Tensor compiler conformance

Preserve the declared computation while changing its execution. Distinguish approved semantic IR from the target graph: decompositions, fusion and folding are legal when the contract permits them.

Work from the affected contract

1. State the input domain, observable outputs/state, dtype/device/layout and numerical budget. Require a bespoke canonical IR only when approved graph identity is an actual system requirement.
2. Trace capture specialization, guards, graph breaks and passes. Preserve source identity/provenance while recording transformations that explain the produced artifact.
3. Build independent reference checks from the declared domain, including boundaries, noncontiguous inputs and aliasing where supported. Test gradients for differentiable/training obligations, and only promised devices/layouts.
4. Derive tolerances from the numerical contract and measurements. A justified contract revision is explicit; widening a tolerance solely to hide a failure is not a repair.
5. Gate artifact reuse on complete cache identity and conformance evidence. Distinguish compile failure, semantic drift, conformance failure and performance regression; bisect pass/node divergence.

Scope and completion

Deliver the affected transformation/contract, manifest or reproduction and observed conformance/cost evidence. Adopting torch.compile for an existing model does not by itself require a new compiler or IR. A review may be clean with explicit sweep coverage and unavailable checks.

Use the user’s existing intent and authorization. Ask only for missing information that materially changes the result; use additional reviewers when they address a concrete uncertainty. Treat unavailable checks as gaps rather than successful verification.

Focused references