text
stringlengths
0
59.1k
key: hf_token
volumeMounts:
- mountPath: /dev/shm
name: dshm
volumes:
- name: dshm
emptyDir:
medium: Memory
# Node selectors are the main difference among the cloud providers,
# making sure vLLM pods land on Nodes with the correct GPU. The
# following are node selector examples for three cloud providers.
#
# - GKE
# nodeSelector:
# cloud.google.com/gke-accelerator: nvidia-l4
# cloud.google.com/gke-gpu-driver-version: default
#
# - EKS
# nodeSelector:
# node.kubernetes.io/instance-type: p4d.24xlarge
#
# - AKS
# nodeSelector:
# agentpiscasi.com/gpu: "true" # Common label for AKS GPU nodes
<|endoftext|>
# source: k8s_examples/AI/vllm-deployment/README.md type: docs
# AI Inference with vLLM on Kubernetes
## Purpose / What You'll Learn
This example demonstrates how to deploy a server for AI inference using [vLLM](https://docs.vllm.ai/en/latest/) on Kubernetes. You’ll learn how to:
- Set up vLLM inference server with a model downloaded from [Hugging Face](https://huggingface.co/).
- Expose the inference endpoint using a Kubernetes `Service`.
- Set up port forwarding from your local machine to the inference `Service` in the Kubernetes cluster.
- Send a sample prediction request to the server using `curl`.
---
## 📚 Table of Contents
- [Prerequisites](#prerequisites)
- [Detailed Steps & Explanation](#detailed-steps--explanation)
- [Verification / Seeing it Work](#verification--seeing-it-work)
- [Configuration Customization](#configuration-customization)
- [Platform-Specific Configuration](#platform-specific-configuration)
- [Cleanup](#cleanup)
- [Further Reading / Next Steps](#further-reading--next-steps)
---
## Prerequisites
- A Kubernetes cluster with access to NVIDIA GPUs. This example was tested on GKE, but can be adapted for other cloud providers like EKS and AKS by ensuring you have a GPU-enabled node pool and have deployed the Nvidia device plugin.
- Hugging Face account token with permissions for model (example model: `google/gemma-3-1b-it`)
- `kubectl` configured to communicate with cluster and in PATH
- `curl` binary in PATH
**Note for GKE users:** To target specific GPU types, you can uncomment the GKE-specific `nodeSelector` in `vllm-deployment.yaml`.
---
## Detailed Steps & Explanation
1. Create a namespace. This example uses `vllm-example`, but you can choose any name:
```bash
kubectl create namespace vllm-example
```
2. Ensure Hugging Face permissions to retrieve model:
```bash
# Env var HF_TOKEN contains hugging face account token
# Make sure to use the same namespace as in the previous step
kubectl create secret generic hf-secret -n vllm-example \
--from-literal=hf_token=$HF_TOKEN
```
3. Apply vLLM server:
```bash
# Make sure to use the same namespace as in the previous steps
kubectl apply -f vllm-deployment.yaml -n vllm-example
```
- Wait for deployment to reconcile, creating vLLM pod(s):
```bash
kubectl wait --for=condition=Available --timeout=900s deployment/vllm-gemma-deployment -n vllm-example
kubectl get pods -l app=gemma-server -w -n vllm-example
```
- View vLLM pod logs:
```bash
kubectl logs -f -l app=gemma-server -n vllm-example
```