diff --git a/vllm_ascend/distributed/kvpool/kv_transfer.py b/vllm_ascend/distributed/kvpool/kv_transfer.py
index cd5245364c4..0265d6a320c 100644
--- a/vllm_ascend/distributed/kvpool/kv_transfer.py
+++ b/vllm_ascend/distributed/kvpool/kv_transfer.py
@@ -117,7 +117,6 @@ def _handle_request(self, req_meta: dict[str, Any]):
             addr_list.append(addr)
             size_list.append(size)
         if self.dcp_size > 1:
-            torch.npu.current_stream().synchronize()
             self.m_store.put(key_list, addr_list, size_list)
         else:
             key_list_tp = key_list[self.tp_rank % self.put_step::self.put_step]
@@ -126,7 +125,6 @@ def _handle_request(self, req_meta: dict[str, Any]):
             size_list_tp = size_list[self.tp_rank %
                                      self.put_step::self.put_step]
             if key_list_tp:
-                torch.npu.current_stream().synchronize()
                 self.m_store.put(key_list_tp, addr_list_tp, size_list_tp)
         if is_last_chunk:
             self.set_finished_request(req_id)
@@ -205,7 +203,6 @@ def _handle_request(  # type: ignore[override]
             addr_list.append(addr)
             size_list.append(size)
         if self.dcp_size > 1:
-            torch.npu.current_stream().synchronize()
             self.m_store.put(key_list, addr_list, size_list)
         else:
             key_list_tp = key_list[self.tp_rank % self.put_step::self.put_step]
@@ -214,7 +211,6 @@ def _handle_request(  # type: ignore[override]
             size_list_tp = size_list[self.tp_rank %
                                      self.put_step::self.put_step]
             if key_list_tp:
-                torch.npu.current_stream().synchronize()
                 self.m_store.put(key_list_tp, addr_list_tp, size_list_tp)
         if req_meta.layer_id == self.final_layer_id and req_meta.is_last_chunk:
             self.set_finished_request(req_meta.req_id)
diff --git a/vllm_ascend/worker/model_runner_v1.py b/vllm_ascend/worker/model_runner_v1.py
index 2e7c4ea299b..ce5848b3495 100644
--- a/vllm_ascend/worker/model_runner_v1.py
+++ b/vllm_ascend/worker/model_runner_v1.py
@@ -2339,7 +2339,6 @@ def execute_model(
                     attn_metadata, self.with_prefill, maybe_padded_num_tokens,
                     input_ids, positions, intermediate_tensors, inputs_embeds)
 
-            self.maybe_wait_for_kv_save()
             finished_sending, finished_recving = self.get_finished_kv_transfer(
                 scheduler_output)
 
@@ -2603,7 +2602,7 @@ def propose_draft_token_ids(sampled_token_ids):
                     # ngram and other speculative decoding methods use the sampled
                     # tokens on the CPU, so they are run after bookkeeping.
                     propose_draft_token_ids(valid_sampled_token_ids)
-
+            self.maybe_wait_for_kv_save()
             if has_kv_transfer_group():
                 get_kv_transfer_group().clear_connector_metadata()