Skip to content

Commit 421d557

Browse files
update
Signed-off-by: eric-higgins-ai <[email protected]>
1 parent e290594 commit 421d557

File tree

1 file changed

+5
-3
lines changed

1 file changed

+5
-3
lines changed

vllm/executor/ray_utils.py

Lines changed: 5 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -16,6 +16,7 @@
1616
from vllm.sequence import ExecuteModelRequest, IntermediateTensors
1717
from vllm.utils import get_ip
1818
from vllm.worker.worker_base import WorkerWrapperBase
19+
import vllm.envs as envs
1920

2021
if TYPE_CHECKING:
2122
from vllm.v1.core.sched.output import SchedulerOutput
@@ -338,6 +339,7 @@ def initialize_ray_cluster(
338339
else:
339340
logger.info("No current placement group found. "
340341
"Creating a new placement group.")
342+
device_resource_request = envs.VLLM_RAY_PER_WORKER_GPUS
341343
num_devices_in_cluster = ray.cluster_resources().get(device_str, 0)
342344
# Log a warning message and delay resource allocation failure response.
343345
# Avoid immediate rejection to allow user-initiated placement group
@@ -349,7 +351,7 @@ def initialize_ray_cluster(
349351
device_str)
350352
# Create a new placement group
351353
placement_group_specs: List[Dict[str, float]] = ([{
352-
device_str: 1.0
354+
device_str: device_resource_request
353355
} for _ in range(parallel_config.world_size)])
354356

355357
# vLLM engine is also a worker to execute model with an accelerator,
@@ -358,11 +360,11 @@ def initialize_ray_cluster(
358360
current_ip = get_ip()
359361
current_node_id = ray.get_runtime_context().get_node_id()
360362
current_node_resource = available_resources_per_node()[current_node_id]
361-
if current_node_resource.get(device_str, 0) < 1:
363+
if current_node_resource.get(device_str, 0) < device_resource_request:
362364
raise ValueError(
363365
f"Current node has no {device_str} available. "
364366
f"{current_node_resource=}. vLLM engine cannot start without "
365-
f"{device_str}. Make sure you have at least 1 {device_str} "
367+
f"{device_str}. Make sure you have at least {device_resource_request} {device_str} "
366368
f"available in a node {current_node_id=} {current_ip=}.")
367369
# This way, at least bundle is required to be created in a current
368370
# node.

0 commit comments

Comments
 (0)