Skip to content

vllm.entrypoints.cohere.api_router

FastAPI router for the Cohere Chat v2 API (POST /cohere/v2/chat).

The Cohere v2 protocol models are sourced from the official cohere Python SDK (pip install cohere). To keep that an optional dependency for vLLM, the SDK-dependent imports - and the route handler itself - are gated on a one-shot probe at module load. If the SDK isn't installed, :func:attach_router becomes a no-op (with an info log) and vLLM continues to boot normally.

Even when the SDK is installed, :func:attach_router also requires VLLM_ENABLE_COHERE_API=1 in the environment before it will expose the route. This keeps non-Cohere deployments that pull in the SDK for unrelated reasons (e.g. test dependencies) from accidentally exposing the api.

Note: the handler must live at module scope (not inside attach_router) so that FastAPI's typing.get_type_hints resolves the CohereChatV2Request body annotation against the module's globals. Defining it locally inside attach_router would hide the type from get_type_hints, causing FastAPI to silently degrade the body parameter into a query parameter and reject every request with 422.

Classes:

Functions:

CohereErrorEnvelopeMiddleware

Bases: BaseHTTPMiddleware

Rewrite vLLM error bodies into the Cohere {message, id} shape.

The endpoint handler above already returns :class:CohereError for errors it owns, but globally-registered exception handlers (e.g. :func:validation_exception_handler for pydantic body errors, :func:http_exception_handler, engine error handlers) fire before the handler runs and produce vLLM's internal ErrorResponse shape ({"error": {"message": ...}}). That shape doesn't match the CohereError schema advertised on the route's OpenAPI responses, so clients (and schema-conformance tests like test_openai_schema.py) would see a mismatch on those paths. This middleware normalises any error body on /cohere/* responses to :class:CohereError.

Source code in vllm/entrypoints/cohere/api_router.py 
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
class CohereErrorEnvelopeMiddleware(BaseHTTPMiddleware):
    """Rewrite vLLM error bodies into the Cohere ``{message, id}`` shape.

    The endpoint handler above already returns :class:`CohereError` for
    errors it owns, but globally-registered exception handlers (e.g.
    :func:`validation_exception_handler` for pydantic body errors,
    :func:`http_exception_handler`, engine error handlers) fire *before*
    the handler runs and produce vLLM's internal
    ``ErrorResponse`` shape (``{"error": {"message": ...}}``). That
    shape doesn't match the ``CohereError`` schema advertised on the
    route's OpenAPI ``responses``, so clients (and schema-conformance
    tests like ``test_openai_schema.py``) would see a mismatch on
    those paths. This middleware normalises any error body on
    ``/cohere/*`` responses to :class:`CohereError`.
    """

    async def dispatch(self, request: Request, call_next):
        response = await call_next(request)
        if not request.url.path.startswith(_COHERE_PATH_PREFIX):
            return response
        if response.status_code < 400:
            return response
        content_type = response.headers.get("content-type", "")
        if not content_type.startswith("application/json"):
            return response

        body = b"".join([chunk async for chunk in response.body_iterator])
        translated = _translate_vllm_error_body(body, request)
        if translated is not None:
            return translated
        passthrough_headers = {
            k: v
            for k, v in response.headers.items()
            if k.lower() != "content-length"
        }
        return Response(
            content=body,
            status_code=response.status_code,
            headers=passthrough_headers,
            media_type=content_type,
        )

_error_response(error, raw_request, *, fallback_status=HTTPStatus.BAD_REQUEST)

Translate vLLM's internal error envelope into Cohere's shape.

Source code in vllm/entrypoints/cohere/api_router.py 
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
def _error_response(
    error: ErrorResponse,
    raw_request: Request | None,
    *,
    fallback_status: int = HTTPStatus.BAD_REQUEST,
) -> JSONResponse:
    """Translate vLLM's internal error envelope into Cohere's shape."""
    info = error.error
    status = info.code or fallback_status
    return JSONResponse(
        status_code=status,
        content=CohereError(
            message=sanitize_message(info.message),
            id=_request_id(raw_request),
        ).model_dump(exclude_none=True),
    )

_request_id(raw_request)

Best-effort lookup of the active request id.

