| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Unit tests for ``vllm/entrypoints/cohere/api_router.py``. |
| 4 | |
| 5 | Covers: |
| 6 | |
| 7 | * The optional-import guard: ``attach_router`` is a no-op when the |
| 8 | ``cohere`` SDK isn't installed. |
| 9 | * The env-var opt-in gate: ``attach_router`` is a no-op unless |
| 10 | ``VLLM_ENABLE_COHERE_API=1`` is set. |
| 11 | * The router wiring: response shapes (JSON + SSE), error translation, |
| 12 | and the ``cohere_serving_chat_v2 is None`` fallback (501 Not |
| 13 | Implemented). |
| 14 | * The render route (``POST /cohere/v2/chat/render``): that the v2 body |
| 15 | reaches :class:`ServingRender` as a converted ``ChatCompletionRequest`` |
| 16 | and that the resulting ``GenerateRequest`` is returned verbatim. |
| 17 | """ |
| 18 | |
| 19 | import json |
| 20 | from argparse import Namespace |
| 21 | from collections.abc import AsyncGenerator |
| 22 | from http import HTTPStatus |
| 23 | |
| 24 | import pytest |
| 25 | from fastapi import FastAPI, HTTPException |
| 26 | from fastapi.exceptions import RequestValidationError |
| 27 | from fastapi.testclient import TestClient |
| 28 | |
| 29 | from vllm.entrypoints.cohere import api_router as api_router_mod |
| 30 | from vllm.entrypoints.cohere.api_router import attach_router |
| 31 | from vllm.entrypoints.cohere.protocol import ( |
| 32 | AssistantMessageResponse, |
| 33 | CohereChatV2Response, |
| 34 | ) |
| 35 | from vllm.entrypoints.cohere.serving import CohereServingChatV2 |
| 36 | from vllm.entrypoints.scale_out.token_in_token_out.protocol import GenerateRequest |
| 37 | from vllm.entrypoints.serve.engine.protocol import ErrorInfo, ErrorResponse |
| 38 | from vllm.entrypoints.serve.exception_handling.handlers.http import ( |
| 39 | http_exception_handler, |
| 40 | ) |
| 41 | from vllm.entrypoints.serve.exception_handling.handlers.validation import ( |
| 42 | validation_exception_handler, |
| 43 | ) |
| 44 | from vllm.entrypoints.serve.exception_handling.handlers.vllm_error import ( |
| 45 | vllm_error_handler, |
| 46 | ) |
| 47 | from vllm.exceptions import VLLMError |
| 48 | from vllm.sampling_params import SamplingParams |
| 49 | |
| 50 | |
| 51 | @pytest.fixture(autouse=True) |
| 52 | def _enable_cohere_api(monkeypatch): |
| 53 | """Auto-enable the Cohere API gate for every test in this module. |
| 54 | |
| 55 | The endpoint is opt-in in production (``VLLM_ENABLE_COHERE_API=1``); |
| 56 | every test in this file exercises the enabled path *except* the |
| 57 | dedicated gate test in :class:`TestEnvVarGate`, which unsets the |
| 58 | flag inside the test body. |
| 59 | """ |
| 60 | monkeypatch.setenv("VLLM_ENABLE_COHERE_API", "1") |
| 61 | |
| 62 | |
| 63 | # ---------------------------------------------------------------------- |
| 64 | # Fakes |
| 65 | # ---------------------------------------------------------------------- |
| 66 | |
| 67 | |
| 68 | class _Handler: |
| 69 | """Minimal stand-in for :class:`CohereServingChatV2` used by the |
| 70 | router. Each test sets ``self.result`` to either: |
| 71 | |
| 72 | * a :class:`CohereChatV2Response` (non-streaming JSON path); |
| 73 | * an async generator yielding SSE frames (streaming path); |
| 74 | * an :class:`ErrorResponse` (error envelope path); or |
| 75 | * an exception (router-level 500 path). |
| 76 | """ |
| 77 | |
| 78 | def __init__(self, result): |
| 79 | self.result = result |
| 80 | |
| 81 | async def create_chat_v2(self, request, raw_request): |
| 82 | if isinstance(self.result, Exception): |
| 83 | raise self.result |
| 84 | return self.result |
