There are a few reasons I can think of to have offline pip packages:

• A package isn’t able to compile on a friend’s computer since they don’t have the million linear algebra libraries that numpy /scipy require.
• You want to archive everything to run a piece of software
• You want to control the packages available to a closed network

Regardless, to my surprise, setting up a repository of python wheels doesn’t take many steps.

Setup

First I would recommend that you setup a virtual environment. Either through pyenv or python-virtualenv.

Then, install whatever packages you would like. Let us use tensorflow as an example:

pip install tensorflow

We’re going to need the packages pip-chill and pip-tools for the next couple steps

pip install pip-chill pip-tools

After you install all the packages you want to be available, freeze the requirements that aren’t dependencies to a text file

pip-chill --no-version > requirements.in

We will then use pip-compile in pip-tools to resolve our dependencies and make our packages as fresh as possible.

pip-compile requirements.in

To sync the current virtual environment with the requirements.txt file that gets produced

pip-sync

Now we have a fully working and resolved environment.

From here, we need to install the wheel package to make the binary wheels.

pip install wheel

Then to create the wheels,

pip wheel --wheel-dir=wheels -r requirements.txt

With this you have a whole repository of wheels under the wheels folder!

Client Side

Now you can get all fancy with your deployment, though I just assumed that the files were mounted in some shared folder.

The client can install all the wheels

pip install /path/to/wheels/*

Or they can just install the packages they want

pip install --no-index -f /path/to/wheels/wheels package_name

If you don’t want to add flags to every command, check out my post on using configuration files with pip.