Tren⚡Torch

Most people learn ML frameworks by importing them. This one is built by hand.

TrenTorch is a from-scratch machine learning framework, built one module at a time on nothing but NumPy: tensors and autograd, layers and losses, a dataloader, optimizers, a full training loop, CNNs, tokenization, embeddings, attention, transformers, and then the systems layer on top: profiling, quantization, compression, acceleration, memoization, benchmarking. No import torch. Every line is yours.

This is a personal, from-scratch implementation of TinyTorch (Harvard CS249r), rebuilt independently and pushed further, with cleaner internals, a much faster CI pipeline, and an implementation extended past the original spec wherever there was room to.

Where to go

Page	What's there
Getting Started	Clone, install, and run your first module
Curriculum & Milestones	The 20 modules and the 6 historical milestones you unlock along the way
CLI Reference	Every `tren` command, condensed
Architecture	How a student's edit becomes a real, importable symbol, the short version
Contributing	How to set up a dev environment and submit changes

The shape of it

You don't import a framework here, you build one, in order:

Foundations (modules 01-08): tensors, activations, layers, losses, a dataloader, autograd, optimizers, a real training loop
Vision (module 09): Conv2d, pooling, CNNs
Language (modules 10-13): tokenization, embeddings, attention, transformer blocks
Optimization (modules 14-20): profiling, quantization, compression, acceleration, memoization, benchmarking, and a capstone that ties it together

Along the way, six milestones let you re-run landmark moments in ML history: the 1958 perceptron, the 1969 XOR crisis, 1986 backpropagation, the 1998 CNN revolution, the 2017 transformer, and an MLPerf-style optimization round, on the framework you actually wrote.

Repository layout (current)

TrenTorch/
├── data/
│   ├── src/                # Curriculum source: 20_capstone/, 09_convolutions/, etc., edit here
│   ├── modules/             # Generated stub notebooks (what a student opens and solves)
│   ├── solutions/           # Reference implementations (maintainer/CI-only, gitignored)
│   ├── milestones/          # The six historical milestone scripts
│   └── trentorch/           # The real, importable package, built from your solved modules
├── platforms/
│   └── cli/                 # The `tren` CLI itself (setup, module workflow, dev/CI tooling)
├── tests/                   # Integration, e2e, and environment tests
├── docs/                    # Deep technical docs: design, system architecture, CLI reference
└── bin/tren                 # Entry point

Everything under data/src/ is what you actually write; everything else is generated from it or drives the workflow around it.

Status

All 20 modules implemented, with unit, integration, CLI, and milestone tests
The tren CLI drives the entire student and maintainer workflow
CI runs the full curriculum, every stage, on every push; see Architecture for how that pipeline works