comment

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

Author: comaniac

vllm/executor/ray_utils.py

@@ -284,8 +284,9 @@ def initialize_ray_cluster(
     assert_ray_available()
     from vllm.platforms import current_platform
 
-    # Connect to a ray cluster.
-    if current_platform.is_rocm() or current_platform.is_xpu():
+    if ray.is_initialized():
+        logger.info("Ray is already initialized. Skipping Ray initialization.")
+    elif current_platform.is_rocm() or current_platform.is_xpu():
         # Try to connect existing ray instance and create a new one if not found
         try:
             ray.init("auto", ignore_reinit_error=True)
@@ -312,6 +313,8 @@ def initialize_ray_cluster(
         current_placement_group = ray.util.get_current_placement_group()
 
     if current_placement_group:
+        logger.info("Using the existing placement group")
+
         # We are in a placement group
         bundles = current_placement_group.bundle_specs
         # Verify that we can use the placement group.
@@ -331,6 +334,8 @@ def initialize_ray_cluster(
                 f"Required number of devices: {parallel_config.world_size}. "
                 f"Total number of devices: {device_bundles}.")
     else:
+        logger.info("No current placement group found. "
+                    "Creating a new placement group.")
         num_devices_in_cluster = ray.cluster_resources().get(device_str, 0)
         # Log a warning message and delay resource allocation failure response.
         # Avoid immediate rejection to allow user-initiated placement group