VLLM_USE_LIGHTOP and VLLM_USE_OPT_CAT

add shared_output and routed_scaling_factor of CompressedTensorsW8A8Int8MoEMethod

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py

@@ -1097,7 +1097,7 @@ class CompressedTensorsW8A8Int8MoEMethod(CompressedTensorsMoEMethod):
         expert_load_view: Optional[torch.Tensor] = None,
         logical_to_physical_map: Optional[torch.Tensor] = None,
         logical_replica_count: Optional[torch.Tensor] = None,
-        
+        shared_output: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
         if enable_eplb:
             raise NotImplementedError(
@@ -1137,7 +1137,9 @@ class CompressedTensorsW8A8Int8MoEMethod(CompressedTensorsMoEMethod):
             w2_scale=layer.w2_weight_scale,
             a1_scale=layer.w13_input_scale,
             a2_scale=layer.w2_input_scale,
-            use_nn_moe=False)
+            use_nn_moe=False,
+            shared_output=shared_output,
+            routed_scaling_factor=routed_scaling_factor)
 
 
 class CompressedTensorsWNA16MarlinMoEMethod(CompressedTensorsMoEMethod):

vllm/model_executor/model_loader/utils.py

@@ -244,7 +244,7 @@ def get_model_architecture(
             else:
                 os.environ['LM_NN'] = '1'
             
-            if (architectures == ['DeepseekV3ForCausalLM'] or architectures == ['DeepSeekMTPModel']):
+            if architectures in [['DeepseekV3ForCausalLM'], ['DeepSeekMTPModel']]:
                 if not envs.is_set("VLLM_USE_LIGHTOP"):
                     os.environ['VLLM_USE_LIGHTOP'] = '1'
                 if not envs.is_set("VLLM_USE_OPT_CAT"):
@@ -254,6 +254,12 @@ def get_model_architecture(
                 os.environ['GEMM_PAD'] = '0'
             if os.getenv('FA_PAD') != '1': 
                 os.environ['FA_PAD'] = '0'
+        else:
+            if architectures in [['DeepseekV3ForCausalLM'], ['DeepSeekMTPModel']]:
+                if not envs.is_set("VLLM_USE_LIGHTOP"):
+                    os.environ['VLLM_USE_LIGHTOP'] = '1'
+                if not envs.is_set("VLLM_USE_OPT_CAT"):
+                    os.environ['VLLM_USE_OPT_CAT'] = '1'
         # awq相关配置
         try:
             if os.getenv('AWQ_MOE_SZ') == None: