[grpo] fix off-policy check (#4852)

Author: hjh0119

swift/trainers/rlhf_trainer/grpo_trainer.py

@@ -1433,7 +1433,7 @@ def _process_infer_requests_images(self, infer_requests: InputsType):
         return
 
     def old_policy(self):
-        return self.num_iterations > 1 or self.args.steps_per_generation > self.args.gradient_accumulation_steps
+        return self.num_iterations > 1 or self.args.gradient_accumulation_steps % self.args.steps_per_generation != 0
 
     @property
     def _queue(self):