Hardware-Aware Python Packages ~ PyTorch and...
summary
Goal: make pip install torch pick the right binary by reading machine hardware (CUDA version, GPU SMs, CPU ISA like AVX/ARMv8/ARMv9). Implementation: new Wheel Variants metadata plus augmented installer (special UV build) and wheel registry; demonstrated installing different CUDA 12.8 vs 13 builds and different dependency sets live. Outcome: working prototype that delivers smaller single-arch wheels, supports NVIDIA/AMD/Intel and older SM fallbacks, and aims to land in the packaging standard via WheelNext.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 1:03Python packaging hardware problems
- 3:30Wheel Next solution
- 5:40Live demo wheel selection
- 8:54Wheel Next benefits
- 9:35Wheel Next community
- 10:19Q&A
related talks
24mHow To Write C++ Extensions in 2026
19mIncreasing the Impact of AI Through Better Software (Pytorch Updates)
40mBuilding a binary extension
27mPython in Bazel
24mAdvancements in High-Performance AI/ML through PyTorch's Python Compiler
44mDeinventing the Wheel
05mTowards Device-agnostic PyTorch: Building Unified Infrastructure for a Multi-Backend Ecos...
29mTaking wheels mobile