[Misc] Better RayExecutor and multiprocessing compatibility (#14705)

Signed-off-by: Cody Yu <hao.yu.cody@gmail.com>

vllm/engine/arg_utils.py

@@ -26,7 +26,7 @@ from vllm.plugins import load_general_plugins
 from vllm.test_utils import MODEL_WEIGHTS_S3_BUCKET, MODELS_ON_S3
 from vllm.transformers_utils.utils import check_gguf_file
 from vllm.usage.usage_lib import UsageContext
-from vllm.utils import FlexibleArgumentParser, StoreBoolean
+from vllm.utils import FlexibleArgumentParser, StoreBoolean, is_in_ray_actor
 
 if TYPE_CHECKING:
     from vllm.transformers_utils.tokenizer_group import BaseTokenizerGroup
@@ -1245,6 +1245,18 @@ class EngineArgs:
             cpu_offload_gb=self.cpu_offload_gb,
             calculate_kv_scales=self.calculate_kv_scales,
         )
+
+        # Get the current placement group if Ray is initialized and
+        # we are in a Ray actor. If so, then the placement group will be
+        # passed to spawned processes.
+        placement_group = None
+        if is_in_ray_actor():
+            import ray
+
+            # This call initializes Ray automatically if it is not initialized,
+            # but we should not do this here.
+            placement_group = ray.util.get_current_placement_group()
+
         parallel_config = ParallelConfig(
             pipeline_parallel_size=self.pipeline_parallel_size,
             tensor_parallel_size=self.tensor_parallel_size,
@@ -1257,6 +1269,7 @@ class EngineArgs:
                 self.tokenizer_pool_extra_config,
             ),
             ray_workers_use_nsight=self.ray_workers_use_nsight,
+            placement_group=placement_group,
             distributed_executor_backend=self.distributed_executor_backend,
             worker_cls=self.worker_cls,
             worker_extension_cls=self.worker_extension_cls,