[BUG]: some param and grad of critic_optim are not in the same device after backward

[TyrionZK]

🐛 Describe the bug

After self.strategy.backward(critic_loss, self.critic, self.critic_optim), some param and grad of critic_optim are not in the same device . I don't know what happen

Environment

device: 8A80080G

[flybird11111]

May I take a look at the traceback？