推論、デプロイ、ランタイム

Hugging Face Transformers

huggingface/transformers

A model-definition framework for state-of-the-art machine learning models across text, vision, audio, and multimodal domains, supporting both inference and training.

★ 163.3Kスター
⑂ 34.1Kフォーク数
2343未解決の問題
Python言語
Apache-2.0ライセンス
Q98編集部スコア

プロジェクトのスクリーンショット

Hugging Face Transformersのスクリーンショット Hugging Face Transformersのスクリーンショット

概要

Transformers serves as a centralized model-definition framework that ensures compatibility across the broader machine learning ecosystem. By standardizing model definitions, it allows seamless integration with various training frameworks like Axolotl, Unsloth, DeepSpeed, and PyTorch-Lightning, as well as inference engines such as vLLM, SGLang, and TGI. The library provides a unified API with minimal abstractions, making it accessible for researchers and developers to leverage over 1 million pretrained model checkpoints available on the Hugging Face Hub.

主な機能

  • Unified API for pretrained models across text, vision, audio, and multimodal tasks
  • Compatibility with major training frameworks and inference engines
  • Access to over 1 million model checkpoints on the Hugging Face Hub
  • High-level Pipeline API for easy inference across modalities
  • Support for model customization and exposure of model internals

要件、インストール、クイックスタート

Create and activate a virtual environment using venv or uv. Install Transformers with PyTorch support via pip using 'pip install "transformers[torch]"' or via uv using 'uv pip install "transformers[torch]"'. For the latest changes, install from source by cloning the repository and running 'pip install ".[torch]"'.

使用方法

Instantiate a pipeline by specifying a task and model, then pass input data to get results. For example, use 'pipeline(task="text-generation", model="Qwen/Qwen2.5-1.5B")' for text generation. For chat-based interactions, construct a chat history and pass it to the pipeline.

モデルの互換性とユースケース

Supports a wide range of models including Qwen, Llama, Whisper, DINOv2, BLIP, Mixtral, BART, T5, Gemma, and many others across various modalities.

ライセンスとリスクに関する注意事項

Licensed under the Apache License, Version 2.0.

Editorial verification: repository URL, owner, description, license and repository statistics were reviewed. License metadata: Apache-2.0. README was fetched for the channel draft; re-check repository dependencies, releases and model terms before production use.

リリースとメンテナンス

Not stated in the repository metadata.

vllm

vllm-project/vllm

★ 89.5KPython

sglang

sgl-project/sglang

★ 32.2KPython

OpenVINO

openvinotoolkit/openvino

★ 10.6KC++