Skip to content

Keep gpu_4_queue at 10 minimum with demand scaling to 40 #471

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
khluu wants to merge 3 commits into
base: main
Choose a base branch
from
Draft

Keep gpu_4_queue at 10 minimum with demand scaling to 40 #471

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
ea90602
Pin gpu_4_queue at 40 persistent instances
khluu Aug 13, 2026
3cd34a2
Scale gpu_4_queue down to 10 persistent instances
khluu Aug 14, 2026
474266f
Allow gpu_4_queue to scale from 10 to 40
khluu Aug 14, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 7 additions & 4 deletions terraform/aws/cloudformation_stack.tf
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -213,14 +213,15 @@ locals {
BuildkiteAgentTokenParameterStorePath = data.aws_ssm_parameter.bk_agent_token_cluster_ci.name
BuildkiteQueue = "gpu_4_queue"
InstanceTypes = "g6.12xlarge" # 4 Nvidia L4 GPUs, 192GB memory
MaxSize = 64
MinSize = 10
MaxSize = 40
ECRAccessPolicy = "readonly"
InstanceOperatingSystem = "linux"
OnDemandPercentage = 100
ImageId = "ami-040f1b73b7a7c7453" # Custom AMI with Nvidia driver 570.133.20
BootstrapScriptUrl = "https://vllm-ci.s3.us-west-2.amazonaws.com/bootstrap.sh"
# Recycle each GPU instance after a single job completes.
BuildkiteTerminateInstanceAfterJob = true
# Keep a persistent baseline while allowing demand-based scale-out.
BuildkiteTerminateInstanceAfterJob = false
}
}

Expand Down Expand Up @@ -373,7 +374,9 @@ resource "aws_cloudformation_stack" "bk_queue_ci_gpu" {
name = "bk-${each.key}"
parameters = { for k, v in each.value : k => v if k != "elastic_ci_stack_version" }

template_url = "https://s3.amazonaws.com/buildkite-aws-stack/v${each.value["elastic_ci_stack_version"]}/aws-stack.yml"
# v6.21.0 forces a blue/green ASG replacement for parameter updates. That
# cannot converge for gpu_4_queue when scarce capacity is already in use.
template_url = each.key == "gpu-4-queue-ci" ? "https://vllm-ci.s3.us-west-2.amazonaws.com/infra/buildkite-aws-stack/v6.21.0/aws-stack-fixed-capacity-b77222719ba6.yml" : "https://s3.amazonaws.com/buildkite-aws-stack/v${each.value["elastic_ci_stack_version"]}/aws-stack.yml"
capabilities = ["CAPABILITY_IAM", "CAPABILITY_NAMED_IAM", "CAPABILITY_AUTO_EXPAND"]

lifecycle {
Expand Down