add shared_output and routed_scaling_factor of w4a8

vllm/model_executor/layers/fused_moe/fused_moe.py

@@ -1771,9 +1771,8 @@ def fused_experts_impl(
                                 use_nn_moe=use_nn_moe)
 
         if envs.VLLM_USE_LIGHTOP and not dpsk_fp16_quick: 
-            if shared_output is not None:
-                op.moe_sum(intermediate_cache3.view(*intermediate_cache3.size()),
-                        out_hidden_states[begin_chunk_idx:end_chunk_idx], shared_output[begin_chunk_idx:end_chunk_idx], None, routed_scaling_factor)
+            op.moe_sum(intermediate_cache3.view(*intermediate_cache3.size()),
+                    out_hidden_states[begin_chunk_idx:end_chunk_idx], shared_output[begin_chunk_idx:end_chunk_idx], None, routed_scaling_factor)
         # else:
         #     ops.moe_sum(intermediate_cache3.view(*intermediate_cache3.size()),
         #                 out_hidden_states[begin_chunk_idx:end_chunk_idx]) 

vllm/model_executor/layers/quantization/slimquant_w4a8.py

@@ -358,6 +358,7 @@ class SlimQuantW4A8Int8MoEMethod:
         use_nn_moe: Optional[bool] = False,
         routed_scaling_factor: Optional[float] = None,
         use_fused_gate: Optional[bool] = False,
+        shared_output: Optional[torch.Tensor] = None,
         **_  
     ) -> torch.Tensor:
         from vllm.model_executor.layers.fused_moe import fused_experts
@@ -398,4 +399,6 @@ class SlimQuantW4A8Int8MoEMethod:
             a1_scale=layer.w13_input_scale,
             a2_scale=layer.w2_input_scale,
             use_nn_moe=use_nn_moe,
+            shared_output=shared_output,
+            routed_scaling_factor=routed_scaling_factor,
         )