feat: deploy vllm-render as standalone service; use simulated render in e2e

DEVELOPMENT.md

@@ -447,7 +447,7 @@ The `deploy/components/` directory contains all reusable Kustomize components:
 - `vllm-decode/` — Decode pod with routing sidecar
 
 **Deployment overlays** — applied on top of the base components:
-- `overlays/simulator/` — adds `--mode=${VLLM_SIM_MODE}`, vllm renderer sidecar, KV cache args, and `--zmq-endpoint` on Decode. Included by default in all dev scenario overlays.
+- `overlays/simulator/` — adds `--mode=${VLLM_SIM_MODE}`, KV cache args, and `--zmq-endpoint` on Decode. Included by default in all dev scenario overlays.
 - `overlays/real-vllm/` — adds `--kv-events-config` on Decode, `--ec-transfer-config` on Encode, and a shared PVC for encoder embeddings.
 
 **Infrastructure components** — shared cluster infrastructure:
@@ -523,7 +523,7 @@ a shared PVC (`ec-cache-pvc`) for encoder embeddings transfer.
 
 | Component | What it adds | When to use |
 |---|---|---|
-| `overlays/simulator/` | `--mode=${VLLM_SIM_MODE}`, vLLM renderer, KV cache args, `--zmq-endpoint` on Decode | Dev/test with simulator image |
+| `overlays/simulator/` | `--mode=${VLLM_SIM_MODE}`, KV cache args, `--zmq-endpoint` on Decode | Dev/test with simulator image |
 | `overlays/real-vllm/` | `--kv-events-config` on Decode (per-pod ZMQ publisher), `--ec-transfer-config` on Encode, ec-cache PVC | Production with real vLLM image |
 
 | Variable | Default | Description |

Makefile

@@ -40,6 +40,8 @@ export VLLM_IMAGE ?= $(VLLM_SIMULATOR_TAG_BASE):$(VLLM_SIMULATOR_TAG)
 # CPU-only vLLM image that exposes `vllm launch render` for the token-producer
 # plugin's HTTP backend.
 export VLLM_RENDER_IMAGE ?= vllm/vllm-openai-cpu:v0.21.0
+export VLLM_RENDER_PORT ?= 8082
+export VLLM_RENDER_URL ?= http://vllm-render:$(VLLM_RENDER_PORT)
 
 BUILDER_TAG ?= dev
 BUILDER_TAG_BASE ?= $(IMAGE_REGISTRY)/$(BUILDER_IMAGE_NAME)

deploy/components/inference-gateway/deployment.yaml

@@ -60,40 +60,9 @@ spec:
           mountPath: /etc/epp
         - name: cache
           mountPath: /cache
-      # vLLM render sidecar — exposes /v1/completions/render and
-      # /v1/chat/completions/render over loopback for the token-producer plugin.
-      - name: vllm-render
-        image: ${VLLM_RENDER_IMAGE}
-        imagePullPolicy: IfNotPresent
-        command: ["vllm", "launch", "render"]
-        args: ["${MODEL_NAME}", "--port=8000"]
-        ports:
-        - name: render-http
-          containerPort: 8000
-        startupProbe:
-          httpGet:
-            path: /health
-            port: 8000
-          initialDelaySeconds: 2
-          periodSeconds: 1
-          timeoutSeconds: 1
-          failureThreshold: 600
-        readinessProbe:
-          httpGet:
-            path: /health
-            port: 8000
-          periodSeconds: 5
-          timeoutSeconds: 1
-          failureThreshold: 1
-        volumeMounts:
-        - name: model-cache
-          mountPath: /root/.cache/huggingface
       volumes:
       - name: epp-config
         configMap:
           name: epp-config
       - name: cache
         emptyDir: {}
-      # HuggingFace cache for the render sidecar's model weights.
-      - name: model-cache
-        emptyDir: {}

