How IBM built Torch Spyre on PyTorch's CRCR to scale accelerator CI testing
PyTorch · x · 2026-09-30
Engineers from IBM and Red Hat published a PyTorch community deep dive on integrating Torch Spyre, their backend for the Spyre accelerator, with PyTorch's Cross-Repository CI Relay (CRCR), reaching L2 integration.
Key points:
- CRCR closes the coordination gap between PyTorch and out-of-tree accelerator repos: upstream changes trigger downstream CI automatically, results surface to PyTorch reviewers, and regressions appear in a unified HUD.
- Torch Spyre uses an agentic pipeline that selects and re-buckets tests from code and execution evidence, deciding which of PyTorch's tens of thousands of tests actually matter for the hardware.
- A declarative YAML framework adapts upstream tests without patching them, keeping the setup current as PyTorch, the backend, and the test suite all evolve.
- Most of the approach applies to any generic privateuse1 device, and the team aims to upstream reusable parts with the community.
More from coding & agent
- Free hands-on guide: turning AI coding into a full-lifecycle 'Agentic Software Factory' — Pavan_Belagatti · 2026-09-30
- Hugging Face Engineer Shares Full Slides on How the Company Uses Agents Internally — lhoestq · 2026-09-30
- DynaRobotics' full-stack approach to workflow automation wins industry praise — chris_j_paxton · 2026-09-30
- STILL, a game that's 100% vibe coded, shows fully interactive city gameplay — Daniel_Farinax · 2026-09-30
- Open-source BabyCoder gives small local LLM agents a dreaming sleep loop that generates new questions — Illustrious_Matter_8 · 2026-09-30
- Dev says latest SOTA coding model feels like a senior engineer collaborator, not a junior — Liu_eroteme · 2026-09-30