[Model] Fix bailing_moe accuracy problem (#28277)

Signed-off-by: zhaozx-cn <zhaozx2116@163.com>

vllm/model_executor/models/bailing_moe.py

@@ -39,7 +39,6 @@ from vllm.distributed import (
     get_pp_group,
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
-    tensor_model_parallel_all_reduce,
 )
 from vllm.model_executor.layers.activation import SiluAndMul
 from vllm.model_executor.layers.fused_moe import SharedFusedMoE
@@ -330,7 +329,9 @@ class BailingMoE(nn.Module):
             final_hidden_states = final_hidden_states + shared_output
 
         if self.tp_size > 1:
-            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
+            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
+                final_hidden_states
+            )
         return final_hidden_states.view(num_tokens, hidden_size)