diff --git a/paddle/fluid/operators/lookup_table_v2_op_npu.cc b/paddle/fluid/operators/lookup_table_v2_op_npu.cc
index c75ea537216f3524ffec53fd199986021a096e49..020dbad53076d20a73f8aeadd8c20ca9157982a1 100644
--- a/paddle/fluid/operators/lookup_table_v2_op_npu.cc
+++ b/paddle/fluid/operators/lookup_table_v2_op_npu.cc
@@ -40,6 +40,7 @@ class LookupTableV2NPUKernel : public framework::OpKernel<T> {
         platform::errors::InvalidArgument("npu only accept LoDTensor"));
     output_t->mutable_data<T>(ctx.GetPlace());
 
+    // add copy ids to ensure ids_t is prepared.
     std::vector<int> ids;
     TensorToVector(*ids_t, ctx.device_context(), &ids);
 
diff --git a/python/paddle/distributed/parallel.py b/python/paddle/distributed/parallel.py
index efe747408428a68772726c28af469b975836511e..1a4fe21afa577fa2e6a1f82528f45acdd977d6b7 100644
--- a/python/paddle/distributed/parallel.py
+++ b/python/paddle/distributed/parallel.py
@@ -193,6 +193,12 @@ def init_parallel_env():
     elif core.is_compiled_with_xpu():
         parallel_helper._set_parallel_ctx(
             core.BKCLParallelContext(strategy, place))
+
+    other_endpoints = strategy.trainer_endpoints[:]
+    other_endpoints.remove(strategy.current_endpoint)
+    if strategy.local_rank == 0:
+        wait_server_ready(other_endpoints)
+
     parallel_helper._init_parallel_ctx()
 
     # 5: init gloo context (step 2: gloo init)