AI accelerator integration gains shared testing and reference tools
A working group details first-half improvements to cross-repository testing, hardware-neutral tests and profiling support.
AI illustrationThe Accelerator Integration Working Group has outlined work completed in the first half of 2026 to make new computing hardware easier to integrate with the AI software ecosystem. The changes include a shared testing relay, less hardware-specific tests and reference implementations for accelerator developers.
One addition, Cross-Repository CI Relay, lets registered downstream projects receive dispatch events when upstream changes are opened or committed. Those projects run their own tests and return results through an authenticated callback. Maintainers can then view upstream and downstream test status together, helping them identify compatibility problems before changes are merged.
Participation is organised into four levels. Projects can begin with notifications and progress through reporting toward non-blocking and eventually blocking checks. The system uses identity verification, authorisation, rate limiting and other validation measures to separate trusted information from results supplied by downstream projects.
The working group also described changes to tests that previously contained fixed references to particular accelerators. Contributors migrated over 276 test files during the first half of the year. The work is intended to let a single test produce variants for different registered devices without vendors maintaining their own patches.
Classification metadata distinguishes tests tied to specific hardware from those that are generic or hardware-neutral. A linter requires new test classes to declare that classification, while an allowlist handles existing unclassified files during the transition.
Profiling work supplies a minimal reference implementation for developers bringing up new accelerator backends. The reference demonstrates integration mechanics rather than serving as a production profiler. Priorities for the second half include reducing the unclassified test allowlist, extending hardware-neutral coverage and connecting hardware classification to automated test selection.