Labs v0.1.1 — Publish tested Labs updates for Volumes I and II2026-11-07
Labs 0.1.1 updates the browser-based Co-Labs for Volumes I and II.
- Introduce shared learning paths, onboarding, and lab helpers for comparing a baseline with a design alternative, recording a prediction, examining tradeoffs, and exporting a reasoned recommendation.
- Connect the notebooks to the updated MLSysIM modeling engine and build the matching browser wheels from source. Check notebook wheel references before export.
- Improve notebook styling, browser save/restore checks, evidence downloads, and live-site paths. Replace the deployed Labs subtree so retired pages do not linger.
The live release excludes Volume III and Volume IV material. Lab outputs are modeled results or recorded student work; an exported report is not proof of mastery or an automatic submission to an instructor.
Site: https://mlsysbook.ai/labs/
Slides v0.1.1: Publish updated Volume I and II lecture materials2026-11-06
Slides 0.1.1 updates the lecture materials and portal for Volumes I and II.
- Rebuild and revise decks around the book's chapter sequence, with selected Volume I topics split into paired lectures.
- Revise figures, typography, speaker notes, activities, and lecture-to-lecture connections; connect calculated examples to shared modeling inputs and retain source-specific assumptions.
- Add mechanical checks for deck structure and visual fit, improve chapter-to-slide links, and carry speaker notes into PowerPoint exports.
The live portal and MLSysBook slide archives include Volumes I and II only. Volume III and Volume IV decks remain outside this live release. PowerPoint slides are images of PDF pages, with speaker notes in the notes pane; slide text is not editable.
Portal: https://mlsysbook.ai/slides/
PDF and PowerPoint downloads: https://github.com/harvard-edge/cs249r_book/releases/tag/slides-latest
TinyTorch v0.2.2 - Publish TinyTorch updates and portable exercise timing2026-11-06
TinyTorch 0.2.2 updates teaching exercises and the Module 01 broadcasting explanation.
- Use a high-resolution interval clock in activation and data-loader timing exercises, avoiding zero-duration wall-clock readings on short runs such as those reported on Windows.
- Report nonpositive intervals as unmeasurable instead of dividing by zero when calculating timing comparisons.
- Clarify vector broadcasting and organize the example headings.
Site: https://mlsysbook.ai/tinytorch/
These timing changes affect the exercises; they do not establish a framework speedup or benchmark result.
MLSYSIM v0.3.1 — Publish modeling corrections and updated reference documentation2026-11-06
MLSys·im 0.3.1 updates the modeling engine, reference documentation, and calibration artifacts.
- Correct training sequence-length accounting, serving compute/memory limits, speculative-decoding memory and verification work, and optimizer-state calculations.
- Share resource accounting across public APIs, validate finite inputs and design-search constraints, and organize scenarios and result schemas by modeling domain.
- Expand modeling conventions, API guidance, executable examples, and calibration documentation. Refresh paper calculations and their supporting receipts.
Numerical compatibility: public call signatures remain compatible, but calculated values and invalid-input behavior can change. Serialized quantities now use canonical unit names and ASCII exponents. Review affected downstream calculations before updating prose, figures, labs, or benchmark tables. These analytical models do not establish a universal error bound against hardware measurements.
Documentation: https://mlsysbook.ai/mlsysim/
The documentation-site release is separate from PyPI package publication.
StaffML v0.1.4 — Publish StaffML updates and service hardening2026-11-06
StaffML 0.1.4 updates the question vault, taxonomy navigation, and service source.
- Correct hardware assumptions and worked reasoning in questions about mobile memory bandwidth, batch-one decoding weight reads, and edge-device rooflines.
- Update taxonomy labels and relationships, preserve subject/concept filtering in practice links, and revise vault authoring/schema checks.
- Replace provider response bodies and arbitrary exceptions in interviewer diagnostics with bounded error codes and HTTP statuses.
- Harden analytics origin checks, submitted-event validation, aggregate key handling, and summary authorization; document service and storage boundaries.