deploy/components/overlays/simulator/kustomization.yaml

@@ -40,99 +40,3 @@ patches:
     - op: add
       path: /spec/template/spec/containers/0/args/4
       value: "--zmq-endpoint=tcp://${EPP_NAME}.${NAMESPACE}.svc.cluster.local:5557"
-    - op: add
-      path: /spec/template/spec/initContainers/0
-      value:
-        name: vllm-render
-        image: ${VLLM_RENDER_IMAGE}
-        imagePullPolicy: IfNotPresent
-        restartPolicy: Always
-        command: ["vllm"]
-        args: ["launch", "render", "${MODEL_NAME}", "--port=8082"]
-        env:
-        - name: HF_TOKEN
-          value: "${HF_TOKEN}"
-        ports:
-        - name: http
-          containerPort: 8082
-        startupProbe:
-          httpGet:
-            path: /health
-            port: 8082
-          initialDelaySeconds: 2
-          periodSeconds: 1
-          failureThreshold: 600
-          timeoutSeconds: 1
-        livenessProbe:
-          httpGet:
-            path: /health
-            port: 8082
-          periodSeconds: 10
-          failureThreshold: 5
-          timeoutSeconds: 1
-        readinessProbe:
-          httpGet:
-            path: /health
-            port: 8082
-          periodSeconds: 5
-          failureThreshold: 1
-          timeoutSeconds: 1
-        volumeMounts:
-        - name: model-cache
-          mountPath: /root/.cache/huggingface
-    - op: add
-      path: /spec/template/spec/volumes/-
-      value:
-        name: model-cache
-        emptyDir: {}
-# --- Prefill and Encode deployments (vllm-p, vllm-e) ---
-# initContainers: [] in base, so append with /-
-- target:
-    kind: Deployment
-    name: "vllm-[pe]"
-  patch: |
-    - op: add
-      path: /spec/template/spec/initContainers/-
-      value:
-        name: vllm-render
-        image: ${VLLM_RENDER_IMAGE}
-        imagePullPolicy: IfNotPresent
-        restartPolicy: Always
-        command: ["vllm"]
-        args: ["launch", "render", "${MODEL_NAME}", "--port=8082"]
-        env:
-        - name: HF_TOKEN
-          value: "${HF_TOKEN}"
-        ports:
-        - name: http
-          containerPort: 8082
-        startupProbe:
-          httpGet:
-            path: /health
-            port: 8082
-          initialDelaySeconds: 2
-          periodSeconds: 1
-          failureThreshold: 600
-          timeoutSeconds: 1
-        livenessProbe:
-          httpGet:
-            path: /health
-            port: 8082
-          periodSeconds: 10
-          failureThreshold: 5
-          timeoutSeconds: 1
-        readinessProbe:
-          httpGet:
-            path: /health
-            port: 8082
-          periodSeconds: 5
-          failureThreshold: 1
-          timeoutSeconds: 1
-        volumeMounts:
-        - name: model-cache
-          mountPath: /root/.cache/huggingface
-    - op: add
-      path: /spec/template/spec/volumes/-
-      value:
-        name: model-cache
-        emptyDir: {}

deploy/components/vllm-decode/deployment.yaml

@@ -67,6 +67,7 @@ spec:
         - "--port=8200"
         - "--model=${MODEL_NAME}"
         - "--data-parallel-size=${VLLM_DATA_PARALLEL_SIZE}"
+        - "--render-url=${VLLM_RENDER_URL}"
         - "${VLLM_EXTRA_ARGS_D}"
         ports:
         - name: http

deploy/components/vllm-encode/deployment.yaml

@@ -30,6 +30,7 @@ spec:
         - "--data-parallel-size=${VLLM_DATA_PARALLEL_SIZE}"
         - "--no-enable-prefix-caching"
         - "--mm-encoder-only"
+        - "--render-url=${VLLM_RENDER_URL}"
         - "${VLLM_EXTRA_ARGS_E}"
         ports:
         - name: encode-http

deploy/components/vllm-prefill/deployment.yaml

@@ -28,6 +28,7 @@ spec:
         - "--model=${MODEL_NAME}"
         - "--port=8000"
         - "--data-parallel-size=${VLLM_DATA_PARALLEL_SIZE}"
+        - "--render-url=${VLLM_RENDER_URL}"
         - "${VLLM_EXTRA_ARGS_P}"
         ports:
         - name: prefill-http

deploy/environments/dev/README.md

@@ -9,7 +9,7 @@ scenario-specific patches. The atomic components live in `deploy/components/`:
 | `vllm-decode/` | Decode pod — always deployed, includes routing sidecar (removed in EPD scenario) |
 | `vllm-prefill/` | Prefill pod — deployed when `DISAGG_P=true` |
 | `vllm-encode/` | Encoder pod — deployed when `DISAGG_E=true` |
-| `overlays/simulator/` | Adds `--mode=${VLLM_SIM_MODE}`, vLLM render, KV cache and ZMQ args (included by all scenario overlays) |
+| `overlays/simulator/` | Adds `--mode=${VLLM_SIM_MODE}`, KV cache and ZMQ args (included by all scenario overlays) |
 
