TinyTorch/docs/student-workflow.md

Student Workflow

This guide explains the actual day-to-day workflow for building your ML framework with TinyTorch.

The Core Workflow

TinyTorch follows a simple three-step cycle:

graph LR
    A[Work in Notebooks<br/>modules/NN_name.ipynb] --> B[Export to Package<br/>tito module complete N]
    B --> C[Validate with Milestones<br/>Run milestone scripts]
    C --> A

    style A fill:#e3f2fd
    style B fill:#f0fdf4
    style C fill:#fef3c7

Step 1: Edit Modules

Work on module notebooks in modules/:

# Example: Working on Module 03 (Layers)
cd modules/03_layers
jupyter lab 03_layers.ipynb

Each module is a Jupyter notebook that you edit interactively. You'll:

• Implement the required functionality
• Add docstrings and comments
• Run and test your code inline
• See immediate feedback

Step 2: Export to Package

Once your module implementation is complete, export it to the main TinyTorch package:

tito module complete MODULE_NUMBER

This command:

• Converts your source files to the tinytorch/ package
• Validates NBGrader metadata
• Makes your implementation available for import

Example:

tito module complete 03  # Export Module 03 (Layers)

After export, your code is importable:

from tinytorch.layers import Linear  # YOUR implementation!

Step 3: Validate with Milestones

Run milestone scripts to prove your implementation works:

cd milestones/01_1957_perceptron
python 01_rosenblatt_forward.py  # Uses YOUR Tensor (M01)
python 02_rosenblatt_trained.py  # Uses YOUR layers (M01-M07)

Each milestone has a README explaining:

• Required modules
• Historical context
• Expected results
• What you're learning

See Milestones Guide for the full progression.

Module Progression

TinyTorch has 20 modules organized in three tiers:

Foundation (Modules 01-07)

Core ML infrastructure - tensors, autograd, training loops

Milestones unlocked:

• M01: Perceptron (after Module 07)
• M02: XOR Crisis (after Module 07)

Architecture (Modules 08-13)

Neural network architectures - data loading, CNNs, transformers

Milestones unlocked:

• M03: MLPs (after Module 08)
• M04: CNNs (after Module 09)
• M05: Transformers (after Module 13)

Optimization (Modules 14-19)

Production optimization - profiling, quantization, benchmarking

Milestones unlocked:

• M06: Torch Olympics (after Module 18)

Capstone Competition (Module 20)

Apply all optimizations in the Torch Olympics Competition

Typical Development Session

Here's what a typical session looks like:

# 1. Work on a module
cd modules/05_autograd
jupyter lab autograd_dev.ipynb
# Edit your implementation interactively

# 2. Export when ready
tito module complete 05

# 3. Validate with existing milestones
cd ../milestones/01_1957_perceptron
python 01_rosenblatt_forward.py  # Should still work!

# 4. Continue to next module or milestone

TITO Commands Reference

The most important commands you'll use:

# Export module to package
tito module complete MODULE_NUMBER

# Check module status (optional capability tracking)
tito checkpoint status

# System information
tito system info

# Join community and benchmark
tito community join
tito benchmark baseline

For complete command documentation, see TITO CLI Reference.

Checkpoint System (Optional)

TinyTorch includes an optional checkpoint system for tracking progress:

tito checkpoint status  # View completion tracking

This is helpful for self-assessment but not required for the core workflow. The essential cycle remains: edit → export → validate.

Notebook Platform Options

TinyTorch notebooks work with multiple platforms, but important distinction:

Online Notebooks (Viewing & Exploration)

• Jupyter/MyBinder: Click "Launch Binder" on any notebook page - great for viewing
• Google Colab: Click "Launch Colab" for GPU access - good for exploration
• Marimo: Click "🍃 Open in Marimo" for reactive notebooks - excellent for learning

⚠️ Important: Online notebooks are for viewing and learning. They don't have the full TinyTorch package installed, so you can't:

• Run milestone validation scripts
• Import from tinytorch.* modules
• Execute full experiments
• Use the complete CLI tools

Local Setup (Required for Full Package)

To actually build and experiment, you need a local installation:

# Clone and setup locally
git clone https://github.com/mlsysbook/TinyTorch.git
cd TinyTorch
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
pip install -e .  # Install TinyTorch package

Why local?

• ✅ Full tinytorch.* package available
• ✅ Run milestone validation scripts
• ✅ Use tito CLI commands
• ✅ Execute complete experiments
• ✅ Export modules to package
• ✅ Full development workflow

Note for NBGrader assignments: Submit .ipynb files (not Marimo's .py format) to preserve grading metadata.

Community & Benchmarking

Join the Community

After completing setup, join the global TinyTorch community:

# Join with optional information
tito community join

# View your profile and progress
tito community profile

# Update your information
tito community update

Privacy: All information is optional. Data is stored locally in .tinytorch/ directory. See Community Guide for details.

Benchmark Your Progress

Validate your setup and track performance:

# Quick baseline benchmark (after setup)
tito benchmark baseline

# Full capstone benchmarks (after Module 20)
tito benchmark capstone --track all

Baseline Benchmark: Quick validation that your setup works correctly - your "Hello World" moment!

Capstone Benchmark: Full performance evaluation across speed, compression, accuracy, and efficiency tracks.

See Community Guide for complete community and benchmarking features.

Instructor Integration

TinyTorch supports NBGrader for classroom use. See the Instructor Guide for complete setup and grading workflows.

For now, focus on the student workflow: building your implementations and validating them with milestones.

What's Next?

1. Start with Module 01: See Getting Started
2. Follow the progression: Each module builds on previous ones
3. Run milestones: Prove your implementations work
4. Build intuition: Understand ML systems from first principles

The goal isn't just to write code - it's to understand how modern ML frameworks work by building one yourself.