set VLLM_USE_OPT_MOE_SUM=1 and VLLM_USE_LIGHTOP_MOE_SUM=1

vllm/envs.py

@@ -1112,11 +1112,11 @@ environment_variables: dict[str, Callable[[], Any]] = {
                  ("true", "1")),  
     # vLLM will use lightop moe_sum 
     "VLLM_USE_LIGHTOP_MOE_SUM":
-        lambda: (os.environ.get("VLLM_USE_LIGHTOP_MOE_SUM", "False").lower() in
+        lambda: (os.environ.get("VLLM_USE_LIGHTOP_MOE_SUM", "True").lower() in
                  ("true", "1")),  
     # vLLM will use lightop moe_align_block_size 
     "VLLM_USE_LIGHTOP_MOE_ALIGN":
-        lambda: (os.environ.get("VLLM_USE_LIGHTOP_MOE_ALIGN", "False").lower() in
+        lambda: (os.environ.get("VLLM_USE_LIGHTOP_MOE_ALIGN", "True").lower() in
                  ("true", "1")),     
     # vLLM will use opt merge_aatn_states, not triton
     "VLLM_USE_MERGE_ATTN_STATES_OPT":