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 14, 05:30 UTC
[Bugfix][DP] DP distribution does not require ray[default] (#23822)
Signed-off-by: Kebe <mail@kebe7jun.com>
Kebe committed
Sep 4, 2025 at 05:21 UTC
a43a3f1770525b7ac88151667da76bc2f15ec50d
1 file changed
+10
-14
vllm/v1/engine/utils.py
+10
-14
@@ -315,7 +315,6 @@ class CoreEngineActorManager:
315
316
import ray
317
from ray._private.state import available_resources_per_node
318
- from ray.util.state import list_nodes
318
319
logger.info("Creating placement groups for data parallel")
320
dp_master_ip = \
@@ -324,31 +323,28 @@ class CoreEngineActorManager:
323
local_engine_count = \
324
vllm_config.parallel_config.data_parallel_size_local
325
327
- nodes = sorted(list_nodes(filters=[("state", "=", "ALIVE")]),
328
- key=lambda node: node.node_ip != dp_master_ip)
329
- assert nodes[0].node_ip == dp_master_ip, (
330
- "The head node is missing or dead")
331
- assert len(nodes) == 1 or nodes[1].node_ip != dp_master_ip, (
332
- "There can only be one head node")
333
-
326
available_resources = available_resources_per_node()
327
world_size = vllm_config.parallel_config.world_size
328
placement_groups: list[PlacementGroup] = []
329
local_dp_ranks: list[int] = []
338
-
339
- for node in nodes:
340
- node_ip = node.node_ip
341
- node_resources = available_resources[node.node_id]
330
+ dp_master_ip_key = f'node:{dp_master_ip}'
331
+ nodes = sorted(available_resources.values(),
332
+ key=lambda x: dp_master_ip_key not in x)
333
+ assert len(nodes) > 0, (
334
+ "No nodes with resources found in Ray cluster.")
335
+ assert dp_master_ip_key in nodes[0], (
336
+ "The DP master node (ip: %s) is missing or dead", dp_master_ip)
337
+ for node_resources in nodes:
338
if "GPU" not in node_resources:
339
continue
340
# For now, each DP rank can only be assigned to one node
341
# TODO(rui): support allocating a single DP rank
342
# to multiple nodes
343
available_engine_count = int(node_resources["GPU"]) // world_size
348
- if node_ip == dp_master_ip:
344
+ if dp_master_ip_key in node_resources:
345
assert available_engine_count >= local_engine_count, (
346
"Not enough resources to allocate DP ranks "
351
- f"on DP master node {node_ip}")
347
+ f"on DP master node {dp_master_ip}")
348
for i in range(local_engine_count):
349
bundles = [{
350
"GPU": 1.0,