| 85 | |
| 86 | |
| 87 | class _RenderChatHandler: |
| 88 | """Stand-in for :class:`CohereServingChatV2` on the render route. |
| 89 | |
| 90 | Only ``to_chat_completion_request`` is reachable from |
| 91 | ``/cohere/v2/chat/render``, and it delegates to the real conversion |
| 92 | classmethod (which needs no engine or tokenizer) so the tests |
| 93 | exercise the actual v2 -> ``ChatCompletionRequest`` translation |
| 94 | rather than a stub of it. |
| 95 | """ |
| 96 | |
| 97 | def to_chat_completion_request(self, request): |
| 98 | return CohereServingChatV2._convert_v2_to_chat_completion(request) |
| 99 | |
| 100 | |
| 101 | class _RenderHandler: |
| 102 | """Minimal stand-in for :class:`ServingRender`. |
| 103 | |
| 104 | Records the ``ChatCompletionRequest`` it was handed so tests can |
| 105 | assert on the conversion, and returns ``self.result`` - either a |
| 106 | :class:`GenerateRequest`, an :class:`ErrorResponse`, or an exception |
| 107 | to raise. |
| 108 | """ |
| 109 | |
| 110 | def __init__(self, result): |
| 111 | self.result = result |
| 112 | self.seen_request = None |
| 113 | |
| 114 | async def render_chat_request(self, request): |
| 115 | self.seen_request = request |
| 116 | if isinstance(self.result, Exception): |
| 117 | raise self.result |
| 118 | return self.result |
| 119 | |
| 120 | |
| 121 | def _generate_request() -> GenerateRequest: |
| 122 | return GenerateRequest( |
| 123 | request_id="chatcmpl-abc", |
| 124 | token_ids=[5, 6, 7], |
| 125 | sampling_params=SamplingParams(temperature=0.25), |
| 126 | model="m", |
| 127 | ) |
| 128 | |
| 129 | |
| 130 | def _build_app(handler: _Handler | None, *, enable_scale_out: bool = True) -> FastAPI: |
| 131 | app = FastAPI() |
| 132 | app.state.args = Namespace(enable_scale_out=enable_scale_out, tokens_only=False) |
| 133 | attach_router(app) |
| 134 | app.state.cohere_serving_chat_v2 = handler |
| 135 | return app |
| 136 | |
| 137 | |
| 138 | def _build_render_app( |
| 139 | chat_handler, render_handler, *, enable_scale_out: bool = True |
| 140 | ) -> FastAPI: |
| 141 | app = FastAPI() |
| 142 | app.state.args = Namespace(enable_scale_out=enable_scale_out, tokens_only=False) |
| 143 | attach_router(app) |
| 144 | app.state.cohere_serving_chat_v2 = chat_handler |
| 145 | app.state.serving_render = render_handler |
| 146 | return app |
| 147 | |
| 148 | |
| 149 | def _build_app_with_vllm_handlers(handler: _Handler | None) -> FastAPI: |
| 150 | """Build a FastAPI app that mirrors the real vLLM setup by installing |
| 151 | validation, HTTP, and vLLM exception handlers. The |
| 152 | :class:`CohereErrorEnvelopeMiddleware` registered by ``attach_router`` |
| 153 | is expected to translate any resulting vLLM ``ErrorResponse`` body |
| 154 | into the ``CohereError`` wire shape. |
| 155 | """ |
| 156 | app = FastAPI() |
| 157 | # ``validation_exception_handler`` reads ``req.app.state.args``; the |
| 158 | # real cli builds this via argparse. |
| 159 | app.state.args = Namespace( |
| 160 | log_error_stack=False, enable_scale_out=True, tokens_only=False |
| 161 | ) |
| 162 | attach_router(app) |
| 163 | app.state.cohere_serving_chat_v2 = handler |
| 164 | app.exception_handler(RequestValidationError)(validation_exception_handler) |
| 165 | app.exception_handler(HTTPException)(http_exception_handler) |
| 166 | app.exception_handler(VLLMError)(vllm_error_handler) |
| 167 | return app |
| 168 | |
| 169 | |
| 170 | def _minimal_request_body() -> dict: |
| 171 | return { |
| 172 | "model": "m", |
| 173 | "messages": [{"role": "user", "content": "hi"}], |
| 174 | } |
