🐛 Fix e2e tests by adding DECODE_REPLICAS override (llm-d#648)

PR llm-d/llm-d#619 changed decode.replicas from 2 to 8 and the default
model from Qwen/Qwen3-0.6B to Qwen/Qwen3-32B. This breaks e2e tests.

Changes:
- Add DECODE_REPLICAS environment variable support to install.sh
- Set DECODE_REPLICAS=1 in the OpenShift e2e workflow
- Fix gaie-sim-epp patch to only run when the deployment exists
- Update DEFAULT_MODEL_ID to Qwen/Qwen3-32B to match llm-d values.yaml
  (fixes silent model substitution failure)
- Increase model-storage size from 30Gi to 100Gi to handle larger models

Signed-off-by: Andrew Anderson <andy@clubanderson.com>
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>

Author: clubanderson

.github/workflows/ci-e2e-openshift.yaml

@@ -475,6 +475,8 @@ jobs:
           CONTROLLER_INSTANCE: ${{ env.WVA_NAMESPACE }}
           # vLLM max-num-seqs for e2e testing (lower = easier to saturate)
           VLLM_MAX_NUM_SEQS: ${{ env.MAX_NUM_SEQS }}
+          # Decode replicas for e2e testing (start with 1 replica, let HPA scale)
+          DECODE_REPLICAS: "1"
         run: |
           echo "Deploying WVA and llm-d infrastructure..."
           echo "  MODEL_ID: $MODEL_ID"
@@ -485,6 +487,7 @@ jobs:
           echo "  WVA_IMAGE_TAG: $WVA_IMAGE_TAG"
           echo "  CONTROLLER_INSTANCE: $CONTROLLER_INSTANCE"
           echo "  VLLM_MAX_NUM_SEQS: $VLLM_MAX_NUM_SEQS"
+          echo "  DECODE_REPLICAS: $DECODE_REPLICAS"
           echo "  HF token configuration: ✓"
           ./deploy/install.sh --model "$MODEL_ID" --accelerator "$ACCELERATOR_TYPE" --release-name "$WVA_RELEASE_NAME" --environment openshift
 
@@ -530,11 +533,14 @@ jobs:
           DEPLOY_HPA: "false"
           # vLLM max-num-seqs for e2e testing (lower = easier to saturate)
           VLLM_MAX_NUM_SEQS: ${{ env.MAX_NUM_SEQS }}
+          # Decode replicas for e2e testing (start with 1 replica, let HPA scale)
+          DECODE_REPLICAS: "1"
         run: |
           echo "Deploying Model B infrastructure in $LLMD_NAMESPACE_B..."
           echo "  MODEL_ID: $MODEL_ID"
           echo "  ACCELERATOR_TYPE: $ACCELERATOR_TYPE"
           echo "  VLLM_MAX_NUM_SEQS: $VLLM_MAX_NUM_SEQS"
+          echo "  DECODE_REPLICAS: $DECODE_REPLICAS"
 
           # Deploy llm-d infrastructure only (no WVA controller, no VA/HPA)
           ./deploy/install.sh --model "$MODEL_ID" --accelerator "$ACCELERATOR_TYPE" --environment openshift

deploy/install.sh

@@ -64,7 +64,7 @@ INSTALL_GATEWAY_CTRLPLANE_ORIGINAL="${INSTALL_GATEWAY_CTRLPLANE:-}"
 INSTALL_GATEWAY_CTRLPLANE="${INSTALL_GATEWAY_CTRLPLANE:-false}"
 
 # Model and SLO Configuration
-DEFAULT_MODEL_ID=${DEFAULT_MODEL_ID:-"Qwen/Qwen3-0.6B"}
+DEFAULT_MODEL_ID=${DEFAULT_MODEL_ID:-"Qwen/Qwen3-32B"}
 MODEL_ID=${MODEL_ID:-"unsloth/Meta-Llama-3.1-8B"}
 ACCELERATOR_TYPE=${ACCELERATOR_TYPE:-"H100"}
 SLO_TPOT=${SLO_TPOT:-10}  # Target time-per-output-token SLO (in ms)
@@ -94,6 +94,8 @@ SKIP_CHECKS=${SKIP_CHECKS:-false}
 E2E_TESTS_ENABLED=${E2E_TESTS_ENABLED:-false}
 # vLLM max-num-seqs (max concurrent sequences per replica, lower = easier to saturate for testing)
 VLLM_MAX_NUM_SEQS=${VLLM_MAX_NUM_SEQS:-""}
+# Decode replicas override (useful for e2e testing with limited GPUs)
+DECODE_REPLICAS=${DECODE_REPLICAS:-""}
 
 # Environment-related variables
 SCRIPT_DIR=$(cd $(dirname "${BASH_SOURCE[0]}") && pwd)
@@ -737,7 +739,7 @@ deploy_llm_d_infrastructure() {
 
         # Increase model-storage volume size
         log_info "Increasing model-storage volume size for model: $MODEL_ID"
-        yq eval '.modelArtifacts.size = "30Gi"' -i "$LLM_D_MODELSERVICE_VALUES"
+        yq eval '.modelArtifacts.size = "100Gi"' -i "$LLM_D_MODELSERVICE_VALUES"
     fi
 
     # Configure llm-d-inference-simulator if needed
@@ -758,6 +760,12 @@ deploy_llm_d_infrastructure() {
       yq eval ".decode.containers[0].args += [\"--max-num-seqs=$VLLM_MAX_NUM_SEQS\"]" -i "$LLM_D_MODELSERVICE_VALUES"
     fi
 
+    # Configure decode replicas if set (useful for e2e testing with limited GPUs)
+    if [ -n "$DECODE_REPLICAS" ]; then
+      log_info "Setting decode replicas to $DECODE_REPLICAS"
+      yq eval ".decode.replicas = $DECODE_REPLICAS" -i "$LLM_D_MODELSERVICE_VALUES"
+    fi
+
     # Deploy llm-d core components
     log_info "Deploying llm-d core components"
     helmfile apply -e $GATEWAY_PROVIDER -n ${LLMD_NS}
@@ -772,8 +780,8 @@ deploy_llm_d_infrastructure() {
     #     -p '{"spec":{"kube":{"service":{"type":"NodePort"}}}}'
     # fi
 
-    # Patch llm-d-inference-simulator deployment if scale-to-zero is enabled
-    if [ "$ENABLE_SCALE_TO_ZERO" == "true" ]; then
+    # Patch llm-d-inference-simulator deployment if scale-to-zero is enabled and simulator is deployed
+    if [ "$ENABLE_SCALE_TO_ZERO" == "true" ] && kubectl get deployment gaie-sim-epp -n $LLMD_NS &>/dev/null; then
         # Patch llm-d-inference-simulator deployment to use the correct image
         log_info "Patching llm-d-inference-simulator deployment to enable flowcontrol and use a new image"
         export DEPLOYMENT_NAME="gaie-sim-epp"