Features

Every kernel output is checked against a high-precision CPU reference, per dtype.

Caught 9/9 LLM-style bugs across 5 GPU classes with 0 false positives on 15/15 controls.

A p95-of-controls × 1.5 envelope fit per op and dtype, not one hand-picked atol/rtol.

Raises kernel-bug recall from 65% to 82% over a single fixed atol=1e-5, rtol=1e-2 — at zero precision cost.

Targets wrong-output kernels that compile, run, and pass the standard check.

The bug classes torch.allclose on one shape structurally cannot see: tail-mask leaks, accumulator scale, missing normalisation, online-softmax rescale.

A per-op shape generator with boundary, regular, and adversarial value distributions.

Adversarial value sampling reaches 99% bug recall — +28 pp over the field-standard default oracle.

Deliberately generates shapes where dimensions are not multiples of the block size.

“Regular shape only” sampling misses 100% of tail-mask bugs; adversarial shapes surface them.

A shared xorshift128+ generator produces the same inputs in Rust and Python.

Every flagged failure reports its seed, dtype, and shape, and reproduces byte-for-byte with gpuemu reproduce <seed>.

The correctness step compares against a CPU fp64 reference — no GPU required.

Run failing seeds on a laptop; a GPU is only needed for the optional PTX/SASS artifact step.

Reads register pressure, spills, and instruction counts from compiled artifacts.

Structural Δregs predicts Δperf% across H100/A100/L40S/A10/3060 — pair it with the oracle, since semantic bugs compile to identical PTX.

gpuemu ci fuzzes, lints, and baseline-diffs, emitting SARIF for code scanning.

The in-house correctness CI that vLLM/SGLang maintainers keep rebuilding by hand, as a GitHub Action.

An ed25519-signed evidence artefact customers verify offline.

An SLA-grade correctness attestation for inference-as-a-service vendors (enterprise tier).