diff --git a/lightllm/common/basemodel/layer_weights/meta_weights/fused_moe/impl/deepgemm_impl.py b/lightllm/common/basemodel/layer_weights/meta_weights/fused_moe/impl/deepgemm_impl.py
index aa080a19f..cfc82face 100644
--- a/lightllm/common/basemodel/layer_weights/meta_weights/fused_moe/impl/deepgemm_impl.py
+++ b/lightllm/common/basemodel/layer_weights/meta_weights/fused_moe/impl/deepgemm_impl.py
@@ -146,7 +146,7 @@ def select_experts_and_quant_input(
         n_group: int,
         scoring_func: str,
     ):
-        topk_weights, topk_idx = self._select_experts(
+        topk_weights, topk_idx, _ = self._select_experts(
             input_tensor=hidden_states,
             router_logits=router_logits,
             correction_bias=e_score_correction_bias,
diff --git a/lightllm/distributed/communication_op.py b/lightllm/distributed/communication_op.py
index f15badde2..07a69db0f 100644
--- a/lightllm/distributed/communication_op.py
+++ b/lightllm/distributed/communication_op.py
@@ -152,6 +152,13 @@ def new_deepep_group(
 
         global_world_size = get_global_world_size()
         deepep_group = dist.new_group(list(range(global_world_size)))
+        # DeepEP reuses this group's NCCL communicator via _comm_ptr(). Because the
+        # group is created without device_id, warm it up first to avoid reading a null
+        # communicator. The default process group's warmup does not cover this group.
+        dist.barrier(
+            group=deepep_group,
+            device_ids=[torch.cuda.current_device()],
+        )
         self.ll_num_tokens = prefill_num_max_dispatch_tokens_per_rank
         self.ll_decode_num_tokens = decode_num_max_dispatch_tokens_per_rank
         self.ll_hidden = hidden_size