update VLLM_USE_CAT_MLA

vllm/envs.py

@@ -1198,7 +1198,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
         
     # vllm will use fused cat and mla
     "VLLM_USE_CAT_MLA":
-        lambda: (os.getenv('VLLM_USE_CAT_MLA', 'True').lower() in
+        lambda: (os.getenv('VLLM_USE_CAT_MLA', 'False').lower() in
                  ("true", "1")),     
 }
 

vllm/model_executor/model_loader/utils.py

@@ -253,6 +253,8 @@ def get_model_architecture(
                     os.environ['VLLM_USE_OPT_CAT'] = '1'
                 # if not envs.is_set("VLLM_USE_LIGHTOP_FILL_MOE_ALIGN"):
                 #     os.environ['VLLM_USE_LIGHTOP_FILL_MOE_ALIGN'] = '1'
+                if not envs.is_set("VLLM_USE_CAT_MLA"):
+                    os.environ['VLLM_USE_CAT_MLA'] = '1'
             else:
                 if not envs.is_set("VLLM_USE_PD_SPLIT"):
                     os.environ['VLLM_USE_PD_SPLIT'] = '1'
@@ -272,6 +274,8 @@ def get_model_architecture(
                     os.environ['VLLM_USE_OPT_CAT'] = '1'
                 # if not envs.is_set("VLLM_USE_LIGHTOP_FILL_MOE_ALIGN"):
                 #     os.environ['VLLM_USE_LIGHTOP_FILL_MOE_ALIGN'] = '1'
+                if not envs.is_set("VLLM_USE_CAT_MLA"):
+                    os.environ['VLLM_USE_CAT_MLA'] = '1'
             else:
                 if not envs.is_set("VLLM_USE_PD_SPLIT"):
                     os.environ['VLLM_USE_PD_SPLIT'] = '1'