optimize the implementation of moe_sum (lightop)

vllm/model_executor/layers/fused_moe/fused_moe.py

@@ -1920,6 +1920,10 @@ def fused_experts_impl(
         else:
             if envs.VLLM_USE_OPT_MOE_SUM:
                 moe_reduce_dispatch(intermediate_cache3.view(*intermediate_cache3.size()), out_hidden_states[begin_chunk_idx:end_chunk_idx], begin_chunk_idx, end_chunk_idx)
+            elif envs.VLLM_USE_LIGHTOP:
+                op.moe_sum(input=intermediate_cache3.view(*intermediate_cache3.size()),
+                    output=out_hidden_states[begin_chunk_idx:end_chunk_idx], bias=None, 
+                    expert_mask=None, num_local_tokens=None, factor=None)
             else:
                 ops.moe_sum(intermediate_cache3.view(*intermediate_cache3.size()),
                                 out_hidden_states[begin_chunk_idx:end_chunk_idx])