x).Subsequently,anotherroundofall-to-allcommunicationisexecutedtogathertheoutputsfromallexperts,whichproducestheMoElayer’soutput(representedasf(x),Equation

2)

.Subsequentoperationsinvolveexecutingthedata-parallelnon-MoElayers.

WefirstprofiledthetrainingprocessofthreepopularMoEmodelsemployingexpertparallelism(detailedinTable

1)onaclustercomprisedoffourA100machines,eachequippedwithaninter

-connectRDMAbandwidthof200Gb/s.Theproportionofall-to-allcommunicationtimerelativetothetotaltrainingdurationisillustratedinFigure

3(a).

Wethendoublethenumberofmachines,andthenumberofexpertstoincreasethemodelscale.TheresultsareshowninFigure

3(b)

and

3(c),respectively.

Ourfindingsrevealthatall-to-allcommunicationaccountedforasubstantialportionofthetotaltime:approximately30%inGPT-MoE(15B),40%inRoBERTa-MoE,and70%inSwin-MoE-L.Andthisoverheadremainsnearlyconstantinlargermodelsandatlargermachinescales.Theseresultshighlightasignificantbottleneckthathampersthescalabilityofthetrainingprocess.Consequently,thedurationofall-to-allcommunicationsubstantiallyconstrainstrainingwithexpertparallelism,leadingtoreducedoverallthroughputandlimitingthepotentialtoscaleupthenumberofexpertseffectively.

2.3Locality-SensitiveHashingAlgorithms

Locality-SensitiveHashing(LSH)isaprobabilisticmethodprimarilyusedtoapproximatenearestneighborsearchinhigh-dimensionalspaces,whichreducesthedimensionalityofdatabymappingsimilardatatothesame“buckets”withhighprobabilityusinghashfunctions.Thisapproachoffersasubstantialreductionincomputationalcomplexity,particularlybeneficialforlarge-scaledataapplications.ThekeyoperationsinLSHincluding:

MappingDataintoBuckets:ThecoreofLSHisafamilyofhashfunctionsthatmaximizetheprobabilityofnearbypointsintheoriginalspacestayingcloseinthehashedspace,whiledistantpointsarelikelytoendupindifferentbuckets.Eachhashfunctionhischaracterizedbytheproperty:P[h(x)=h(y)]=1?d(x,y)/D,whered(x,y)isthedistancebetweenpointsxandy,andDdenotesthediameterofthespace.Tomapsimilardataintothesamebucket,multiplehashfunctionsfromthisfamilyareselectedbasedonthespecificattributesofthedata(e.g.,Euclideandistance,cosinesimilarity)andthedesiredgranularityofthebuckets.Datapointsarethenhashedbythese

5

residuals

device0

廠

expert0

tokenscentroidsE(centroids)E(tokens)

3

4

1

2

expert1

All-To-All

All-To-All

LSH-BasedClustering

device1

Residual-basedErrorCompensation

expert2

tokenscentroidsresiduals

?=

E(centroids)residualsE(tokens)

+=

calculatethecentroids

average(1,,5)=

a

2

device2

buckets

centroids

a

b

hashfunction

tokenshash

1

2

3

4

5

6

Figure5:SchematicofMoEtrainingwithLocality-SensitiveHashing(LSH-MoE).

functions,andeachpointisassignedtobucketsaccordingtoitshashvalues,effectivelycategorizingsimilaritemstogetherforclustering.

CalculatingClusterCentroids:Bygroupingdatapointsintobucketsasdeterminedbytheirhash

values,datapointsareeffectivelyclustered.Eachbucketrepresentsaclusterofdatapointsandthe

centroidofeachclusteristhencalculatedasthemeanofallpointswithinthatcluster,formulatedas

Cj=ε1xi,whereCjisthecentroidofthej-thbucket,njisthenumberofpointsinthej-th

bucket,andxiarethedatapointsinthebucket.

3Methodology

3.1TheMotivationofTokenSimilarity

Toexplorethepotentialoptimizationforall-to-allcommunicationsinMoEtraining,weconductedanin-depthanalysisofthedatainvolvedintheseall-to-allcommunications,identifyingahighdegreeofsimilarity,termedtokensimilarity.Specifically,weappliedPrincipalComponentAnalysis(PCA)toreducethedimensionalityoftheinputtokensofall-to-allcommunicationsandobservedadistinctclusteringphenomenon,asillustratedintheFigure

4.

