在 AWS Trainium 上掌握基础模型的训练后工作流程.pdf

1、 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.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 6Samir AraujoDiwakar BansalMaster Post-Training workflows for foundation models on AWS T

2、rainiumPrincipal Solutions ArchitectAnnapurna Labs,AWSPrincipal GenAI SpecialistAWS 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Generative AI trendsWhat is post-trainingThe DeepSeek moment fueling post-training Post-training techniques:SFT,Reinforcement Learning,Knowledge Dist

3、illation Post-training on AWS Trainium with NeuronOpen discussionAgenda 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Unlabeled DataTraining Build new foundation modelsAdapt models to proprietary or domain specific dataProprietary dataFine-TuningAI ApplicationsInferenceDeploy AI

4、 models and agentsGenAI workloads 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.What will drive AI in 2026?AI agents is a tectonic shift in how we build,deploy,and interact with AI Domain Specific Models build for specific applications,private data or personalized to the individ

5、ual Lower Training CostPost-training rather than pre-training models can reduce the cost by up to 80%Higher demand for domain specific models produced by more efficient post training techniquesAI Agents40%of enterprise apps will be integrated with task-specific AI agents by the end of 2026,up from l

6、ess than 5%today(Gartner)2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Post-training is broader definition of ML model training which includes fine-tuning plus other techniques like Reinforcement Learning(RLHF),instruction tuning,and alignment methodsIn many cases,an out-of-box