Setting Up Your Environment

What you install and where files go

transformers is a Python library that needs a deep-learning backend to run models; for current releases that means PyTorch. Install both in a virtual environment so project versions do not clash. The first time you call from_pretrained with a model id, the library downloads the config, tokenizer files and weights from the Hub and stores them in a local cache, by default under ~/.cache/huggingface/hub. Later calls reuse the cache, so only the first run is slow. You can move the cache with the HF_HOME environment variable, which is useful on college lab machines with a small home directory. Public models download without an account, but gated models (whose authors ask you to accept a licence) and private repositories need an access token, which you create in your Hub settings and supply with the hf auth login command or the HF_TOKEN environment variable.

Install into a virtual environment

Create an isolated environment, then install the library with PyTorch and the companion packages used in this course.

python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install torch transformers datasets accelerate evaluate peft
# Optional: move the download cache to a bigger disk
export HF_HOME=/data/hf-cache

Check the install and log in

Print the installed versions and whether a GPU is visible. hf auth login stores your token locally; on older huggingface_hub versions the command is huggingface-cli login.

python -c "import transformers, torch; print(transformers.__version__, torch.cuda.is_available())"
hf auth login        # paste a token from your Hub settings
hf auth whoami       # confirms which account is active

Never hard-code your token

Do not paste a Hub token into a notebook you will share or a script you will push to GitHub. Read it from the HF_TOKEN environment variable or a secrets manager, and give it read scope unless you really need to upload.

त्वरित जाँच: You run the same `from_pretrained("some-model")` line twice on one machine. Why is the second run much faster?

• The second run downloads a smaller version of the weights
• transformers compiles the model to C++ after the first run
• The Hub caches the model in your browser
• The files are reused from the local Hugging Face cache