diff --git a/src/aiperf/endpoints/vllm_generate.py b/src/aiperf/endpoints/vllm_generate.py
new file mode 100644
index 0000000000..4fa01d2592
--- /dev/null
+++ b/src/aiperf/endpoints/vllm_generate.py
@@ -0,0 +1,142 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES.
+# SPDX-License-Identifier: Apache-2.0
+
+"""vLLM token-in/token-out ``/inference/v1/generate`` endpoint."""
+
+from __future__ import annotations
+
+from typing import Any
+
+from aiperf.common.models import (
+    BaseResponseData,
+    ExtractedPayload,
+    InferenceServerResponse,
+    ParsedResponse,
+    RequestInfo,
+    RequestRecord,
+)
+from aiperf.endpoints.base_endpoint import BaseEndpoint
+
+
+class VllmGenerateEndpoint(BaseEndpoint):
+    """Send and measure vLLM ``GenerateRequest`` token arrays.
+
+    The endpoint is intentionally non-streaming so the same payload works with
+    vLLM and Dynamo's vLLM-compatible engine API. Use ``--endpoint-path`` only
+    when the server mounts the API at a custom path.
+    """
+
+    def format_payload(self, request_info: RequestInfo) -> dict[str, Any]:
+        if len(request_info.turns) != 1:
+            raise ValueError(
+                "vLLM generate endpoint requires one token payload per turn"
+            )
+
+        turn = request_info.turns[0]
+        extra_body = dict(turn.extra_body or {})
+        token_ids = extra_body.pop("token_ids", None)
+        if not self._valid_token_ids(token_ids):
+            raise ValueError(
+                "turn.extra_body.token_ids must be a non-empty list of integers"
+            )
+
+        sampling_params = dict(extra_body.pop("sampling_params", {}) or {})
+        if turn.max_tokens is not None:
+            sampling_params.setdefault("max_tokens", turn.max_tokens)
+        if extra_body.pop("stream", False):
+            raise ValueError("vLLM generate endpoint does not support streaming")
+
+        payload: dict[str, Any] = {
+            "model": turn.model or request_info.model_endpoint.primary_model_name,
+            "token_ids": token_ids,
+            "sampling_params": sampling_params,
+            "stream": False,
+        }
+        if request_info.x_request_id:
+            payload["request_id"] = request_info.x_request_id
+        payload.update(dict(self.model_endpoint.endpoint.extra or []))
+        payload.update(extra_body)
+        if payload.get("stream") is not False:
+            raise ValueError("vLLM generate endpoint requires stream=false")
+        return payload
+
+    def extract_payload_inputs(self, payload: dict[str, Any]) -> ExtractedPayload:
+        result = ExtractedPayload()
+        token_ids = payload.get("token_ids")
+        if self._valid_token_ids(token_ids):
+            result.pretokenised_token_count = len(token_ids)
+        return result
+
+    def parse_response(
+        self, response: InferenceServerResponse
+    ) -> ParsedResponse | None:
+        return self._parse_response(response, prompt_tokens=None)
+
+    def extract_response_data(self, record: RequestRecord) -> list[ParsedResponse]:
+        prompt_tokens = self._prompt_token_count(record)
+        return [
+            parsed
+            for response in record.responses or []
+            if (parsed := self._parse_response(response, prompt_tokens)) is not None
+        ]
+
+    def _parse_response(
+        self,
+        response: InferenceServerResponse,
+        prompt_tokens: int | None,
+    ) -> ParsedResponse | None:
+        payload = response.get_json()
+        if not isinstance(payload, dict):
+            return None
+        choices = payload.get("choices")
+        if (
+            not isinstance(choices, list)
+            or not choices
+            or not isinstance(choices[0], dict)
+        ):
+            return None
+        completion_ids = choices[0].get("token_ids")
+        if not isinstance(completion_ids, list) or not all(
+            isinstance(token_id, int) and not isinstance(token_id, bool)
+            for token_id in completion_ids
+        ):
+            return None
+
+        completion_tokens = len(completion_ids)
+        usage = {
+            "prompt_tokens": prompt_tokens,
+            "completion_tokens": completion_tokens,
+            "total_tokens": (
+                prompt_tokens + completion_tokens if prompt_tokens is not None else None
+            ),
+        }
+        return ParsedResponse(
+            perf_ns=response.perf_ns,
+            data=BaseResponseData(),
+            usage=usage,
+            metadata={
+                "request_id": payload.get("request_id"),
+                "finish_reason": choices[0].get("finish_reason"),
+                "completion_token_ids": completion_ids,
+            },
+        )
+
+    @classmethod
+    def _prompt_token_count(cls, record: RequestRecord) -> int | None:
+        if not record.turns:
+            return None
+        turn = record.turns[-1]
+        payload = turn.raw_payload or turn.extra_body or {}
+        token_ids = payload.get("token_ids") if isinstance(payload, dict) else None
+        return len(token_ids) if cls._valid_token_ids(token_ids) else None
+
+    @staticmethod
+    def _valid_token_ids(value: Any) -> bool:
+        return (
+            isinstance(value, list)
+            and bool(value)
+            and all(
+                isinstance(token_id, int) and not isinstance(token_id, bool)
+                for token_id in value
+            )
+        )
diff --git a/src/aiperf/plugin/plugins.yaml b/src/aiperf/plugin/plugins.yaml
index d72121a9d0..5fe3c5782f 100644
--- a/src/aiperf/plugin/plugins.yaml
+++ b/src/aiperf/plugin/plugins.yaml
@@ -198,6 +198,19 @@ endpoint:
       supports_videos: true
       metrics_title: LLM Metrics
 
+  vllm_generate:
+    class: aiperf.endpoints.vllm_generate:VllmGenerateEndpoint
+    description: |
+      Non-streaming vLLM token-in/token-out generate endpoint. Accepts exact
+      token_ids and reconstructs usage from request and response token arrays.
