Your benchmark shows a 10× speedup over PyTorch on a si...

Before claiming a 10× speedup, you must first verify correctness by asserting that both implementations produce numerically equivalent outputs, as a kernel returning garbage can appear arbitrarily fast. Second, check input sizes; for very small inputs, PyTorch’s dispatch overhead may dominate both paths, rendering the ratio meaningless. Third, use torch.profiler to verify that PyTorch is dispatching to the expected kernel path for the specific shape tested. Finally, measure memory traffic: a 10× speedup on a memory-bound operation is only legitimate if you reduced HBM round-trips by 10×, while a similar speedup on a compute-bound operation would require 10× more FLOPs/second than the hardware provides, which is a red flag.