Advanced AI Infrastructure Design and Optimization
Chapter 1: Chapter 1: Architectural Patterns for AI Platforms
MLOps Principles at Scale
Compute Selection: CPU, GPU, and TPU Architectures
High-Bandwidth Interconnects for Distributed Systems
Storage Solutions for Large-Scale AI Datasets
Networking Topologies for ML Clusters
Hands-on Practical: Environment and Tooling Setup
Chapter 2: Chapter 2: Engineering Distributed Model Training
Data Parallelism with Synchronous and Asynchronous Updates
Model and Pipeline Parallelism for Large Models
Implementing Training with Horovod
Leveraging Microsoft DeepSpeed for ZeRO and Offloading
Fault Tolerance and Checkpointing in Long-Running Jobs
Hands-on Practical: Distributed Training with PyTorch FSDP
Chapter 3: Chapter 3: Advanced Resource Orchestration with Kubernetes
Managing ML Workflows with KubeFlow Pipelines
Advanced GPU Scheduling and Sharing
Cluster Autoscaling for Dynamic ML Workloads
Strategies for Using Spot and Preemptible Instances
Multi-Tenancy with Namespaces, Quotas, and Priority Classes
Practice: Configure a GPU-Aware Autoscaling Group
Chapter 4: Chapter 4: High-Performance Model Inference and Serving
Architecting Inference Services for Latency and Throughput
Model Optimization with TensorRT and ONNX Runtime
Model Quantization Techniques: INT8 and FP8
Serving Multiple Models with NVIDIA Triton Inference Server
A/B Testing and Canary Deployments for Models
Hands-on Practical: Deploying an Optimized Model on Triton
Chapter 5: Chapter 5: Scalable Data Management and Feature Engineering
Designing and Implementing a Feature Store
Real-time vs. Batch Feature Computation
Data Versioning and Lineage with DVC and Pachyderm
High-Throughput Data Processing with Spark and Ray
Managing Data Lakes and Data Warehouses for AI
Practice: Build a Basic Feature Ingestion Pipeline
Chapter 6: Chapter 6: Financial Operations and Governance for AI
Applying FinOps Principles to ML Workloads
Cost Attribution and Showback Models for ML Teams
Optimizing Cloud Storage Costs for Datasets
Right-Sizing Compute for Training and Inference
Automating Cost Anomaly Detection
Governance Policies for Resource Consumption
Practice: Analyzing a Cloud Cost and Usage Report
Model and Pipeline Parallelism for Large Models
Was this section helpful?
- GPipe: Efficient Training of Giant Models using Pipeline Parallelism, Yanping Huang, Youlong Cheng, Ankur Bapna, Orhan Firat, Dehao Chen, Mia Chen, HyoukJoong Lee, Jiquan Ngiam, Quoc V Le, Yonghui Wu, Zhifeng Chen, 2019 Advances in Neural Information Processing Systems, Vol. 32 (NeurIPS) DOI: 10.5555/3454287.3455115 - Introduces pipeline parallelism with micro-batching to improve device utilization, addressing the 'pipeline bubble' for training large models.
- Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism, Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, Bryan Catanzaro, 2019 arXiv preprint arXiv:1909.08053 DOI: 10.48550/arXiv.1909.08053 - Describes how to implement tensor (inter-layer) and pipeline parallelism for very large language models, including discussions on communication and hybrid approaches.
- DeepSpeed: Pipeline Parallelism, DeepSpeed Team, 2023 (Microsoft) - Provides practical guidance and advanced scheduling techniques for implementing pipeline parallelism within the DeepSpeed framework.