Interactive lab redesign
Activity-Level MLSysIM Lab Blueprint
This page summarizes the implemented labs as ML systems trade-off environments. Each lab opens with a chapter anchor and track-specific scenario, then walks through one or more learning nuggets. Each nugget follows the same rhythm: predict, find the binding constraint, explore the trade-off surface, decide, and reflect.
Depth And Narrative Contract
The labs are designed as substantial guided engineering activities, not quick demos. A compact run should take about 30 minutes; a full report-oriented assignment can reasonably take 45 to 60 minutes.
Opening narrative
Every lab should start with volume, chapter, title, chapter recap, learning objectives, track context, scenario brief, and lab map.
Chapter anchor
Each lab maps to a specific book chapter and turns that chapter into a systems decision: constraint, knob, evidence, and defensible choice.
Time-on-task
Default path: 3 to 5 minutes orientation, 20 to 35 minutes prediction/evidence/decision work, 7 to 15 minutes report and reflection.
Student artifact
The target artifact is a local Markdown report with track, scenario, predictions, evidence summary, final decision, residual risk, provenance/caveats, and result snapshot.
Canonical Track Materialization
Students pick one track and carry it through the labs. The track is not just a label: it changes stakeholder, device family, model family, metrics, guardrails, dominant constraints, narrative, and report framing.
Volume Track Arcs
The same chapter goals repeat across all tracks, but the long-form narrative changes. Volume I teaches deployment feasibility for a selected context. Volume II scales that same context into fleet, operations, responsibility, and synthesis questions.
Implemented Volume Coverage
This table shows what each lab is trying to teach, which book chapter it anchors to, and how the four tracks materialize the same pedagogy differently.
| Volume | Lab | Chapter anchor | Core concept | Track materialization |
|---|
Core Concepts
These are the cross-cutting ML systems ideas the labs should reinforce repeatedly.
| Concept | What students need to learn | MLSysIM responsibility |
|---|
Package Boundary
MLSysIM is the reusable engine for both the textbook and labs. The lab helper layer owns pedagogy and UI.
MLSysIM should contain
- Constants and registries used by the book and labs.
- Hardware, model, workload, and scenario data.
- Domain formulas, solvers, sweeps, and result schemas.
- Book-friendly helpers for reproducible callouts and figures.
Lab helper package should contain
- Marimo UI components and layout.
- Track selector and curriculum mapping.
- Design ledger and browser persistence.
- CSS, Plotly themes, prompts, and instructor-facing metadata.
from mlsysim import Engine, Hardware, Models, Scenarios
and
from mlsysbook_labs import DesignLedger, track_selector, tradeoff_poster, decision_card
Experience System
The labs need a consistent academic shell, not a dark dashboard banner. The interaction should support chapter reading, repeated learning nuggets, prediction, trade-off exploration, reflection, and decision-making.
Visual Direction
- Light academic interface with restrained borders and readable spacing.
- Volume I accent: Harvard Crimson
#A51C30. - Volume II accent: ETH Blue
#1F407A. - Replace inline dark gradients with a shared
lab_header(). - Use status colors only for feasible, caution, violated, and selected states.
Pedagogical Cycle
- Lab wrapper: volume, chapter, mini recap, track, scenario, and nugget map.
- Each nugget names one chapter idea and why it matters.
- Prediction before instruments are revealed.
- Binding constraint, trade-off surface, engineering decision.
- Reflection prompt and design ledger update.
Interaction Devices
The device catalog defines which Marimo controls and visuals to use for each pedagogical purpose, so widgets are selected intentionally.
Control devices
Slider cards, range sliders, number predictions, radio/segmented strategy selectors, dropdowns, checklists, switches, and commit forms.
Visual devices
Constraint budgets, stacked bars, curves, Pareto frontiers, heatmaps, rooflines, queueing charts, timelines, world maps, topology diagrams, and tables.