Ouranalysissuggeststhattheobservedsimilarityamongtokensmaystemfromtwoprimaryfactors:

?DataRelatedInfluences:Thesimilarityispartiallyduetothenatureofreal-worlddata,whichoftenadherestoZipf’s

Law[18]

.Thisresultsinaskeweddistribution,withcertaindataelementsappearmorefrequentlythanothers.

?ModelStructureRelatedInfluences:ThedesignofTrans-

formerarchitecture[34],especiallyitsattentionmecha

-nisms,significantlyimpactstokensimilarity.Inmodels

likeBERT[7],attentionlayersaredesignedtocaptureand

integratecontextinformationacrosstokens,thushomoge-nizingtokenrepresentationsandemphasizingtheirsharedsemanticrelationshipsatthesentencelevel.

Figure4:PrincipalCom-ponentAnalysis(PCA)Visu-alizationofinputtokensin-volvedinall-to-allcommuni-cation.

6

3.2LSH-MoE

MotivatedbytheTokenSimilarityobservedinSection

3.1,weintroduce

LSH-MoE,anovelMoE

trainingframeworkthatintegrateslocality-sensitivehashing(LSH)forrapidclusteringofinputtokens.Ourmethodtransmitsonlytheclusteringcentroids,significantlyreducingcommunicationvolumes.Tocounteractthenegativeeffectsofcompression,wealsoimplementaresidual-basederrorcompensationscheme.

AsdepictedinFigure

5,

LSH-MoEinitiallyemploys(1)anLSH-basedclusteringmethodtocompresstokensintocentriodsforsubsequentprocessing,effectivelyreducingcommunicationoverhead.Itthensequentiallyexecutes(2)all-to-allcommunication,expertcomputation,andanother(3)all-to-allcommunicationtoproducetheprocessedoutputsE(centriods).Finally,itintroduces(4)aresidual-basederrorcompensationmethodtoapproximatetheexpert-processedresultsE(tokens),byintegratingE(centriods)withresiduals.Meanwhile,wealsooutlinetheworkflowofourLSH-MoEframeworkintheAlgorithm

1

ofAppendix

A.1.

ThekeycomponentsofourLSH-MoEframeworkincludesanefficientLSH-basedclusteringalgorithmforrapidprocessingandanresidual-basederrorcompensationschemetominimizequalitydegradation.

EfficientLSH-basedClusteringAlgorithm.Sincethedatatobecompressed(theinputdataforall-to-allcommunication)isgenerateddynamicallyandinrealtime,pre-compressingitoroverlappingcompressiontimewithotherprocessingtasksisnotfeasible.Consequently,selectinganefficientonlinecompressionalgorithmiscrucial.Traditionalclusteringalgorithms,suchasK-Means,oftenencountercomputationalchallengesandefficiencylimitations.Locality-sensitivehashing(LSH)addresstheseissuesbyhashingsimilardatapointsintothesamebuckets,enablingfastersimilaritydetectioninhigh-dimensionalspaces.

NumerousLSHalgorithmshavebeendeveloped,eachemployingauniquehashingapproachformappingdataontobuckets.Weconductedexperimentstoevaluateseveralpopularhashingalgorithms,includingcross-polytopehashingandsphericalhashing.BasedonourevaluationsinSection

4.5,we

selectedcross-polytopehashingastheoptimalalgorithmforourapplication.Cross-polytopehashingstandsoutforitsmethodofmappinginputvectorstothenearestvertexonacross-polytope.Thisprocessisfacilitatedbyapplyingrandomlyrotatedcross-polytopes,whicheffectivelysegmentthesurfaceoftheunitsphere.Thealgorithmcanbemathematicallyrepresentedasfollows:

LSH(x)=argmaxi∈{±1,±2,...,±d}|Rx|i(3)

whereRisarandomrotationmatrix,disthedimensionalityofthespace,and|Rx|idenotestheabsolutevalueofthei-thcomponentoftherotatedvectorRx.

ThisformulaencapsulateshowtheinputvectorxistransformedbytherotationmatrixRandthenmappedtothenearestvertexofthecross-polytopebyselectingthedimensionithatmaximizestheabsolutevalueofthecomponentsofRx.Thismethodeffectivelysegmentsthehigh-dimensionalspaceandenhancestheclusteringefficiencybyrapidlyidentifyingsimilardatapoints.

