Python Quickstart

This is the fastest way to start using Hugging Face in real code.

Install the Main Packages

pip install -U transformers datasets tokenizers evaluate accelerate peft huggingface_hub

Optional extras depend on your workflow:

• torch for most PyTorch model usage
• tensorflow if using TensorFlow-backed models
• sentencepiece for some tokenizers
• bitsandbytes for quantization workflows

The Simplest Useful Example: Pipelines

Pipelines are high-level task wrappers.

from transformers import pipeline

classifier = pipeline("sentiment-analysis")
print(classifier("This library saves a huge amount of setup time."))

Why this is useful:

• Minimal code
• Good for fast validation
• Lets you test tasks before building custom code

Text Generation Example

from transformers import pipeline

generator = pipeline("text-generation", model="gpt2")
result = generator("Three practical uses for Hugging Face are", max_new_tokens=40)
print(result[0]["generated_text"])

Using Auto Classes

When you need more control, use AutoTokenizer and AutoModel... classes.

from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_id = "distilbert/distilbert-base-uncased-finetuned-sst-2-english"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

This pattern is common because it loads the right config automatically for the chosen repo.

Tokenize Input

inputs = tokenizer(
    ["I love practical libraries.", "This output is disappointing."],
    padding=True,
    truncation=True,
    return_tensors="pt",
)

Important options:

• padding=True for batching
• truncation=True to avoid overlong input errors
• return_tensors="pt" for PyTorch tensors

Run Inference

import torch

with torch.no_grad():
    outputs = model(**inputs)
    predictions = outputs.logits.argmax(dim=-1)

print(predictions.tolist())

Using the Hub in Code

from huggingface_hub import login

# Better: set HF_TOKEN in your environment and call login(token=...)
# login(token="...")

For scripts and servers, prefer environment variables over interactive login.

Caching

Downloaded models are cached locally.

This means:

• First load is slower
• Later loads are faster
• Disk usage grows over time

That is normal.

Device Placement

You may run on:

• CPU for simple testing
• GPU for larger or faster workloads
• Specialized accelerators in managed environments

Rule of thumb: Start on CPU with a small model. Move to GPU only when the workload justifies it.

Common Beginner Pattern

Use this sequence:

1. Start with a pipeline
2. Prove the task is viable
3. Switch to auto classes for more control
4. Add batching, metrics, and evaluation
5. Fine-tune only if zero-shot or prompting is not enough

Frequent Errors