Astral, in collaboration with PyTorch, NVIDIA, and Quansight, has released an experimental build of uv that supports 'wheel variants' — a new packaging concept from the WheelNext project. Wheel variants allow package authors to encode hardware-specific build metadata (e.g., CUDA, ROCm, XPU) so that installers can automatically select the correct wheel for a user's machine. Currently, Python wheel filenames can only encode OS, architecture, and Python version, leaving GPU and accelerator compatibility ambiguous. The new design introduces vendor-published provider plugins for feature detection, package-level variant metadata, and installer-side resolution logic. The experimental build can already install the correct PyTorch 2.8.0 variant based on detected GPU hardware. A formal PEP is planned for later in 2025.
Sort: