Lab 7: Run SGLang on HAMi GPU Shares

进阶时长： about 45 minutes环境： Kubernetes cluster with NVIDIA GPUs

This lab demonstrates how to deploy SGLang, a high-performance LLM serving framework optimized for RadixAttention, on a Kubernetes cluster using HAMi for GPU memory and compute isolation. Upon completion, you will have an OpenAI-compatible model service running SGLang on a partitioned GPU.

Learning Objectives

• Understand the benefits of SGLang for high-throughput inference
• Run SGLang using HAMi's nvidia.com/gpu, nvidia.com/gpumem, and nvidia.com/gpucores resources
• Test the SGLang OpenAI-compatible API via port forwarding

Lab Overview

HAMi + SGLang Lab Flowchart

Deployment Architecture

HAMi + SGLang Deployment Architecture

Prerequisites

• A working Kubernetes cluster with HAMi installed
• At least 1 NVIDIA GPU node with sufficient memory for the target model
• kubectl connected to the cluster
• The cluster can pull SGLang images (lmsysorg/sglang:latest) and access Hugging Face models

Step 1: Create the SGLang Deployment

We will create a Kubernetes Deployment for SGLang, requesting a specific partition of the GPU using HAMi's extended resources.

Save the following YAML as sglang-deployment.yaml:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: sglang-llama3
  labels:
    app: sglang
spec:
  replicas: 1
  selector:
    matchLabels:
      app: sglang
  template:
    metadata:
      labels:
        app: sglang
    spec:
      containers:
        - name: sglang
          image: lmsysorg/sglang:latest
          command:
            - python3
            - "-m"
            - "sglang.launch_server"
            - "--model-path"
            - "meta-llama/Meta-Llama-3-8B-Instruct"
            - "--host"
            - "0.0.0.0"
            - "--port"
            - "30000"
          ports:
            - containerPort: 30000
          resources:
            limits:
              nvidia.com/gpu: 1
              nvidia.com/gpumem: 24000
              nvidia.com/gpucores: 50
          env:
            - name: HUGGING_FACE_HUB_TOKEN
              value: "YOUR_HF_TOKEN" # Replace with your Hugging Face token

Key Resource Constraints:

• nvidia.com/gpu: 1: Requests 1 virtual GPU slot.
• nvidia.com/gpumem: 24000: Allocates exactly 24GB of GPU memory to the SGLang worker.
• nvidia.com/gpucores: 50: Allocates 50% of the physical GPU's compute capacity.

Apply the deployment:

kubectl apply -f sglang-deployment.yaml

Step 2: Create the Service

Expose the SGLang deployment internally using a ClusterIP service.

Save as sglang-service.yaml:

apiVersion: v1
kind: Service
metadata:
  name: sglang-service
spec:
  selector:
    app: sglang
  ports:
    - protocol: TCP
      port: 30000
      targetPort: 30000

Apply the service:

kubectl apply -f sglang-service.yaml

Step 3: Verify Inference

Wait for the SGLang pod to become Running. Since it needs to download the Llama 3 weights, this might take several minutes depending on your network speed.

kubectl get pods -l app=sglang -w

Once running, port-forward the service to your local machine:

kubectl port-forward svc/sglang-service 30000:30000

Open a new terminal and test the OpenAI-compatible endpoint:

curl http://localhost:30000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "meta-llama/Meta-Llama-3-8B-Instruct",
    "messages": [{"role": "user", "content": "Explain GPU virtualization in one sentence."}],
    "temperature": 0.7
  }'

If successful, SGLang will return a generated response powered by your HAMi-partitioned GPU!

Step 4: Cleanup

To remove the lab resources from your cluster, run:

kubectl delete -f sglang-service.yaml
kubectl delete -f sglang-deployment.yaml