+      Override the endpoint path for Dynamo's vLLM-compatible engine API.
+    metadata:
+      endpoint_path: /inference/v1/generate
+      supports_streaming: false
+      produces_tokens: true
+      tokenizes_input: false
+      metrics_title: LLM Metrics
+
   cohere_rankings:
     class: aiperf.endpoints.cohere_rankings:CohereRankingsEndpoint
     description: |
diff --git a/tests/unit/common/enums/test_endpoints_enums.py b/tests/unit/common/enums/test_endpoints_enums.py
index de01bd6541..8becefada5 100644
--- a/tests/unit/common/enums/test_endpoints_enums.py
+++ b/tests/unit/common/enums/test_endpoints_enums.py
@@ -69,6 +69,14 @@ class TestEndpointType:
                 "/generate",
                 "LLM Metrics",
             ),
+            (
+                EndpointType.VLLM_GENERATE,
+                "vllm_generate",
+                False,
+                True,
+                "/inference/v1/generate",
+                "LLM Metrics",
+            ),
         ],
     )
     def test_endpoint_type_metadata(
@@ -106,6 +114,7 @@ def test_endpoint_type_metadata(
             "hf_tei_rankings",
             "cohere_rankings",
             "huggingface_generate",
+            "vllm_generate",
         ],
     )
     def test_enum_string_comparison(self, tag_value):
diff --git a/tests/unit/endpoints/test_vllm_generate.py b/tests/unit/endpoints/test_vllm_generate.py
new file mode 100644
index 0000000000..905e5d42f7
--- /dev/null
+++ b/tests/unit/endpoints/test_vllm_generate.py
@@ -0,0 +1,83 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES.
+# SPDX-License-Identifier: Apache-2.0
+
+import orjson
+import pytest
+
+from aiperf.common.models import RequestRecord, TextResponse, Turn
+from aiperf.endpoints.vllm_generate import VllmGenerateEndpoint
+from aiperf.plugin import plugins
+from aiperf.plugin.enums import EndpointType
+from tests.unit.endpoints.conftest import create_model_endpoint, create_request_info
+
+
+@pytest.fixture
+def endpoint():
+    return VllmGenerateEndpoint(
+        create_model_endpoint(EndpointType.VLLM_GENERATE, model_name="test-model")
+    )
+
+
+def test_metadata():
+    metadata = plugins.get_endpoint_metadata(EndpointType.VLLM_GENERATE)
+    assert metadata.endpoint_path == "/inference/v1/generate"
+    assert metadata.supports_streaming is False
+    assert metadata.produces_tokens is True
+    assert metadata.tokenizes_input is False
+
+
+def test_format_payload(endpoint):
+    request = create_request_info(
+        model_endpoint=endpoint.model_endpoint,
+        max_tokens=17,
+        extra_body={"token_ids": [1, 2, 3], "sampling_params": {"temperature": 0}},
+    )
+
+    payload = endpoint.format_payload(request)
+
+    assert payload == {
+        "model": "test-model",
+        "token_ids": [1, 2, 3],
+        "sampling_params": {"temperature": 0, "max_tokens": 17},
+        "stream": False,
+        "request_id": "test-request-id",
+    }
+
+
+def test_format_payload_rejects_missing_tokens(endpoint):
+    request = create_request_info(model_endpoint=endpoint.model_endpoint)
+    with pytest.raises(ValueError, match="token_ids"):
+        endpoint.format_payload(request)
+
+
+def test_extract_payload_inputs_counts_exact_ids(endpoint):
+    extracted = endpoint.extract_payload_inputs({"token_ids": [10, 11, 12, 13]})
+    assert extracted.pretokenised_token_count == 4
+
+
+def test_extract_response_data_reconstructs_usage(endpoint):
+    response = TextResponse(
+        perf_ns=123,
+        text=orjson.dumps(
+            {
+                "request_id": "req-1",
+                "choices": [
+                    {"index": 0, "token_ids": [20, 21], "finish_reason": "stop"}
+                ],
+            }
+        ).decode(),
+        content_type="application/json",
+    )
+    record = RequestRecord(
+        model_name="test-model",
+        responses=[response],
+        turns=[Turn(role="user", raw_payload={"token_ids": [1, 2, 3]})],
+    )
+
+    parsed = endpoint.extract_response_data(record)
+
+    assert len(parsed) == 1
+    assert parsed[0].usage.prompt_tokens == 3
+    assert parsed[0].usage.completion_tokens == 2
+    assert parsed[0].usage.total_tokens == 5
+    assert parsed[0].metadata["completion_token_ids"] == [20, 21]