Segmentation fault when defining a custom loss in lightgbm

Question

I've been trying the following custom loss for a regression task in Lightgbm and boosting type = "goss", and I'm getting a segmentation fault:

def hm(y_true, y_pred):
    residual = (y_true - y_pred).astype("float")
    residual_abs = np.fabs(residual)
    grad = np.where(residual_abs<=300, 
                    np.copysign(1, -residual)* 0.5,
                    np.where(residual_abs>=900, 
                             np.copysign(1, -residual)* 1,
                             np.copysign(1, -residual)* 2))
    hess = np.ones(y_true.shape[0]) 
    return grad, hess

Any ideas why it might be failing? Im using goss and default hyperparameters