|
| 1 | +# WVA Stack Deployment Script |
| 2 | + |
| 3 | +This directory contains a deployment script for setting up the Workload-Variant-Autoscaler (WVA) stack for E2E testing. |
| 4 | + |
| 5 | +## Table of Contents |
| 6 | + |
| 7 | +- [Overview](#overview) |
| 8 | +- [Prerequisites](#prerequisites) |
| 9 | +- [Quick Start](#quick-start) |
| 10 | +- [Configuration](#configuration) |
| 11 | +- [Usage Examples](#usage-examples) |
| 12 | +- [Verification](#verification) |
| 13 | +- [Viewing Logs](#viewing-logs) |
| 14 | +- [Cleanup](#cleanup) |
| 15 | +- [Kind Cluster Configuration](#kind-cluster-configuration) |
| 16 | + |
| 17 | +## Overview |
| 18 | + |
| 19 | +The `deploy-wva-stack.sh` script automates the deployment of: |
| 20 | +- WVA Controller |
| 21 | +- llm-d Infrastructure |
| 22 | +- Prometheus Monitoring |
| 23 | +- Prometheus Adapter |
| 24 | +- Optional: VariantAutoscaling (VA) |
| 25 | +- Optional: HorizontalPodAutoscaler (HPA) |
| 26 | + |
| 27 | +The script clones the official WVA repository from GitHub and uses the deployment scripts provided there. |
| 28 | + |
| 29 | +## Prerequisites |
| 30 | + |
| 31 | +- `kubectl` configured with cluster access |
| 32 | +- `helm` installed |
| 33 | +- `git` installed |
| 34 | +- `HF_TOKEN` environment variable set (for llm-d deployment) |
| 35 | + |
| 36 | +## Quick Start |
| 37 | + |
| 38 | +### Basic Deployment (WVA + llm-d + Prometheus) |
| 39 | + |
| 40 | +**On existing Kubernetes cluster:** |
| 41 | +```bash |
| 42 | +export HF_TOKEN="hf_xxxxx" |
| 43 | +./deploy-wva-stack.sh |
| 44 | +``` |
| 45 | + |
| 46 | +**On Kind cluster (emulated GPUs):** |
| 47 | +```bash |
| 48 | +export KIND_CLUSTER_NAME="kind-fma-cluster" |
| 49 | +./deploy-wva-stack.sh --kind |
| 50 | +``` |
| 51 | + |
| 52 | +This deploys: |
| 53 | +- ✅ WVA Controller |
| 54 | +- ✅ llm-d Infrastructure (with simulator for Kind) |
| 55 | +- ✅ Prometheus |
| 56 | +- ✅ Prometheus Adapter |
| 57 | +- ❌ VariantAutoscaling (opt-in) |
| 58 | +- ❌ HPA (opt-in) |
| 59 | + |
| 60 | +### Deployment with HPA and VA |
| 61 | + |
| 62 | +**On existing cluster:** |
| 63 | +```bash |
| 64 | +export HF_TOKEN="hf_xxxxx" |
| 65 | +./deploy-wva-stack.sh --with-hpa --with-va |
| 66 | +``` |
| 67 | + |
| 68 | +**On Kind cluster:** |
| 69 | +```bash |
| 70 | +export KIND_CLUSTER_NAME="kind-fma-cluster" |
| 71 | +./deploy-wva-stack.sh --kind --with-hpa --with-va |
| 72 | +``` |
| 73 | + |
| 74 | +### Create Kind Cluster and Deploy |
| 75 | + |
| 76 | +```bash |
| 77 | +# Create Kind cluster and deploy full stack |
| 78 | +./deploy-wva-stack.sh --create-kind --with-hpa --with-va |
| 79 | +``` |
| 80 | + |
| 81 | +### Create Kind Cluster Only (No Deployment) |
| 82 | + |
| 83 | +```bash |
| 84 | +# Create Kind cluster with custom configuration (no deployment) |
| 85 | +export KIND_CLUSTER_NAME="kind-fma-cluster" |
| 86 | +export KIND_CLUSTER_NODES=4 |
| 87 | +export KIND_CLUSTER_GPUS=2 |
| 88 | +export KIND_CLUSTER_GPU_TYPE="nvidia" |
| 89 | +./deploy-wva-stack.sh --create-kind-only |
| 90 | + |
| 91 | +# Later, deploy to the existing cluster |
| 92 | +# IMPORTANT: Must export KIND_CLUSTER_NAME to match your cluster |
| 93 | +export KIND_CLUSTER_NAME="kind-fma-cluster" |
| 94 | +./deploy-wva-stack.sh --kind --with-hpa |
| 95 | +``` |
| 96 | + |
