[#17016][feat] Add selectable OpenEngine gRPC server

.github/CODEOWNERS

@@ -501,6 +501,7 @@
 /pyproject.toml @NVIDIA/trt-llm-oss-compliance
 /requirements-dev.txt @NVIDIA/trt-llm-oss-compliance
 /requirements-grpc-smg.txt @NVIDIA/trt-llm-oss-compliance
+/requirements-openengine.txt @NVIDIA/trt-llm-oss-compliance
 /requirements.txt @NVIDIA/trt-llm-oss-compliance
 /setup.py @NVIDIA/trt-llm-oss-compliance
 /tests/unittest/api_stability/ @NVIDIA/trt-llm-noncommitted-api-review-committee

docker/Dockerfile.multi

@@ -130,7 +130,7 @@ COPY scripts scripts
 COPY tensorrt_llm tensorrt_llm
 COPY triton_kernels triton_kernels
 COPY 3rdparty 3rdparty
-COPY setup.py requirements.txt requirements-dev.txt requirements-grpc-smg.txt constraints.txt LICENSE README.md ./
+COPY setup.py requirements.txt requirements-dev.txt requirements-grpc-smg.txt requirements-openengine.txt constraints.txt LICENSE README.md ./
 
 ENV CCACHE_DIR=/root/.cache/ccache
 # Build the TRT-LLM wheel

jenkins/current_image_tags.properties

@@ -13,8 +13,8 @@
 #     images are adopted from PostMerge pipelines, the abbreviated commit hash is used instead.
 IMAGE_NAME=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm
 
-LLM_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:pytorch-26.05-py3-x86_64-ubuntu24.04-skip-tritondevel-202608191543-17891
-LLM_SBSA_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:pytorch-26.05-py3-sbsa-ubuntu24.04-skip-tritondevel-202608191543-17891
-LLM_ROCKYLINUX8_PY310_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:cuda-13.2.1-devel-rocky8-x86_64-rocky8-py310-skip-tritondevel-202608191543-17891
-LLM_ROCKYLINUX8_PY312_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:cuda-13.2.1-devel-rocky8-x86_64-rocky8-py312-skip-tritondevel-202608191543-17891
-LLM_SBSA_WHEEL_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:cuda-13.2.1-devel-ubuntu24.04-sbsa-ubuntu24.04-py312-skip-tritondevel-202608191543-17891
+LLM_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:pytorch-26.05-py3-x86_64-ubuntu24.04-skip-tritondevel-202608271702-17084
+LLM_SBSA_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:pytorch-26.05-py3-sbsa-ubuntu24.04-skip-tritondevel-202608271702-17084
+LLM_ROCKYLINUX8_PY310_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:cuda-13.2.1-devel-rocky8-x86_64-rocky8-py310-skip-tritondevel-202608271702-17084
+LLM_ROCKYLINUX8_PY312_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:cuda-13.2.1-devel-rocky8-x86_64-rocky8-py312-skip-tritondevel-202608271702-17084
+LLM_SBSA_WHEEL_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:cuda-13.2.1-devel-ubuntu24.04-sbsa-ubuntu24.04-py312-skip-tritondevel-202608271702-17084

requirements-openengine.txt

@@ -0,0 +1,12 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+# Source: https://github.com/ai-dynamo/openengine/releases/tag/v0.1.0
+# BSR module: https://buf.build/openengine/openengine
+# Immutable BSR commit: 768a93c7b44e40f28c692ad0b471a8f2
+--extra-index-url https://buf.build/gen/python
+openengine-openengine-grpc-python==1.67.1.2.20260730172104+768a93c7b44e
+openengine-openengine-protocolbuffers-python==31.1.0.2.20260730172104+768a93c7b44e
+openengine-openengine-protocolbuffers-pyi==31.1.0.2.20260730172104+768a93c7b44e
+grpcio>=1.67.1,<2
+protobuf>=6.31.1,<7

