biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Distributed] Allow the placement group more time to wait for resources to be ready (#11138)

Browse Source 
Signed-off-by: Jiaxin Shan <seedjeffwan@gmail.com>
This commit is contained in:
Jiaxin Shan
2024-12-13 12:17:37 -08:00
committed by GitHub
parent 0a56bcc03d
commit 0d8451c3a4
1 changed files with 7 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
vllm/executor/ray_utils.py
View File

@@ -277,10 +277,14 @@ def initialize_ray_cluster(
f"Total number of devices: {device_bundles}.") f"Total number of devices: {device_bundles}.")
else: else:
num_devices_in_cluster = ray.cluster_resources().get(device_str, 0) 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
# created and wait cluster to be ready
if parallel_config.world_size > num_devices_in_cluster: if parallel_config.world_size > num_devices_in_cluster:
raise ValueError( logger.warning(
f"The number of required {device_str}s exceeds the total " "The number of required %ss exceeds the total "
f"number of available {device_str}s in the placement group.") "number of available %ss in the placement group.", device_str,
device_str)
# Create a new placement group # Create a new placement group
placement_group_specs: List[Dict[str, float]] = ([{ placement_group_specs: List[Dict[str, float]] = ([{
device_str: 1.0 device_str: 1.0