Skip to main content

​
Overview

Use this guide to change the set of models served by a running inference release: adding a new model, replacing a model’s checkpoint, or removing a model. You edit your inference_values.yaml file and run helm upgrade; the chart reconciles the model Deployments, Services, and Ingress objects to match. You can make these changes on their own against the current chart version, or apply them as part of a chart upgrade to a new Poolside bundle. To upgrade the chart, see Upgrade on Amazon EKS; make the model edits described here in the same inference_values.yaml file before you run helm upgrade.

​
Prerequisites

  • A working deployment completed with the Install on Amazon EKS guide.
  • The customized inference_values.yaml file you used to install.
  • The new model checkpoint, provided by Poolside.
  • Workstation tools:
    • helm 3.12 or later
    • kubectl, configured for your EKS cluster
    • aws CLI, to upload checkpoints to S3
    • jq, to parse JSON responses from the inference API

​
Downtime

Adding a model does not affect models that are already serving. Updating a checkpoint rolls that model’s Deployment, and the model server re-downloads the checkpoint from S3 on restart, so expect a delay before it becomes ready again. Plan a maintenance window for single-replica models.

​
Add a model

Extract the new checkpoint archive as described in Upload model checkpoints to S3 so its files sit at the prefix root, then upload it to your S3 bucket. Use a distinct prefix per model: 
aws s3 cp ./checkpoints/<new-model> s3://<bucket-name>/checkpoints/<new-model> \
  --recursive \
  --region <aws-region>
For checkpoint upload details such as concurrency throttling, see Upload model checkpoints to S3. Add a new key under models in your inference_values.yaml file. Give the model its own ingressHost, covered by the ACM certificate referenced in ingress.annotations:
inference_values.yaml
models:
  # ...existing models...
  malibu:
    model: s3://<bucket-name>/checkpoints/<new-model>
    modelName: Malibu
    modelType: agent
    gpus: 2
    ingressHost: <malibu-hostname>
Apply the change with helm upgrade: 
helm upgrade inference ./charts/inference \
  --namespace poolside-models \
  -f ./inference_values.yaml
The chart creates a new Deployment, Service, and Ingress named inference-<model-key> for the model. Confirm the new pod starts and the ingress is created: 
kubectl get pods -n poolside-models
kubectl get ingress inference-<model-key> -n poolside-models
Create a DNS record for the new ingressHost, pointing it at the load balancer address.

​
Update a model checkpoint

Upload the new checkpoint to a new, versioned prefix rather than overwriting the existing one. A new path lets helm upgrade detect the change and roll the Deployment automatically, and it lets you roll back by pointing at the previous path. Extract the archive first, as in Step 3, so the files sit at the prefix root: 
aws s3 cp ./checkpoints/<model-key>-<version> s3://<bucket-name>/checkpoints/<model-key>-<version> \
  --recursive \
  --region <aws-region>
Point the model’s model field at the new path in your inference_values.yaml file. Update modelName only if the served model name changes:
inference_values.yaml
models:
  laguna-m:
    model: s3://<bucket-name>/checkpoints/laguna-m-<version>
    modelName: Laguna
    modelType: agent
    gpus: 4
    ingressHost: <laguna-m-hostname>
Apply the change: 
helm upgrade inference ./charts/inference \
  --namespace poolside-models \
  -f ./inference_values.yaml
The model’s Deployment rolls, and the init container downloads the new checkpoint on startup. Watch the rollout: 
kubectl rollout status deploy/inference-<model-key> -n poolside-models
If you reuse the same S3 path instead of a versioned one, helm upgrade detects no change to the values and does not restart the model. Force a restart so the init container re-downloads the checkpoint:
kubectl rollout restart deploy/inference-<model-key> -n poolside-models

​
Remove a model

Delete the model’s key from models in your inference_values.yaml file, then apply the change: 
helm upgrade inference ./charts/inference \
  --namespace poolside-models \
  -f ./inference_values.yaml
The chart removes that model’s Deployment, Service, and Ingress. Confirm the resources are gone: 
kubectl get deploy,svc,ingress -n poolside-models -l app.kubernetes.io/component=inference
If you no longer need the model’s checkpoint, delete it from the bucket: 
aws s3 rm s3://<bucket-name>/checkpoints/<model-key> --recursive --region <aws-region>
You can also remove the model’s DNS record once the ingress is gone.

​
Verification

Confirm a model serves traffic, where <model-hostname> is the ingressHost of that model: 
curl -s https://<model-hostname>/v1/models \
  -H "Authorization: Bearer <vllm-api-key>" \
  | jq -r '.data[].id'
If API key authentication is off, omit the Authorization header. For questions about model checkpoints or hardware requirements, contact Poolside support.