gpuemu for frontier-lab kernel teams

The fastest-moving kernel teams in the world are shipping LLM-generated CUDA and Triton at a pace no human reviewer can keep up with. The correctness gate hasn’t kept up.

The pain, in their own data

PyTorch’s own triage shows 19.2% of high-priority torch.compile issues are silent wrong-output bugs — the second-largest category after outright crashes. These are the bugs that pass review, pass CI, and pass the benchmark, because the benchmark is a one-shape allclose. They surface as a model that is quietly worse.

When a frontier lab posts a dozen simultaneous kernel/inference roles, none of them mention “correctness” — the discipline is assumed, not owned. That gap is exactly where silent regressions live.

The workflow gpuemu gives you

• Pre-merge gate. Every kernel PR runs gpuemu ci against an fp64 oracle with op-schema-aware adversarial inputs. A regression blocks the merge.
• Scaled to hundreds of ops. Per-op schemas and calibrated tolerances mean coverage grows with a config entry, not a hand-written test per shape.
• Replay-seed links. Every failure comes with a seed and an input snapshot. The author reproduces it byte-for-byte on a laptop — no GPU, no flaky repro.
• PR-blocking SARIF. Findings land as code-scanning annotations on the exact diff.

The result: the silent wrong-output category stops being invisible and starts being a red check.