| 175 | |
| 176 | |
| 177 | # ---------------------------------------------------------------------- |
| 178 | # Optional-import guard |
| 179 | # ---------------------------------------------------------------------- |
| 180 | |
| 181 | |
| 182 | class TestOptionalCohereImport: |
| 183 | """``attach_router`` probes for the SDK once at module load (because |
| 184 | the route handler uses types imported from ``cohere``) and stashes |
| 185 | the result in ``_SDK_AVAILABLE``. Tests simulate the "SDK missing" |
| 186 | state by flipping that flag for the duration of the test. |
| 187 | |
| 188 | ``attach_router`` checks the env-var gate before the SDK probe, so |
| 189 | the SDK-missing branch is only reachable when the operator opts in |
| 190 | via ``VLLM_ENABLE_COHERE_API=1``. The flag-off-and-SDK-missing case |
| 191 | below exists to pin down that ordering — the flag-off short-circuits |
| 192 | """ |
| 193 | |
| 194 | def test_flag_off_and_sdk_missing_stays_silent_about_sdk(self, monkeypatch, caplog): |
| 195 | """Flag off doesn't do SDK-missing check. |
| 196 | |
| 197 | When the operator hasn't opted in, ``attach_router`` must not |
| 198 | warn about the ``cohere`` SDK being missing: they never asked |
| 199 | for the endpoint, so surfacing the SDK gap is misleading noise. |
| 200 | Only the flag-off DEBUG message should fire. |
| 201 | """ |
| 202 | monkeypatch.delenv("VLLM_ENABLE_COHERE_API", raising=False) |
| 203 | monkeypatch.setattr(api_router_mod, "_SDK_AVAILABLE", False) |
| 204 | |
| 205 | with caplog.at_level("DEBUG", logger="vllm.entrypoints.cohere.api_router"): |
| 206 | app = FastAPI() |
| 207 | attach_router(app) |
| 208 | |
| 209 | paths = [getattr(r, "path", None) for r in app.routes] |
| 210 | assert "/cohere/v2/chat" not in paths |
| 211 | # The flag-off short-circuit ran; the SDK check never did. |
| 212 | assert not any( |
| 213 | "SDK is not installed" in rec.message for rec in caplog.records |
| 214 | ), "SDK-missing log leaked despite the flag being off" |
| 215 | |
| 216 | def test_flag_on_but_sdk_missing_logs_warning(self, monkeypatch, caplog): |
| 217 | """Misconfiguration path: the operator explicitly opted into the |
| 218 | endpoint via ``VLLM_ENABLE_COHERE_API=1`` (already set by the |
| 219 | autouse fixture) but forgot to install ``cohere``. |
| 220 | """ |
| 221 | monkeypatch.setattr(api_router_mod, "_SDK_AVAILABLE", False) |
| 222 | |
| 223 | with caplog.at_level("DEBUG", logger="vllm.entrypoints.cohere.api_router"): |
| 224 | app = FastAPI() |
| 225 | attach_router(app) |
| 226 | |
| 227 | paths = [getattr(r, "path", None) for r in app.routes] |
| 228 | assert "/cohere/v2/chat" not in paths |
| 229 | warn_sdk_records = [ |
| 230 | rec |
| 231 | for rec in caplog.records |
| 232 | if "VLLM_ENABLE_COHERE_API=1" in rec.message |
| 233 | and "SDK is not installed" in rec.message |
| 234 | ] |
| 235 | assert warn_sdk_records, ( |
| 236 | "expected a WARNING that pairs the opt-in flag with the " |
| 237 | "missing SDK so operators notice the misconfiguration" |
| 238 | ) |
| 239 | assert all(rec.levelname == "WARNING" for rec in warn_sdk_records) |
| 240 | |
| 241 | def test_attach_router_registers_route_when_cohere_present(self): |
| 242 | app = _build_app(handler=None) |
| 243 | paths = [getattr(r, "path", None) for r in app.routes] |
| 244 | assert "/cohere/v2/chat" in paths |
| 245 | |
| 246 | |
| 247 | # ---------------------------------------------------------------------- |
| 248 | # VLLM_ENABLE_COHERE_API gate |
| 249 | # ---------------------------------------------------------------------- |
