Train ML models at scale with Amazon SageMaker Training
summary
Shows how Amazon SageMaker Training spins ephemeral GPU clusters, downloads code from ECR and data from S3, logs to CloudWatch, and writes artifacts back to S3 so you pay only while jobs run. Demonstrates PyTorch and Hugging Face estimators, claims the SageMaker Training compiler speeds kernels with no extra charge, and reports ~20–40% internode gains using SM Distributed Data Parallel and the model-parallel library with pipeline schedules. Covers checkpointing and managed spot training to cut costs and resume after interruptions.
AI-written summary · about
chapters (17)
- 0:00Intro
- 4:19SageMaker training ephemeral clusters
- 5:36Many ways to train models on SageMaker
- 8:40View training job in AWS Console
- 10:19SageMaker Training Compiler
- 12:07Distributed training - Datasets
- 12:49Distributed training - Models
- 16:59SageMaker distributed training libraries
- 18:07Distributed training - Data Parallel (DDP)
- 19:44SageMaker DDP-Training script updates
- 22:57Distributed training - Model Parallel (SMP)
- 24:151. Use SMP to automate your model partitioning
- 25:342. Pipeline execution to stabilize GPU utilization
- 27:22Pipeline execution schedule
- 32:16Hyperparameter tuning methods
- 37:09Cost reduction with managed Spot Instances
- 38:29Getting started
related talks
31mTrain and tune ML models to the highest accuracy using Amazon SageMaker
33mBuild, train, and deploy ML models using Amazon SageMaker
44mScaling ML from 0 to millions of users
27mScaling ML to the next level: Hosting thousands of models on SageMaker
37mStreamlining Machine Learning Operations
34mBuild, train and deploy Machine Learning models with Amazon SageMaker
37mStreamlining Machine Learning Operations
29mTrain billion-parameter models with model parallelism on Amazon SageMaker