Prefers the id the underlying chat handler stamped onto raw_request.state.request_metadata (if it got that far before failing), falling back to the X-Request-Id HTTP header. May return None if neither is available, in which case the field is omitted from the response.

Source code in vllm/entrypoints/cohere/api_router.py 
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
def _request_id(raw_request: Request | None) -> str | None:
    """Best-effort lookup of the active request id.

    Prefers the id the underlying chat handler stamped onto
    ``raw_request.state.request_metadata`` (if it got that far before
    failing), falling back to the ``X-Request-Id`` HTTP header. May
    return ``None`` if neither is available, in which case the field
    is omitted from the response.
    """
    if raw_request is None:
        return None
    meta = getattr(raw_request.state, "request_metadata", None)
    if meta is not None and getattr(meta, "request_id", None):
        return meta.request_id
    return raw_request.headers.get("X-Request-Id")

_translate_vllm_error_body(raw, request)

Translate a vLLM ErrorResponse body to a CohereError body.

Returns None if raw does not match the vLLM error envelope (which signals the middleware to pass the body through unchanged).

Source code in vllm/entrypoints/cohere/api_router.py 
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
def _translate_vllm_error_body(raw: bytes, request: Request) -> JSONResponse | None:
    """Translate a vLLM ``ErrorResponse`` body to a ``CohereError`` body.

    Returns ``None`` if ``raw`` does not match the vLLM error envelope
    (which signals the middleware to pass the body through unchanged).
    """
    try:
        data = json.loads(raw)
    except (json.JSONDecodeError, TypeError, ValueError):
        return None
    if not (
        isinstance(data, dict)
        and isinstance(data.get("error"), dict)
        and "message" in data["error"]
    ):
        return None
    try:
        err = ErrorResponse.model_validate(data)
    except Exception:  # noqa: BLE001 - malformed envelope; pass through
        return None
    return _error_response(err, request)

attach_router(app)

Register POST /cohere/v2/chat on app.

No-op when either:

  • the VLLM_ENABLE_COHERE_API env var isn't set to 1. The Cohere v2 endpoint is opt-in because it carries Cohere-specific request/response semantics (grounding citations, tool_plan, PLAN/THINKING_CONTENT blocks) that are only meaningful when serving a Cohere Command-family model.
  • the optional cohere SDK isn't installed (the v2 protocol models live there)

The two skip paths log at different levels: an operator who set VLLM_ENABLE_COHERE_API=1 but forgot to install cohere sees a WARNING (they explicitly asked for the endpoint and it's silently absent), whereas the default-off skip logs at debug.

Source code in vllm/entrypoints/cohere/api_router.py 
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
def attach_router(app: FastAPI) -> None:
    """Register ``POST /cohere/v2/chat`` on ``app``.

    No-op when either:

    * the ``VLLM_ENABLE_COHERE_API`` env var isn't set to ``1``. The
      Cohere v2 endpoint is opt-in because it carries Cohere-specific
      request/response semantics (grounding citations, tool_plan,
      PLAN/THINKING_CONTENT blocks) that are only meaningful when
      serving a Cohere Command-family model.
    * the optional ``cohere`` SDK isn't installed (the v2 protocol
      models live there)

    The two skip paths log at different levels: an operator who set
    ``VLLM_ENABLE_COHERE_API=1`` but forgot to install ``cohere`` sees
    a WARNING (they explicitly asked for the endpoint and it's silently
    absent), whereas the default-off skip logs at debug.
    """
    enabled = envs.VLLM_ENABLE_COHERE_API
    if not enabled:
        logger.debug(
            "VLLM_ENABLE_COHERE_API is not set; /cohere/v2/chat endpoint "
            "disabled. Set VLLM_ENABLE_COHERE_API=1 to enable it."
        )
        return
    if not _SDK_AVAILABLE:
        logger.warning(
            "VLLM_ENABLE_COHERE_API=1 but the `cohere` SDK is not "
            "installed; /cohere/v2/chat will not be exposed. Install "
            "with `pip install cohere` to enable the endpoint."
        )
        return
    app.include_router(router)
    app.add_middleware(CohereErrorEnvelopeMiddleware)