set VLLM_USE_V1=1

vllm/engine/arg_utils.py

@@ -1745,7 +1745,8 @@ class AsyncEngineArgs(EngineArgs):
 
 
 def _raise_or_fallback(feature_name: str, recommend_to_remove: bool):
-    if envs.is_set("VLLM_USE_V1") and envs.VLLM_USE_V1:
+    # if envs.is_set("VLLM_USE_V1") and envs.VLLM_USE_V1:
+    if envs.VLLM_USE_V1:
         raise NotImplementedError(
             f"VLLM_USE_V1=1 is not supported with {feature_name}.")
     msg = f"{feature_name} is not supported by the V1 Engine. "
@@ -1757,7 +1758,8 @@ def _raise_or_fallback(feature_name: str, recommend_to_remove: bool):
 
 
 def _warn_or_fallback(feature_name: str) -> bool:
-    if envs.is_set("VLLM_USE_V1") and envs.VLLM_USE_V1:
+    # if envs.is_set("VLLM_USE_V1") and envs.VLLM_USE_V1:
+    if envs.VLLM_USE_V1:
         logger.warning(
             "Detected VLLM_USE_V1=1 with %s. Usage should "
             "be considered experimental. Please report any "

vllm/envs.py

@@ -665,7 +665,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
 
     # If set, use the V1 code path.
     "VLLM_USE_V1":
-    lambda: bool(int(os.getenv("VLLM_USE_V1", "0"))),
+    lambda: bool(int(os.getenv("VLLM_USE_V1", "1"))),
 
     # Disable aiter ops unless specifically enabled.
     # Acts as a parent switch to enable the rest of the other operations.