[t-vi]

Note that PyTorch's kernels are somewhat generic in shape. It has always been relatively easy to get speedups by specializing the shape, e.g. Apache TVM had that (back before it was "Apache" even).