Site: https://mlsysbook.ai/staffml/
This release includes service-source changes; a site release does not by itself establish deployment of every Worker. Analytics filtering does not guarantee that submitted data contains no personal information.
Kits v0.1.1 — Publish Kits updates and exclude stale nested build output2026-11-06
Kits 0.1.1 updates the hardware-lab site and its publication workflow.
- Revise lab instructions, titles, and tables, and repair hardware-resource links.
- Improve phone layouts and dark-mode styling, including the hardware carousel, shared navigation, and code captions.
- Exclude stale nested build output from deployment and use the current repository paths for rendering and release manifests.
Site: https://mlsysbook.ai/kits/
vol2-v0.3.1: Publish SocratiQ and Volume II updates2026-11-06
Volume II 0.3.1 restores SocratiQ in the online book and updates the text, figures, and printed layout.
- Select a passage for an explanation, ask follow-up questions, and open quizzes or flashcards alongside the reading. A skippable first-use tour and optional learning preferences explain and adjust the workflow.
- The updated SocratiQ guide explains keyboard and touch controls, browser-local study records, and the passage and preferences sent for AI help. Reader conversations remain in page memory.
- Chapter revisions improve paragraph flow and connections between systems topics. The reference training job now uses a shared step definition; figure, equation, and appendix changes address presentation and printed-page fit.
PDF and EPUB editions are attached. AI explanations and quiz feedback can be mistaken; they are study aids, not evidence of mastery.
vol1-v0.8.1: Publish SocratiQ and Volume I updates2026-11-06
Volume I 0.8.1 restores SocratiQ in the online book and updates the text, figures, and printed layout.
- Select a passage for an explanation, ask follow-up questions, and open quizzes or flashcards alongside the reading. A skippable first-use tour and optional learning preferences explain and adjust the workflow.
- The updated SocratiQ guide explains keyboard and touch controls, browser-local study records, and the passage and preferences sent for AI help. Reader conversations remain in page memory.
- Chapter revisions clarify scope and narrative connections, qualify overstated claims, and move derived values into rerunnable calculations. Figure and equation changes address presentation and printed-page fit.
PDF and EPUB editions are attached. AI explanations and quiz feedback can be mistaken; they are study aids, not evidence of mastery.
MLSys·im 0.3.02026-10-25
MLSys·im 0.3.0
First-principles infrastructure modeling for the Machine Learning Systems textbook.
MLSys·im 0.3.0 reorganizes the package around nested registries and a dedicated
mlsysim.engine layer, adds Physical AI and agentic modeling (embodied
platforms, sensors, actuators, robotics and drivetrain physics), and corrects
several distributed-training, serving, and hardware-registry numbers. It
removes a number of top-level convenience imports, so code written against
0.1.1 needs the import updates listed below.
> PyPI note: 0.1.2 and 0.2.0 were tagged but never published to PyPI, so
> on PyPI 0.3.0 directly follows 0.1.1. If you are upgrading from the PyPI
> 0.1.1 release, the v0.1.2 entry below also applies to you.
Breaking Changes
Top-level imports. The following names are no longer importable from the
top-level mlsysim package. Import them from their defining module instead.
