Helm Chart Lacks Clear Support for Multi-Node vLLM Deployment

[shohamyamin]

The current Helm chart does not explicitly support deploying vLLM across multiple vllm nodes on Kubernetes, or it's unclear how to configure it. Improved documentation or multi-node support is needed for deploying LLM that require multi-node

[ahg-g]

For reference, and if someone would like to add that to helm charts, the LeaderWorkerSet API can be used to deploy multi-node vllm on k8s: https://docs.vllm.ai/en/latest/deployment/frameworks/lws.html

Also check examples on the LWS repo: https://github.com/kubernetes-sigs/lws/tree/main/docs/examples/vllm

[YuhanLiu11]

Thanks for submitting the issue!

Multi-node deployment should be supported. If you encounter any issues running a multi-node deployment, feel free to let us know.

We will improve the documentation to clarify how to configure multi-node deployment on Kubernetes.

[ApostaC]

@shohamyamin we recently did some local tests and will update the helm charts and the docs soon!

[shohamyamin]

@ApostaC great to hear that. Did you know if this configuration can be run on a rootless environment(like Openshift or other k8s rootless environments)?

[ApostaC]

Yeah, you should be able to do that.

I tried to setup a rootless k8s environment with kubeadm + 2 physical nodes recently and successfully got it run.

[shohamyamin]

Great I will try at the moment that the chart will be updated

[ApostaC]

@shohamyamin Hey, the tensor parallelism support is added in PR #105

[simon-mo]

On LWS, it doesn't work out of the box on vLLM container image. Please submit a PR to vLLM repo so the vLLM container has the ability to start ray cluster.

[0xThresh]

While this is specific to EKS, it looks like AWS has a Dockerfile that can support it for AWS here: https://aws-ia.github.io/terraform-aws-eks-blueprints/patterns/machine-learning/multi-node-vllm/#dockerfile

It also looks like there is a PR already open to help address this, so I'll keep an eye on that: vllm-project/vllm#12913

[ahg-g]

I think we can make the multi-node deployment generic if we have the following script as part of the vllm image: vllm-project/vllm#12913