| 250 | |
| 251 | |
| 252 | class TestEnvVarGate: |
| 253 | """The Cohere v2 endpoint is opt-in via ``VLLM_ENABLE_COHERE_API``. |
| 254 | |
| 255 | Even with the SDK installed, :func:`attach_router` must skip route |
| 256 | registration and middleware installation unless the env flag is |
| 257 | set. The autouse fixture on this module enables the flag by |
| 258 | default, so each test here explicitly disables it. |
| 259 | """ |
| 260 | |
| 261 | def test_attach_router_noop_when_flag_unset(self, monkeypatch, caplog): |
| 262 | monkeypatch.delenv("VLLM_ENABLE_COHERE_API", raising=False) |
| 263 | |
| 264 | # The flag-off skip logs at DEBUG on purpose: this is the default |
| 265 | # state for every non-Cohere vLLM deployment, so an INFO log on |
| 266 | # every server startup would be pointless noise. The test raises |
| 267 | # caplog's level accordingly. |
| 268 | with caplog.at_level("DEBUG", logger="vllm.entrypoints.cohere.api_router"): |
| 269 | app = FastAPI() |
| 270 | attach_router(app) |
| 271 | |
| 272 | paths = [getattr(r, "path", None) for r in app.routes] |
| 273 | assert "/cohere/v2/chat" not in paths |
| 274 | debug_flag_records = [ |
| 275 | rec |
| 276 | for rec in caplog.records |
| 277 | if "VLLM_ENABLE_COHERE_API is not set" in rec.message |
| 278 | ] |
| 279 | assert debug_flag_records, ( |
| 280 | "expected a DEBUG message that the cohere flag is off" |
| 281 | ) |
| 282 | assert all(rec.levelname == "DEBUG" for rec in debug_flag_records) |
| 283 | |
| 284 | def test_attach_router_noop_when_flag_zero(self, monkeypatch): |
| 285 | monkeypatch.setenv("VLLM_ENABLE_COHERE_API", "0") |
| 286 | |
| 287 | app = FastAPI() |
| 288 | attach_router(app) |
| 289 | |
| 290 | paths = [getattr(r, "path", None) for r in app.routes] |
| 291 | assert "/cohere/v2/chat" not in paths |
| 292 | |
| 293 | |
| 294 | # ---------------------------------------------------------------------- |
| 295 | # Endpoint behavior |
| 296 | # ---------------------------------------------------------------------- |
| 297 | |
| 298 | |
| 299 | class TestEndpoint: |
| 300 | def test_501_when_handler_missing(self): |
| 301 | app = _build_app(handler=None) |
| 302 | with TestClient(app) as client: |
| 303 | r = client.post("/cohere/v2/chat", json=_minimal_request_body()) |
| 304 | assert r.status_code == HTTPStatus.NOT_IMPLEMENTED |
| 305 | body = r.json() |
| 306 | assert "does not support" in body["message"] |
| 307 | assert "id" not in body # excluded by ``exclude_none=True`` |
| 308 | |
| 309 | def test_non_streaming_response_is_json(self): |
| 310 | msg = AssistantMessageResponse(content=[{"type": "text", "text": "hello"}]) |
| 311 | result = CohereChatV2Response(id="r1", finish_reason="COMPLETE", message=msg) |
| 312 | app = _build_app(handler=_Handler(result)) |
| 313 | with TestClient(app) as client: |
| 314 | r = client.post("/cohere/v2/chat", json=_minimal_request_body()) |
| 315 | assert r.status_code == HTTPStatus.OK |
| 316 | assert r.headers["content-type"].startswith("application/json") |
| 317 | body = r.json() |
| 318 | assert body["id"] == "r1" |
| 319 | assert body["finish_reason"] == "COMPLETE" |
| 320 | assert body["message"]["content"][0]["text"] == "hello" |
| 321 | |
| 322 | def test_streaming_response_is_sse(self): |
| 323 | async def _gen() -> AsyncGenerator[str, None]: |
| 324 | yield 'data: {"type":"message-start"}\n\n' |
| 325 | yield "data: [DONE]\n\n" |
| 326 | |
| 327 | app = _build_app(handler=_Handler(_gen())) |
| 328 | with TestClient(app) as client: |
| 329 | r = client.post( |
| 330 | "/cohere/v2/chat", |
