Hugging Face CLI

Downloads individual files or entire repository snapshots from the Hub with built-in caching layer that stores files locally, supports resumable downloads via HTTP range requests, and implements smart cache invalidation. Uses a content-addressed cache structure where files are stored by their blob hash, enabling deduplication across multiple model versions and automatic cleanup of unused files.

Downloads entire repository snapshots with optional filtering by file patterns, allowing developers to exclude large files (e.g., safetensors, ONNX variants) and download only needed components. Implements a two-pass strategy: first fetches repository metadata to enumerate files, then downloads only selected files in parallel, with automatic handling of symlinks and LFS pointers.

Provides utilities for inspecting and managing the local Hub cache directory, including cache size calculation, file listing by age/size, and automatic cleanup of old or unused files. Implements cache strategy with configurable retention policies (LRU, size-based, age-based). Monitors available disk space and warns before cache exceeds thresholds.

Provides a comprehensive CLI (huggingface-cli) with subcommands for all major Hub operations (login, download, upload, repo management). Implements progress bars for file operations, colored output for readability, and structured error messages. Uses argparse for command parsing with automatic help generation and shell completion support.

Implements MCP server that exposes Hub functionality (search, download, upload, inference) as tools callable by LLMs and AI agents. Provides structured tool definitions with JSON schemas for parameter validation. Enables LLMs to autonomously search for models, download files, and run inference without human intervention.

Provides a low-level commit API (create_commit) for atomic multi-file operations on Hub repositories. Implements conflict detection and resolution strategies (abort, overwrite, merge), file deletion via commit operations, and support for both HTTP and Git backends. Enables transactional semantics where multiple files are committed together or not at all.

Uploads files to Hub repositories via HTTP multipart requests with automatic routing to appropriate storage backend (standard Git, Git-LFS for large files, or Xet for deduplication). Implements chunked upload for large files, automatic LFS pointer generation, and conflict resolution via commit-based versioning. Supports both single-file and batch folder uploads with progress tracking.

Provides a Python API (HfApi class) for repository lifecycle management including creation, deletion, visibility changes, and branch/tag operations. Implements REST API calls to Hub backend with automatic error handling, retry logic, and permission validation. Supports both model and dataset repositories with identical interface patterns.

Implements search functionality across Hub repositories using server-side filtering and ranking. Supports filtering by task type, library, language, license, and custom metadata fields. Returns paginated results with metadata including model size, downloads, and last update time. Implements efficient pagination via cursor-based offsets rather than page numbers.

Provides a unified Python interface (InferenceClient) for running inference on 35+ ML tasks across multiple providers (Hugging Face Inference API, Replicate, Together AI, Fal AI, SambaNova). Automatically routes requests to appropriate provider based on model availability and user configuration. Supports both synchronous and asynchronous execution, streaming responses for text generation, and structured output parsing.

Provides Python classes (ModelCard, DatasetCard, SpaceCard) for creating and managing repository documentation with structured YAML frontmatter. Automatically validates metadata against Hub schema, generates markdown templates, and syncs card content to repository. Supports programmatic metadata updates without manual YAML editing.

Implements a mixin pattern (ModelHubMixin, PyTorchModelHubMixin, etc.) that adds push_to_hub() and from_pretrained() methods to any ML model class. Automatically handles model serialization, config file generation, and repository management. Supports framework-specific implementations for PyTorch, TensorFlow, Keras, and custom frameworks.

Implements a POSIX-like file system interface (HfFileSystem) that wraps Hub repositories as virtual file systems, enabling use with fsspec-compatible tools. Supports standard operations (ls, cat, open, glob) on Hub files without downloading entire repositories. Integrates with Pandas, Polars, and other data tools that accept fsspec URLs.

Manages Hugging Face API tokens with automatic detection from environment variables, config files, and interactive login. Implements secure token storage in platform-specific locations (keyring on Linux/macOS, credential manager on Windows). Provides token validation and permission checking before operations.