Wheelhouse

Offline Python package downloader with full dependency resolution.

Wheelhouse downloads Python wheel packages and all their transitive dependencies for a target platform — without needing to run on that platform. Perfect for preparing offline installations on air-gapped servers.

Features

• Cross-platform downloads — Download wheels for Linux x86_64 or Windows AMD64 from any machine
• Full dependency resolution — Recursively resolves dependencies with backtracking and conflict detection
• PEP 658 metadata resolution — Fetches lightweight .metadata (~5KB) to resolve dependencies without full wheel downloads
• PEP 508 marker evaluation — Correctly filters platform-specific and optional dependencies
• PyTorch Hardware Backends — Built-in support for Intel GPU (--torch-xpu), NVIDIA CUDA (--torch-cuda cu118, cu124, cu126, cu128), CPU, and ROCm
• Dry-run mode — Preview resolved packages and version tree without downloading wheels (--dry-run)
• Index caching & resume — Caches package index locally as JSON and automatically resumes partial downloads (HTTP 206 / Range)
• Configurable — TOML configuration file + CLI overrides

Installation

pip install wheelhouse

Or install from source:

git clone https://github.com/crimson-and-clover/wheelhouse.git
cd wheelhouse
pip install -e .

Quick Start

Download packages via CLI

# Download numpy and all its dependencies for Linux + Python 3.10
wheelhouse download numpy --python 3.10 --platform linux_x86_64

# Download from a requirements file
wheelhouse download -r requirements.txt --python 3.10 --platform linux_x86_64

# Download PyTorch with Intel XPU backend
wheelhouse download torch torchvision --torch-xpu

# Download PyTorch with CUDA 11.8 backend
wheelhouse download "torch==2.1.2+cu118" --torch-cuda cu118

# Preview dependency tree without downloading (dry run)
wheelhouse download -r requirements.txt --dry-run

# Fast concurrent download with 8 worker threads
wheelhouse download -r requirements.txt -j 8

# List supported platforms, Python versions, and PyTorch backends
wheelhouse list

# Specify custom output directory
wheelhouse download numpy -o ./my_wheels/

Use as a library

from wheelhouse.config import Config
from wheelhouse.models import PackageRequirement
from wheelhouse.resolver import Resolver
from packaging.specifiers import SpecifierSet

config = Config(
    python_version="3.10",
    platform="linux_x86_64",
    torch_backend="cu118",
)

requirements = [
    PackageRequirement(package_name="numpy"),
    PackageRequirement(package_name="requests", version_spec=SpecifierSet(">=2.28")),
]

with Resolver(config) as resolver:
    resolved = resolver.resolve(requirements)

for name, pkg in resolved.items():
    print(f"{name} == {pkg.version} -> {pkg.filepath}")

Install on target machine

Copy the wheels/ directory to the target machine, then:

pip install --no-index --find-links=wheels/ -r requirements.txt

Supported Platforms

Platform	Identifier
Linux x86_64	linux_x86_64 (default)
Linux ARM64	linux_aarch64
Windows AMD64	win_amd64
macOS ARM64 (Apple Silicon)	macos_arm64
macOS x86_64 (Intel)	macos_x86_64

To add a new platform, run wheelhouse dump-tags on a machine with the target OS and copy the generated JSON file into wheelhouse/data/platforms/.

Supported Python Versions

• Python 3.10
• Python 3.11
• Python 3.12
• Python 3.13

Development

# Install with dev dependencies
pip install -e ".[dev]"

# Run tests
pytest

# Lint
ruff check wheelhouse/ tests/

License

MIT