Would calling torch.sort(…, stable=False) ever fail reproducibility? For example, due to race conditions in the parallelization. Obviously stable=True has to be reproducible. However stable=False doesn’t have to be reproducible unless it is specified in the documentation.