I debuted on PyPI with Harvard

Hello everyone,

I'm happy to share I've published my first python package as a contributor to the CS249r project led by the Harvard-Edge Lab. Gonna code more :)

Check the package here: https://pypi.org/project/mlsysim/0.1.1/

Just pip install mlsysim

Read the whole story here:

ML systems today are designed largely on intuition. Senior engineers do back-of-envelope math to figure out whether a workload fits on a GPU, what the memory bandwidth bottleneck will be, whether a 64-node H100 cluster can hit a 50ms latency SLA, what the total cost of ownership looks like over 14 days in Quebec. mlsysim formalises that math.

It's a first-principles analytical framework for ML infrastructure. Five layers, from workload representation down to execution and design space search. You describe your cluster in a declarative YAML file, define your constraints, and the engine returns a three-lens scorecard: Feasibility, Performance, and Macro economics. If a constraint is violated, the CLI exits with a semantic error code. It's built to run in CI/CD pipelines and talk to AI agents, not just humans.

The accuracy sits within 2 to 5x of measured performance for well-characterised workloads. That's not a gap, that's the point. This is the math you do before you benchmark, not after.

What I contributed sits inside this framework and inside TinyTorch, the from-scratch PyTorch reimplementation that CS249r students build through 21 modules. The Tensor API additions I shipped in PR #1392, view(), masked_fill(), contiguous(), ndim, numel(), are now part of a package that Harvard uses to teach machine learning systems to the next generation of engineers.

Thanks for your attention to this matter.
Rocky