[V1] Implement vLLM V1 [1/N] (#9289)

vllm/envs.py

@@ -68,6 +68,7 @@ if TYPE_CHECKING:
     VLLM_TORCH_COMPILE_LEVEL: int = 0
     VLLM_CUSTOM_OPS: List[str] = []
     VLLM_DISABLED_KERNELS: List[str] = []
+    VLLM_USE_V1: bool = False
 
 
 def get_default_cache_root():
@@ -450,6 +451,10 @@ environment_variables: Dict[str, Callable[[], Any]] = {
     "VLLM_DISABLED_KERNELS":
     lambda: [] if "VLLM_DISABLED_KERNELS" not in os.environ else os.environ[
         "VLLM_DISABLED_KERNELS"].split(","),
+
+    # If set, use the V1 code path.
+    "VLLM_USE_V1":
+    lambda: bool(int(os.getenv("VLLM_USE_V1", "0"))),
 }
 
 # end-env-vars-definition