EP8: Training Models at Scale | AWS for AI Podcast

02/09/2025 1h 4min Episodio 8
EP8: Training Models at Scale  | AWS for AI Podcast

Listen "EP8: Training Models at Scale | AWS for AI Podcast"

Episode Synopsis


Join us for an enlightening conversation with Anton Alexander, AWS's Senior Specialist for Worldwide Foundation Models, as we delve into the complexities of training and scaling large foundation models. Anton brings his unique expertise from working with the world's top model builders, along with his fascinating journey from Trinidad and Tobago to becoming a leading AI infrastructure expert.Discover practical insights on managing massive GPU clusters, optimizing distributed training, and handling the critical challenges of model development at scale. Learn about cutting-edge solutions in GPU failure detection, checkpointing strategies, and the evolution of inference workloads. Get an insider's perspective on emerging trends like GRPO, visual LLMs, and the future of AI model development.Don't miss this technical deep dive where we explore real-world solutions for building and deploying foundational AI models, featuring discussions on everything from low-level infrastructure optimization to high-level AI development strategies.Learn more: http://go.aws/47yubYqAmazon SageMaker HyperPod : https://aws.amazon.com/fr/sagemaker/ai/hyperpod/The Llama 3 Herd of Models paper : https://arxiv.org/abs/2407.21783Chapters:00:00:00 : Introduction and Guest Background00:01:18 : Anton Journey from Caribbean to AI 00:05:52 : Mathematics in AI 00:07:20 : Large Model Training Challenges00:09:54 : GPU failures : Lama Herd of models 00:13:40 : Grey failures 00:15:05 : Model training trends00:17:40 : Managing Mixture of Experts Models00:21:50 : Estimate how many GPUs you need.00:25:12 : Monitoring loss function 00:27:08 : Crashing trainings 00:28:10 : SageMaker Hyperpod story00:32:15 : How we automate managing grey failures00:37:28 : which metrics to optimize for 00:40:23 : Checkpointing Strategies 00:44:48 : USE Utilization, Saturation, Errors 00:50:11 : SageMaker Hyperpod for Inferencing 00:54:58 : Resiliency in Training vs Inferencing workloads 00:56:44 : NVIDIA NeMo Ecosystem and Agents00:59:49 : Future Trends in AI 01:03:17 : Closing Thoughts

More episodes of the podcast AWS For AI