| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """End-to-end integration tests for the Cohere Chat v2 endpoints |
| 4 | (``POST /cohere/v2/chat`` and ``POST /cohere/v2/chat/render``). |
| 5 | |
| 6 | These tests spin up a real ``vllm serve`` process via |
| 7 | :class:`tests.utils.RemoteOpenAIServer` and exercise the Cohere Chat v2 |
| 8 | contract over HTTP. They mirror the pattern used by |
| 9 | :mod:`tests.entrypoints.anthropic.test_messages`. |
| 10 | |
| 11 | Two layers of integration are covered: |
| 12 | |
| 13 | 1. Raw HTTP via :mod:`httpx` — always runs, verifies the wire contract. |
| 14 | 2. Cohere SDK (``pip install cohere``) — auto-skipped when the optional |
| 15 | dependency isn't installed, verifies SDK-level interop. |
| 16 | |
| 17 | We use a tiny generic chat model (``HuggingFaceTB/SmolLM2-135M-Instruct``, |
| 18 | ≈135M params) so the fixture stays runnable on CPU-only laptops. The |
| 19 | Cohere v2 endpoint is model-agnostic and just performs the v2 ↔ OpenAI |
| 20 | chat-completion translation, so the choice of model only matters for |
| 21 | response *shape* — the translation logic itself is unit-tested |
| 22 | elsewhere. |
| 23 | """ |
| 24 | |
| 25 | import json |
| 26 | |
| 27 | import httpx |
| 28 | import pytest |
| 29 | import pytest_asyncio |
| 30 | |
| 31 | from tests.utils import RemoteOpenAIServer |
| 32 | |
| 33 | # Tiny chat-tuned model so the fixture is cheap to boot on CPU-only |
| 34 | # hosts (Mac, CI runners without GPUs). The Cohere v2 translation is |
| 35 | # completely independent of which underlying chat model is loaded. |
| 36 | MODEL_NAME = "HuggingFaceTB/SmolLM2-135M-Instruct" |
| 37 | SERVED_MODEL_NAME = "command-r-plus-08-2024" |
| 38 | |
| 39 | |
| 40 | @pytest.fixture(scope="module") |
| 41 | def server(): |
| 42 | args = [ |
| 43 | # 1k tokens is plenty for the prompts these tests send; trimming |
| 44 | # it from the default keeps the KV-cache footprint tiny. |
| 45 | "--max-model-len", |
| 46 | "1024", |
| 47 | "--max-num-seqs", |
| 48 | "4", |
| 49 | "--dtype", |
| 50 | "bfloat16", |
| 51 | "--enforce-eager", |
| 52 | # Advertise a Cohere model name so Cohere SDK ``model=`` calls |
| 53 | # round-trip without the ``model not found`` check. |
| 54 | "--served-model-name", |
| 55 | SERVED_MODEL_NAME, |
| 56 | # Disable the reasoning-model path so the test doesn't require |
| 57 | # the conversation to surface a thinking block (SmolLM2 doesn't |
| 58 | # emit Cohere-style reasoning tokens out of the box). |
| 59 | "--no-cohere-is-reasoning-model", |
| 60 | # Registers /cohere/v2/chat/render alongside /cohere/v2/chat. |
| 61 | "--enable-scale-out", |
| 62 | ] |
| 63 | |
| 64 | # Cap the CPU KV-cache pool the vLLM CPU backend reserves at |
| 65 | # startup. The default (4 GB) trips Mac's RAM check on smaller |
| 66 | # machines; 1 GB is more than enough for max_model_len=1024 and |
| 67 | # max_num_seqs=4. |
| 68 | # ``VLLM_ENABLE_COHERE_API=1`` flips the opt-in gate that |
| 69 | # ``vllm.entrypoints.cohere.api_router.attach_router`` reads to |
| 70 | # decide whether to register ``POST /cohere/v2/chat`` at server |
| 71 | # startup. Without it the fixture would boot a server that returns |
| 72 | # 404 for every request the tests below make. |
| 73 | env_dict = { |
| 74 | "VLLM_CPU_KVCACHE_SPACE": "1", |
| 75 | "VLLM_ENABLE_COHERE_API": "1", |
| 76 | } |
| 77 | |
| 78 | with RemoteOpenAIServer(MODEL_NAME, args, env_dict=env_dict) as remote_server: |
