[Bugfix] Skip dead and non-GPU nodes for Ray DP engine allocation (#22275)

Signed-off-by: Rui Qiao <ruisearch42@gmail.com>

vllm/v1/engine/utils.py

@@ -297,10 +297,10 @@ class CoreEngineActorManager:
         local_engine_count = \
             vllm_config.parallel_config.data_parallel_size_local
 
-        nodes = sorted(list_nodes(),
+        nodes = sorted(list_nodes(filters=[("state", "=", "ALIVE")]),
                        key=lambda node: node.node_ip != dp_master_ip)
         assert nodes[0].node_ip == dp_master_ip, (
-            "The first node must be the head node")
+            "The head node is missing or dead")
         assert len(nodes) == 1 or nodes[1].node_ip != dp_master_ip, (
             "There can only be one head node")
 
@@ -312,6 +312,8 @@ class CoreEngineActorManager:
         for node in nodes:
             node_ip = node.node_ip
             node_resources = available_resources[node.node_id]
+            if "GPU" not in node_resources:
+                continue
             # For now, each DP rank can only be assigned to one node
             # TODO(rui): support allocating a single DP rank
             # to multiple nodes
@@ -346,6 +348,13 @@ class CoreEngineActorManager:
                     )
                     placement_groups.append(pg)
                     local_dp_ranks.append(i)
+        if len(placement_groups) < num_pg_to_create:
+            raise ValueError(
+                f"Not enough resources to allocate {num_pg_to_create} "
+                "placement groups, only created "
+                f"{len(placement_groups)} placement groups. "
+                "Available resources: "
+                f"{available_resources}")
         return placement_groups, local_dp_ranks
 
     @staticmethod