Set up the environment

Tracks 0 and 1 built the understanding. Track 2 turns it into a working fine-tune you write yourself — no platform, nothing hidden. We'll use the standard open-source stack: PyTorch plus Hugging Face's libraries. This lesson gets that stack installed and verified so the next eight lessons just run.

What you need

• Python 3.10+ and pip.
• A GPU for the training lessons (a single consumer card is plenty for SmolLM2-135M). You can read along on CPU; training will just be slow.
• Disk space for the model (~0.3 GB) and a cache directory.

Create an isolated environment

Always work in a virtual environment (venv) so this project's packages don't collide with anything else on your machine.

# create and activate a venv
python -m venv .venv
source .venv/bin/activate         # Windows: .venv\Scripts\activate
python -m pip install --upgrade pip

Install the libraries

Each library has one job: torch is the tensor/autograd engine; transformers gives us models, tokenizers, and the Trainer; peft provides LoRA; trl adds SFT helpers; datasets handles data; accelerate manages device placement under the hood.

# install torch first, matching your CUDA (see pytorch.org for the exact command)
pip install torch

# then the Hugging Face stack
pip install transformers peft trl datasets accelerate

Verify your setup

Run this once. It confirms the imports work and tells you whether PyTorch can see a GPU.

import torch, transformers, peft
print("torch:", torch.__version__)
print("transformers:", transformers.__version__)
print("peft:", peft.__version__)
print("CUDA available:", torch.cuda.is_available())
if torch.cuda.is_available():
    print("GPU:", torch.cuda.get_device_name(0))

If CUDA available prints True and you see your GPU's name, you're ready. If it prints False but you have a GPU, your torch build is CPU-only — reinstall torch with the CUDA build from pytorch.org. With the environment working, the next lesson loads the model we'll fine-tune.