setup.py

@@ -147,6 +147,7 @@ def has_ext_modules(self):
 devel_deps, _ = parse_requirements(
     Path("requirements-dev-windows.txt"
          if on_windows else "requirements-dev.txt"))
+openengine_deps, _ = parse_requirements(Path("requirements-openengine.txt"))
 mx_deps = ["modelexpress>=0.4.1,<0.6.0"]
 # Gateway protocol adapters are opt-in extras: the default installation must
 # not carry any gateway protobuf package. Each gateway owns a dedicated
@@ -563,6 +564,7 @@ def get_build_state_options():
     scripts=['tensorrt_llm/llmapi/trtllm-llmapi-launch'],
     extras_require={
         "devel": devel_deps + grpc_smg_deps,
+        "openengine": openengine_deps,
         "mx": mx_deps,
         "grpc-smg": grpc_smg_deps,
     },

tensorrt_llm/commands/serve.py

@@ -1205,10 +1205,13 @@ def launch_visual_gen_server(
     "--grpc",
     is_flag=True,
     default=False,
-    help="Run gRPC server instead of OpenAI HTTP server. "
-    "gRPC server accepts pre-tokenized requests and returns raw token IDs. "
-    "Requires the tensorrt_llm[grpc-smg] extra.",
+    help="Run the selected gRPC protocol instead of the OpenAI HTTP server.",
     status="prototype")
+@stability_option("--grpc-protocol",
+                  type=click.Choice(["smg", "openengine"]),
+                  default="smg",
+                  help="Protocol used when --grpc is enabled.",
+                  status="prototype")
 @stability_option(
     "--served_model_name",
     type=str,
@@ -1276,9 +1279,9 @@ def serve(model: str, tokenizer: Optional[str], custom_tokenizer: Optional[str],
           agent_types: Optional[str], video_pruning_rate: Optional[float],
           telemetry: bool, custom_module_dirs: list[Path],
           chat_template: Optional[str], allow_request_chat_template: bool,
-          middleware: tuple[str, ...], grpc: bool, enable_visual_gen: bool,
-          served_model_name: Optional[str], visual_gen_args: Optional[str],
-          report_addr: Optional[str]):
+          middleware: tuple[str, ...], grpc: bool, grpc_protocol: str,
+          enable_visual_gen: bool, served_model_name: Optional[str],
+          visual_gen_args: Optional[str], report_addr: Optional[str]) -> None:
     """Running an OpenAI API compatible server
 
     MODEL: model name | HF checkpoint path | TensorRT engine path
@@ -1292,6 +1295,9 @@ def serve(model: str, tokenizer: Optional[str], custom_tokenizer: Optional[str],
             "to the 'pytorch' backend. See "
             "https://github.com/NVIDIA/TensorRT-LLM/issues/15638 for details.")
 
+    if not grpc and grpc_protocol != "smg":
+        raise click.UsageError("--grpc-protocol requires --grpc")
+
     if moe_cluster_parallel_size is not None:
         logger.warning(
             "--moe_cluster_parallel_size / --cluster_size is deprecated and "
@@ -1439,6 +1445,10 @@ def _serve_llm():
             media_io_kwargs=parsed_media_io_kwargs)
 
         if grpc:
+            if num_serve_frontends != 1:
+                raise click.UsageError(
+                    "--num_serve_frontends must be 1 when --grpc is enabled.")
+
             # gRPC mode: launch gRPC server instead of OpenAI HTTP server
             # Check for unsupported arguments that are silently ignored in gRPC mode
             unsupported_args = {
@@ -1466,18 +1476,32 @@ def _serve_llm():
                         f"Argument '{name}' is not supported when running in gRPC mode. "
                         f"The gRPC server is designed for use with external routers that handle "
                         f"these features (e.g., tool parsing, chat templates).")
-            if find_spec("smg_grpc_proto") is None:
-                raise ValueError(
-                    "gRPC serving with the SMG protocol requires the optional "
-                    "'smg-grpc-proto' package. Install it with: "
-                    'pip install "tensorrt_llm[grpc-smg]"')
-
-            from tensorrt_llm.grpc.smg.server import launch_smg_server
-
-            launch_smg_server(host,
-                              port,
-                              llm_args,
-                              served_model_name=served_model_name)
+            if grpc_protocol == "smg":
+                if find_spec("smg_grpc_proto") is None:
+                    raise ValueError(
+                        "gRPC serving with the SMG protocol requires the optional "
+                        "'smg-grpc-proto' package. Install it with: "
+                        'pip install "tensorrt_llm[grpc-smg]"')
+
+                from tensorrt_llm.grpc.smg.server import launch_smg_server
+
+                launch_smg_server(host,
+                                  port,
+                                  llm_args,
+                                  served_model_name=served_model_name)
+            else:
+                try:
+                    from tensorrt_llm.grpc.openengine.server import \
+                        launch_server as launch_grpc_server
+                except ImportError as error:
+                    raise click.ClickException(
+                        f"Failed to import OpenEngine support: {error}. "
+                        "Install the optional Python bindings with `python -m "
+                        "pip install --extra-index-url "
+                        "https://buf.build/gen/python "
+                        "\"tensorrt_llm[openengine]\"`.") from error
+
+                launch_grpc_server(host, port)
         else:
             # Default: launch OpenAI HTTP server
             launch_server(
@@ -1519,6 +1543,9 @@ def _serve_visual_gen():
             "--report_addr is only supported for the OpenAI HTTP server, not "
             f"the {'gRPC' if grpc else 'VisualGen'} server.")
     if is_visual_gen:
+        if grpc:
+            raise click.UsageError(
+                "--grpc is not supported by the VisualGen server")
         _serve_visual_gen()
     else:
         _serve_llm()

