diff --git a/vllm/executor/ray_gpu_executor.py b/vllm/executor/ray_gpu_executor.py
index 261fcfb7dad9..888a8056448f 100644
--- a/vllm/executor/ray_gpu_executor.py
+++ b/vllm/executor/ray_gpu_executor.py
@@ -393,6 +393,7 @@ async def _run_workers_async(
         self,
         method: str,
         *args,
+        use_ray_compiled_dag: bool = False,
         driver_args: Optional[List[Any]] = None,
         driver_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs,
@@ -409,12 +410,27 @@ async def _run_workers_async(
         driver_executor = make_async(getattr(self.driver_worker, method))
         coros.append(driver_executor(*driver_args, **driver_kwargs))
 
-        # Run the ray workers asynchronously.
-        for worker in self.workers:
-            coros.append(worker.execute_method.remote(method, *args, **kwargs))
+        if use_ray_compiled_dag:
+            # Right now, compiled DAG can only accept a single
+            # input. TODO(sang): Fix it.
+            output_channels = self.forward_dag.execute(1)
+            try:
+                ray_worker_outputs = [
+                    pickle.loads(chan.begin_read()) for chan in output_channels
+                ]
+            finally:
+                # Has to call end_read in order to reuse the DAG.
+                for chan in output_channels:
+                    chan.end_read()
+            return [await asyncio.gather(*coros)] + ray_worker_outputs
+        else:
+            # Run the ray workers asynchronously.
+            for worker in self.workers:
+                coros.append(
+                    worker.execute_method.remote(method, *args, **kwargs))
 
-        all_outputs = await asyncio.gather(*coros)
-        return all_outputs
+            all_outputs = await asyncio.gather(*coros)
+            return all_outputs
 
     async def execute_model_async(
         self,