[executor] feat: add supports_pp attr to executors (#21786)

Signed-off-by: Haibin Lin <[email protected]>

Author: eric-haibin-lin

vllm/engine/arg_utils.py

@@ -1490,14 +1490,18 @@ def _is_v1_supported_oracle(self, model_config: ModelConfig) -> bool:
                 and _warn_or_fallback("Engine in background thread")):
             return False
 
-        if (self.pipeline_parallel_size > 1
-                and self.distributed_executor_backend
-                not in (ParallelConfig.distributed_executor_backend, "ray",
-                        "mp", "external_launcher")):
-            name = "Pipeline Parallelism without Ray distributed executor " \
-                    "or multiprocessing executor or external launcher"
-            _raise_or_fallback(feature_name=name, recommend_to_remove=False)
-            return False
+        if self.pipeline_parallel_size > 1:
+            supports_pp = getattr(self.distributed_executor_backend,
+                                  'supports_pp', False)
+            if not supports_pp and self.distributed_executor_backend not in (
+                    ParallelConfig.distributed_executor_backend, "ray", "mp",
+                    "external_launcher"):
+                name = "Pipeline Parallelism without Ray distributed " \
+                        "executor or multiprocessing executor or external " \
+                        "launcher"
+                _raise_or_fallback(feature_name=name,
+                                   recommend_to_remove=False)
+                return False
 
         # The platform may be supported on V1, but off by default for now.
         if not current_platform.default_v1(  # noqa: SIM103

vllm/executor/executor_base.py

@@ -35,6 +35,7 @@ class ExecutorBase(ABC):
     """
 
     uses_ray: bool  # whether the executor uses Ray for orchestration.
+    supports_pp: bool = False  # whether the executor supports PP
 
     def __init__(
         self,

vllm/v1/executor/multiproc_executor.py

@@ -41,6 +41,8 @@
 
 class MultiprocExecutor(Executor):
 
+    supports_pp: bool = True
+
     def _init_executor(self) -> None:
         # Call self.shutdown at exit to clean up
         # and ensure workers will be terminated.

vllm/v1/executor/ray_distributed_executor.py

@@ -43,6 +43,8 @@ def result(self, timeout=None):
 class RayDistributedExecutor(RayDistributedExecutorV0, Executor):
     """Ray distributed executor using Ray Compiled Graphs."""
 
+    supports_pp: bool = True
+
     def _init_executor(self) -> None:
         super()._init_executor()