tensorrt_llm/grpc/openengine/README.md

@@ -0,0 +1,46 @@
+<!-- SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. -->
+<!-- SPDX-License-Identifier: Apache-2.0 -->
+
+# TensorRT-LLM OpenEngine stub server
+
+`trtllm-serve` can expose an experimental OpenEngine gRPC server instead of its normal OpenAI HTTP server. SMG remains the default gRPC protocol.
+
+Install the optional Python bindings from the Buf Schema Registry:
+
+```bash
+python -m pip install \
+  --extra-index-url https://buf.build/gen/python \
+  "tensorrt_llm[openengine]"
+```
+
+Then select OpenEngine when starting the gRPC server:
+
+```bash
+trtllm-serve <model> \
+  --grpc \
+  --grpc-protocol openengine \
+  --host 0.0.0.0 \
+  --port 50051
+```
+
+Existing `--grpc` invocations continue to select SMG. OpenEngine and VisualGen cannot be enabled together.
+
+This initial integration is a protocol stub. Every OpenEngine RPC returns gRPC status `UNIMPLEMENTED`; no request reaches the TensorRT-LLM engine. OpenEngine and SMG are independent protocol integrations. This integration does not make a replacement or convergence decision between them.
+
+## Dependency provenance
+
+The schema source is the Apache-2.0-licensed [`ai-dynamo/openengine`](https://github.com/ai-dynamo/openengine) repository at signed Git tag [`v0.1.0`](https://github.com/ai-dynamo/openengine/releases/tag/v0.1.0). That release maps to the public [`buf.build/openengine/openengine`](https://buf.build/openengine/openengine) module at immutable BSR commit `768a93c7b44e40f28c692ad0b471a8f2`.
+
+The BSR generated the pinned wheels from that module commit:
+
+| Package | Generator | Version | SHA-256 |
+| --- | --- | --- | --- |
+| `openengine-openengine-grpc-python` | [`grpc/python`](https://buf.build/grpc/python) | `1.67.1.2.20260730172104+768a93c7b44e` | `1485aed9799c4eb9367d1a261ca5cc5319f1e9b8d950ac98a26f3cb3641b8cf6` |
+| `openengine-openengine-protocolbuffers-python` | [`protocolbuffers/python`](https://buf.build/protocolbuffers/python) | `31.1.0.2.20260730172104+768a93c7b44e` | `6eae12c3d8d06147fccf608da9772d6391139031fabdafdb7cf4c71a19c1f25e` |
+| `openengine-openengine-protocolbuffers-pyi` | [`protocolbuffers/pyi`](https://buf.build/protocolbuffers/pyi) | `31.1.0.2.20260730172104+768a93c7b44e` | `8b0a054dbdaaa67459b3fa4786f13d8f6f4d30cf30be325f5416dbd97aba46a6` |
+
+Buf documents the package naming and version format in its [Python-generated SDK guide](https://buf.build/docs/bsr/generated-sdks/python/). The final version segment is the BSR commit prefix. The exact requirements are pinned in `requirements-openengine.txt`.
+
+## Maintenance boundary
+
+The OpenEngine contributor community owns this adapter, its tests, protocol version updates, and integration bugs. TensorRT-LLM internal APIs do not provide compatibility guarantees to protocol adapters. Adapter updates must follow core runtime changes and must not block normal TensorRT-LLM development or releases.