 These overlays are used by both `scripts/kind-dev-env.sh` (for local KIND clusters)
 and e2e tests (via `kustomize build` + env var substitution).
@@ -70,6 +70,8 @@ Variables substituted at deploy time via `envsubst` or Go test `substituteMany`:
 | `VLLM_IMAGE` | vLLM container image (simulator or real) | `ghcr.io/llm-d/llm-d-inference-sim:v0.10.2` |
 | `SIDECAR_IMAGE` | Routing sidecar image | `ghcr.io/llm-d/llm-d-router-disagg-sidecar:dev` |
 | `VLLM_RENDER_IMAGE` | vLLM render sidecar image | `vllm/vllm-openai-cpu:v0.21.0` |
+| `VLLM_RENDER_PORT` | Port the vllm-render Service listens on | `8082` |
+| `VLLM_RENDER_URL` | URL of the vllm-render Service used by decode/encode/prefill pods | `http://vllm-render:${VLLM_RENDER_PORT}` |
 | `MODEL_NAME` | Model name passed to vLLM. Can be a real HuggingFace model (e.g. `TinyLlama/TinyLlama-1.1B-Chat-v1.0`, `Qwen/Qwen3-VL-2B-Instruct`) or an arbitrary name when using the simulator (e.g. `food-review`) | `food-review` |
 | `POOL_NAME` | InferencePool name | `food-review-inference-pool` |
 | `VLLM_REPLICA_COUNT_E` | Encode deployment replicas | `1` |

deploy/environments/dev/base-kind-istio/kustomization.yaml

@@ -16,6 +16,7 @@ kind: Kustomization
 resources:
 - destination-rules.yaml
 - services.yaml
+- vllm-render.yaml
 - ../../../components/istio-control-plane/
 - ../../../components/inference-gateway/
 

deploy/environments/dev/base-kind-istio/vllm-render.yaml

@@ -0,0 +1,76 @@
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: vllm-render
+  labels:
+    app: vllm-render
+spec:
+  replicas: 1
+  selector:
+    matchLabels:
+      app: vllm-render
+  template:
+    metadata:
+      labels:
+        app: vllm-render
+    spec:
+      containers:
+      - name: vllm-render
+        image: ${VLLM_RENDER_IMAGE}
+        imagePullPolicy: IfNotPresent
+        command: ["vllm", "launch", "render"]
+        args: ["${MODEL_NAME}", "--port=${VLLM_RENDER_PORT}"]
+        # For gated models, inject HF_TOKEN via a Secret reference:
+        # env:
+        # - name: HF_TOKEN
+        #   valueFrom:
+        #     secretKeyRef:
+        #       name: hf-token
+        #       key: token
+        ports:
+        - name: render-http
+          containerPort: ${VLLM_RENDER_PORT}
+        startupProbe:
+          httpGet:
+            path: /health
+            port: ${VLLM_RENDER_PORT}
+          initialDelaySeconds: 2
+          periodSeconds: 1
+          failureThreshold: 600
+          timeoutSeconds: 1
+        livenessProbe:
+          httpGet:
+            path: /health
+            port: ${VLLM_RENDER_PORT}
+          periodSeconds: 10
+          failureThreshold: 5
+          timeoutSeconds: 1
+        readinessProbe:
+          httpGet:
+            path: /health
+            port: ${VLLM_RENDER_PORT}
+          periodSeconds: 5
+          failureThreshold: 1
+          timeoutSeconds: 1
+        volumeMounts:
+        - name: model-cache
+          mountPath: /root/.cache/huggingface
+      volumes:
+      - name: model-cache
+        emptyDir: {}
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: vllm-render
+  labels:
+    app: vllm-render
+spec:
+  selector:
+    app: vllm-render
+  ports:
+  - name: http
+    protocol: TCP
+    port: ${VLLM_RENDER_PORT}
+    targetPort: ${VLLM_RENDER_PORT}
+  type: ClusterIP

deploy/environments/dev/e2e-infra/vllm-render.yaml

@@ -0,0 +1,66 @@
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: vllm-render
+  labels:
+    app: vllm-render
+spec:
+  replicas: 1
+  selector:
+    matchLabels:
+      app: vllm-render
+  template:
+    metadata:
+      labels:
+        app: vllm-render
+    spec:
+      containers:
+      - name: vllm-render
+        image: ${VLLM_RENDER_IMAGE}
+        imagePullPolicy: IfNotPresent
+        args:
+        - "--model=${MODEL_NAME}"
+        - "--port=${VLLM_RENDER_PORT}"
+        - "--mode=echo"
+        - "--force-dummy-tokenizer"
+        ports:
+        - name: http
+          containerPort: ${VLLM_RENDER_PORT}
+        startupProbe:
+          httpGet:
+            path: /health
+            port: ${VLLM_RENDER_PORT}
+          initialDelaySeconds: 2
+          periodSeconds: 1
+          failureThreshold: 600
+          timeoutSeconds: 1
+        livenessProbe:
+          httpGet:
+            path: /health
+            port: ${VLLM_RENDER_PORT}
+          periodSeconds: 10
+          failureThreshold: 5
+          timeoutSeconds: 1
+        readinessProbe:
+          httpGet:
+            path: /health
+            port: ${VLLM_RENDER_PORT}
+          periodSeconds: 5
+          failureThreshold: 1
+          timeoutSeconds: 1
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: vllm-render
+  labels:
+    app: vllm-render
+spec:
+  selector:
+    app: vllm-render
+  ports:
+  - name: http
+    protocol: TCP
+    port: ${VLLM_RENDER_PORT}
+    targetPort: ${VLLM_RENDER_PORT}
+  type: ClusterIP