| 331 | json={**_minimal_request_body(), "stream": True}, |
| 332 | ) |
| 333 | assert r.status_code == HTTPStatus.OK |
| 334 | assert r.headers["content-type"].startswith("text/event-stream") |
| 335 | body = r.text |
| 336 | assert "message-start" in body |
| 337 | assert body.rstrip().endswith("[DONE]") |
| 338 | |
| 339 | def test_error_response_translated_to_cohere_envelope(self): |
| 340 | err = ErrorResponse( |
| 341 | error=ErrorInfo( |
| 342 | message="bad request", |
| 343 | type="bad_request", |
| 344 | code=400, |
| 345 | ) |
| 346 | ) |
| 347 | app = _build_app(handler=_Handler(err)) |
| 348 | with TestClient(app) as client: |
| 349 | r = client.post("/cohere/v2/chat", json=_minimal_request_body()) |
| 350 | assert r.status_code == HTTPStatus.BAD_REQUEST |
| 351 | body = r.json() |
| 352 | assert body == {"message": "bad request"} |
| 353 | |
| 354 | def test_handler_exception_returns_500_envelope(self): |
| 355 | app = _build_app(handler=_Handler(RuntimeError("kaboom"))) |
| 356 | with TestClient(app) as client: |
| 357 | r = client.post("/cohere/v2/chat", json=_minimal_request_body()) |
| 358 | assert r.status_code == HTTPStatus.INTERNAL_SERVER_ERROR |
| 359 | body = r.json() |
| 360 | assert body == {"message": "kaboom"} |
| 361 | |
| 362 | def test_non_json_content_type_rejected(self): |
| 363 | """The ``validate_json_request`` dependency raises |
| 364 | ``RequestValidationError`` (HTTP 422) for non-JSON content |
| 365 | types, matching the behavior of the other vLLM API routers. |
| 366 | """ |
| 367 | app = _build_app(handler=None) |
| 368 | with TestClient(app) as client: |
| 369 | r = client.post( |
| 370 | "/cohere/v2/chat", |
| 371 | content=json.dumps(_minimal_request_body()), |
| 372 | headers={"content-type": "text/plain"}, |
| 373 | ) |
| 374 | assert r.status_code == HTTPStatus.UNPROCESSABLE_ENTITY |
| 375 | |
| 376 | def test_invalid_body_returns_422(self): |
| 377 | # ``model`` is required; omit it to trip Pydantic validation. |
| 378 | app = _build_app(handler=None) |
| 379 | with TestClient(app) as client: |
| 380 | r = client.post( |
| 381 | "/cohere/v2/chat", |
| 382 | json={"messages": [{"role": "user", "content": "hi"}]}, |
| 383 | ) |
| 384 | assert r.status_code == HTTPStatus.UNPROCESSABLE_ENTITY |
| 385 | |
| 386 | |
| 387 | # ---------------------------------------------------------------------- |
| 388 | # Render endpoint |
| 389 | # ---------------------------------------------------------------------- |
| 390 | |
| 391 | |
| 392 | class TestRenderEndpoint: |
| 393 | """``POST /cohere/v2/chat/render`` is the Cohere counterpart to |
| 394 | ``POST /v1/chat/completions/render``: it converts the v2 body to a |
| 395 | ``ChatCompletionRequest`` and hands it to the shared |
| 396 | ``ServingRender``, returning the resulting ``GenerateRequest`` |
| 397 | instead of running generation. |
| 398 | |
| 399 | The route therefore spans *two* app-state handlers, and these tests |
| 400 | pin both halves: that the conversion preserves the Cohere-specific |
| 401 | request surface, and that the render handler's output (or error) |
| 402 | reaches the client unchanged. |
| 403 | """ |
| 404 | |
| 405 | def test_route_registered(self): |
| 406 | app = _build_render_app(_RenderChatHandler(), _RenderHandler(None)) |
| 407 | paths = [getattr(r, "path", None) for r in app.routes] |
| 408 | assert "/cohere/v2/chat/render" in paths |
| 409 | |
| 410 | def test_route_not_registered_when_flag_unset(self): |
| 411 | app = _build_app(handler=None, enable_scale_out=False) |
| 412 | paths = [getattr(r, "path", None) for r in app.routes] |
