Build evidence, not just familiarity
These are experiment briefs. Use a small model, pin the environment, and distinguish measured results from estimates. Each lab specifies a concrete deliverable.
1 · Find a single GPU’s limiting resource
Explain a baseline using memory accounting, a roofline estimate, and a short trace.
2 · Protect chat during document bursts
Tune scheduling and cache reuse while keeping separate latency targets for each traffic class.
3 · Decide whether disaggregation is worth it
Compare at the same GPU budget and find a workload where the complicated design loses.
4 · Survive a bad day
Bound overload and retry amplification; demonstrate cancellation, cleanup, and recovery.
5 · Defend an end-to-end service
Show the contract, design, quality gate, performance evidence, cost model, and runbook.