from torch import Tensor
def forward(input: Tensor, scale: float) -> Tensor:
...
def backward(output_grads: Tensor, softmax_results: Tensor, scale: float) -> Tensor: