How to Build backward for Sparse tensor dot Multiplication

[LuCheng2000]

I attempted to achieve the effect of dot multiplication on sparse tensors, based on the spatial information of the indices in the sparse tensor corresponding to multiplication. Although the network model was successfully trained, the results were very unsatisfactory. So I checked the network gradients to determine the problem and found that all gradients before the dot multiplication operation were None, which means there are some issues with the backward. I have no idea to know how the backward in spconv is implemented. Is the data returned by grad_output in the ordinary tensor form? If so, how can I complete the corresponding spatial rule operation without the indices data of grad_output?

class z_sparse_dot_F(Function):
@staticmethod
def forward(ctx,x,y):
ctx.save_for_backward(x, y)
return z_sparse_dot_f(x,y)

@staticmethod
def backward(ctx, grad_output):
    pdb.set_trace()
    x,y=ctx.saved_tensors
    grad_x = z_sparse_dot_f(grad_output,y) 
    grad_y = z_sparse_dot_f(grad_output,x)
    return grad_x, grad_y