使用 Amazon SageMaker AI 自定义和扩展基础模型.pdf

1、A I M 3 6 3Customize&scale foundation models using Amazon SageMaker AISumedha Swamyhe/himSr.Manager,Product ManagementAWS AI PlatformsGiuseppe A.Porcellihe/himPrincipal,ML Solutions ArchitectAWS AI Platforms 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.AgendaThe need for custom

2、izationFoundation model customization techniquesHow to customize foundation models using SageMaker AIDemoQ&A 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Why customize modelsFoundation models are powerfulbut genericDifferentiated value comes from customizing models to YOUR data

3、,YOUR workflowsModel customization=Competitive advantageCustomizationModel customization encodes domain expertise beyond what prompt engineering and RAG can capture.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Why customize modelsFoundation models are powerfulbut genericDiffere

4、ntiated value comes from customizing models to YOUR data,YOUR workflowsModel customization=Competitive advantageCustomizationModel customization encodes domain expertise beyond what prompt engineering and RAG can capture.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Most teams n

5、ever escape prototype phaseThe customization tax:Why teams get stuckThe reality of DIY customization:Infrastructure wrestlingFragmented tools Version tracking chaosGovernance blind spotsWeeks on plumbing vs.thinkingResultYAML configsGPU clustersVersion trackingDeploymentGovernance 2025,Amazon Web Se

6、rvices,Inc.or its affiliates.All rights reserved.How we eliminated the customization taxServerless InfrastructureNo cluster managementAutomatic distributed trainingPay for resources usedProduction ready techniquesSFT,DPO,RL on a broad choice of models:Nova,gpt-oss,llama,Qwen,and moreConfigure,dont b