Hugging Face Repository Downloader

This Python script allows you to download repositories from Hugging Face, including support for fast transfer mode when available. It can resume interrupted downloads and skip already downloaded files.

Features

• Download entire repositories from Hugging Face
• Support for private repositories with authentication
• Resume interrupted downloads
• Skip already downloaded files
• Optional fast transfer mode for high-bandwidth connections
• Ignore specific files or patterns using glob syntax
• Download multiple repositories using a URL file

Installation

pip install huggingface-repo-downloader

From source:

git clone https://github.com/nmehran/huggingface-repo-downloader.git
cd huggingface-repo-downloader
pip install -e .

Usage

huggingface-repo-downloader <repository_url> [options]
huggingface-repo-downloader --url-file <file_path> [options]

The python huggingface_downloader.py ... form still works from a clone.

Options

• -o, --output: Specify output directory (optional, default: script directory)
• --use-auth-token: Use the Hugging Face auth token for private or access-restricted repos
• --revision: Specific revision to download (default: main)
• --fast: Enable fast transfer mode (requires hf_transfer package)
• --ignore-file: Specify a file containing glob patterns to ignore
• --url-file: Specify a file containing repository URLs to download
• --force: Force re-download of all files, ignoring existing metadata
• --help: Show help message and exit
• --auth-help: Show instructions for setting up authentication

Examples

1. Download a public repository:

   python huggingface_downloader.py meta-llama/Meta-Llama-3.1-8B-Instruct/tree/main
   

2. Download a private repository with authentication:

   python huggingface_downloader.py your-private-repo/model --use-auth-token
   

3. Download a specific revision:

   python huggingface_downloader.py facebook/bart-large --revision v1.0
   

4. Use fast transfer mode:

   python huggingface_downloader.py google/t5-v1_1-large --fast
   

5. Use an ignore file to skip certain files or patterns:

   python huggingface_downloader.py meta-llama/Meta-Llama-3.1-8B-Instruct --ignore-file .hfignore
   

6. Force re-download of all files:

   python huggingface_downloader.py meta-llama/Meta-Llama-3.1-8B-Instruct --force
   

7. Download multiple repositories using a URL file:

   python huggingface_downloader.py --url-file repos.txt --use-auth-token
   

Fast Transfer Mode

The --fast flag enables fast transfer mode, which can significantly increase download speeds on high-bandwidth connections. However, it's important to note:

1. It requires the hf_transfer package to be installed.
2. It lacks progress bars, making it less user-friendly for monitoring download progress.
3. It's primarily a power user tool, designed for scenarios where standard Python downloads can't fully utilize available bandwidth (typically beyond ~500MB/s).
4. It is not applicable to most users, even at most higher-speed internet connections.

By default, fast transfer mode is disabled to ensure a more user-friendly experience with progress bars and compatibility across different network conditions.

Ignore File

You can use an ignore file to specify patterns of files or directories that should be skipped during the download. The ignore file uses glob syntax for pattern matching.

To use an ignore file:

1. Create a text file (e.g., .hfignore) with one pattern per line.
2. Use the --ignore-file option when running the script, pointing to your ignore file.

Example .hfignore file content:

*.txt
test_data/*
logs/

This would ignore all .txt files, everything in the test_data directory, and the logs directory.

Note:

• Lines starting with # (including those with leading whitespace) are treated as comments and ignored.
• Empty lines are skipped.

URL File

You can use a URL file to specify multiple repository URLs to download in a single run. This is useful for batch downloading repositories.

To use a URL file:

1. Create a text file (e.g., repos.txt) with one repository URL per line.
2. Use the --url-file option when running the script, pointing to your URL file.

Example repos.txt file content:

meta-llama/Meta-Llama-3.1-8B-Instruct
facebook/bart-large
google/t5-v1_1-large

This would download all three repositories in sequence.

Note:

• Lines starting with # are treated as comments and ignored.
• Empty lines are skipped.
• You can specify different revisions for each repository by appending /tree/<revision> to the URL.

Metadata Handling

The script uses metadata to keep track of downloaded files and their versions. This helps to avoid unnecessary re-downloads of unchanged files. The metadata is stored in a JSON file in the output directory.

When you run the script:

• It compares the current repository metadata with the stored metadata.
• Only files that are new or have changed since the last download are downloaded.
• The --force option overrides this behavior and re-downloads all files.

Authentication

For private or access-restricted repositories, you need to set up authentication. Follow these steps:

1. Visit https://huggingface.co/settings/tokens to generate a token.
2. Run 'huggingface-cli login' and enter your token when prompted.

For detailed instructions, run:

python huggingface_downloader.py --auth-help

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.