Merlin ： GPU 加速的推荐系统框架.pdf

1、NVIDIAMerlin:AGPUAcceleratedRecommendation Framework王泽衰（Joey Wang），Dec 17th 2020#page#Industrial Recommendation ChallengesFeatureDataloadingScaling TrainingDeploymentEngineeringTabular data scalesThe path to productionIterating over featureEmbedding tablesfrom research ispoorly using thedont fit eas

2、ily on GPUcombinations takescommon deep learningcomplex and requiresmethod of item bylonger than training.andare hard to scale.significant engineeringitemeffort.#page#Merlin Recommendation FrameworkTritonNVTabularHugeCTRFeatureDataloadingScaling TrainingDeploymentEngineeringInference time dataPrepar

3、e massiveEasy to use data andtransformsand multidatasets in minutesAsynchronous batchmodel support providemodel parallel trainingallowing for moredataloading means theallow you to scale tomaximumthroughputexploration and betterGPU is always utilizedTB sized embeddingswithlatencymodels.constraints#pa

4、ge#Merlin is in Open Beta!NVIDIA MERLIN OPEN BETADemocratizing Large-Scare Deep Learning RecommendersOpen SourceETLDATATRAININGINFERENCELOADEREasy to Use01000010GPU AcceleratedHugeCTRNVTabulaNVTabiBillions）TensorflowEnd to EndOPyTorchEWBEDCINSSUser QueryRAPIDSRAPIDSCUDNNTritonDATALAKE#page#NVTabular

5、GPU Accelerated ETL#page#The average data scientist spends 75% of their time in ETLas opposed to training modelstart ETLrs03Lmrestart ETLeh，forgottoadd afeatur1212/test model113orkflowgetacoffeegetacof8repeaCPUGPUPOWEREDPOWEREDWORKFLOWWORKFLOWdatasetcollectionFIngUneanalysisovemightETLovernightresta

6、rtETLworkflowagaintraininferencestaylateontiime#page#NVTabular: Recommender System ETL on GPUhttps:/ Transforms accelerated on GPUFully compatible with (and built upon!） the RAPIDS ecosystem CDask-cuDF）No limit on dataset size (not bound by GPU or CPU memory）A Higher level abstraction- What you want