| 97 | +**Note:** When deploying to an existing Kind cluster with `--kind`, you must export `KIND_CLUSTER_NAME` to match your cluster name (default: `kind-wva-gpu-cluster`). |
| 98 | + |
| 99 | +### WVA Only (No llm-d) |
| 100 | + |
| 101 | +```bash |
| 102 | +./deploy-wva-stack.sh --wva-only |
| 103 | +``` |
| 104 | + |
| 105 | +### llm-d Only (No WVA) |
| 106 | + |
| 107 | +```bash |
| 108 | +export HF_TOKEN="hf_xxxxx" |
| 109 | +./deploy-wva-stack.sh --llmd-only |
| 110 | +``` |
| 111 | + |
| 112 | +## Configuration |
| 113 | + |
| 114 | +### Environment Variables |
| 115 | + |
| 116 | +You can configure the deployment using environment variables: |
| 117 | + |
| 118 | +```bash |
| 119 | +# Deployment flags |
| 120 | +export DEPLOY_WVA=true # Deploy WVA controller (default: true) |
| 121 | +export DEPLOY_LLM_D=true # Deploy llm-d infrastructure (default: true) |
| 122 | +export DEPLOY_VA=false # Deploy VariantAutoscaling (default: false) |
| 123 | +export DEPLOY_HPA=false # Deploy HPA (default: false) |
| 124 | +export DEPLOY_PROMETHEUS=true # Deploy Prometheus (default: true) |
| 125 | +export DEPLOY_PROMETHEUS_ADAPTER=true # Deploy Prometheus Adapter (default: true) |
| 126 | + |
| 127 | +# WVA repository configuration |
| 128 | +export WVA_REPO_BRANCH=main # WVA repository branch (default: main) |
| 129 | + |
| 130 | +# HuggingFace token (required for llm-d) |
| 131 | +export HF_TOKEN="hf_xxxxx" |
| 132 | + |
| 133 | +# Run deployment |
| 134 | +./deploy-wva-stack.sh |
| 135 | +``` |
| 136 | + |
| 137 | +### Command-Line Options |
| 138 | + |
| 139 | +```bash |
| 140 | +./deploy-wva-stack.sh [options] |
| 141 | + |
| 142 | +Options: |
| 143 | + -h, --help Show help message |
| 144 | + -c, --cleanup Clean up existing deployment before installing |
| 145 | + -d, --delete-repo Delete WVA repository after deployment |
| 146 | + --wva-only Deploy only WVA (skip llm-d) |
| 147 | + --llmd-only Deploy only llm-d (skip WVA) |
| 148 | + --with-hpa Deploy HPA (default: false) |
| 149 | + --with-va Deploy VariantAutoscaling (default: false) |
| 150 | + --kind Deploy to Kind cluster (uses kind-emulator scripts) |
| 151 | + --create-kind Create Kind cluster before deployment |
| 152 | + --create-kind-only Create Kind cluster only (skip deployment) |
| 153 | + --delete-kind Delete Kind cluster after deployment |
| 154 | +``` |
| 155 | + |
| 156 | +## Usage Examples |
| 157 | + |
| 158 | +### Example 1: Full Stack with HPA for Testing |
| 159 | + |
| 160 | +```bash |
| 161 | +export HF_TOKEN="hf_xxxxx" |
| 162 | +export DEPLOY_VA=true |
| 163 | +export DEPLOY_HPA=true |
| 164 | +./deploy-wva-stack.sh |
| 165 | +``` |
| 166 | + |
| 167 | + |
| 168 | +### Example 2: Infrastructure Only |
| 169 | + |
| 170 | +```bash |
| 171 | +export HF_TOKEN="hf_xxxxx" |
| 172 | +export DEPLOY_VA=false |
| 173 | +export DEPLOY_HPA=false |
| 174 | +./deploy-wva-stack.sh |
| 175 | +``` |
| 176 | + |
| 177 | + |
| 178 | +### Example 3: Kind Cluster with Custom Configuration |
| 179 | + |
| 180 | +```bash |
| 181 | +# Create Kind cluster with custom settings |
| 182 | +export KIND_CLUSTER_NAME="kind-fma-cluster" |
| 183 | +export KIND_CLUSTER_NODES=4 |
| 184 | +export KIND_CLUSTER_GPUS=2 |
| 185 | +export KIND_CLUSTER_GPU_TYPE="nvidia" |
| 186 | +./deploy-wva-stack.sh --create-kind --with-hpa |
| 187 | +``` |
| 188 | + |
| 189 | +### Example 4: Custom WVA Branch |
| 190 | + |
| 191 | +```bash |