Residual-basedErrorCompensationScheme.InourLSH-MoEframework,wecompresstheintermediateactivationvalueswithinthemodelnetwork.Unlikegradientcompression,thisprocessdoesnottolerateerrorswell.Therefore,itisessentialtominimizecompression-inducederrorstoensureminimalimpactonmodelperformance.Toaddressthis,weimplementanovelresidual-basedgradientcompensationstrategy,outlinedasfollows:

1.Wefirstcapturetheresidualforeachdatapointrelativetoitsclustercentroid,definedbytheequation:

2.Aftertheexpertnetworkcomputesoutputsfortheclustercenters,thefinalstepistorestoretheprocessedresultforeachtokenbyaddingbackthepreviouslyrecordedresidual:

Yij←{E(clusterj)+?Clusterjk|k=1,2,...,Nj}.(5)

Thiserrorcompensationschemeeffectivelymitigatespotentialaccuracylosscausedbydatacompres-sioninall-to-allcommunication,ensuringthefidelityandrobustnessoftheLSH-MoEframework.TheexperimentalresultsinSection

4

showthatimplementingthiscompensationmechanismenables

7

Table1:Modelsforevaluation,where“-”indicatesthatthevaluesaredifferentacrosslayers.

Model

#Layer

dmodel

dffn

#Experts

#Params.(MoE)

#Params.(Total)

RoBERTa-MoE

12

768

3072

16

302M

394M

T5-MoE

16

1024

16384

16

8594M

9288M

GPT-MoE(15B)

12

768

3072

512

14507M

14629M

GPT-MoE(52B)

24

1024

4096

512

51539M

51740M

Swin-MoE-L

24

-

-

32

-

946M

themodeltrainedwithLSH-MoEtoachieveanaccuracycomparabletothatofamodeltrainedwithoutcompression.Thisoutcomehighlightstheeffectivenessofourproposederrorcompensationstrategyinpreservingmodelperformancedespitethechallengesposedbydatacompressioninall-to-allcommunication.

3.3ScalabilityAnalysisofLSH-MoE

Toeffectivelydemonstratethescalabilityofourapproach,particularlyintermsofitsapplicabilitytobothlargermodelsandlargercomputationalclusters,weconductedatheoreticalanalysis.ThisanalysisprimarilyfocusesonthecomputationoverheadandthecommunicationcostsassociatedwithMixtureofExperts(MoE),specificallyconsideringall-to-allcommunicationoverhead.Wederivedtheratioofcommunicationtimetocomputationtime,highlightinghowthisratioevolvesasboththescaleoftheserversandthemodelsizeincrease.Thisrelationshipiscrucialforunderstandingscalabilityandcanbeformallyexpressedasfollows:

wherekrepresentsthenumberofexpertsactivatedpertoken,FLOPsandBinterdenotetheGPU’scomputationabilityandthenetworkperformance,wisthenumberofGPUservers,andhisthehiddensizeofmodel.Notably,thefirstterm,.Additionally,scalingMoEmodelstypicallyemphasizesincreasingthenumberoflayersandexperts,

whilethegrowthinhiddensize(h)tendstobegradual,asseeninmodelslikeSwitch-Transformer[9]

.Consequently,whenboththemodelscaleandthenumberoftrainingserversgrow,theproportionofall-to-allcommunicationtimeremainsnearlyunchanged.ThisinsightunderpinsthescalabilityoftheLSH-MoEmethod,demonstratingitsrobustnessinlarger-scalesettingsandsupportingitspotentialinfuturelarge-scaleapplications.Foradetailedderivation,pleaserefertoAppendix

A.2.

4Experiment

4.1Implementation

OurLSH-MoEcomprisesadatacompression/restorationcomponentandacommunicationcompo-nent.WeutilizePyTorch1.11fordevelopingtheLSHclusteringandNCCLforimplementingthecommunication.Additionally,ourmethodisframework-independentandcanbeeasilyappliedto

otherMoEtrainingframeworkssuchasHetu-MoE[21,

26],DeepSpeed-MoE[29],andTutel[12]

.

4.2BenchmarksandDatasets

Ourevaluationsareconductedbyscalingpre-trainedmodelsequippedwithMoEarchitectureacrossvariousapplicationdomains.ThisincludesmodelslikeRoBERTa-MoE,T5-MoEandGPT-MoEinnaturallanguageprocessing(NLP),aswellasSwin-MoEincomputervision(CV).Amongthesemodels,RoBERTa-MoEandT5-MoEareevaluatedonpre-trainingtask,whileGPT-MoEandSwin-MoEundergofine-tuningevaluationbasedontheirofficialopen-sourcedmodelcheckpoints

1

2.

Wealsoevaluatedthezero-shotaccuracyofthepre-trainedT5-MoE.ModelconfigurationsaredetailedinTable

1.

1/facebookr