Hugging Face Hub Documentation

Hugging Face and AWS maintain Deep Learning Containers (DLCs) to provide optimized Docker images for machine learning workloads. These containers come pre-installed with libraries like transformers, datasets, and tokenizers, alongside high-performance serving engines. They are designed to be used within Amazon SageMaker AI to skip the manual process of building and optimizing environments.

Key details:

• Availability: Images are hosted in the Amazon Elastic Container Registry (ECR).
• Licensing: Open source under Apache 2.0.
• Usage: Can be used for both training and inference of models from the Hugging Face Hub.

Resource Groups are a feature available in Team & Enterprise plans that allow organizations to enforce fine-grained access control to their repositories. They enable administrators to group related repositories, manage member access at a group level instead of an individual repository level, and manage costs by attributing them to specific groups.

Key capabilities include:

• Permission Roles: Assign no_access, read, contributor, write, or admin roles to team members within a group.
• Privacy: Keep private repositories visible only to authorized group members.
• Team Isolation: Enable multiple teams to work independently within a single organization.
• Governance: Configure which member roles are permitted to create new resource groups.
• Budgeting: Attribute costs to specific resource groups for better management.

Hugging Face Inference Providers allows you to use state-of-the-art open models across dozens of different providers using a single Hugging Face (HF) token. This integration works with a wide variety of existing tools, including coding agents, LLM frameworks, and evaluation suites, without requiring changes to your existing workflow.

Key Benefits:

• Tool Compatibility: Use your existing preferred tools.
• Provider Agility: Switch between different model providers without changing your code.
• Direct Pricing: Access zero-markup pricing, matching direct provider rates.
• Unified Authentication: Use one single HF token for all providers and models.

Hugging Face Spaces allow you to host Machine Learning (ML) demo applications directly on your personal or organization profile. They are used to create ML portfolios, showcase projects, and collaborate within the ML ecosystem.

Spaces support several hosting methods:

• Gradio SDK: Build Python-based apps quickly.
• Docker: Host arbitrary applications using a Dockerfile.
• Static HTML/JS: Create static web-based Spaces using JavaScript and HTML.

You can also upgrade Spaces to run on GPUs or other accelerated hardware.

The Model Hub is a central platform for the Hugging Face community to host, store, discover, and share model checkpoints. It serves as a repository for pre-trained models that can be integrated into various workflows.

Key ways to interact with models on the Hub:

• Download and use models: Use the huggingface_hub client library for direct access, or use 🤗 Transformers for fine-tuning and inference.
• Library Integration: Models are compatible with over 15 integrated libraries.
• Production Deployment: Leverage Inference Providers or Inference Endpoints to run models in production environments.

Hugging Face Jobs provides managed compute infrastructure for AI and data workflows. It is designed for tasks such as fine-tuning AI models, running inference with GPUs, and performing data ingestion or processing.

Key Features:

• UV & Docker-like Interface: Run workloads using a familiar syntax.
• Hardware Flexibility: Supports various hardware flavors including CPUs, GPUs (e.g., A100s), and TPUs.
• Workload Versatility: Supports UV scripts, Docker images (from Hugging Face Spaces or Docker Hub), and any command-based workload.
• Scalability: Supports running many jobs in parallel for tasks like parameter tuning or parallel data processing.
• Pay-as-you-go: Billing is based on the seconds used.

Argilla is an open-source tool for building and iterating on AI data. When deployed on Hugging Face Spaces with Hugging Face OAuth enabled, it allows community members to join your Argilla server to annotate datasets.

Key capabilities include:

• Dataset Configuration: Set up questions for human feedback (Label, NER, Ranking, Rating, free text, etc.).
• Model Integration: Use model predictions to evaluate outputs or accelerate annotation.
• Data Exploration: Use search and semantic similarity to find critical data subsets.
• HF Hub Integration: Pull and push datasets to the Hugging Face Hub for versioning and training.

The default Gradio Spaces environment includes several pre-installed libraries and system packages. You do not need to manually install these unless you require a specific version.

Python Libraries

• huggingface_hub: Used to manage repositories and files on the Hub and programmatically access Inference Providers.
• requests: Used for calling third-party APIs.
• datasets: Used to fetch or display datasets from the Hub.
• gradio: The core framework. You can specify a specific version using the sdk_version key in your README.md file.

Hugging Face Team & Enterprise subscriptions provide advanced capabilities for managing models and teams:

• Access Control: Use resource groups to manage permissions for specific teams or users.
• Storage Region: Select between US or EU data storage regions for compliance.
• Advanced Analytics: Access Publisher Analytics for deep insights into usage and adoption.
• Extended Storage: Additional private storage for large artifacts and growing portfolios.
• Organization Blog Posts: Publish research updates and announcements directly on Hugging Face from your organization's profile.