Full-Stack LLM Systems

This track builds one connected mental model. Complete modules in order unless a diagnostic self-test shows both conceptual and applied mastery.

Prerequisites

• Python fluency sufficient to inspect tensor code.
• Algebra, derivatives, probability, and basic Linux tooling.
• Willingness to make estimates and record failed predictions.

Outcomes

By the end, you can trace how data, architecture, kernels, communication, serving policy, and evaluation jointly determine an LLM system’s behavior and cost.

Ordered modules

1. Foundations
2. Deep Learning
3. Transformers
4. LLM Training
5. Inference and Serving
6. GPU Kernels and Compilers
7. Distributed Systems
8. Evaluation

Track checkpoints

• After Module 3: derive and implement a small attention block.
• After Module 5: produce a memory-and-latency budget for a serving workload.
• After Module 7: defend a parallelism and collective-communication plan.
• After Module 8: publish an evaluation report with limitations.

Self-test

1. Which layers would you inspect when throughput drops but model outputs stay stable?
2. Why can a benchmark improvement fail to translate to product quality?
3. What evidence distinguishes remembering a definition from applied understanding?

Capstone

Design a constrained LLM service. Specify model and data assumptions, training strategy, kernel and distributed bottlenecks, serving SLOs, and an evaluation matrix. Include estimates, risks, and at least two experiments that could invalidate the design.