NVIDIA introduces Wheel Variants, a new Python packaging format that automatically selects optimal GPU-accelerated packages based on your hardware. This eliminates the need to manually navigate PyTorch, JAX, or RAPIDS websites to find CUDA-compatible versions. The format uses variant properties like 'nvidia::cuda_version_lower_bound::12.0' to specify hardware requirements, with plugins detecting local capabilities and selecting the best wheel automatically. PyTorch 2.8.0 will include experimental support, developed in collaboration with Meta, Astral, and Quansight as part of the WheelNext initiative.
Table of contents
What are the technical challenges with CUDA compatibility?What is the Wheel Variant format?Example NVIDIA GPU-specific implementationWhat are the ecosystem benefits for Python package end users?What are the ecosystem benefits for package maintainers?What are the ecosystem benefits for NVIDIA GPU users?What are the broader applications of Wheel Variants beyond PyTorch and NVIDIA GPU computing?How to build Wheel VariantsWhat is the implementation road map for Wheel Variants?ConclusionGet involvedSort: