feat(trtllm): add OpenEngine sidecar

.github/codeowners/areas.yaml

@@ -231,6 +231,7 @@ areas:
   - lib/backend-common/examples/
   - lib/sidecar/README.md
   - lib/sidecar/common/
+  - lib/sidecar/openengine/
   - lib/bindings/
   - lib/bindings/c/
   - lib/bindings/python/

CODEOWNERS

@@ -216,6 +216,7 @@
 /lib/llm/src/discovery/                                                      @ai-dynamo/dynamo-runtime-codeowners
 /lib/runtime/src/utils/                                                      @ai-dynamo/dynamo-runtime-codeowners
 /lib/runtime/src/config/                                                     @ai-dynamo/dynamo-runtime-codeowners
+/lib/sidecar/openengine/                                                     @ai-dynamo/dynamo-runtime-codeowners
 /examples/custom_backend/                                                    @ai-dynamo/dynamo-runtime-codeowners
 /lib/llm/src/local_model/                                                    @ai-dynamo/dynamo-runtime-codeowners
 /lib/runtime/src/compute/                                                    @ai-dynamo/dynamo-runtime-codeowners

Cargo.toml

@@ -27,6 +27,7 @@ members = [
     "lib/sidecar/vllm",
     "lib/sidecar/sglang",
     "lib/sidecar/trtllm",
+    "lib/sidecar/openengine",
     "lib/bindings/c",
     "lib/bindings/python/codegen",
     "deploy/inference-gateway/ext-proc",
@@ -83,6 +84,7 @@ velo = { version = "0.1.0" }
 # External dependencies
 aiconfigurator-core = { git = "https://github.com/ai-dynamo/aiconfigurator.git", rev = "ffbab15797d3a3f14c382e937c0ab0c1c5cbc530", package = "aiconfigurator-core" }
 anyhow = { version = "1" }
+async-openai = { version = "=0.34.0", default-features = false }
 lru = { version = "0.16" }
 async-nats = { version = "0.49.1", features = ["service"] }
 async-stream = { version = "0.3" }

examples/backends/trtllm/deploy/README.md

@@ -1,3 +1,8 @@
+<!--
+SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+SPDX-License-Identifier: Apache-2.0
+-->
+
 # TensorRT-LLM Kubernetes Deployment Configurations
 
 This directory contains Kubernetes Custom Resource Definition (CRD) templates for deploying TensorRT-LLM inference graphs using the **DynamoGraphDeployment** resource.
@@ -68,6 +73,53 @@ TensorRT-LLM worker. This example uses the conservative engine configuration in
   `experimental.checkpoint.enabled: true` and
   `startupPolicy: WaitForCheckpoint`
 
+### 8. **OpenEngine Sidecar Examples** (`v1beta1/openengine-*.yaml`)
+
+The OpenEngine examples keep TRT-LLM in the GPU container and run
+`dynamo-openengine-sidecar` as the CPU-only `main` container in the same pod.
+Loopback OpenEngine gRPC connects the pair; TRT-LLM's HTTP listener remains
+available and uses the same `LLM` instance.
+
+- `openengine-agg.yaml`: aggregate text
+- `openengine-disagg.yaml`: context-first 1P1D text and LoRA
+- `openengine-multimodal-agg.yaml`: aggregate image/video
+- `openengine-multimodal-disagg.yaml`: image/video context-first 1P1D
+- `openengine-audio.yaml`: aggregate Phi-4 audio
+- `openengine-lora.yaml`: aggregate lazy LoRA lifecycle
+
+Replace both placeholder images with builds containing the sidecar and the
+local TRT-LLM fork. The TRT-LLM image must also contain this repository's
+`/workspace/examples/backends/trtllm/engine_configs` directory. Every engine is
+started with the PyTorch backend and immutable OpenEngine schema release
+`df3a9be24a2a36a4ff7a6d4fef9f1d7480ae210d`.
+
+LoRA manifests expect a PVC named `trtllm-openengine-lora-cache`. For 1P1D it
+must support `ReadWriteMany`, because context and generation are different pods
+and both mount `/lora-cache`. Create it before the DGD, for example:
+
+```yaml
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: trtllm-openengine-lora-cache
+spec:
+  accessModes: [ReadWriteMany]
+  resources:
+    requests:
+      storage: 20Gi
+```
+
+OpenEngine workers publish model cards with no frontend media decoder. This is
+the URL/data passthrough mode: ordered media stays encoded until TRT-LLM fetches
+and decodes it. Do not add a frontend media decoder to these deployments;
+decoded/RDMA media is intentionally rejected by the sidecar.
+
+The Phi-4 audio manifest additionally expects a PVC named
+`trtllm-openengine-phi4-model`. Populate its root with a raw git clone of
+`microsoft/Phi-4-multimodal-instruct` before creating the DGD. The mounted
+directory must contain the bundled `speech-lora/` subdirectory; TRT-LLM's
+Phi-4-MM loader does not support the Hugging Face snapshot layout.
+
 ## CRD Structure
 
 All templates use the **DynamoGraphDeployment** CRD:

examples/backends/trtllm/deploy/v1beta1/openengine-agg.yaml

@@ -0,0 +1,97 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+apiVersion: nvidia.com/v1beta1
+kind: DynamoGraphDeployment
+metadata:
+  name: trtllm-openengine-agg
+spec:
+  backendFramework: trtllm
+  components:
+  - name: Frontend
+    podTemplate:
+      spec:
+        containers:
+        - args:
+          - python3 -m dynamo.frontend --router-mode round-robin --http-port 8000
+          command: [/bin/sh, -c]
+          env:
+          - name: DYN_NAMESPACE
+            value: dynamo
+          image: my-registry/dynamo-openengine-sidecar:my-tag
+          name: main
+    replicas: 1
+    type: frontend
+  - name: OpenEngineWorker
+    podTemplate:
+      spec:
+        containers:
+        - args:
+          - --openengine-endpoint
+          - http://127.0.0.1:50051
+          - --expected-engine
+          - tensorrt_llm
+          command: [dynamo-openengine-sidecar]
+          env:
+          - name: DYN_NAMESPACE
+            value: dynamo
+          - name: DYN_SYSTEM_PORT
+            value: "8081"
+          - name: DYN_LORA_ENABLED
+            value: "false"
+          image: my-registry/dynamo-openengine-sidecar:my-tag
+          name: main
+          resources:
+            limits:
+              cpu: "2"
+              memory: 4Gi
+            requests:
+              cpu: "1"
+              memory: 2Gi
+        - args:
+          - Qwen/Qwen3-0.6B
+          - --backend
+          - pytorch
+          - --host
+          - 127.0.0.1
+          - --port
+          - "8001"
+          - --config
+          - /workspace/examples/backends/trtllm/engine_configs/qwen3/agg.yaml
+          - --openengine-host
+          - 127.0.0.1
+          - --openengine-port
+          - "50051"
+          command: [trtllm-serve]
+          env:
+          - name: HF_HOME
+            value: /model-cache
+          - name: OPENENGINE_SCHEMA_RELEASE
+            value: df3a9be24a2a36a4ff7a6d4fef9f1d7480ae210d
+          image: my-registry/tensorrtllm-openengine:my-tag
+          name: engine
+          resources:
+            limits:
+              nvidia.com/gpu: "1"
+            requests:
+              nvidia.com/gpu: "1"
+          volumeMounts:
+          - mountPath: /model-cache
+            name: model-cache
+        volumes:
+        - emptyDir: {}
+          name: model-cache
+    replicas: 1
+    type: worker