| 79 | yield remote_server |
| 80 | |
| 81 | |
| 82 | # ---------------------------------------------------------------------- |
| 83 | # Layer 1: raw HTTP contract (no optional deps) |
| 84 | # ---------------------------------------------------------------------- |
| 85 | |
| 86 | |
| 87 | @pytest_asyncio.fixture |
| 88 | async def httpx_client(server): |
| 89 | async with httpx.AsyncClient( |
| 90 | base_url=server.url_root, timeout=httpx.Timeout(120.0) |
| 91 | ) as client: |
| 92 | yield client |
| 93 | |
| 94 | |
| 95 | @pytest.mark.asyncio |
| 96 | async def test_cohere_v2_chat_non_streaming(httpx_client: httpx.AsyncClient): |
| 97 | resp = await httpx_client.post( |
| 98 | "/cohere/v2/chat", |
| 99 | json={ |
| 100 | "model": SERVED_MODEL_NAME, |
| 101 | "messages": [{"role": "user", "content": "Say hi."}], |
| 102 | "max_tokens": 16, |
| 103 | "stream": False, |
| 104 | }, |
| 105 | ) |
| 106 | assert resp.status_code == 200, resp.text |
| 107 | payload = resp.json() |
| 108 | # The response envelope follows Cohere v2's schema. |
| 109 | assert "message" in payload |
| 110 | assert payload["message"]["role"] == "assistant" |
| 111 | # ``content`` is a list of content blocks; the synthesized |
| 112 | # ``CohereServingChatV2`` should always emit at least one ``text`` |
| 113 | # block (it falls back to an empty block when the model returned |
| 114 | # nothing). |
| 115 | content = payload["message"]["content"] |
| 116 | assert isinstance(content, list) and len(content) >= 1 |
| 117 | assert content[0]["type"] == "text" |
| 118 | # ``usage`` is always populated by the translator. |
| 119 | assert "usage" in payload |
| 120 | assert "billed_units" in payload["usage"] |
| 121 | assert "tokens" in payload["usage"] |
| 122 | # ``finish_reason`` is one of Cohere's enum values. |
| 123 | assert payload["finish_reason"] in { |
| 124 | "COMPLETE", |
| 125 | "MAX_TOKENS", |
| 126 | "STOP_SEQUENCE", |
| 127 | "TOOL_CALL", |
| 128 | "ERROR", |
| 129 | } |
| 130 | |
| 131 | |
| 132 | @pytest.mark.asyncio |
| 133 | async def test_cohere_v2_chat_streaming(httpx_client: httpx.AsyncClient): |
| 134 | """Streaming returns SSE frames in the v2 message-lifecycle shape.""" |
| 135 | events: list[dict] = [] |
| 136 | async with httpx_client.stream( |
| 137 | "POST", |
| 138 | "/cohere/v2/chat", |
| 139 | json={ |
| 140 | "model": SERVED_MODEL_NAME, |
| 141 | "messages": [{"role": "user", "content": "Hi"}], |
| 142 | "max_tokens": 8, |
| 143 | "stream": True, |
| 144 | }, |
| 145 | ) as resp: |
| 146 | assert resp.status_code == 200, await resp.aread() |
| 147 | assert resp.headers["content-type"].startswith("text/event-stream") |
| 148 | async for line in resp.aiter_lines(): |
| 149 | if not line.startswith("data: "): |
| 150 | continue |
| 151 | data = line[len("data: ") :] |
| 152 | if data == "[DONE]": |
| 153 | events.append({"type": "_DONE_"}) |
| 154 | continue |
| 155 | events.append(json.loads(data)) |
| 156 | |
| 157 | types = [ev["type"] for ev in events] |
| 158 | # The lifecycle always starts with message-start and ends with [DONE] |
| 159 | # preceded by message-end. |
| 160 | assert types[0] == "message-start" |
| 161 | assert types[-1] == "_DONE_" |
| 162 | assert types[-2] == "message-end" |
| 163 | # message-start must carry the chunk/message id. |
| 164 | assert events[0].get("id") |
| 165 | |
| 166 | |
| 167 | @pytest.mark.asyncio |
| 168 | async def test_cohere_v2_chat_validation_error_returns_400( |
