This is a PEP 517 build backend to act as a "soft link" between a package released on a 3rd party package index and pypi.org. This allows users to keep the simple UX of pip install foo while hosting the actual wheel files on a non-PyPi index.
- Install hatch: https://hatch.pypa.io/latest/install/
- Run
hatch shell - Run
hatch build -t wheel cd demo- Grab your favorite wheel file on https://pypi.nvidia.com, e.g.
wget https://pypi.nvidia.com/nvidia-cuda-runtime-cu12/nvidia_cuda_runtime_cu12-12.4.99-py3-none-manylinux2014_x86_64.whl - Build the sdist:
There should now be a new
PIP_FIND_LINKS=`pwd`/../dist PIP_NO_INDEX=1 python -m build --sdist --config-setting source_wheel=nvidia_cuda_runtime_cu12-12.4.99-py3-none-manylinux2014_x86_64.whldist/folder with e.g.nvidia_cuda_runtime_cu12-12.4.99.tar.gz - Install the sdist:
Feel free to test with
pip install --no-cache ./dist/<my_sdist>.tar.gz--extra-index-url https://pypi.nvidia.comas well. - Done! You should see everything install correctly
There are a few configuration options available for wheel_stub. These can be set in your pyproject.toml file:
-
index_url: The url to the 3rd party index to install wheels from. This is the only required configuration setting. -
stub_only: Do not install the wheel from the 3rd party package repository, instead erroring with instructions of how to use the repository directly. This can be used in cases such as a project which depends on non-PyPi dependencies, which would get rejected from pypi.org. Set automatically based on project dependencies.trueorfalse(defaultfalse) -
include_cuda_debuginfo: Whether to query information about the available CUDA runtime when reporting an error.trueorfalse(defaultfalse)
This project is released under the Apache-2 license. It vendors the packaging project, which is BSD 2-clause and Apache 2 dual licensed. It also vendors tomli, which is MIT licensed. For more details see the LICENSE file.