Use standalone_compile by default in torch >= 2.8.0 (#18846)

Signed-off-by: rzou <zou3519@gmail.com>

vllm/compilation/backends.py

@@ -16,7 +16,7 @@ import vllm.envs as envs
 from vllm.config import CompilationConfig, VllmConfig
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
-from vllm.utils import resolve_obj_by_qualname
+from vllm.utils import is_torch_equal_or_newer, resolve_obj_by_qualname
 
 from .compiler_interface import (CompilerInterface, EagerAdaptor,
                                  InductorAdaptor, InductorStandaloneAdaptor)
@@ -29,7 +29,8 @@ logger = init_logger(__name__)
 
 def make_compiler(compilation_config: CompilationConfig) -> CompilerInterface:
     if compilation_config.use_inductor:
-        if envs.VLLM_TEST_STANDALONE_COMPILE:
+        if envs.VLLM_USE_STANDALONE_COMPILE and is_torch_equal_or_newer(
+                "2.8.0"):
             logger.info("Using InductorStandaloneAdaptor")
             return InductorStandaloneAdaptor()
         else:

vllm/compilation/compiler_interface.py

@@ -155,7 +155,7 @@ class InductorStandaloneAdaptor(CompilerInterface):
     This is not on by default yet, but we plan to turn it on by default for
     PyTorch 2.8.
 
-    Use VLLM_TEST_STANDALONE_COMPILE to toggle this on or off.
+    Use VLLM_USE_STANDALONE_COMPILE to toggle this on or off.
     """
     name = "inductor_standalone"
 

vllm/envs.py

@@ -143,10 +143,10 @@ def maybe_convert_int(value: Optional[str]) -> Optional[int]:
 
 def get_vllm_port() -> Optional[int]:
     """Get the port from VLLM_PORT environment variable.
-    
+
     Returns:
         The port number as an integer if VLLM_PORT is set, None otherwise.
-        
+
     Raises:
         ValueError: If VLLM_PORT is a URI, suggest k8s service discovery issue.
     """
@@ -308,9 +308,11 @@ environment_variables: dict[str, Callable[[], Any]] = {
     lambda: bool(
         os.environ.get("VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE", "1") != "0"),
 
-    # Internal flag to enable/disable Inductor standalone compile
-    "VLLM_TEST_STANDALONE_COMPILE":
-    lambda: os.environ.get("VLLM_TEST_STANDALONE_COMPILE", "0") != "0",
+    # Feature flag to enable/disable Inductor standalone compile.
+    # In torch <= 2.7 we ignore this flag; in torch >= 2.8 this is
+    # enabled by default.
+    "VLLM_USE_STANDALONE_COMPILE":
+    lambda: os.environ.get("VLLM_USE_STANDALONE_COMPILE", "1") == "1",
 
     # local rank of the process in the distributed setting, used to determine
     # the GPU device id
@@ -892,7 +894,7 @@ def compute_hash() -> str:
         "VLLM_USE_TRITON_AWQ",
         "VLLM_DP_RANK",
         "VLLM_DP_SIZE",
-        "VLLM_TEST_STANDALONE_COMPILE",
+        "VLLM_USE_STANDALONE_COMPILE",
     ]
     for key in environment_variables_to_hash:
         if key in environment_variables: