Transformers Cheatsheet

Q: How do I install Transformers?

Run `pip install transformers`. For deep learning backends: `pip install transformers[torch]` for PyTorch, `pip install transformers[tf-cpu]` for TensorFlow. The library auto-detects which backend is installed. For GPU support, install PyTorch or TensorFlow with CUDA separately.

Q: What is the pipeline API?

Pipeline is the easiest way to use Transformers. `from transformers import pipeline; pipe = pipeline('sentiment-analysis'); pipe('I love this!')` returns a list of dicts with label and score. Available tasks include: text-generation, summarization, translation, sentiment-analysis, image-classification, object-detection, question-answering, and more.

Q: How do I load a specific pretrained model?

Use `AutoModel.from_pretrained('model-name')` and `AutoTokenizer.from_pretrained('model-name')`. The model name can be any HuggingFace Hub ID like `bert-base-uncased`, `gpt2`, `t5-small`, `meta-llama/Llama-3.1-8b`, or a local path.

Q: How do I manage the model cache?

Models are cached at `~/.cache/huggingface/hub/` by default. Set `export TRANSFORMERS_CACHE=/path/to/cache` to change it. Use `huggingface-cli scan-cache` and `delete-cache` to manage disk space. A single large model like Llama-3.1-8B takes ~15GB.

Transformers is the core library of the HuggingFace ecosystem, providing a unified interface to thousands of pretrained models. Supports PyTorch, TensorFlow, and JAX backends. The `pipeline` API enables one-liner inference for text classification, generation, summarization, translation, QA, image classification, and more.

Updated: 2026-07-20·21 commands

Quick Start

``bash # Install pip install transformers

# Check environment transformers-cli env

# Download a model (without loading) transformers-cli download bert-base-uncased

# One-liner sentiment analysis python -c "from transformers import pipeline; print(pipeline('sentiment-analysis')('I love Transformers!'))"

# Text generation with GPT-2 python -c "from transformers import pipeline; print(pipeline('text-generation', model='gpt2')('Once upon a time')[0]['generated_text'])"`

`Pipeline API`

The pipeline function is the easiest entry point. It handles model loading, tokenization, and output decoding automatically. Supported tasks: sentiment-analysis, text-generation, summarization, translation_xx_to_yy, fill-mask, question-answering, image-classification, image-segmentation, object-detection, zero-shot-classification, and more.

`Auto Classes`

AutoModel, AutoTokenizer, AutoModelForCausalLM, AutoModelForSeq2SeqLM, and other Auto classes automatically select the correct model architecture based on the model name. This is the recommended way to load models programmatically.

`Model Loading`

Models are identified by their HuggingFace Hub ID. Popular models include: BERT (bert-base-uncased), GPT-2 (gpt2), T5 (t5-small), BART (facebook/bart-large), Llama (meta-llama/Llama-3.1-8b), Mistral (mistralai/Mistral-7B-v0.1), and thousands more.

`Environment`

TRANSFORMERS_CACHE controls the model cache directory. HF_HOME is the root HuggingFace directory. The transformers-cli env` command prints all relevant environment info including installed backends, CUDA version, and cache locations.

Startup & Modes(15)

Command	Level	Example
`transformers-cli --help` Show transformers-cli help	Basic	`transformers-cli --help`
`transformers-cli env` Show Transformers environment info	Basic	`transformers-cli env`
`pip install transformers` Install Transformers library	Basic	`pip install transformers`
`pip install transformers[torch]` Install Transformers (with PyTorch)	Basic	`pip install transformers[torch]`
`pip install transformers[tf-cpu]` Install Transformers (with TensorFlow CPU)	Basic	`pip install transformers[tf-cpu]`
`pipeline text-generation` Use pipeline API for text generation	Intermediate	`python -c 'from transformers import pipeline; pipe = pipeline("text-generation", model="gpt2"); print(pipe("Hello")[0]["generated_text"])'`
`module env` Run transformers-cli env as a module	Intermediate	`python -m transformers.commands.run transformers-cli env`
`pipeline sentiment-analysis` Sentiment analysis pipeline	Basic	`python -c 'from transformers import pipeline; print(pipeline("sentiment-analysis")("I love this!"))'`
`pipeline image-classification` Image classification pipeline	Intermediate	`python -c 'from transformers import pipeline; print(pipeline("image-classification")("photo.jpg"))'`
`pipeline summarization` Text summarization pipeline	Intermediate	`python -c 'from transformers import pipeline; print(pipeline("summarization")("Long article text here..."))'`
`pipeline translation` Translation pipeline (English to French)	Intermediate	`python -c 'from transformers import pipeline; print(pipeline("translation_en_to_fr")("Hello"))'`
`export TRANSFORMERS_CACHE` Set Transformers cache directory	Basic	`export TRANSFORMERS_CACHE=/data/hf-cache`
`export HF_HOME` Set HuggingFace home directory	Basic	`export HF_HOME=/data/hf-home`
`pipeline fill-mask` Mask filling pipeline (BERT type)	Intermediate	`python -c 'from transformers import pipeline; print(pipeline("fill-mask")("I love [MASK]!"))'`
`pipeline question-answering` Question answering pipeline	Intermediate	`python -c 'from transformers import pipeline; qa = pipeline("question-answering"); print(qa(question="What is AI?", context="AI is the simulation of intelligence."))'`

ai-model(6)

Command	Level	Example
`transformers-cli download` Download pretrained model	Intermediate	`transformers-cli download bert-base-uncased`
`transformers-cli download --cache-dir` Specify download cache directory	Intermediate	`transformers-cli download bert-base-uncased --cache-dir /data/models`
`AutoModel from_pretrained` Load pretrained model with AutoModel	Intermediate	`python -c 'from transformers import AutoModel; AutoModel.from_pretrained("bert-base-uncased")'`
`AutoTokenizer from_pretrained` Load matching tokenizer for a model	Intermediate	`python -c 'from transformers import AutoTokenizer; tok = AutoTokenizer.from_pretrained("bert-base-uncased"); print(tok("Hello world"))'`
`AutoModelForCausalLM` Load causal language model (GPT/Llama type)	Intermediate	`python -c 'from transformers import AutoModelForCausalLM; AutoModelForCausalLM.from_pretrained("gpt2")'`
`AutoModelForSeq2SeqLM` Load sequence-to-sequence model (T5/BART type)	Intermediate	`python -c 'from transformers import AutoModelForSeq2SeqLM; AutoModelForSeq2SeqLM.from_pretrained("t5-small")'`

FAQ

How do I install Transformers?

What is the pipeline API?

How do I load a specific pretrained model?

How do I manage the model cache?

This cheatsheet is compiled from official tool documentation. Last updated: 2026-07-20.