gpuemu is a GPU-less correctness oracle for deep-learning kernels. It catches silently-wrong CUDA and Triton kernels that a single-shape torch.allclose check accepts as correct — tail-mask leaks, accumulator-scale errors, missing normalisation, and online-softmax rescale bugs. Measured across five GPU classes, it catches 10/10 LLM-style buggy kernels with zero false positives on the correct controls.
Catch silently-wrong LLM-generated GPU kernels in CI without access to a physical GPU.
gpuemu is one option in a category that includes torch.allclose, KernelBench , and TritonBench. Our Compare page has the full side-by-side.