| 0.1.1 import | 0.3.0 import |
|---|---|
| from mlsysim import Workload, TransformerWorkload, CNNWorkload | from mlsysim.models.types import Workload, TransformerWorkload, CNNWorkload |
| from mlsysim import HardwareNode | from mlsysim.hardware.types import HardwareNode |
| from mlsysim import Fleet, Node, NetworkFabric | from mlsysim.systems.types import Fleet, Node, NetworkFabric |
| from mlsysim import DeploymentTier | from mlsysim.systems.types import DeploymentTier |
| from mlsysim import PerformanceProfile | from mlsysim.engine.engine import PerformanceProfile |
| from mlsysim import SystemEvaluator, SystemEvaluation | from mlsysim.engine.evaluation import SystemEvaluator, SystemEvaluation |
| from mlsysim import SimulationConfig, loadconfig | from mlsysim.engine.config import SimulationConfig, loadconfig |
| from mlsysim import SingleNodeModel (and every other solver or optimizer class) | from mlsysim.solvers import SingleNodeModel |
| from mlsysim import Infra | from mlsysim import Infrastructure (Infra.Grids, Infra.Racks become Infrastructure.Grids, Infrastructure.Racks; Infra.Quebec becomes Infrastructure.Grids.Quebec, likewise US_Avg, Poland) |
| from mlsysim import infra (subpackage) | from mlsysim import infrastructure |
| from mlsysim import Applications | from mlsysim import Scenarios (Applications.Doorbell becomes Scenarios.SmartDoorbell, AutoDrive becomes AutonomousVehicle, Workstation becomes LocalTraining, Frontier becomes FrontierTraining, Chatbot becomes ChatbotServing, KWS becomes KeywordSpotting) |
| from mlsysim import Archetypes | from mlsysim import Scenarios (same mapping as Applications; Archetypes.CloudV100 becomes Scenarios.FrontierTraining, Archetypes.TinyMLM7 becomes Scenarios.SmartDoorbell) |
| from mlsysim import show | import mlsysim.show or from mlsysim import show still works as an explicit submodule import; it is no longer imported eagerly |
| mlsysim.viz attribute access without import | import mlsysim.viz (top-level plotroofline and plotevaluation_scorecard are unchanged) |
| from mlsysim import ureg | Unchanged; ureg is still exported (now from mlsysim.core.units) |
| from mlsysim import Tiers | Removed, no replacement registry. The DeploymentTier type remains in mlsysim.systems.types. |
| from mlsysim import sim (Persona, BaseSimulation, SystemLedger, ...) | Removed, no replacement. |
| from mlsysim import constants / mlsysim.core.constants | Removed. Units and Q_ live in mlsysim.core.units; physical constants in mlsysim.physics.constants; specs in the registries. |
Moved modules. mlsysim.core.engine, mlsysim.core.solver,
mlsysim.core.evaluation, mlsysim.core.scenarios, and mlsysim.core.config
moved to mlsysim.engine.*. Use mlsysim.solvers as the stable import path
for solver classes. mlsysim.core.defaults is removed (see Internal).
Registry paths. Flat registry aliases are removed; use the nested paths.
| 0.1.1 | 0.3.0 |
|---|---|
| Hardware.H100, H200, B200, A100, V100, T4, MI250X, MI300X, Gaudi2, Gaudi3, TPUv5p, TPUv6, Trainium2 | Hardware.Cloud.<same name> |
| Hardware.NVL72 | Hardware.Cloud.GB200_NVL72 |
| Hardware.CerebrasCS3 | Hardware.Cloud.Cerebras_CS3 |
| Hardware.DGXSpark, Hardware.MacBook | Hardware.Workstation.DGX_Spark, Hardware.Workstation.MacBookM3Max |
| Hardware.Jetson | Hardware.Edge.JetsonOrinNX |
| Hardware.iPhone, Hardware.Snapdragon | Hardware.Mobile.iPhone15Pro, Hardware.Mobile.Snapdragon8Gen3 |
| Hardware.ESP32, Hardware.Himax | Hardware.Tiny.ESP32_S3, Hardware.Tiny.HimaxWE1 |
| Hardware.Networks. | Systems.Fabrics. |
| Models.GPT2, GPT3, GPT4, Llama270B, Llama38B, Llama3_70B | Models.Language.<same name> |
| Models.ResNet50, MobileNetV2, AlexNet | Models.Vision.<same name> |
| Models.WakeVision | Models.Tiny.WakeVision |
| Models.DLRM | Models.Recommendation.DLRM |
| Models.Mamba28B | Models.StateSpace.Mamba28B |
| Models.StableDiffusionv15 | Models.GenerativeVision.StableDiffusionv15 |
| Systems.Cloud, Edge, Mobile, Tiny, Systems.Tiers | Removed; use the Hardware.* entries directly |
Behavior changes.
- Hardware.TPUv4 and Hardware.Cloud.TPUv4 were aliases of TPU v5p in
0.1.1. Hardware.Cloud.TPUv4 is now a distinct TPU v4 record (275 TFLOP/s,
32 GiB), so results computed against it change. To reproduce 0.1.1 numbers,
use Hardware.Cloud.TPUv5p; the flat Hardware.TPUv4 alias is removed.
