PyTorch launches Cross-Repository CI Relay for seamless downstream testing
New system automatically triggers and tracks CI across hardware backend repos
PyTorch’s ecosystem relies on downstream hardware backends (Intel XPU, AMD ROCm, Apple MPS, Qualcomm AI Engine) and foundation libraries (vLLM, SGLang, Hugging Face). These repos run custom CI, but until now had no standard way to automatically test against upstream PyTorch changes. The new Cross-Repository CI Relay (CRCR) closes this gap. When a PR or commit hits pytorch/pytorch, CRCR dispatches CI triggers to all registered downstream repos. Those repos run workflows and report results back via an authenticated OIDC callback. Status appears on the PyTorch CI HUD within seconds, giving maintainers a single dashboard for both in-tree and cross-repo CI health.
The relay uses a four-tier allowlist for incremental onboarding. L1 (Notify) – downstream receivs dispatch but doesn't report; L2 (Report) – results appear on the HUD; L3 (Signal) – adds a non-blocking check run on the upstream PR; L4 (Gate) – a blocking check run reserved for critical projects. The architecture leverages AWS Lambda, Redis for state, DynamoDB for storage, and ClickHouse for analytics. This system ensures upstream changes don’t break downstream repos without visibility, reducing coordination overhead and enabling data-driven merge decisions. Future support for L3/L4 will let downstream teams signal regressions directly in the PR workflow.
- CRCR dispatches CI to registered downstream repos whenever a PR is opened against pytorch/pytorch
- Results appear on PyTorch CI HUD within seconds via authenticated OIDC callbacks
- Four participation levels: L1 (notification), L2 (reporting), L3 (non-blocking signal), L4 (blocking gate)
Why It Matters
Ensures upstream changes don't break critical hardware backends without visibility, reducing coordination overhead.