#cumsum

PyTorch CUDA cumsum reads past buffer past 2^30 elements

A ones-tensor repro shows an out-of-bounds global load in the GPU prefix-sum path once the length exceeds roughly a billion floats.