README.md

MLPerf EDU Example Path

Set up first if you have not already. The project README has the quickstart, and these examples additionally need the dev extra:

uv sync --locked --extra dev

Run everything below from the mlperf-edu/ project directory. The classroom guide covers teaching context and grading.

The numbered examples form one classroom sequence. They use registered workloads and product CLI artifacts rather than standalone toy measurements.

Example	Question	Primary Artifact
01 health check	Is this installation ready for benchmark work?	Suite health HTML
02 inference tradeoff	What did a controlled batch-size change do?	Pro condition report
03 training tradeoff	Did a training change produce an acceptable checkpoint?	Training and inference lineage
04 result comparison	Which comparisons are valid, and why?	Compatibility-checked HTML
05 assignment package	Can the result be verified and graded elsewhere?	Portable ZIP and grade JSON

Each README states the learning goal, hardware expectations, allowed changes, report sections to inspect, interpretation questions, and a suggested rubric. Unless an instructor supplies a different template, students should submit an answers.md file with numbered responses beside the generated artifacts. The CLI grader checks the benchmark contract; the rubric also grades those written responses. The separate research/pro-collection example uses the same plan mechanism for a research-facing study.

Before choosing a max workload, read the fourteen-workload readiness matrix. Every workload runs locally, but they separate into those that reproduce their inherited target and those recorded as a miss, and the matrix states the next quality task for each. Instructors should publish course-machine max runtime, memory, download, and disk budgets because those costs are hardware dependent. The initial course-image budget already covers every functional min path on CPU and the available MPS paths.

The three legacy lab*.py files remain standalone teaching experiments. They do not emit canonical benchmark artifacts and should not be presented as registered MLPerf EDU results.