feat: add plain Service baseline without Gateway or routing proxy

README.md

@@ -414,7 +414,7 @@ llmdbenchmark --version
 | `-p NS` | `LLMDBENCH_NAMESPACE` | Namespace(s) to render into the plan |
 | `-m MODELS` | `LLMDBENCH_MODELS` | Model to render the plan for |
 | `-t METHODS` | `LLMDBENCH_METHODS` | Deployment method (`standalone`, `modelservice`) |
-| `--gateway-class CLASS` | `LLMDBENCH_GATEWAY_CLASS` | Override the scenario's `gateway.className`. Accepted on the modelservice path: `epponly`, `istio`, `agentgateway`, `gke`, `data-science-gateway-class`. Ignored (any value accepted, including `none`) when the active deploy method is `kustomize`, `standalone`, or `fma`. |
+| `--gateway-class CLASS` | `LLMDBENCH_GATEWAY_CLASS` | Override the scenario's `gateway.className`. Accepted on the modelservice path: `none`, `epponly`, `istio`, `agentgateway`, `gke`, `data-science-gateway-class`. `none` exposes decode vLLM directly through a plain Service with no Gateway, EPP, Envoy, or routing proxy. Ignored when the active deploy method is `kustomize`, `standalone`, or `fma`. |
 | `-f` / `--monitoring` | | Enable monitoring in rendered templates (PodMonitor, EPP verbosity) |
 | `-k FILE` | `LLMDBENCH_KUBECONFIG` / `KUBECONFIG` | Kubeconfig path (used for cluster resource auto-detection) |
 

config/templates/jinja/08_httproute.yaml.j2

@@ -35,13 +35,15 @@
    `siblingStacks` and `stackIndex` are injected by
    llmdbenchmark.parser.render_plans._process_stack.
 #}
-{# Modelservice-only template. Two skip conditions:
+{# Modelservice-only template. Skip when:
      1. -t kustomize -- upstream guide manifests define their own routing.
      2. gateway.className == 'epponly' -- no Gateway exists, so an HTTPRoute
-        would be unresolvable; clients hit the EPP service directly. #}
+      would be unresolvable; clients hit the EPP service directly.
+    3. gateway.className == 'none' -- clients hit the model server's plain
+      Service directly; no Gateway, router, or HTTPRoute is deployed. #}
 {% if standalone.enabled is defined and not standalone.enabled
       and not (kustomize.enabled | default(false))
-      and gateway.className | default('') != 'epponly' %}
+    and gateway.className | default('') not in ['epponly', 'none'] %}
 {% set _mode = (httpRoute.mode if httpRoute is defined and httpRoute.mode is defined else 'per-stack') %}
 {% if _mode == 'shared' %}
 {% set _owner_index = sharedInfraStackIndex | default(1) | int %}

config/templates/jinja/10_helmfile-main.yaml.j2

@@ -20,13 +20,16 @@ helmDefaults:
    ghcr.io/llm-d/charts. The scenario layer uses the chart's `router.*`
    keys directly -- see 12_router-values.yaml.j2 for the renderer. #}
 {% set _epponly = (gateway.className | default('') == 'epponly') %}
+{% set _direct = (gateway.className | default('') == 'none') %}
 {% set _gaie_chart = 'oci://ghcr.io/llm-d/charts/llm-d-router-standalone' if _epponly else 'oci://ghcr.io/llm-d/charts/llm-d-router-gateway' %}
 
 repositories:
   - name: llm-d-modelservice
     url: {{ helmRepositories.llmDModelservice.url }}
+{% if not _direct %}
   - name: llm-d-infra
     url: {{ helmRepositories.llmDInfra.url }}
+{% endif %}
 
 releases:
 {# `infra-llmdbench` (llm-d-infra chart -> Gateway resource + CRDs) is scenario-
@@ -44,7 +47,7 @@ releases:
    `epponly` skips this release entirely - there is no Gateway resource. #}
 {% set _is_shared_infra_owner = (stackIndex is not defined or stackIndex | int == sharedInfraStackIndex | default(1) | int) %}
 {% set _is_first_stack = _is_shared_infra_owner %}
-{% if _is_first_stack and not _epponly %}
+{% if _is_first_stack and not _epponly and not _direct %}
   - name: infra-llmdbench
     namespace: {{ gateway.namespace }}
     chart: llm-d-infra/llm-d-infra
@@ -62,16 +65,19 @@ releases:
     chart: llm-d-modelservice/llm-d-modelservice
     version: {{ chartVersions.llmDModelservice }}
     installed: true
+    {% if not _direct %}
     needs:
-{% if _is_first_stack and not _epponly %}
+    {% if _is_first_stack and not _epponly and not _direct %}
       - {{ gateway.namespace }}/infra-llmdbench
 {% endif %}
       - {{ gateway.namespace }}/{{ model_id_label }}-router
+    {% endif %}
     values:
       - ms-values.yaml
     labels:
       kind: inference-stack
 
