A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 2, 09:45 UTC
[Bugfix] Fix Ray placement group allocation with grouped nodes (#43998)
Signed-off-by: <conway.zhu@cohere.com> Signed-off-by: root <conway.zhu@cohere.com>
czhu-cohere committed
May 29, 2026 at 11:51 UTC
8b9deeec4b7ad889c62f31e2eec2d2537c780685
1 file changed
+6
-1
vllm/v1/engine/utils.py
+6
-1
@@ -575,7 +575,9 @@ class CoreEngineActorManager:
575
node_ip_keys = [
576
key
577
for key in node_resources
578
- if key != "node:__internal_head__" and key.startswith("node:")
578
+ if key != "node:__internal_head__"
579
+ and key.startswith("node:")
580
+ and "_group_" not in key
581
]
582
assert len(node_ip_keys) == 1, (
583
f"Zero or multiple node IP keys found in node resources: {node_ip_keys}"
@@ -654,6 +656,9 @@ class CoreEngineActorManager:
656
if len(placement_groups) == dp_size:
657
break
658
659
+ if len(placement_groups) == dp_size:
660
+ break
661
+
662
if len(placement_groups) < dp_size:
663
raise ValueError(
664
f"Not enough resources to allocate {dp_size} "