Custom DiceLoss grad Implementation doesnt match diceloss gradient

I am working on implementing my own diceloss gradient, I have coded the dice loss gradient equation using python/numpy but it doesn't match the gradient computed by pyTorch

can someone help me?

The derivate equation I am using is :

def dice_loss_grad(y, label):
        sum_squares = np.sum(y**2)+np.sum(label**2)
        num =  (label * sum_squares) - (2*y*np.sum(y*label))
        dem =   (sum_squares)**2 
        res = -2*(num/dem)
        return res

c = DiceLoss()
loss = c(input, output)
loss.backward(retain_graph=True)

result_torch = grad(outputs=loss, inputs=input, retain_graph=True)[0].detach().numpy()
loss_j = dice_loss_grad(input.detach().numpy(), output.detach().numpy())
print(result_torch)
print(loss_j)

class DiceLoss(nn.Module):
    def __init__(self, weight=None, size_average=True):
        super(DiceLoss, self).__init__()

    def forward(self, inputs, targets, smooth=0):
        
        #comment out if your model contains a sigmoid or equivalent activation layer
        # inputs = F.sigmoid(inputs)       
        
        #flatten label and prediction tensors
        
        intersection = (inputs * targets).sum()                            
        dice = (2.*intersection + smooth)/(inputs.sum() + targets.sum() + smooth)  
        
        return 1-dice