+{% if not _direct %}
   - name: {{ model_id_label }}-router
     namespace: {{ gateway.namespace }}
     chart: {{ _gaie_chart }}
@@ -89,3 +95,4 @@ releases:
     labels:
       kind: inference-stack
 {% endif %}
+{% endif %}

config/templates/jinja/11_infra.yaml.j2

@@ -2,8 +2,9 @@
 {% if standalone.enabled is defined and not standalone.enabled and not (kustomize.enabled | default(false)) %}
 {% set gw_class = gateway.className %}
 {# `epponly` renders no infra values: no Gateway is installed, the standalone
-   router chart provides HTTP service directly via the EPP Envoy sidecar. #}
-{% if gw_class == 'epponly' %}
+  router chart provides HTTP service directly via the EPP Envoy sidecar.
+  `none` also renders no infra values because clients use a plain Service. #}
+{% if gw_class in ['epponly', 'none'] %}
 {% elif gw_class == 'gke' %}
 {# ── GKE: managed gateway with destinationRule ── #}
 gateway:

config/templates/jinja/13a_modelservice-direct-service.yaml.j2

@@ -0,0 +1,31 @@
+{# Modelservice direct baseline: expose decode vLLM without Gateway, EPP,
+   Envoy, or the routing proxy.  RenderPlans forces routing.proxy.enabled=false
+   for this mode, so vLLM binds directly to routing.servicePort. #}
+{% set decode_create = (decode.enabled if decode.enabled is defined else (decode.replicas | default(0) | int > 0)) %}
+{% if standalone.enabled is defined and not standalone.enabled
+      and not (kustomize.enabled | default(false))
+      and gateway.className | default('') == 'none'
+      and decode_create %}
+apiVersion: v1
+kind: Service
+metadata:
+  name: {{ model_id_label }}-direct
+  namespace: {{ namespace.name }}
+  labels:
+    app.kubernetes.io/managed-by: llm-d-benchmark
+    llm-d.ai/model: {{ model_id_label }}
+    llm-d.ai/role: decode
+    stood-up-from: llm-d-benchmark
+    stood-up-via: modelservice
+spec:
+  type: ClusterIP
+  selector:
+    llm-d.ai/model: {{ model_id_label }}
+    llm-d.ai/role: decode
+  ports:
+    - name: http
+      appProtocol: http
+      port: {{ routing.servicePort }}
+      targetPort: {{ routing.servicePort }}
+      protocol: TCP
+{% endif %}

config/templates/values/defaults.yaml

@@ -641,6 +641,8 @@ gateway:
   #                                Kubernetes Gateway is deployed, the EPP
   #                                runs with an Envoy sidecar that serves
   #                                HTTP directly (single-stack only).
+  #   none                       - plain ClusterIP Service to decode vLLM;
+  #                                no Gateway, EPP, Envoy, or routing proxy.
   className: istio
   providerNamespace: istio-system
   # Gateway version is derived from chartVersions.istiod at plan time

docs/standup.md

@@ -142,18 +142,22 @@ The scenario parameters can be roughly categorized in four groups:
 
 | Variable                                     | Meaning                                                                | Note                                                                                                     |
 | -------------------------------------------- | ---------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------- |
-| LLMDBENCH_VLLM_MODELSERVICE_GATEWAY_CLASS_NAME | Gateway implementation used for the inference gateway                 | Default=`istio`. Supported: `istio`, `agentgateway`, `gke`, `data-science-gateway-class`, `epponly`.     |
+| LLMDBENCH_VLLM_MODELSERVICE_GATEWAY_CLASS_NAME | Gateway implementation used for the inference gateway                 | Default=`istio`. Supported: `none`, `istio`, `agentgateway`, `gke`, `data-science-gateway-class`, `epponly`.     |
 
 Gateway class options (set via `gateway.className` in the scenario YAML):
 
 | `className`                  | What it deploys                                                                                                  | Use when                                                          |
 |------------------------------|------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------|
+| `none`                       | ModelService decode pods plus a plain ClusterIP Service; **no** Gateway, HTTPRoute, EPP, Envoy, or routing proxy | Measuring direct model-server performance and routing overhead    |
 | `istio` (default)            | istio-base + istiod control plane, a Gateway + HTTPRoute, the `llm-d-router-gateway-dev` chart                       | Default; most flexible / production deployments                   |
 | `agentgateway`               | agentgateway-crds + agentgateway controller, a Gateway + HTTPRoute, the `llm-d-router-gateway-dev` chart             | Want agentgateway's data plane instead of Envoy/Istio             |
 | `gke`                        | Uses GKE-managed Gateway controller; same `llm-d-router-gateway-dev` chart                                           | Running on GKE                                                    |
 | `data-science-gateway-class` | OpenDataHub / OpenShift AI managed Gateway                                                                           | Running on OpenShift AI                                           |
 | `epponly`                    | **No** Kubernetes Gateway, **no** HTTPRoute, the `llm-d-router-standalone-dev` chart (EPP with an Envoy sidecar serving HTTP) | You want llm-d's standalone router topology without any gateway   |
 
