huggingface/swift-xet

Requirements

• Swift 6.0+ / Xcode 16+
• macOS 13+ / iOS 15+ / tvOS 15+ / watchOS 8+ / visionOS / Linux

Installation

Swift Package Manager

Add the following to your Package.swift file:

dependencies: [
    .package(url: "https://github.com/huggingface/swift-xet.git", from: "0.2.0")
]

Then add the dependency to your target:

.target(
    name: "YourTarget",
    dependencies: [
        .product(name: "Xet", package: "swift-xet")
    ]
)

Usage

Downloading a File

To download a file, you need:

• A file ID: the 64-character hex hash from the X-Xet-Hash response header
• A refresh URL: the Hub endpoint for obtaining CAS access tokens

import Xet

try await Xet.withDownloader(
    refreshURL: refreshURL,
    hubToken: "hf_..."  // optional, required for private repos
) { downloader in
    // Download to memory
    let data = try await downloader.data(for: fileID)

    // Download to disk
    try await downloader.download(fileID, to: destinationURL)
}

Partial Downloads

Both methods support partial downloads via the byteRange parameter:

// Download first 1MiB only
let data = try await downloader.data(
    for: fileID,
    byteRange: 0..<(1024 * 1024)
)

The file ID comes from the X-Xet-Hash header when resolving a file URL without following redirects:

// Construct URLs for a Hugging Face repository
let repoType = "datasets"  // or "models", "spaces"
let repoID = "username/repo-name"
let revision = "main"
let filePath = "path/to/file.bin"

let resolveURL = URL(string:
    "https://huggingface.co/\(repoType)/\(repoID)/resolve/\(revision)/\(filePath)"
)!

let refreshURL = URL(string:
    "https://huggingface.co/api/\(repoType)/\(repoID)/xet-read-token/\(revision)"
)!

// Get the file ID by making a request that doesn't follow redirects
// and reading the X-Xet-Hash header from the response

Tuning HTTP Performance

This package uses AsyncHTTPClient under the hood for CAS and xorb downloads. The downloader manages a small pool of HTTP clients and shuts them down automatically when you use Xet.withDownloader. Tuning can help when you need to balance throughput, memory, and connection limits for your network environment.

You can configure the client pool and timeouts through XetDownloader.Configuration:

var configuration = XetDownloader.Configuration.default
configuration.connectionsPerHost = 8
configuration.poolSize = 2
configuration.readTimeout = 300

try await Xet.withDownloader(
    refreshURL: refreshURL,
    hubToken: "hf_...",
    configuration: configuration
) { downloader in
    try await downloader.download(fileID, to: destinationURL)
}

How It Works

The Xet protocol reconstructs files from deduplicated, compressed chunks:

1. Token Refresh: Obtain a short-lived CAS access token from the Hub
2. Reconstruction Query: Fetch metadata describing which chunks comprise the file
3. Chunk Download: Fetch compressed chunk data from xorb storage
4. Decompression: Decompress chunks using LZ4 or BG4+LZ4
5. Reassembly: Concatenate chunks in order to reconstruct the file

Xorb Format

Files are stored as xorbs (Xet Orbs)—sequences of compressed chunks. Each chunk has an 8-byte header specifying:

• Version (1 byte)
• Compressed size (3 bytes, little-endian)
• Compression scheme (1 byte): none, LZ4, or BG4+LZ4
• Uncompressed size (3 bytes, little-endian)

BG4 (Byte Grouping 4) is a preprocessing step that improves compression for floating-point and structured data by grouping bytes by position.

Contributing

This is a community project and we welcome contributions. Please check out [Issues tagged with good first issue][good-first-issues] if you are looking for a place to start!

Please ensure your code passes the build and test suite before submitting a pull request. You can run the tests with swift test.

[good-first-issues]: https://github.com/huggingface/swift-xet/issues?q=is%3Aissue%20state%3Aopen%20label%3A%22good%20first%20issue%22

License

Apache 2.

Package Metadata

Repository: huggingface/swift-xet

Default branch: main

README: README.md