Before you start

This chapter prepares the local environment you will use from chapter 1 onward — Python and a virtual environment right away, PyTorch a bit later. Every chapter has a browser half (JavaScript, nothing to install) and a local half that writes real files into your own my-llm/ project. The local half starts immediately: chapter 1 already has you running python -m scripts.train_bigram.

Skip ahead if python3 -m venv .venv && source .venv/bin/activate && pip install torch is muscle memory. You will not learn anything new on this page.

In a hurry? Steps 1 to 3 (Python + venv, ~5 minutes) are all you need for chapters 1 to 10 — they run on the standard library alone. Step 4 installs PyTorch, which only becomes necessary at chapter 12. Doing it now costs you one coffee and saves you an interruption later.

The whole setup takes 10–20 minutes, mostly waiting for the PyTorch download.

1. Install Python 3.11 or newer

The book uses type-hint features that need Python 3.11+. 3.12 or 3.13 is recommended.

brew install python@3.13

# Install from https://www.python.org/downloads/ and check 'Add Python to PATH' during install.

sudo apt update && sudo apt install python3.13 python3.13-venv python3-pip

Check that the install worked:

python3 --version

py --version

python3 --version

You should see Python 3.13.x (or 3.11.x / 3.12.x — anything from 3.11 up is fine).

2. Create the project folder

The local artifact you build across the book lives in my-llm/. Chapter 1 repeats these exact commands, so creating it now or there makes no difference — just don't do it twice.

mkdir -p my-llm/llm my-llm/scripts my-llm/data && cd my-llm

mkdir my-llm; mkdir my-llm\\llm,my-llm\\scripts,my-llm\\data; cd my-llm

mkdir -p my-llm/llm my-llm/scripts my-llm/data && cd my-llm

3. Create and activate the virtual environment

A venv isolates this project's package versions from your system Python. A clean rm -rf .venv (or Remove-Item .venv) is enough to undo every install.

python3 -m venv .venv && source .venv/bin/activate

py -m venv .venv; .\\.venv\\Scripts\\Activate.ps1

python3 -m venv .venv && source .venv/bin/activate

After activation your shell prompt should show (.venv) somewhere. From now on, every python and pip command in this book assumes the venv is active. If you open a new terminal, you have to activate again.

4. Install PyTorch

Chapters 12–21 need PyTorch; chapters 1–10 do not (they use the standard library only). Install the CPU build now anyway — it is around 200 MB, and you will not want to wait for it mid-chapter:

pip install torch

pip install torch

pip install torch

This is the longest single download in the book. Have a coffee.

If you have a recent NVIDIA GPU and want CUDA, see the PyTorch install selector. The CPU build is enough for every chapter; CUDA / MPS just make training faster.

5. Verify hardware acceleration

Run a quick check from the activated venv:

python -c "import torch; print(torch.__version__); print(f\"mps={torch.backends.mps.is_available()} cuda={torch.cuda.is_available()}\")"

python -c "import torch; print(torch.__version__); print(f\"mps={torch.backends.mps.is_available()} cuda={torch.cuda.is_available()}\")"

python -c "import torch; print(torch.__version__); print(f\"mps={torch.backends.mps.is_available()} cuda={torch.cuda.is_available()}\")"

Expected result depends on your machine:

• Apple Silicon Mac (M1 through M5): mps=True cuda=False. The book uses MPS — 5-20× faster than CPU on matrix multiplies.
• NVIDIA GPU on Linux/Windows with the CUDA PyTorch build: mps=False cuda=True. Fastest option.
• Anything else: mps=False cuda=False. CPU. Every chapter still works; training in chapter 13 takes ~10 minutes instead of ~2.

6. Common errors and what to do

7. Time budget for the rest of the book

Rough wall-clock estimate per part, assuming you also run and inspect the code:

• Part I — Start the project (chapters 1–4): ~1.5 hours. Browser cells plus plain-Python files in my-llm/ — no PyTorch yet.
• Part II — Make it learn (5–7): ~1.5 hours. Same mix.
• Part III — Build the transformer (8–10): ~1.5 hours. Same mix.
• Part IV — Train and use the LLM (11–16): ~4 hours. PyTorch kicks in, and the centre of gravity shifts to your machine. The training run in chapter 13 is ~10 minutes on CPU, ~2 on MPS/CUDA. Chapter 15 downloads ~500 MB of GPT-2 weights.
• Part V — Make it useful, cheaper, and usable (17–21): ~4 hours. The SFT run in chapter 17 is ~2 minutes; the chat REPL in chapter 20 is interactive; the chapter 21 capstone fine-tune is ~20 minutes on CPU.
• Part VI — Appendices (optional): ~30 minutes if you read the backprop derivation and the RLHF/DPO conceptual walk-through. Skippable.

Total: ~13 hours including waits. You can do it in one weekend, or one chapter per evening for two weeks.

Going further

• Python's official venv tutorial — the canonical reference if you want to understand virtual environments more deeply.
• PyTorch install selector — pick CPU / MPS / CUDA / specific Python version and get the right pip command.

Next up: chapter 1 — the dumbest model that exists. Start building.