Hugging Face

Hugging Face is the GitHub of AI. It hosts 1M+ models. 2025 sees massive growth in Multimodal models and Robotics (LeRobot).

When to Use

• Model Discovery: Finding the SOTA open-source model for any task.
• Inference: transformers library is the standard way to run models in Python.
• Datasets: Accessing standard datasets (load_dataset('squad')).

Core Concepts

Transformers Library

The API to download and run models. pipeline('sentiment-analysis').

Hugging Face Hub (Hugging Face CLI)

Versioning, git-based storage for large model weights (git lfs).

Spaces

Hosting simple Gradio/Streamlit apps for model demos.

Best Practices (2025)

Do:

• Use bitsandbytes: Load 70B models in 4-bit precision easily.
• Use accelerate: For multi-GPU training/inference distributed across devices.
• Push to Hub: Share your fine-tunes.

Don't:

• Don't hardcode paths: Use from_pretrained("repo/id") to auto-cache models.

References

• Hugging Face Documentation