+`none` is a baseline lane, not a routing topology. It requires at least one
+decode replica and does not support P/D disaggregation.
+
 ### Overriding `gateway.className` from the CLI
 
 Every subcommand that renders templates (`plan`, `standup`, `experiment`,

llmdbenchmark/interface/experiment.py

@@ -47,7 +47,8 @@ def add_subcommands(
         default=env("LLMDBENCH_GATEWAY_CLASS"),
         help=(
             "Override the scenario's gateway.className. Supported values: "
-            "epponly, istio, agentgateway, gke, data-science-gateway-class. "
+            "none, epponly, istio, agentgateway, gke, "
+            "data-science-gateway-class. "
             "Only takes effect on the modelservice deploy path."
         ),
     )

llmdbenchmark/interface/plan.py

@@ -49,7 +49,8 @@ def add_subcommands(
         default=env("LLMDBENCH_GATEWAY_CLASS"),
         help=(
             "Override the scenario's gateway.className. Supported values: "
-            "epponly, istio, agentgateway, gke, data-science-gateway-class. "
+            "none, epponly, istio, agentgateway, gke, "
+            "data-science-gateway-class. "
             "Only takes effect on the modelservice deploy path."
         ),
     )

llmdbenchmark/interface/run.py

@@ -44,7 +44,8 @@ def add_subcommands(
         help=(
             "Override the scenario's gateway.className when the run phase "
             "re-renders templates for setup overrides. Supported values: "
-            "epponly, istio, agentgateway, gke, data-science-gateway-class."
+            "none, epponly, istio, agentgateway, gke, "
+            "data-science-gateway-class."
         ),
     )
     run_parser.add_argument(

llmdbenchmark/interface/standup.py

@@ -52,7 +52,8 @@ def add_subcommands(
         default=env("LLMDBENCH_GATEWAY_CLASS"),
         help=(
             "Override the scenario's gateway.className. Supported values: "
-            "epponly, istio, agentgateway, gke, data-science-gateway-class. "
+            "none, epponly, istio, agentgateway, gke, "
+            "data-science-gateway-class. "
             "Only takes effect on the modelservice deploy path -- ignored "
             "by kustomize/standalone/fma."
         ),

llmdbenchmark/parser/README.md

@@ -95,6 +95,11 @@ chain runs, then pops the nested copy so `config.yaml` has a single home per
 section. Templates, resolvers and standup steps always read the **top-level**
 keys, so the two spellings render identically -- pick one per section.
 
+Set `gateway.className: none` to skip Gateway and router resources and expose
+decode vLLM directly through a plain Service. The renderer also disables the
+modelservice routing proxy so the resulting lane measures the model server
+without Gateway, EPP, or Envoy overhead.
+
 The hoist runs **before** setup overrides so the precedence stays
 `defaults < scenario < treatment/CLI`: DoE experiment treatments and CLI
 overrides target the **top-level** dotted path (e.g.

llmdbenchmark/parser/render_plans.py

@@ -667,6 +667,7 @@ def _resolve_deploy_method(self, values: dict) -> dict:
     # at render time so a typo doesn't silently produce a broken Gateway /
     # InferencePool chart configuration.
     _SUPPORTED_GATEWAY_CLASSES: tuple[str, ...] = (
+        "none",
         "epponly",
         "istio",
         "agentgateway",
@@ -791,6 +792,71 @@ def _validate_epponly_constraints(
 
