gary149/huggingface-hub-skill

Overview

This skill provides expert guidance for developing with the Hugging Face Hub ecosystem, covering HTTP/curl APIs, Python and JavaScript SDKs, Spaces apps (Gradio, Docker), model formats, datasets, automation, and security. It helps you upload/download repositories, run inference, configure Spaces, and automate workflows with practical examples and patterns. Use it when integrating Hub features into apps, CI/CD, or production pipelines.

How this skill works

The skill explains concrete API calls (curl and HTTP), SDK usage (huggingface_hub in Python and @huggingface/hub in JavaScript), and inference via InferenceClient. It shows repository operations (hf_hub_download, upload_file, snapshot_download), Space configuration (YAML and app files), dataset access (pandas, Polars, DuckDB), and formats like GGUF and PEFT/LoRA. It also covers automation (webhooks, Jobs), security (tokens, OAuth, scoping), and deployment considerations for reproducibility and resource selection.

When to use it

• You need to upload or download models, weights, or datasets programmatically.
• You want to call Hugging Face inference endpoints from server or curl.
• You are building or configuring a Space (Gradio or Docker) for deployment.
• You must load large datasets efficiently with Parquet via pandas/Polars/DuckDB.
• You plan LoRA/PEFT workflows, GGUF conversion, or model quantization.
• You need automation: webhooks, scheduled Jobs, or CI/CD integration.

Best practices

• Store HF tokens in environment variables and use fine-grained tokens for production.
• Prefer .safetensors for model weights and pin SDK versions in Spaces.
• Add thorough model and dataset cards documenting use, limitations, and data sources.
• Preload heavy models with preload_from_hub to reduce Space startup time.
• Choose hardware to match workload (avoid OOM by selecting appropriate GPU/CPU).
• Use Parquet for large dataset storage and query via DuckDB for interactive analysis.

Example use cases

• Use hf_hub_download to fetch a specific model file for local inference in a script.
• Create a Gradio Space that loads a pretrained transformer and serves a chat interface.
• Invoke the InferenceClient with curl for chat completions or image generation in CI.
• Deploy a LoRA adapter: find, load, merge, and upload the adapted model to the Hub.
• Trigger retraining with a webhook that launches a Job or GitHub Actions workflow.