Building Efficient PyTorch Models on Vertex AI
summary
Shows how to run PyTorch with custom Docker containers + multi-node, multi‑GPU configs, how to register model artifacts and deploy them behind Vertex AI endpoints, and how to build reproducible pipelines so stages (data prep, training, aggregation, deploy) run independently. Demonstrates a Vertex reduction server reduces training time ~45% vs NCCL all‑reduce by doing weight aggregation on a smaller CPU instance, and explains hyperparameter tuning via Vertex AI trials and using archived model artifacts for inference.
AI-written summary · about
chapters (8)
- 0:00Introduction to distributed
- 1:06Distributed training basics
- 3:01Setting up custom containers
- 4:53Multi-node configuration
- 7:40Vertex reduction server
- 10:51Hyperparameter tuning
- 12:40Model inference and registry
- 14:51Vertex AI pipelines
related talks
33mFaster model training and experimentation with Vertex AI
45mVertex AI: Pipelines for your MLOps workflows
51mHands on with Cloud AI
12mBuild end-to-end solutions with Vertex AI
26mBuilding Portable ML Pipelines for Development and Production
40mVertex AI: Pipelines for your MLOps workflows
02mFuel your custom models on the Cloud AI Platform
22mQuickly deploy NVIDIA optimized generative AI on Azure AI Studio