diff --git a/recipe/one_step_off_policy/grpo_qwen3_8b_gsm8k_fsdp2_8_8_npu.sh b/recipe/one_step_off_policy/grpo_qwen3_8b_gsm8k_fsdp2_8_8_npu.sh
index 6a39b87fead..109a685f64d 100644
--- a/recipe/one_step_off_policy/grpo_qwen3_8b_gsm8k_fsdp2_8_8_npu.sh
+++ b/recipe/one_step_off_policy/grpo_qwen3_8b_gsm8k_fsdp2_8_8_npu.sh
@@ -60,6 +60,7 @@ python3 -m recipe.one_step_off_policy.main_ppo \
     actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=1 \
     actor_rollout_ref.rollout.tensor_model_parallel_size=4 \
     actor_rollout_ref.rollout.name=vllm \
+    actor_rollout_ref.rollout.mode=sync \
     actor_rollout_ref.rollout.gpu_memory_utilization=0.8 \
     actor_rollout_ref.rollout.max_num_batched_tokens=$((max_prompt_length + max_response_length)) \
     actor_rollout_ref.rollout.n=8 \