| 413 | assert "/cohere/v2/chat/render" not in paths |
| 414 | assert "/cohere/v2/chat" in paths |
| 415 | |
| 416 | def test_returns_generate_request_json(self): |
| 417 | render_handler = _RenderHandler(_generate_request()) |
| 418 | app = _build_render_app(_RenderChatHandler(), render_handler) |
| 419 | with TestClient(app) as client: |
| 420 | r = client.post("/cohere/v2/chat/render", json=_minimal_request_body()) |
| 421 | assert r.status_code == HTTPStatus.OK |
| 422 | assert r.headers["content-type"].startswith("application/json") |
| 423 | body = r.json() |
| 424 | assert body["request_id"] == "chatcmpl-abc" |
| 425 | assert body["token_ids"] == [5, 6, 7] |
| 426 | assert body["sampling_params"]["temperature"] == 0.25 |
| 427 | # The body must be a wire-valid GenerateRequest so callers can |
| 428 | # feed it straight back to /v1/generate. |
| 429 | assert GenerateRequest.model_validate(body).token_ids == [5, 6, 7] |
| 430 | |
| 431 | def test_cohere_request_surface_survives_conversion(self): |
| 432 | """The point of routing through ``CohereServingChatV2`` rather |
| 433 | than asking callers to pre-convert: Cohere-only fields |
| 434 | (``documents``, ``safety_mode``, ``citation_options``) must land |
| 435 | in ``chat_template_kwargs`` so the rendered prompt matches what |
| 436 | ``/cohere/v2/chat`` would have produced. |
| 437 | """ |
| 438 | render_handler = _RenderHandler(_generate_request()) |
| 439 | app = _build_render_app(_RenderChatHandler(), render_handler) |
| 440 | body = { |
| 441 | **_minimal_request_body(), |
| 442 | "documents": [{"id": "doc-1", "data": {"text": "Paris is in France."}}], |
| 443 | "safety_mode": "STRICT", |
| 444 | "max_tokens": 32, |
| 445 | } |
| 446 | with TestClient(app) as client: |
| 447 | r = client.post("/cohere/v2/chat/render", json=body) |
| 448 | assert r.status_code == HTTPStatus.OK |
| 449 | |
| 450 | seen = render_handler.seen_request |
| 451 | assert seen is not None |
| 452 | assert seen.model == "m" |
| 453 | assert seen.messages[0]["content"] == "hi" |
| 454 | # ``max_tokens`` matters because ServingRender feeds it to |
| 455 | # ``get_max_tokens`` when building sampling params. |
| 456 | assert seen.max_completion_tokens == 32 |
| 457 | kwargs = seen.chat_template_kwargs or {} |
| 458 | assert kwargs["documents"][0]["id"] == "doc-1" |
| 459 | # The conversion lower-cases safety_mode for the renderer. |
| 460 | assert kwargs["safety_mode"] == "strict" |
| 461 | |
| 462 | def test_501_when_chat_handler_missing(self): |
| 463 | app = _build_render_app(None, _RenderHandler(_generate_request())) |
| 464 | with TestClient(app) as client: |
| 465 | r = client.post("/cohere/v2/chat/render", json=_minimal_request_body()) |
| 466 | assert r.status_code == HTTPStatus.NOT_IMPLEMENTED |
| 467 | assert "does not support" in r.json()["message"] |
| 468 | |
| 469 | def test_501_when_render_handler_missing(self): |
| 470 | # A server without ``serving_render`` in app state (e.g. a |
| 471 | # non-generate task set) can't render, even though the Cohere |
| 472 | # chat handler exists. |
| 473 | app = _build_render_app(_RenderChatHandler(), None) |
| 474 | with TestClient(app) as client: |
| 475 | r = client.post("/cohere/v2/chat/render", json=_minimal_request_body()) |
| 476 | assert r.status_code == HTTPStatus.NOT_IMPLEMENTED |
| 477 | assert "does not support" in r.json()["message"] |
| 478 | |
| 479 | def test_error_response_translated_to_cohere_envelope(self): |
| 480 | err = ErrorResponse( |
| 481 | error=ErrorInfo(message="unknown model", type="NotFound", code=404) |
