diff --git a/fastdeploy/cache_manager/transfer_factory/rdma_cache_transfer.py b/fastdeploy/cache_manager/transfer_factory/rdma_cache_transfer.py
index 94abbb3b8..6a0c0ac36 100644
--- a/fastdeploy/cache_manager/transfer_factory/rdma_cache_transfer.py
+++ b/fastdeploy/cache_manager/transfer_factory/rdma_cache_transfer.py
@@ -61,18 +61,12 @@ class RDMACommManager:
         Connect to remote gpu and write cache.
         """
         assert self.splitwise_role == "prefill", "only prefill can call this method"
-        addr = f"{ip}:{port!s}"
-        if addr in self.connected_rdma:
-            return True
         ret = self.messager.is_connected(ip, str(port))
         if ret:
-            self.connected_rdma.add(addr)
             return True
 
         ret = self.messager.connect(ip, str(port))
         logger.info(f"connect to remote rdma address {ip}:{port} status is {ret}")
-        if ret == 0:
-            self.connected_rdma.add(addr)
         return ret == 0
 
     def write_cache(self, ip, port, local_block_ids, remote_block_ids, layer_idx):
diff --git a/fastdeploy/output/token_processor.py b/fastdeploy/output/token_processor.py
index a9bf68723..fbb978407 100644
--- a/fastdeploy/output/token_processor.py
+++ b/fastdeploy/output/token_processor.py
@@ -265,13 +265,13 @@ class TokenProcessor:
                         llm_logger.info(f"finished_task_id: {finished_task_id}")
                         self.prefill_result_status[finished_task_id[0]] = finished_task_id[1]
                 if task_id in self.prefill_result_status:
-                    self.split_connector.send_first_token(task.disaggregate_info, [result])
                     self.resource_manager.stop_flags[index] = True
                     self.resource_manager.tasks_list[index] = None
                     self.resource_manager._recycle_block_tables(task)
                     if self.prefill_result_status[task_id] != "finished":
                         result.error_code = 400
                         result.error_message = f"{task_id} failed to {self.prefill_result_status[task_id]}"
+                    self.split_connector.send_first_token(task.disaggregate_info, [result])
                     del self.resource_manager.req_dict[task_id]
                     break
                 else: