Hugging Face Transformers

Hugging Face's open-source Python library for loading pretrained models with a few lines of code.

Transformers is Hugging Face's open-source library, bundling the architecture definitions and loading code for a huge number of pretrained models, spanning language models, vision models, and vision-language models (VLMs — models that can look at an image and describe it in words). Its core interfaces are AutoModel, AutoProcessor, and from_pretrained: give it a model name and it automatically downloads the weights, configuration, and tokenizer (the tool that splits text into tokens). In embodied AI, many VLA backbones — PaliGemma, Qwen-VL, Llama, and others — are loaded directly through it, and models like OpenVLA also publish their weights in a Transformers-compatible format. That makes it a foundational library that anyone reading code or fine-tuning models in this space has to know.

ExampleOpenVLA's official example loads the 7B weights with AutoModelForVision2Seq.from_pretrained, then calls predict_action to output robot arm actions.

Also called

transformers, 🤗 Transformers

Related

Hugging Face · Transformer · Vision-Language Model · OpenVLA · Hugging Face Diffusers · Hugging Face Mirror (hf-mirror.com)

Sources

huggingface/transformers GitHub
Transformers 文档 (Chinese)