[Engineering Tracker] Variant PyTorch 2.8.0 Experiment

[DEKHTIARJonathan]

July Sprint !

Deadline: Release date is looking like July 30th - We should roughly aim for this date.

Overall I'm reasonably confident - most of the work is done already.

variantlib

• dynamic plugin support @DEKHTIARJonathan @mgorny
  - Almost done - need to square up and verify.

• variant dependency specifiers @mgorny

foo; "x86_64" in variant_namespaces
bar; "x86_64::level" not in variant_features
baz; "x86_64::level::v2" in variant_properties

CC: @charliermarsh @konstin

• variantlib should not install plugins / create venv. That's the job of the installer @mgorny

CC: @charliermarsh @konstin

NVIDIA

• Make the NVIDIA variant plugin dynamic @DEKHTIARJonathan

• Think a little bit more about the features we will need to expose for torch and how we expose them @vyasr @msarahan @DEKHTIARJonathan

Astral

• Detect, download and parse variants.json file (example) and install in an independent (temporary?) venv the required plugins
• Plug into variantlib and test "end-to-end" that simple use case are functioning.
• Provide feedback on what had to change on uv.lock, what we can use to evolve pylock.toml
• Overall design feedback from the installer perspective (even if long term you will probably re-implement variantlib)

Meta @seemethere @atalman

• Provide alpha/beta wheels of PyT 2.8.0
• Re-package all the wheels using make-variant tooling in the variant format (@DEKHTIARJonathan will definitely help here).
• Prepare a separate & isolated index to host these artifacts

WheelNext

• Start working on a proto-PEP / DPO thread
• Marketing & communication around this release: META / Astral / NVIDIA (blog posts, youtube video?) @seemethere @charliermarsh @DEKHTIARJonathan @warsaw