| 482 | ) |
| 483 | app = _build_render_app(_RenderChatHandler(), _RenderHandler(err)) |
| 484 | with TestClient(app) as client: |
| 485 | r = client.post("/cohere/v2/chat/render", json=_minimal_request_body()) |
| 486 | assert r.status_code == HTTPStatus.NOT_FOUND |
| 487 | assert r.json() == {"message": "unknown model"} |
| 488 | |
| 489 | def test_render_exception_returns_500_envelope(self): |
| 490 | app = _build_render_app( |
| 491 | _RenderChatHandler(), _RenderHandler(RuntimeError("tokenizer exploded")) |
| 492 | ) |
| 493 | with TestClient(app) as client: |
| 494 | r = client.post("/cohere/v2/chat/render", json=_minimal_request_body()) |
| 495 | assert r.status_code == HTTPStatus.INTERNAL_SERVER_ERROR |
| 496 | assert r.json() == {"message": "tokenizer exploded"} |
| 497 | |
| 498 | def test_conversion_exception_returns_500_envelope(self): |
| 499 | """Conversion runs inside the same ``try`` as the render call, so |
| 500 | a bad v2 body that slips past Pydantic still yields the Cohere |
| 501 | error envelope rather than an unhandled 500 with vLLM's shape. |
| 502 | """ |
| 503 | |
| 504 | class _Boom: |
| 505 | def to_chat_completion_request(self, request): |
| 506 | raise ValueError("unconvertible message") |
| 507 | |
| 508 | app = _build_render_app(_Boom(), _RenderHandler(_generate_request())) |
| 509 | with TestClient(app) as client: |
| 510 | r = client.post("/cohere/v2/chat/render", json=_minimal_request_body()) |
| 511 | assert r.status_code == HTTPStatus.INTERNAL_SERVER_ERROR |
| 512 | assert r.json() == {"message": "unconvertible message"} |
| 513 | |
| 514 | def test_invalid_body_returns_422(self): |
| 515 | app = _build_render_app(_RenderChatHandler(), _RenderHandler(None)) |
| 516 | with TestClient(app) as client: |
| 517 | r = client.post( |
| 518 | "/cohere/v2/chat/render", |
| 519 | json={"messages": [{"role": "user", "content": "hi"}]}, |
| 520 | ) |
| 521 | assert r.status_code == HTTPStatus.UNPROCESSABLE_ENTITY |
| 522 | |
| 523 | |
| 524 | # ---------------------------------------------------------------------- |
| 525 | # CohereErrorEnvelopeMiddleware |
| 526 | # ---------------------------------------------------------------------- |
| 527 | |
| 528 | |
| 529 | class TestCohereErrorEnvelope: |
| 530 | """When the app installs vLLM's global exception handlers, validation |
| 531 | and HTTP errors escape as ``ErrorResponse`` bodies before the route |
| 532 | handler runs. The middleware installed by ``attach_router`` must |
| 533 | normalise those bodies to the ``CohereError`` shape declared on the |
| 534 | endpoint's OpenAPI ``responses`` map so schema-conformance tests |
| 535 | (``test_openai_schema.py``) don't see a mismatch on ``/cohere/*`` |
| 536 | responses. |
| 537 | """ |
| 538 | |
| 539 | def test_validation_error_body_is_cohere_shaped(self): |
| 540 | # ``model=""`` and ``messages=[]`` trip our custom field |
| 541 | # validators, which raise VLLMValidationError and are routed |
| 542 | # through ``vllm_error_handler`` in the real vLLM |
| 543 | # server (producing the ``{"error": {...}}`` shape). |
| 544 | app = _build_app_with_vllm_handlers(handler=None) |
| 545 | with TestClient(app) as client: |
| 546 | r = client.post("/cohere/v2/chat", json={"messages": [], "model": ""}) |
| 547 | assert r.status_code == HTTPStatus.BAD_REQUEST |
| 548 | body = r.json() |
| 549 | # ``CohereError`` has ``message`` at the top level, not nested |
| 550 | # under an ``error`` envelope. |
| 551 | assert "error" not in body |
| 552 | assert "message" in body |
