WheelNext & Wheel Variants: An update, and a request for feedback!

[jamestwebber (James Webber)]

@charliermarsh

You can still install a wheel without running any code, if you know what wheel you want. You should also be able to define, in advance, the set of features that you want the installer to “assume” are available on your machine, without running any of the provider code. If you want the installer to infer the correct wheel automatically, then it would need to run the provider code.

I’m sure this has been discussed in the WheelNext group, but to throw out a more opt-in flavor of this proposal:

• Wheel variants exist as proposed: a package publishes a JSON of wheels and the features they each require
• A standard is developed for installers to support wheel-variant plugins
• The user must opt to install the plugin for any given feature they care about
  • e.g. if I have a GPU and I want my installer to automatically choose the best wheel for my GPU, I need to install the GPU-variant-plugin^[1].
  • Installers may try to automatically determine the appropriate plugins when they are installed, but this is configurable.
  • Like any other package, these plugins are voluntarily installed and can be audited by those who wish to do so.
• Upon [installer] install foo, the installer will consider the variants available and the plugins installed and choose a wheel accordingly.
  • If I ask for generic pytorch and didn’t install a GPU-variant-plugin then I’m gonna get the CPU version

---

1. and configure it, unless that plugin is going to run on install ↩︎