AIKIDO-2025-10405

TL;DR

NumPy’s default integer type on Windows is int32, which can cause integer overflows in operations like np.prod((60000, 60000)), returning incorrect negative values. In flatten_tensor, Python integers are generally cast to int64, but one instance applies this conversion too late, after the overflow has already occurred. Explicitly setting dtype=np.int64 in np.prod prevents this issue by ensuring calculations use 64-bit integers.