tensorrt_llm/grpc/openengine/__init__.py

@@ -0,0 +1,8 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""OpenEngine gRPC integration for TensorRT-LLM."""
+
+from .server import OpenEngineServer, launch_server
+
+__all__ = ["OpenEngineServer", "launch_server"]

tensorrt_llm/grpc/openengine/server.py

@@ -0,0 +1,100 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""OpenEngine gRPC server lifecycle for TensorRT-LLM."""
+
+import asyncio
+import signal
+
+import grpc
+import uvloop
+from openengine.v1 import openengine_pb2_grpc
+
+from tensorrt_llm.logger import logger
+
+__all__ = ["OpenEngineServer", "launch_server"]
+
+
+def _format_bind_address(host: str, port: int) -> str:
+    """Format a host and port as a gRPC bind address."""
+    if ":" in host and not (host.startswith("[") and host.endswith("]")):
+        host = f"[{host}]"
+    return f"{host}:{port}"
+
+
+class OpenEngineServer:
+    """OpenEngine gRPC server with intentionally unimplemented RPCs.
+
+    Args:
+        host: Interface on which the server listens.
+        port: Port on which the server listens. Use zero to select a free port.
+    """
+
+    def __init__(self, host: str, port: int) -> None:
+        self.host = host
+        self.port = port
+        self._server = grpc.aio.server()
+        openengine_pb2_grpc.add_InferenceServicer_to_server(
+            openengine_pb2_grpc.InferenceServicer(), self._server
+        )
+        openengine_pb2_grpc.add_ControlServicer_to_server(
+            openengine_pb2_grpc.ControlServicer(), self._server
+        )
+        self._bind_address = _format_bind_address(host, port)
+        bound_port = self._server.add_insecure_port(self._bind_address)
+        if bound_port == 0:
+            raise RuntimeError(f"Failed to bind OpenEngine server to {self._bind_address}")
+        if port == 0:
+            self.port = bound_port
+
+    async def start(self) -> None:
+        """Start accepting OpenEngine requests."""
+        await self._server.start()
+        address = _format_bind_address(self.host, self.port)
+        logger.info(f"OpenEngine stub server started on {address}")
+
+    async def stop(self, grace: float = 5.0) -> None:
+        """Stop accepting OpenEngine requests.
+
+        Args:
+            grace: Maximum time in seconds to allow active RPCs to finish.
+        """
+        await self._server.stop(grace=grace)
+        logger.info("OpenEngine stub server stopped")
+
+    async def wait_for_termination(self) -> None:
+        """Wait until the OpenEngine server terminates."""
+        await self._server.wait_for_termination()
+
+
+def launch_server(host: str, port: int) -> None:
+    """Launch the dedicated OpenEngine gRPC server.
+
+    Args:
+        host: Interface on which the server listens.
+        port: Port on which the server listens.
+    """
+
+    async def serve() -> None:
+        server = OpenEngineServer(host=host, port=port)
+        loop = asyncio.get_running_loop()
+        stop_event = asyncio.Event()
+
+        def signal_handler() -> None:
+            logger.info("Received shutdown signal")
+            stop_event.set()
+
+        for sig in (signal.SIGTERM, signal.SIGINT):
+            loop.add_signal_handler(sig, signal_handler)
+
+        try:
+            logger.warning(
+                "OpenEngine protocol support is a stub: no model is loaded and "
+                "all RPCs return UNIMPLEMENTED."
+            )
+            await server.start()
+            await stop_event.wait()
+        finally:
+            await server.stop()
+
+    uvloop.run(serve())

tests/unittest/api_stability/references/trtllm_serve_cli.yaml

@@ -193,6 +193,15 @@ commands:
         is_flag: true
         flags:
         - "--grpc"
+      grpc_protocol:
+        type: Choice(['openengine', 'smg'])
+        default: smg
+        status: prototype
+        required: false
+        multiple: false
+        is_flag: false
+        flags:
+        - "--grpc-protocol"
       host:
         type: str
         default: localhost