diff --git a/vllm/executor/ray_distributed_executor.py b/vllm/executor/ray_distributed_executor.py
index 18ff32155c5f..708bc0ee1b2c 100644
--- a/vllm/executor/ray_distributed_executor.py
+++ b/vllm/executor/ray_distributed_executor.py
@@ -559,6 +559,15 @@ def _compiled_ray_dag(self, enable_asyncio: bool):
                     envs.VLLM_USE_RAY_COMPILED_DAG_NCCL_CHANNEL)
         logger.info("VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM = %s",
                     envs.VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM)
+        # Enlarge the default value of "RAY_CGRAPH_get_timeout" to 300 seconds
+        # (it is 10 seconds by default). This is a Ray environment variable to
+        # control the timeout of getting result from a compiled graph execution,
+        # i.e., the distributed execution that includes model forward runs and
+        # intermediate tensor communications, in the case of vllm.
+        os.environ.setdefault("RAY_CGRAPH_get_timeout", "300")  # noqa: SIM112
+        logger.info("RAY_CGRAPH_get_timeout is set to %s",
+                    os.environ["RAY_CGRAPH_get_timeout"])  # noqa: SIM112
+
         with InputNode() as input_data:
             # Example DAG: PP=2, TP=4
             #