Hugging Face Transformers

Simple and Optimized Inference

Transformers provides a Pipeline class for simple and optimized inference across many machine learning tasks, including text generation, image segmentation, automatic speech recognition, and document question answering.

Quickly run inference on a wide range of tasks with minimal code.

Comprehensive Training Support

The Trainer class supports mixed precision, torch.compile, FlashAttention, and distributed training for PyTorch models, enabling efficient training and fine-tuning.

Train and fine-tune models with advanced features and distributed scaling.

Fast Text Generation

The generate method provides fast text generation for large language models (LLMs) and vision language models (VLMs), with support for streaming and multiple decoding strategies.

Generate high-quality text and multimodal outputs with flexible decoding options.

Over 1M+ Pretrained Checkpoints

There are over 1 million Transformers model checkpoints available on the Hugging Face Hub, covering text, vision, audio, video, and multimodal models.

Access a vast library of pretrained models to reduce compute cost and time.

Model-Definition Standard

Transformers centralizes model definitions so that a supported model is compatible with major training frameworks (Axolotl, Unsloth, DeepSpeed, FSDP, PyTorch-Lightning), inference engines (vLLM, SGLang, TGI), and adjacent libraries (llama.cpp, mlx).

Use the same model definition across the entire AI ecosystem.

Fast and Easy to Use

Every model is implemented from only three main classes (configuration, model, and preprocessor) and can be quickly used for inference or training with Pipeline or Trainer.

Get started quickly with a simple, consistent API across all models.

• Central model-definition frameworkTransformers acts as the model-definition framework for state-of-the-art machine learning models in text, computer vision, audio, video, and multimodal models, for both inference and training. It centralizes the model definition so that this definition is agreed upon across the ecosystem.
• Broad ecosystem compatibilityIf a model definition is supported, it will be compatible with the majority of training frameworks (Axolotl, Unsloth, DeepSpeed, FSDP, PyTorch-Lightning), inference engines (vLLM, SGLang, TGI), and adjacent modeling libraries (llama.cpp, mlx) which leverage the model definition from transformers.
• Comprehensive features for inference and trainingTransformers provides everything you need for inference or training with state-of-the-art pretrained models, including Pipeline for simple and optimized inference, Trainer with mixed precision, torch.compile, FlashAttention, and distributed training, and generate for fast text generation with LLMs and VLMs, including streaming and multiple decoding strategies.
• Simple and efficient designEvery model is implemented from only three main classes (configuration, model, and preprocessor) and can be quickly used for inference or training with Pipeline or Trainer. Pretrained models reduce carbon footprint, compute cost, and time.

• Scope of compatibilityThe documentation states compatibility with the 'majority' of training frameworks and inference engines, but does not list all possible integrations. Do not assume compatibility with frameworks not explicitly mentioned (e.g., Axolotl, Unsloth, DeepSpeed, FSDP, PyTorch-Lightning, vLLM, SGLang, TGI, llama.cpp, mlx).
• Model checkpoint countThe page mentions 'over 1M+ Transformers model checkpoints on the Hugging Face Hub' but this is a general claim and may change over time. Do not treat this as a precise or static number.
• Course recommendationThe documentation recommends the LLM course for learning, but does not provide details about its content or cost. Do not infer that the course is free or that it covers specific topics beyond what is stated.
• Design principlesThe design principles are described as 'Fast and easy to use' and 'Pretrained models', but the page does not provide benchmarks or quantitative evidence for these claims. Do not overstate performance or efficiency without additional data.

Inference with Pipelines

Use Transformers' Pipeline class for simple, optimized inference across tasks like text generation, image segmentation, automatic speech recognition, and document question answering. It abstracts the complexity so you can quickly run state-of-the-art models on your data.

Training with Trainer

Leverage the comprehensive Trainer to fine-tune pretrained models with mixed precision, torch.compile, and FlashAttention. It supports distributed training for PyTorch models, helping you reduce compute cost and time while achieving state-of-the-art performance.

Fast Text Generation

Use the generate method for fast text generation with large language models (LLMs) and vision language models (VLMs). It includes support for streaming and multiple decoding strategies, making it ideal for chat, summarization, and creative writing applications.

Multimodal Model Support

Transformers acts as the model-definition framework for state-of-the-art models in text, computer vision, audio, video, and multimodal domains. If a model definition is supported, it is compatible with major training frameworks, inference engines, and adjacent libraries, enabling broad ecosystem integration.

Define the exact aI Open-source Tools workflow Hugging Face Transformers should support.

Compare it with closely related AI tools in the same category before committing.

Set review rules for accuracy, privacy, brand voice, compliance, and final approval.

Connect useful outputs to the wider stack instead of leaving them inside the AI tool.