The Wire · Showcase
PYTORCH SHIPS L2 CROSS-REPO CI RELAY, PATCHES PYTHON 3.15 WHEEL BUILDS
By RepoJournal · Filed · About PyTorch
The cross-repository CI relay system jumped to L2 implementation overnight, while core PyTorch simultaneously locked down Python 3.15 compatibility across multiple backends.
PyTorch's test infrastructure team shipped the L2 implementation of the cross-repository CI relay [1], a major expansion of the L1 system that enables downstream CI callbacks and result aggregation. This lands in lockstep with ci-infra's L2 CRCR deployment configuration [2], which adds AWS Lambda functions to handle result callbacks from downstream CI. The synchronization matters: L2 cannot merge until both repos ship together. Meanwhile, core PyTorch dropped cuda-bindings from Python 3.15 wheel builds [3] because NVIDIA's cuda-python still doesn't support cp315, making the nightly builds fail on install. XPU and ROCm manywheel builds are also skipped for 3.15 [4] until Triton XPU stabilizes. ExecutorTech fixed a critical partition validation bug in the Arm backend [5] where removing Q/DQ bridge nodes after capability-based partitioning could reintroduce dependency cycles in complex models like MobileViT. The Helion DSL team expanded cutedsl benchmarking to 32 targets [7] and staged the dynamic_persistent scheduler infrastructure [8], part of a stacked series that extends tcgen05 direct-entry paths and adds autotune knobs [9]. Infrastructure shipping cleanly across five repos. One note: the group norm CUDA kernel std::aligned_storage replacement was reverted [6] due to internal ROCM build breakage, so that's back on the roadmap.
Action items
- → Monitor L2 CRCR deployment in ci-infra and test-infra - verify relay callbacks work end-to-end before production promotion pytorch/test-infra [immediate]
- → If building Python 3.15 PyTorch wheels, validate cuda-bindings is stripped and builds complete successfully pytorch/pytorch [plan]
- → Review the Arm backend partition validation fix if you ship models with complex attention blocks (MobileViT, transformers) pytorch/executorch [plan]
- → Track Triton XPU stabilization for Python 3.15 - XPU manywheel builds blocked until that clears pytorch/pytorch [monitor]
References
- [1] [CRCR] Initial implementation of L2 (#7967) pytorch/test-infra
- [2] Add L2 CRCR deployment configuration (#614) pytorch/ci-infra
- [3] Drop cuda-bindings dependency for Python 3.15 wheel builds (#184891) pytorch/pytorch
- [4] Skip XPU manywheel builds for Python 3.15 (#184906) pytorch/pytorch
- [5] Arm backend: fix(arm): validate partitions for dependency cycles after Q/DQ de-tagging (#18191) pytorch/executorch
- [6] Revert "Replace deprecated std::aligned_storage in group/layer norm CUDA kernels (#184474)" pytorch/pytorch
- [7] [cutedsl] Expand hillclimber benchmarking to 32 targets ↗ pytorch/helion
- [8] [cutedsl] stage dynamic_persistent scheduler object infrastructure ↗ pytorch/helion
- [9] [cutedsl] extend tcgen05 direct-entry path to additional targets ↗ pytorch/helion
FAQ
- What changed in PyTorch on May 23, 2026?
- The cross-repository CI relay system jumped to L2 implementation overnight, while core PyTorch simultaneously locked down Python 3.15 compatibility across multiple backends.
- What should PyTorch teams do about it?
- Monitor L2 CRCR deployment in ci-infra and test-infra - verify relay callbacks work end-to-end before production promotion • If building Python 3.15 PyTorch wheels, validate cuda-bindings is stripped and builds complete successfully • Review the Arm backend partition validation fix if you ship models with complex attention blocks (MobileViT, transformers)
- Which PyTorch repositories shipped on May 23, 2026?
- pytorch/test-infra, pytorch/ci-infra, pytorch/pytorch, pytorch/executorch, pytorch/helion