diff --git a/unirl/train/backend/base_backend.py b/unirl/train/backend/base_backend.py
index 936b504c1..549b2e4a2 100644
--- a/unirl/train/backend/base_backend.py
+++ b/unirl/train/backend/base_backend.py
@@ -308,6 +308,8 @@ def optimizer_step(self, *, max_grad_norm: float) -> float:
         if self.ema is not None:
             self.ema.step(self._optimizer_step_count)
         self._optimizer_step_count += 1
+        # Release consumed grads before train-state offload.
+        self.optimizer.zero_grad(set_to_none=True)
         return grad_norm
 
     def on_rollout_end(self) -> None: