I built a pre-flight check tool for PyTorch, because silent failures are the worst kind

Last month I was debugging a training run that produced suspiciously bad results. The loop ran fine....