Skip to content

README.md

Latest commit

 

History

History
75 lines (58 loc) · 2.87 KB

README.md

File metadata and controls

75 lines (58 loc) · 2.87 KB
Raw

05 — Assignment Package

Learning Goal

Produce one portable functional-readiness result whose workload, profile, claim boundary, and provenance can be verified and graded on another machine. This lab isolates artifact portability; Examples 02 and 03 carry the authoritative quality work.

Runtime and Hardware

The assignment uses the fast image-classification min path and requires no authoritative dataset download. It runs on a laptop CPU, Apple Silicon, or CUDA system. The artifact proves setup and execution only; it does not claim that the 85% max quality gate was evaluated.

Student Flow

uv run mlperf health
uv run mlperf run --workload image-classification --profile min \
  --output-dir submissions/05-assignment
uv run mlperf package \
  submissions/05-assignment/image-classification_min.provd.json \
  --output submissions/05-assignment.zip
uv run mlperf grade submissions/05-assignment.zip \
  --assignment examples/05-assignment-package/assignment.yaml \
  --output submissions/05-assignment-grade.json

The last command is the same fail-closed contract check the instructor runs. Before submission, identify the functional decision, why the displayed max target is context only, checkpoint source, executed device, and provenance status in the HTML report.

Instructor Flow

uv run mlperf grade submissions/05-assignment.zip \
  --assignment examples/05-assignment-package/assignment.yaml \
  --output submissions/05-assignment-grade.json

The grader accepts a directory, manifest, or portable ZIP. Package verification rejects traversal, symbolic links, duplicate members, unindexed files, digest or size mismatches, invalid provenance, false quality claims, configuration drift, and unexpected result cardinality.

Allowed Changes

This contract permits no workload or profile changes. Students may change the output location and choose an available device. Score-bearing result packaging remains disabled where doing so would redistribute fetch-only or release-review dataset bytes.

Interpretation Questions

  1. Which artifact binds the report, weights, and dataset evidence?
  2. What does a passing package verification establish, and why does it not establish the 85% quality gate?
  3. Which assignment fields catch an unexpected workload or profile?
  4. Why should an instructor still rerun a sample of submissions?

Suggested Rubric

Item Points Evidence
Setup and execution 2 Complete packaged min artifact
Claim boundary 3 Correct explanation that quality was not evaluated
Systems interpretation 2 Measured region, throughput, and device explanation
Reproducibility 3 Verified package that passes the assignment contract

Submit the ZIP, self-check grade JSON, and answers.md. The instructor should regenerate the authoritative grade JSON rather than trusting the submitted self-check.