- Engine.solve raises ValueError for an unknown precision name instead of
silently using FP16 byte widths.
- Infrastructure.FacilityCooling.UptimeSurvey2022Average is removed; use
Infrastructure.FacilityCooling.Typical (now 1.55).
- ContinuousBatchingModel.solve() and ServingCapacityModel.solve() rename
seqlen to maxseq_len.
- ContinuousBatchingResult.memoryfragmentationpct is replaced by
pagedinternalfragmentation and staticinternalfragmentation
(fractions in [0, 1]). New fields: staticmaxactive_requests,
staticthroughputtokenspersec, statickvcache_size, and
meanrequesttokens.
- Metadata accepts only provenance (the source and source_url fields
are dropped); TraceableConstant is removed in favor of sourced().
Physical AI & Embodied Modeling
- **Embodied, Sensors, and Actuators registries (mlsysim.Embodied,
mlsysim.Sensors, mlsysim.Actuators).** Embodied covers
Vehicle.UberATG_VolvoXC90, Vehicle.Robotaxi, AMR.LogisticsAMR,
AMR.WarehouseAMR, Manipulator.Panda, Quadruped.Spot,
Humanoid.Atlas, Humanoid.UnitreeH1, and Drone.DJIMatrice, with mass,
velocity, braking, sensing latency, and power envelopes. Sensors covers
Camera, IMU, and LiDAR; Actuators covers FramelessMotor,
HarmonicDrive, HubMotor, QuasiDirectDrive, Servo, and the composed
PandaClassJoint. Provenance states what is sourced: the Panda record is a
datasheet; Spot, Atlas, Unitree H1, and the DJI Matrice 350 RTK are
estimates that link their spec pages and name the unsourced fields; the AMR
and robotaxi are estimated class profiles.
- Robotics physics (mlsysim.physics.robotics). Closed-form models for
stopping and safety envelopes (calcsafestopping_distance,
calcmaxpermittedvelocity, calcwatchdogleasebound,
calccbfqporthogonalprojection), contact and impact
(calckineticenergy, calctransientimpact_force,
calctripwirecontactforceaccumulation), robot buses
(calccanfdbusutilization, calcethercatcycletime), perception and
mapping budgets (calcteleopingestion_budget,
calctsdfvoxelgridbudget), validation statistics
(calcclopperpearsonzerofailurebound, calczerofailuresample_size),
and dynamics such as reflected inertia, seam torque, inverted-pendulum fall
tim
TinyTorch v0.2.1 - CUDA matrix-multiply extension, tito venv tool fixes, dark mode and phone layout for the guide2026-10-24
TinyTorch v0.2.1
CUDA matrix-multiply extension, tito venv tool fixes, dark mode and phone layout for the guide
<!-- Release notes generated using configuration in .github/release.yml at tinytorch-v0.2.1 -->
What's Changed
🐛 Bug Fixes
fix(tinytorch): fix venv jupytext resolution so tensor.py exports correctly by @Shashank-Tripathi-07 in https://github.com/harvard-edge/cs249rbook/pull/2152
fix(tinytorch): resolve venv tools by path instead of bare PATH lookup by @Shashank-Tripathi-07 in https://github.com/harvard-edge/cs249rbook/pull/2153
ci: bring main's workflows in line with dev before the October release by @profvjreddi in https://github.com/harvard-edge/cs249rbook/pull/2158
🔥 TinyTorch
TinyTorch: CUDA SGEMM extension (naive, coalesced, tiled) by @SatnamCodes in https://github.com/harvard-edge/cs249rbook/pull/2146
New Contributors
@SatnamCodes made their first contribution in https://github.com/harvard-edge/cs249rbook/pull/2146
Full Changelog: https://github.com/harvard-edge/cs249r_book/compare/tinytorch-v0.2.0...tinytorch-v0.2.1
---
Website: https://mlsysbook.ai/tinytorch/