Skip to content

Reduced GPU requirements #272

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

Closed
wants to merge 1 commit into from
Closed

Reduced GPU requirements #272

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
4 changes: 2 additions & 2 deletions pkg/manifests/vllm/deployment.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ kind: Deployment
metadata:
name: vllm-llama2-7b-pool
spec:
replicas: 3
replicas: 1
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

with only 1 replica you cannot really see the benefits of the request scheduling algorithms. let's keep 3?

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Big plus 1 here. While I understand the desire to minimize the accelerator footprint of a getting started guide. Someone interested in an inference-tailored scheduling tool will have access to multiple accelerators to test this.

We should add comments clarifying why we are reducing the max-loras also. (L4s need this, but A100/H100s can go higher)

Copy link
Contributor

@danehans danehans Feb 3, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I agree with ^. However, we need a simple mechanism, e.g. kustomize, to allow users to change the default 3 to 1 or any other value. I've been using 1 in my test/dev env and can at least see the LoRA load balancing for my single node kind cluster. For example, create a kustomization.yaml file in pkg/manifests/vllm/:

apiVersion: kustomize.config.k8s.io/v1beta1
kind: Kustomization
resources:
  - deployment.yaml
replicas:
  - name: vllm-llama2-7b-pool
    count: 1

Run kubectl apply -k pkg/manifests/vllm/ and the vllm deployment will be run with 1 replica.

selector:
matchLabels:
app: vllm-llama2-7b-pool
Expand All @@ -39,7 +39,7 @@ spec:
- "8000"
- "--enable-lora"
- "--max-loras"
- "4"
- "2"
- "--max-cpu-loras"
- "12"
- "--lora-modules"
Expand Down