Pedagogical devices
Prediction locks, method notes, failure boundary callouts, reflection cards, decision cards, before/after comparisons, and ledger timelines.
Critical lens
Each device must help students identify a knob, constraint, trade-off, decision, assumption, or residual risk.
Design Completeness
Before implementation, each lab should capture the ML systems translation, thematic nuggets, misconceptions, student artifacts, rubrics, and instructor use.
ML to systems
For every chapter: name the ML concept, systems translation, knob, binding resource, failure mode, and defensible decision.
Thematic nuggets
Reinforce durable ideas: D.A.M coupling, Iron Law, resource budgets, bottleneck migration, data movement, tails, scale, reliability, and responsibility.
Misconceptions
Each nugget should correct a specific wrong intuition, such as smaller is always faster or more GPUs always means linear speedup.
Artifacts and rubric
Capture prediction, bottleneck diagnosis, trade-off interpretation, selected configuration, residual risk, chapter principle, and result snapshot.
Downloadable report
Each lab should export a lightweight Markdown report and optional JSON ledger snapshot so students can submit evidence of completion.
Instructor guide
Use one consolidated guide organized by volume and chapter, with optional local notes beside each lab for maintenance.
Instructor Adoption
The labs should be easy to adopt as global teaching material. An instructor should understand why to assign a lab, where it fits, what students submit, and how to evaluate the result.
Mini recap
Every lab opens with the chapter emphasis, key terms, ML-to-systems translation, what to watch, and the common trap the lab corrects.
Assignment-ready
Each lab should include copy-ready assignment text, expected report contents, and a lightweight rubric.
Adoption modes
Standalone homework, lecture companion, module sequence, and capstone preparation should all be supported without redesigning the lab.
Instructor value
Students practice ML systems judgment without requiring a custom cluster, hardware kit, long build project, or instructor-authored simulator.
Student Report
Every lab should end with a simple downloadable artifact that a student can submit to show they completed the lab and made a reasoned ML systems decision.
What it proves
The student selected a track, completed the nuggets, interpreted the binding constraints, and made a defensible engineering decision.
Required fields
Student identifier, lab ID, volume/chapter, track, scenario, predictions, constraints, selected policy, trade-off interpretation, residual risk, timestamp, and MLSysIM version.
Format
Markdown first for easy submission and inspection. JSON ledger snapshot is optional for automated checks. HTML can come later if polish is needed.
Implementation boundary
mlsysbook_labs builds and downloads the report. MLSysIM provides typed, serializable result snapshots from the core engine.
Versioning
The labs should be versioned like instructional infrastructure so instructors can assign stable material and students can submit reports that identify exactly what they completed.
Visible metadata
Each lab should expose lab ID, lab version, updated date, book anchor, MLSysIM version, helper package version, report schema version, and ledger schema version.
Report provenance
Downloaded reports must include lab version, MLSysIM version, report schema version, timestamp, track, scenario, and result snapshot.
Release channels
Stable for courses, preview for redesigned labs, and dev for active work. Stable releases need tests, changelog, instructor guide snapshot, and known limitations.
Change policy
Major changes alter assignment meaning, minor changes add supported capability, and patch changes fix defects without changing learning objectives.
Activity Matrix
Use the controls to filter by volume or search for concepts such as latency, compression, fairness, serving, or collective communication.
| Volume | Chapter / Lab | Activity | What will be covered | How it will be covered | MLSysIM needs |
|---|
Volume 2 Mapping Decision
Decision: Volume 2 should add a dedicated Collective Communication lab and realign downstream lab numbering with the book sequence.
| Option | What it means | Trade-off |
|---|---|---|
| Chosen | Add a dedicated Collective Communication lab after Distributed Training and renumber downstream labs. | Best chapter alignment and cleaner pedagogy. The implementation must update lab IDs, links, index entries, and ledger keys together. |
MLSysIM Engine Plan
These APIs or model families should be added or consolidated so labs and textbook examples share one source of truth.