deploy/environments/dev/epd/patch-decode.yaml

@@ -22,6 +22,7 @@ spec:
         - "--block-size=16"
         - "--data-parallel-size=${VLLM_DATA_PARALLEL_SIZE}"
         - "--zmq-endpoint=tcp://${EPP_NAME}.${NAMESPACE}.svc.cluster.local:5557"
+        - "--render-url=${VLLM_RENDER_URL}"
         - "${VLLM_EXTRA_ARGS_D}"
         ports:
         - $patch: replace

scripts/kind-dev-env.sh

@@ -65,6 +65,8 @@ export SIDECAR_IMAGE
 # Set a default VLLM_RENDER_IMAGE if not provided (CPU-only vLLM image that
 # runs `vllm launch render` for the token-producer plugin's HTTP backend).
 export VLLM_RENDER_IMAGE="${VLLM_RENDER_IMAGE:-vllm/vllm-openai-cpu:v0.21.0}"
+export VLLM_RENDER_PORT="${VLLM_RENDER_PORT:-8082}"
+export VLLM_RENDER_URL="${VLLM_RENDER_URL:-http://vllm-render:${VLLM_RENDER_PORT}}"
 
 # Set the inference pool name for the deployment
 export POOL_NAME="${POOL_NAME:-${MODEL_NAME_SAFE}-inference-pool}"
@@ -414,15 +416,15 @@ export ENABLE_LEADER_ELECTION=false
 # Deploy Istio base (shared infrastructure)
 kubectl kustomize --enable-helm deploy/environments/dev/base-kind-istio \
   | envsubst '${POOL_NAME} ${MODEL_NAME} ${MODEL_NAME_SAFE} ${EPP_NAME} ${EPP_IMAGE} ${VLLM_IMAGE} \
-  ${SIDECAR_IMAGE} ${VLLM_RENDER_IMAGE} ${TARGET_PORTS} ${NAMESPACE} ${METRICS_ENDPOINT_AUTH} \
+  ${SIDECAR_IMAGE} ${VLLM_RENDER_IMAGE} ${VLLM_RENDER_PORT} ${VLLM_RENDER_URL} ${TARGET_PORTS} ${NAMESPACE} ${METRICS_ENDPOINT_AUTH} \
   ${EPP_REPLICA_COUNT} ${VLLM_REPLICA_COUNT_E} ${VLLM_REPLICA_COUNT_P} ${VLLM_REPLICA_COUNT_D} \
   ${VLLM_DATA_PARALLEL_SIZE} ${ENABLE_LEADER_ELECTION}' \
   | kubectl --context ${KUBE_CONTEXT} apply -f -
 
 # Deploy scenario-specific vLLM components
 kubectl kustomize --enable-helm ${KUSTOMIZE_DIR} \
   | envsubst '${POOL_NAME} ${MODEL_NAME} ${MODEL_NAME_SAFE} ${EPP_NAME} ${EPP_IMAGE} ${VLLM_IMAGE} \
-  ${SIDECAR_IMAGE} ${VLLM_RENDER_IMAGE} ${TARGET_PORTS} ${NAMESPACE} \
+  ${SIDECAR_IMAGE} ${VLLM_RENDER_IMAGE} ${VLLM_RENDER_PORT} ${VLLM_RENDER_URL} ${TARGET_PORTS} ${NAMESPACE} \
   ${VLLM_REPLICA_COUNT_E} ${VLLM_REPLICA_COUNT_P} ${VLLM_REPLICA_COUNT_D} ${VLLM_DATA_PARALLEL_SIZE} \
   ${KV_CONNECTOR_TYPE} ${EC_CONNECTOR_TYPE} ${CONNECTOR_TYPE} ${KV_CACHE_ENABLED} ${HF_TOKEN} ${VLLM_SIM_MODE} \
   ${DECODE_ROLE} ${VLLM_EXTRA_ARGS_E} ${VLLM_EXTRA_ARGS_P} ${VLLM_EXTRA_ARGS_D}' \