InfrastructureCurating

AI Infrastructure

GPU systems, distributed training, and efficient serving

A systems-oriented track for understanding the infrastructure that makes model training and inference reliable.

Learning map

The focus is the layer below the model API: memory hierarchy, parallelism, scheduling, data movement, and the operational practices around them.

Resource index

6 links

Official PyTorch material for distributed data parallel, model parallel, and related patterns.

The CS336 course site provides a practical connection between model training and systems constraints.

Personal notes on GPU environments, remote development, networking, and tools.

Lecture notes and transcript-style explanations with an emphasis on system behavior.

Technical writing on GPU systems, evaluation infrastructure, and practical workflows.

Course resources and published material on the learning site.

DataHub workflow

Hardware and memory

Parallelism

Training systems

Inference serving

Operational practice

Syllabus

GPU memory, bandwidth, communication, and why bottlenecks move as models scale.

Data, tensor, pipeline, and sequence parallelism plus the engineering around distributed jobs.

Serving, batching, caching, monitoring, and the practical work of making model systems dependable.