| 169 | httpx_client: httpx.AsyncClient, |
| 170 | ): |
| 171 | # Missing required ``model`` field → FastAPI/Pydantic returns 400. |
| 172 | resp = await httpx_client.post( |
| 173 | "/cohere/v2/chat", |
| 174 | json={"messages": []}, |
| 175 | ) |
| 176 | assert resp.status_code == 400 |
| 177 | |
| 178 | |
| 179 | @pytest.mark.asyncio |
| 180 | async def test_cohere_v2_chat_render(httpx_client: httpx.AsyncClient): |
| 181 | """``/cohere/v2/chat/render`` tokenizes without running the model. |
| 182 | |
| 183 | Beyond the ``GenerateRequest`` shape, this pins the cross-layer |
| 184 | contract that makes the endpoint useful: v2 sampling fields |
| 185 | (``temperature``, ``stop_sequences``) survive the v2 -> |
| 186 | ``ChatCompletionRequest`` -> ``SamplingParams`` hop, so a caller can |
| 187 | render once and generate later without re-deriving them. |
| 188 | """ |
| 189 | resp = await httpx_client.post( |
| 190 | "/cohere/v2/chat/render", |
| 191 | json={ |
| 192 | "model": SERVED_MODEL_NAME, |
| 193 | "messages": [{"role": "user", "content": "Say hi."}], |
| 194 | "max_tokens": 16, |
| 195 | "temperature": 0.125, |
| 196 | "stop_sequences": ["END"], |
| 197 | }, |
| 198 | ) |
| 199 | assert resp.status_code == 200, resp.text |
| 200 | data = resp.json() |
| 201 | |
| 202 | token_ids = data["token_ids"] |
| 203 | assert isinstance(token_ids, list) and len(token_ids) > 0 |
| 204 | assert all(isinstance(t, int) and t >= 0 for t in token_ids) |
| 205 | assert data["request_id"] |
| 206 | |
| 207 | sampling_params = data["sampling_params"] |
| 208 | assert sampling_params["temperature"] == 0.125 |
| 209 | assert sampling_params["stop"] == ["END"] |
| 210 | |
| 211 | # No generation happened, so the response carries none of the chat |
| 212 | # envelope the /cohere/v2/chat endpoint would return. |
| 213 | assert "message" not in data |
| 214 | assert "finish_reason" not in data |
| 215 | |
| 216 | |
| 217 | @pytest.mark.asyncio |
| 218 | async def test_cohere_v2_chat_render_token_ids_decode_to_the_prompt( |
| 219 | httpx_client: httpx.AsyncClient, |
| 220 | ): |
| 221 | """The rendered ids are the real templated prompt, not a debug dump. |
| 222 | |
| 223 | Round-tripping them through ``/detokenize`` and finding the user |
| 224 | message back proves the endpoint applied the chat template and |
| 225 | tokenized it, which is what a caller splitting prompt construction |
| 226 | from generation depends on. |
| 227 | """ |
| 228 | render_resp = await httpx_client.post( |
| 229 | "/cohere/v2/chat/render", |
| 230 | json={ |
| 231 | "model": SERVED_MODEL_NAME, |
| 232 | "messages": [{"role": "user", "content": "What is the capital of France?"}], |
| 233 | "max_tokens": 8, |
| 234 | }, |
| 235 | ) |
| 236 | assert render_resp.status_code == 200, render_resp.text |
| 237 | token_ids = render_resp.json()["token_ids"] |
| 238 | |
| 239 | detok_resp = await httpx_client.post( |
| 240 | "/detokenize", |
| 241 | json={"model": SERVED_MODEL_NAME, "tokens": token_ids}, |
| 242 | ) |
| 243 | assert detok_resp.status_code == 200, detok_resp.text |
| 244 | prompt = detok_resp.json()["prompt"] |
| 245 | assert "What is the capital of France?" in prompt |
| 246 | |
| 247 | |
| 248 | @pytest.mark.asyncio |
| 249 | async def test_cohere_v2_chat_render_validation_error_returns_400( |
| 250 | httpx_client: httpx.AsyncClient, |
| 251 | ): |
| 252 | """Bad bodies get the Cohere error envelope, not vLLM's.""" |
