Hi,
We are running ray 2.9.2 on K8s.
We scheduled this placement group that contains 2 bundles on a deployment. And from ray list placement-groups
it looks like the placement-group is already created.
- placement_group_id: 9b504ae1b1d536d1e9c2242d5d5801000000
name: SERVE_REPLICA::vllm-playground-2#VLLMDeployment:Meta--Llama-2-7b-chat-hf#mWisbK
creator_job_id: '01000000'
state: CREATED
bundles:
- bundle_id:
placement_group_id: 9b504ae1b1d536d1e9c2242d5d5801000000
bundle_index: 0
unit_resources:
accelerator_type_g5_2xl: 0.01
GPU: 1.0
CPU: 1.0
node_id: 6695428f77f9796c5fca6c564674a849724269a7c951389cf97d6de5
- bundle_id:
placement_group_id: 9b504ae1b1d536d1e9c2242d5d5801000000
bundle_index: 1
unit_resources:
accelerator_type_g5_2xl: 0.01
GPU: 1.0
CPU: 2.0
node_id: 848f42d243f2175b7ed8319a2d25679b377ccf47c937d4730ec2ae21
is_detached: true
But in the ray status, looks like the demand is pending being allocated to the second bundle.
Demands:
{'CPU': 2.0, 'GPU': 1.0, 'accelerator_type_g5_2xl': 0.01}: 1+ pending tasks/actors (1+ using placement groups)
I am wondering how we can better debug this?