         return errors
 
+    @staticmethod
+    def _normalize_direct_service_mode(values: dict) -> dict:
+        """Make ``gateway.className=none`` a true direct-vLLM baseline.
+
+        The modelservice chart enables its per-pod routing proxy by default.
+        A plain Service pointing at that port would still put the proxy in the
+        request path, defeating the baseline.  Disable it before templates are
+        rendered so the chart makes vLLM bind directly to ``servicePort``.
+        """
+        gateway_class = (values.get("gateway") or {}).get("className", "")
+        modelservice_enabled = (values.get("modelservice") or {}).get("enabled", True)
+        if gateway_class == "none" and modelservice_enabled:
+            routing = values.setdefault("routing", {})
+            routing.setdefault("proxy", {})["enabled"] = False
+
+            # Accelerator-neutral guides may provide a custom command that
+            # binds decode vLLM to the proxy backend port. Direct mode has no
+            # proxy, so make the custom command follow the chart's normal
+            # proxy-disabled behavior and listen on the Service port instead.
+            decode_vllm = values.setdefault("decode", {}).get("vllm") or {}
+            custom_command = decode_vllm.get("customCommand")
+            if isinstance(custom_command, str):
+                decode_vllm["customCommand"] = custom_command.replace(
+                    "$VLLM_METRICS_PORT",
+                    "$VLLM_INFERENCE_PORT",
+                )
+                values["decode"]["vllm"] = decode_vllm
+        return values
+
+    @staticmethod
+    def _validate_direct_service_constraints(
+        values: dict,
+        stack_name: str,
+    ) -> list[str]:
+        """Reject configurations that require routing in direct mode."""
+        gateway_class = (values.get("gateway") or {}).get("className", "")
+        modelservice_enabled = (values.get("modelservice") or {}).get("enabled", True)
+        if gateway_class != "none" or not modelservice_enabled:
+            return []
+
+        errors: list[str] = []
+        http_route_mode = (values.get("httpRoute") or {}).get("mode")
+        if http_route_mode == "shared":
+            errors.append(
+                f"[{stack_name}] gateway.className=none cannot be used with "
+                "httpRoute.mode=shared (direct mode deploys no Gateway or "
+                "HTTPRoute)."
+            )
+
+        prefill = values.get("prefill") or {}
+        if prefill.get("enabled") and int(prefill.get("replicas", 0) or 0) > 0:
+            errors.append(
+                f"[{stack_name}] gateway.className=none cannot be used with "
+                "prefill replicas (direct mode bypasses P/D routing)."
+            )
+
+        decode = values.get("decode") or {}
+        decode_enabled = decode.get("enabled", int(decode.get("replicas", 0) or 0) > 0)
+        if not decode_enabled or int(decode.get("replicas", 0) or 0) < 1:
+            errors.append(
+                f"[{stack_name}] gateway.className=none requires at least one "
+                "decode replica to back the direct Service."
+            )
+        return errors
+
     def _log_image_overrides(self, values: dict) -> None:
         """Log images that have been explicitly set (not 'auto').
 
@@ -1596,6 +1662,7 @@ def _process_stack(
             merged_values, total_stacks=total_stacks
         )
         merged_values = self._resolve_inference_pool_host(merged_values)
+        merged_values = self._normalize_direct_service_mode(merged_values)
         merged_values = self._normalize_router_block(merged_values)
         merged_values = self._substitute_config_variables(merged_values)
 
@@ -1614,6 +1681,14 @@ def _process_stack(
             self.logger.log_error(msg)
             stack_errors.render_errors.append(msg)
 
+        direct_service_errors = self._validate_direct_service_constraints(
+            merged_values,
+            stack_name=stack_name,
+        )
+        for msg in direct_service_errors:
+            self.logger.log_error(msg)
+            stack_errors.render_errors.append(msg)
+
         kustomize_errors = self._validate_kustomize_patches(
             merged_values,
             stack_name=stack_name,

llmdbenchmark/run/steps/step_03_detect_endpoint.py

@@ -9,6 +9,7 @@
     find_fma_endpoint,
     find_gateway_endpoint,
     find_epponly_endpoint,
+    find_direct_modelservice_endpoint,
     find_custom_endpoint,
     find_kustomize_endpoint,
     discover_hf_token_secret,
@@ -146,6 +147,23 @@ def execute(
                     namespace,
                     model_id_label,
                 )
+            elif gateway_class == "none":
+                model_id_label = plan_config.get("model_id_label", "")
+                direct_port = str(
+                    self._resolve(
+                        plan_config,
+                        "routing.servicePort",
+                        default="8000",
+                    )
+                )
+                service_ip, service_name, gateway_port = (
+                    find_direct_modelservice_endpoint(
+                        cmd,
+                        namespace,
+                        model_id_label,
+                        direct_port,
+                    )
+                )
             else:
                 service_ip, service_name, gateway_port = find_gateway_endpoint(
                     cmd, namespace, release

llmdbenchmark/smoketests/base.py

@@ -15,6 +15,7 @@
     _rand_suffix,
     compute_gateway_path_prefix,
     find_custom_endpoint,
+    find_direct_modelservice_endpoint,
     find_epponly_endpoint,
     find_gateway_endpoint,
     find_kustomize_endpoint,
@@ -144,6 +145,16 @@ def discover_endpoint(
                 namespace,
                 model_id_label,
             )
+        elif gateway_class == "none":
+            direct_port = str(
+                _nested_get(plan_config, "routing", "servicePort") or "8000"
+            )
+            service_ip, _, gateway_port = find_direct_modelservice_endpoint(
+                cmd,
+                namespace,
+                model_id_label,
+                direct_port,
+            )
         else:
             service_ip, _, gateway_port = find_gateway_endpoint(cmd, namespace, release)