set VLLM_USE_TOPK_RENORM=1

vllm/envs.py

@@ -1235,7 +1235,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
     # vLLM will use optimized topk_softmax + renormalize
     "VLLM_USE_TOPK_RENORM":
         lambda:
-        (os.environ.get("VLLM_USE_TOPK_RENORM", "False").lower() in
+        (os.environ.get("VLLM_USE_TOPK_RENORM", "True").lower() in
                 ("true", "1")),
         
 }