| 192 | +export WVA_REPO_BRANCH=feature/new-feature |
| 193 | +export HF_TOKEN="hf_xxxxx" |
| 194 | +./deploy-wva-stack.sh |
| 195 | +``` |
| 196 | + |
| 197 | +## Verification |
| 198 | + |
| 199 | +After deployment, verify the components: |
| 200 | + |
| 201 | +```bash |
| 202 | +# Check WVA controller |
| 203 | +kubectl get pods -n workload-variant-autoscaler-system |
| 204 | + |
| 205 | +# Check llm-d resources |
| 206 | +kubectl get all -n $(kubectl get ns -o name | grep llm-d | head -n1 | cut -d'/' -f2) |
| 207 | + |
| 208 | +# Check VariantAutoscaling resources |
| 209 | +kubectl get variantautoscalings.llmd.ai -A |
| 210 | + |
| 211 | +# Check HPA resources |
| 212 | +kubectl get hpa -A |
| 213 | + |
| 214 | +# Check Prometheus |
| 215 | +kubectl get pods -n workload-variant-autoscaler-monitoring |
| 216 | +``` |
| 217 | + |
| 218 | +## Viewing Logs |
| 219 | + |
| 220 | +```bash |
| 221 | +# WVA controller logs |
| 222 | +kubectl logs -n workload-variant-autoscaler-system \ |
| 223 | + -l control-plane=controller-manager -f |
| 224 | + |
| 225 | +# llm-d logs (example for decode deployment) |
| 226 | +kubectl logs -n llm-d-inference-scheduler \ |
| 227 | + deployment/ms-inference-scheduling-llm-d-modelservice-decode -f |
| 228 | +``` |
| 229 | + |
| 230 | +## Cleanup |
| 231 | + |
| 232 | +### Remove Deployment Only |
| 233 | + |
| 234 | +```bash |
| 235 | +./deploy-wva-stack.sh --cleanup |
| 236 | +``` |
| 237 | + |
| 238 | +OR if HPA and VA are enabled: |
| 239 | + |
| 240 | +```bash |
| 241 | +./deploy-wva-stack.sh --cleanup --with-hpa --with-va |
| 242 | +``` |
| 243 | + |
| 244 | +### Remove Deployment and Kind Cluster |
| 245 | + |
| 246 | +```bash |
| 247 | +./deploy-wva-stack.sh --cleanup --delete-kind |
| 248 | +``` |
| 249 | + |
| 250 | +## Kind Cluster Configuration |
| 251 | + |
| 252 | +When using `--create-kind` or `--create-kind-only`, you can customize the cluster configuration: |
| 253 | + |
| 254 | +### Environment Variables |
| 255 | + |
| 256 | +```bash |
| 257 | +export KIND_CLUSTER_NAME="kind-fma-cluster" # Default: kind-wva-gpu-cluster |
| 258 | +export KIND_CLUSTER_NODES=4 # Default: 3 |
| 259 | +export KIND_CLUSTER_GPUS=2 # Default: 4 (GPUs per node) |
| 260 | +export KIND_CLUSTER_GPU_TYPE="nvidia" # Default: mix (nvidia|amd|intel|mix) |
| 261 | +``` |
| 262 | + |
| 263 | +### Usage Examples |
| 264 | + |
| 265 | +**Create cluster and deploy immediately:** |
| 266 | +```bash |
| 267 | +export KIND_CLUSTER_NAME="kind-fma-cluster" |
| 268 | +export KIND_CLUSTER_NODES=4 |
| 269 | +export KIND_CLUSTER_GPUS=2 |
| 270 | +export KIND_CLUSTER_GPU_TYPE="nvidia" |
| 271 | +./deploy-wva-stack.sh --create-kind --with-hpa |
| 272 | +``` |
| 273 | + |
| 274 | +**Create cluster only (for later deployment):** |
| 275 | +```bash |
| 276 | +export KIND_CLUSTER_NAME="kind-fma-cluster" |
| 277 | +export KIND_CLUSTER_NODES=4 |
| 278 | +export KIND_CLUSTER_GPUS=2 |
| 279 | +export KIND_CLUSTER_GPU_TYPE="nvidia" |
| 280 | +./deploy-wva-stack.sh --create-kind-only |
| 281 | +``` |
| 282 | + |
| 283 | +### Cluster Features |
| 284 | + |
| 285 | +The Kind cluster includes: |
| 286 | +- Emulated GPU resources (nvidia.com/gpu, amd.com/gpu, intel.com/gpu) |
| 287 | +- HPAScaleToZero feature gate enabled (for scale-to-zero testing) |
| 288 | +- llm-d inference simulator (no real model loading) |
0 commit comments