| 253 | resp = await httpx_client.post("/cohere/v2/chat/render", json={"messages": []}) |
| 254 | assert resp.status_code == 400 |
| 255 | body = resp.json() |
| 256 | assert "error" not in body |
| 257 | assert body.get("message") |
| 258 | |
| 259 | |
| 260 | @pytest.mark.asyncio |
| 261 | async def test_cohere_v2_chat_documents_field_accepted( |
| 262 | httpx_client: httpx.AsyncClient, |
| 263 | ): |
| 264 | """The v2 endpoint forwards ``documents`` into chat_template_kwargs. |
| 265 | |
| 266 | We only assert the request is accepted and produces a 200 response — |
| 267 | the renderer-level effect is covered by ``tests/renderers/test_cohere.py``. |
| 268 | """ |
| 269 | resp = await httpx_client.post( |
| 270 | "/cohere/v2/chat", |
| 271 | json={ |
| 272 | "model": SERVED_MODEL_NAME, |
| 273 | "messages": [{"role": "user", "content": "Summarize."}], |
| 274 | "documents": [{"id": "d1", "data": {"title": "T", "snippet": "S"}}], |
| 275 | "max_tokens": 16, |
| 276 | "stream": False, |
| 277 | }, |
| 278 | ) |
| 279 | assert resp.status_code == 200, resp.text |
| 280 | |
| 281 | |
| 282 | # ---------------------------------------------------------------------- |
| 283 | # Layer 2: Cohere SDK round-trip (auto-skipped if SDK absent) |
| 284 | # ---------------------------------------------------------------------- |
| 285 | |
| 286 | |
| 287 | @pytest_asyncio.fixture |
| 288 | async def cohere_async_client(server): |
| 289 | cohere = pytest.importorskip("cohere") |
| 290 | # The vLLM endpoint is mounted at ``/cohere/v2/chat`` while the |
| 291 | # cohere SDK targets ``${base_url}/v2/chat``; point base_url at the |
| 292 | # ``/cohere`` prefix so paths line up. |
| 293 | client = cohere.AsyncClientV2( |
| 294 | api_key="dummy", |
| 295 | base_url=server.url_for("cohere"), |
| 296 | ) |
| 297 | try: |
| 298 | yield client |
| 299 | finally: |
| 300 | # ``AsyncClientV2`` exposes a sync close; if a future version |
| 301 | # adds aclose we still close cleanly. |
| 302 | close = getattr(client, "aclose", None) or getattr(client, "close", None) |
| 303 | if close is not None: |
| 304 | result = close() |
| 305 | if hasattr(result, "__await__"): |
| 306 | await result |
| 307 | |
| 308 | |
| 309 | @pytest.mark.asyncio |
| 310 | async def test_cohere_sdk_non_streaming(cohere_async_client): |
| 311 | resp = await cohere_async_client.chat( |
| 312 | model=SERVED_MODEL_NAME, |
| 313 | messages=[{"role": "user", "content": "Say hi."}], |
| 314 | max_tokens=16, |
| 315 | ) |
| 316 | # SDK parses our JSON into typed objects. |
| 317 | assert resp.message.role == "assistant" |
| 318 | assert resp.message.content is not None |
| 319 | assert len(resp.message.content) >= 1 |
| 320 | assert resp.message.content[0].type == "text" |
| 321 | assert resp.finish_reason in { |
| 322 | "COMPLETE", |
| 323 | "MAX_TOKENS", |
| 324 | "STOP_SEQUENCE", |
| 325 | "TOOL_CALL", |
| 326 | "ERROR", |
| 327 | } |
| 328 | |
| 329 | |
| 330 | @pytest.mark.asyncio |
| 331 | async def test_cohere_sdk_streaming(cohere_async_client): |
| 332 | events: list[str] = [] |
| 333 | stream = cohere_async_client.chat_stream( |
| 334 | model=SERVED_MODEL_NAME, |
| 335 | messages=[{"role": "user", "content": "Hi"}], |
| 336 | max_tokens=8, |
| 337 | ) |
| 338 | async for ev in stream: |
| 339 | events.append(ev.type) |
| 340 | |
| 341 | assert events, "SDK stream yielded no events" |
| 342 | assert events[0] == "message-start" |
| 343 | assert events[-1] == "message-end" |