transformers

Transformers — Model Loading, Inference, and Fine-tuning with Hugging Face

Skills Overview

Use the Hugging Face Transformers library to load pretrained models, run pipeline-based inference, perform text generation, and fine-tune with Trainer. Supports NLP, computer vision, audio, and multimodal tasks.

Use Cases

1. Quick Model Inference

When you need to quickly perform common AI tasks (text classification, question answering, named entity recognition, image classification, etc.), using the pipeline API lets you get optimized inference results directly without manual configuration. Suitable for prototype development, simple inference tasks, and scenarios that don’t require custom preprocessing.

2. Model Fine-tuning and Training

When you need to adapt a pretrained model to a specific domain or task, use the Trainer API for fine-tuning. It supports features such as automatic mixed-precision training, distributed training, and logging—ideal for improving model performance on specific tasks, domain adaptation, or academic research.

3. Advanced Model Management

When you need fine-grained control over model initialization, device placement, precision settings, or want to perform model inspection, manually loading AutoModel and the tokenizer provides full configuration control. Suitable for custom preprocessing pipelines, multi-GPU/TPU deployment, model debugging, or scenarios where you need a deep understanding of model inputs and outputs.

Core Features

Pipeline — Fast Inference

Provides inference interfaces optimized for many tasks, supporting 20+ task types including text generation, text classification, named entity recognition, question answering, summarization, translation, image classification, object detection, audio classification, and more. Model loading and inference can be done with a single line of code, without manually handling preprocessing and postprocessing logic.

Model Loading and Configuration

Use AutoModel* classes (AutoModelForCausalLM, AutoModelForSequenceClassification, etc.) and AutoTokenizer to load pretrained models from the Hugging Face Hub. Supports advanced configuration options such as automatic device mapping (device_map="auto"), precision control (float16, bfloat16), quantization, and more.

Trainer — Training and Fine-tuning

Offers a unified training interface, integrating TrainingArguments configuration, data preprocessing, training loop, evaluation, and logging. Supports training techniques such as single-node multi-GPU, distributed training, mixed-precision training, and gradient accumulation. You can also integrate logging tools like Weights & Biases and TensorBoard.

Frequently Asked Questions

What task types does Transformers support?

Transformers supports NLP (text classification, named entity recognition, question answering, summarization, translation), computer vision (image classification, object detection, image segmentation), audio (audio classification, speech recognition), and multimodal tasks (text-image retrieval, visual question answering). For the specific supported task types, refer to the references/pipelines.md documentation.

How do I handle authentication for gated or private models?

For gated or private models, you must first accept the model license on the Hugging Face Hub, then perform an interactive login via the CLI command hf auth login, or set the environment variable HF_TOKEN=... (recommended to read from a secrets manager and not hard-code in your code). After authentication, set trust_remote_code=True to load models that require custom code.

What important changes are there in Transformers 5.x?

Transformers 5.x supports only the PyTorch backend (TensorFlow and JAX support have been removed). New projects should work with transformers 5.x and huggingface_hub 1.x. When loading gated or custom-architecture models, you need to accept the license on the Hub and use trust_remote_code=True only after reviewing the custom code. Cache location can be configured with the HF_HOME or HF_HUB_CACHE environment variables.

Skill Boundaries

Note: This skill is only applicable to work using the Hugging Face Transformers library, including AutoModel loading, pipeline inference, tokenizers usage, and TrainingArguments training. It does not apply to general machine learning tasks outside the Transformers library.