Skip to content

MLSys·im 0.1.0

Choose a tag to compare

View all tags
@profvjreddi profvjreddi released this 24 Apr 17:57
· 7050 commits to dev since this release
mlsysim-v0.1.0
caa57a4

MLSys·im 0.1.0 — Initial Release

Release date: 2026-04-01

MLSys·im is a first-principles analytical engine for predicting performance, cost, and carbon footprint of ML systems. It is the computational companion to the Machine Learning Systems textbook.

This is the first public release. The API surface, the 22-wall taxonomy, and the solver portfolio are all considered stable for the 0.1.x line.

Install

pip install mlsysim

Verify the install:

python -c "import mlsysim; print(mlsysim.__version__)"
mlsysim eval Llama3_8B H100 --batch-size 32

Requires Python 3.10+. No GPU required — the engine computes from closed-form equations.

Five-line quickstart

import mlsysim
from mlsysim import Engine

profile = Engine.solve(
    model      = mlsysim.Models.ResNet50,
    hardware   = mlsysim.Hardware.Cloud.A100,
    batch_size = 1,
    precision  = "fp16",
)

print(f"Bottleneck: {profile.bottleneck}")              # → Memory
print(f"Latency:    {profile.latency.to('ms'):~.2f}")   # → 0.54 ms
print(f"Throughput: {profile.throughput:.0f}")          # → 1843 / second

Highlights

Core framework

  • 22-wall taxonomy organizing every constraint that bounds ML system performance, across six domains (Node, Data, Algorithm, Fleet, Ops, Analysis).
  • 26 analytical solvers (Model, Solver, Optimizer classes) covering all 22 walls.
  • Pint unit system with dimensional analysis enforced at runtime.
  • TraceableConstant pattern — every default carries a citation.
  • Pipeline composer for chaining solvers with explain() and run().
  • 3-tier evaluation scorecard: Feasibility → Performance → Macro/Economics.
  • Design Space Exploration (DSE) engine with declarative search and constraint evaluation.

Hardware Registry (15+ accelerators)

V100, A100, H100, H200, B200, GB200 NVL72, MI300X, TPUv5p, T4, Cerebras CS-3, Jetson Orin NX, ESP32-S3, nRF52840, Himax WE-I Plus, DGX Spark, MacBook M3 Max, iPhone 15 Pro, Pixel 8.

Full precision support: FP32, TF32, BF16, FP16, FP8, INT8, INT4. Multi-level memory hierarchy with HBM, SRAM, and Flash (TinyML). All specifications verified against manufacturer datasheets.

Model Registry

GPT-2/3/4, Llama-2/3 (7B/8B/70B), BERT Base/Large, ResNet-50, MobileNetV2, AlexNet, Mamba, Stable Diffusion v1.5, DS-CNN, WakeVision. HuggingFace importer included for arbitrary Transformer workloads.

Analytical models

SingleNodeModel, NetworkRooflineModel, EfficiencyModel, ForwardModel, ServingModel, ContinuousBatchingModel, WeightStreamingModel, TailLatencyModel, DataModel, TransformationModel, TopologyModel, ScalingModel, InferenceScalingModel, CompressionModel, DistributedModel, ReliabilityModel, OrchestrationModel, EconomicsModel, SustainabilityModel, CheckpointModel, ResponsibleEngineeringModel, SensitivitySolver, SynthesisSolver, ParallelismOptimizer, BatchingOptimizer, PlacementOptimizer.

CLI

mlsysim eval        Evaluate the analytical physics of an ML system (YAML or CLI flags)
mlsysim serve       Evaluate LLM serving (prefill + decode)
mlsysim optimize    Search the design space for optimal configurations
mlsysim zoo         Explore the built-in registries
mlsysim audit       Profile your local hardware against the Iron Law
mlsysim schema      Export the JSON Schema for the mlsys.yaml configuration file

Testing

367 tests, 100% pass rate. Coverage includes formula unit tests with known answers, full solver suite, physics-bound validation across all registry hardware, wall-taxonomy completeness, pipeline composition, and three optimization backends (exhaustive, OR-tools, scipy).

Documentation

Known limitations & gotchas

  • First-order analytical model. Predictions are typically within 15–30% of measured throughput on well-optimized workloads. Use MLSys·im to compare options and identify bottlenecks; validate with empirical benchmarks before committing to a production SLA. See accuracy.qmd for full validation against MLPerf v4.0.
  • Slide PDFs. Many tutorials cross-link to lecture slides at github.com/harvard-edge/cs249r_book/releases/download/slides-latest/*.pdf. The slides-latest release tag is not yet published; these links will resolve once the slides ship.
  • Hosted notebook launchers. Google Colab and Binder buttons are planned but not wired up for 0.1.0. Tutorials run locally on any Python 3.10+ environment.

Project links

Assets 4
Loading