先进的分布式训练策略以实现最佳的GPU利用率.pdf

1、 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.A I M 3 5 4Advanced distributed training strategies for optimal GPU utilizationKeita Watanabe(he/him)Principal Specialist SA,ML Training&InferenceAman Shanbhag(he/hi

2、m)Specialist SA,ML Training&Inference 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.(Rough)AgendaIntroduction to the Generative AI Stack on AWSML Training:What happens in a training loop?Whats in the accelerators memory?Optimization Strategies:ParallelismInfrastructure Considera

3、tions:Building an ultra-clusterOrchestrationThis is a chalk talk,so you tell us what youd like to talk about!2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.GPUsInferentiaTrainiumSageMakerEC2 Capacity BlocksNeuronUltraClustersEFANitroAmazon QBusiness Amazon QDeveloperAmazon Q in Q

4、uickSightAmazon Q in Connect Amazon BedrockGuardrailsAgentsCustomization CapabilitiesBroadest selection of modelsStudioAmazon Q in AWS Supply ChainAPPLICATIONS THAT LEVERAGE LLMs AND OTHER FMsTOOLS TO BUILD WITH LLMs AND OTHER FMsINFRASTRUCTURE FOR FM TRAINING AND INFERENCE 2025,Amazon Web Services,

5、Inc.or its affiliates.All rights reserved.APPLICATIONS THAT LEVERAGE LLMs AND OTHER FMsTOOLS TO BUILD WITH LLMs AND OTHER FMsINFRASTRUCTURE FOR FM TRAINING AND INFERENCEGPUsInferentiaTrainiumSageMakerEC2 Capacity BlocksNeuronUltraClustersEFANitro 2025,Amazon Web Services,Inc.or its affiliates.All ri

6、ghts reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.ML Training 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.W H Y A C C E L E R A T O R S?(CUDA)CoresOriginally designed for graphics Which is good for lots of+,xSource:https:/ Web Services,Inc.or its