def gradient_penalty(self, batch_x, fake_image):
        batchsz = batch_x.shape[0]
        # [b, w, h, c]
        t = paddle.uniform((batchsz, 1, 1, 1))
        t = paddle.broadcast_to(t, batch_x.shape)

        interplate: paddle.Tensor = t * batch_x + (1 - t) * fake_image
        interplate.stop_gradient = False
        d_interplate_logits = self.discriminator(interplate)
        grads = paddle.grad(
            outputs=d_interplate_logits, 
            inputs=interplate,
            grad_outputs=paddle.ones_like(d_interplate_logits),
            create_graph=True,
            retain_graph=True
            )[0]
        grads = paddle.reshape(grads, [grads.shape[0], -1])
        epsilon = 1e-12
        gp = paddle.sqrt(
            paddle.mean(paddle.square(grads), axis=1) + epsilon
        )
        gp = paddle.mean((gp - 1) ** 2)
        return gp

代码报错：The Op flatten_contiguous_range_grad doesn't have any grad op.

在同样结构网络下运行相同的获取gp的函数，pytorch可以运行，paddle报错，是不支持高阶求导吗？？？？