| 553 | assert isinstance(body["message"], str) and body["message"] |
| 554 | |
| 555 | def test_http_error_body_is_cohere_shaped(self): |
| 556 | # A raised ``HTTPException`` from anywhere in the request cycle |
| 557 | # is routed through ``http_exception_handler`` (producing the |
| 558 | # ``ErrorResponse`` shape) and must be translated. |
| 559 | app = _build_app_with_vllm_handlers(handler=None) |
| 560 | |
| 561 | @app.get("/cohere/v2/boom") |
| 562 | async def _boom(): |
| 563 | raise HTTPException(status_code=418, detail="teapot") |
| 564 | |
| 565 | with TestClient(app) as client: |
| 566 | r = client.get("/cohere/v2/boom") |
| 567 | assert r.status_code == 418 |
| 568 | body = r.json() |
| 569 | assert body == {"message": "teapot"} |
| 570 | |
| 571 | def test_non_cohere_path_is_not_translated(self): |
| 572 | app = _build_app_with_vllm_handlers(handler=None) |
| 573 | |
| 574 | @app.get("/v1/other") |
| 575 | async def _other(): |
| 576 | raise HTTPException(status_code=400, detail="nope") |
| 577 | |
| 578 | with TestClient(app) as client: |
| 579 | r = client.get("/v1/other") |
| 580 | assert r.status_code == HTTPStatus.BAD_REQUEST |
| 581 | body = r.json() |
| 582 | # Non-cohere paths keep the vLLM ``ErrorResponse`` shape. |
| 583 | assert "error" in body |
| 584 | assert body["error"]["message"] == "nope" |
| 585 | |
| 586 | def test_streaming_response_passes_through(self): |
| 587 | # SSE responses have content-type text/event-stream; the |
| 588 | # middleware must never buffer these (which would break |
| 589 | # streaming) even though they're on ``/cohere/*``. |
| 590 | async def _gen() -> AsyncGenerator[str, None]: |
| 591 | yield 'data: {"type":"message-start"}\n\n' |
| 592 | yield "data: [DONE]\n\n" |
| 593 | |
| 594 | app = _build_app_with_vllm_handlers(handler=_Handler(_gen())) |
| 595 | with TestClient(app) as client: |
| 596 | r = client.post( |
| 597 | "/cohere/v2/chat", |
| 598 | json={**_minimal_request_body(), "stream": True}, |
| 599 | ) |
| 600 | assert r.status_code == HTTPStatus.OK |
| 601 | assert r.headers["content-type"].startswith("text/event-stream") |
| 602 | assert "message-start" in r.text |
| 603 | assert r.text.rstrip().endswith("[DONE]") |
| 604 | |
| 605 | def test_already_cohere_shaped_body_passes_through(self): |
| 606 | # When the handler returns an ``ErrorResponse`` the route |
| 607 | # itself translates it to ``CohereError``; the middleware sees |
| 608 | # the ``CohereError`` shape and must leave it alone. |
| 609 | err = ErrorResponse( |
| 610 | error=ErrorInfo(message="already cohere", type="Bad Request", code=400) |
| 611 | ) |
| 612 | app = _build_app_with_vllm_handlers(handler=_Handler(err)) |
| 613 | with TestClient(app) as client: |
| 614 | r = client.post("/cohere/v2/chat", json=_minimal_request_body()) |
| 615 | assert r.status_code == HTTPStatus.BAD_REQUEST |
| 616 | body = r.json() |
| 617 | # No ``error`` wrapper: the route already emitted the wire shape. |
| 618 | assert body == {"message": "already cohere"} |
| 619 | |
| 620 | def test_request_id_preserved_in_translated_body(self): |
| 621 | # Client-provided ``X-Request-Id`` should be echoed as |
| 622 | # ``CohereError.id`` so callers can correlate failures. |
| 623 | app = _build_app_with_vllm_handlers(handler=None) |
| 624 | with TestClient(app) as client: |
| 625 | r = client.post( |
| 626 | "/cohere/v2/chat", |
| 627 | json={"messages": [], "model": ""}, |
| 628 | headers={"X-Request-Id": "req-abc"}, |
| 629 | ) |
| 630 | assert r.status_code == HTTPStatus.BAD_REQUEST |
| 631 | body = r.json() |
| 632 | assert body.get("id") == "req-abc" |