This simple example shows how to use weak dependencies and package extensions in Julia, introduced in Julia 1.9. This incredible feature allows:
- Load packages only when needed (perfect for large dependencies)
- Write functions with multiple dispatch (different implementations for different types)
I don't like the example from the official documentation, because it's not clear how to use it in a real scenario. I've been looking for a simple example, but I couldn't find it. So, I decided to write my own example, and share it with the community.
This example contains a package named BaseAdd, implementing functions that add two arrays. The goal is to implement a function that does the same, but using the GPU. Our problem begins with the fact that CUDA.jl package is a large dependency, increasing the loading time when someone imports it. This gives us the perfect scenario to use weak dependencies and package extensions.
As you can see in Project.toml, there's no dependency pointing to CUDA. It is defined as a weak dependency ([weakdeps] section), and loaded in the extension module BaseAddExt ([extensions] section). In this way, the package BaseAdd can be loaded without loading CUDA, and if needed, CUDA can be loaded later (i.e. calling add!(y::CuArray, x::CuArray)).